Search

The importance of journals for diffusing the results of scientific research has increased considerably. In the digital era, Portable Document Format (PDF) became the established format of electronic journal articles. This structured form, combined with a regular and wide dissemination, spread scientific advancements easily and quickly. However, the rapidly increasing numbers of published scientific articles requires more time and effort on systematic literature reviews, searches and screens. The comprehension and extraction of useful information from the digital documents is also a challenging task, due to the complex structure of PDF. To help a soil science team from the United States Department of Agriculture (USDA) build a queryable journal paper system, we used web crawler to download articles on soil science from the digital library. We applied named entity recognition and table analysis to extract useful information including authors, journal name and type, publish date, abstract, DOI, experiment location in papers and highlight the paper characteristics in a computer queryable format in the system. Text classification is applied on to identify the parts of interest to the users and save their search time. We used traditional machine learning techniques including logistic regression, support vector machine, decision tree, naive bayes, k-nearest neighbors, random forest, ensemble modeling, and neural networks in text classification and compare the advantages of these approaches in the end. Advisor: Stephen D. Scott

Data and artificial intelligence technology can help transport operators such as airlines and railways to increase their operating income. Recognizing high-value passengers is particularly important for transportation operators because they can bring more than 50% of profits. However, unpredictable factors such as the global COVID-19 pandemic epidemic and data privacy leaks have challenged the assessment of high-value passengers. The aim of this thesis is to adequately fill the customer value assessment gap under unpredictable transportation factors. The thesis adopts the form of a journal, and it consists of six papers. Papers 1 to 3 provide an in-depth study on how Big Data machine learning technology can help in the evaluation of passenger value. Paper 1 shows the Bayesian belief network of an airline content recommendation model. In this paper, the static causality model is transformed into a dynamic causal model to show its enhancement effect on the prediction of consumer choice preferences. Paper 2 proposes an optimized airline passenger value assessment model. A fit test is conducted with TravelSky's value score to improve the accuracy of passenger value assessment. Paper 3 adopts a two-stage modelling method. In the first stage, series regression, gradient boosting decision tree and neural network model are used for modelling the passenger value evaluation. These models are integrated to improve the accuracy of passenger value assessment. Papers 4 to 6 cover management theory using federated learning methods. Paper 4 proposes a new classification model based on federated learning, which is used for high-value passenger recognition in the civil aviation industry. This model enhances the accuracy of passenger value assessment. Paper 5 uses federated learning technology to simulate the willingness to pay for upgrade services and based on the results, facilitate the possibility to accurately predict the precise population that is willing to perform aircraft upgrade services, thereby improving the accuracy of customers' willingness to pay for upgrade services. Finally, Paper 6 proposes a method of applying vertical federated learning technology to help airlines to identify target customers. It also extends the discussion from civil aviation to railway and proposes a concept of constructing the mobility as a service (MaaS) platform for the future.

Each human exome contains thousands of protein-altering variants located in more than 19,000 genes. Humans typically have two copies of a gene, and variants that affect one or both gene copies are called heterozygous and homozygous, respectively. If one gene copy is affected by deleterious heterozygous variation and cannot produce normal protein, this could result in a dominant disease. However, some genes can tolerate disruption of one copy, but deleterious homozygous or two heterozygous variants in different copies could still result in a recessive disease. Finally, humans can tolerate the inactivation or deletion of both copies of some genes without developing diseases. Because studied diseases' inheritance patterns are frequently known (e.g. if one of the parents and a child both have a disease, the inheritance pattern is likely to be dominant), clinical researchers want to know a candidate disease-causing variant inheritance pattern to prioritise candidate disease genes for laboratory validation. Although inheritance pattern is a property of disease causing variants, it can be predicted using gene-level properties. The aim of this study was to develop gene-level computational metrics that can be used for this task, and recently created large variant population databases such as Genome Aggregation Database (gnomAD, >137,000 individual exomes/genomes) provided novel data for such studies. This thesis is written in the journal format and consists of three paper-style result chapters. In the first paper, we analysed deviations from Hardy-Weinberg Equilibrium of rare variants in gnomAD to detect potential disease-causing and heterozygous advantageous variants based on homozygous deficiency in the healthy populations. The second paper developed a gene variation intolerance ranking (GeVIR) system by measuring how unevenly variants in gnomAD were distributed in a gene relative to other genes. Finally, in the third paper, we developed multiple supervised machine learning models based on various gene properties (including GeVIR) and combined them into a single continuous gene ranking metric that can be used to measure gene predisposition to disease inheritance patterns (DIP). In conclusion, this thesis contributed to the understanding of variant population data and the application of supervised ML methods to classify candidate disease genes in the context of disease inheritance patterns. The primary outcome of this research was the development of two continuous gene metrics, GeVIR and DIP (available for 19,361 and 15,794 protein-coding genes, respectively), both of which can be used to distinguish dominant, recessive and non-disease genes. We anticipate that these metrics will aid clinical researchers in the prioritisation of candidate disease genes.

This thesis contributes to the study of identification in macroeconomics. The first two chapters combine machine learning techniques with econometrics to provide new insights into this long-standing question. I use natural language processing techniques in the first chapter to derive a novel monetary policy shock series. Interested researchers can readily extend this idea to other areas where large amounts of technical documents are available, for example, tax policies or financial markets. In the second chapter, we develop a new method to estimate state dependent policy effects. While researchers previously had to decide on the variables that determine the state, their interactions and their functional form, our approach nests these decisions in a data-driven way. We hope that our methodology simplifies the estimation of state dependent policy effects and leads to new findings in this area. The final chapter provides a novel approach to identifying expectation shocks in a fiscal policy VAR and discusses the particular case of constructing counterfactual impulse response functions for expectations. Both can be useful for studying other policy transmission mechanisms. The first chapter uses text analysis methods from the linguistic machine learning literature to construct a new monetary policy shock series. Measures of monetary shocks commonly give rise to the puzzling result that a monetary tightening has an expansionary effect. A possible reason is that agents may believe that monetary shocks contain infor- mation regarding the central bank's assessment of the economic environment. Under this hypothesis, the estimated response to monetary policy shocks would contain two conflating effects: the actual effect of monetary policy and the reaction of private agents to the newly acquired information. This paper overcomes this problem by extracting a novel series of monetary shocks using text analysis methods on central bank documents. The resulting text-based variables contain the informational content from changes in the policy rate. Thus, they can be used to extract exogenous changes in monetary policy that are orthogonal to any central bank information. Using this information-free measure of monetary policy shocks reveals that a monetary tightening is not expansionary, even when estimated on more recent periods. The second chapter is co-authored with Christian Rörig and proposes a flexible frame- work to identify state dependent effects of macroeconomic policies. In the literature, it is common to either estimate constant policy effects or introduce state dependency in a parametric fashion. The latter, however, demands prior assumptions about the func- tional form. Our new method allows identifying state dependent effects and possible interactions in a data-driven way. Specifically, we estimate heterogeneous policy effects using semi-parametric varying-coefficient models in an otherwise standard VAR structure. While keeping a parametric reduced form for interpretability and efficiency, we estimate the coefficients as functions of modifying macroeconomic variables, using random forests as the underlying non-parametric estimator. Simulation studies show that this method correctly identifies multiple states even for relatively small sample sizes. To further val- idate our method, we apply the semi-parametric framework to a historical data set by Ramey & Zubairy and offer a more granular perspective on the dependence of the fiscal policy efficacy on unemployment and interest rates. Allowing for interactions between un- employment and interest rates, we show that it is indeed unemployment that is important to explain state dependent fiscal policy effects. The final chapter is co-authored with my supervisor Pontus Rendahl. In our paper, we empirically study the role of expectations in the transmission of fiscal policy. We extend an otherwise standard fiscal policy VAR with inflation and output expectations to construct counterfactual IRFs. Counterfactual IRFs allow us to ask how the economy would have responded to a government spending shock holding inflation or output expectations fixed. This exercise reveals that output expectations are the key driver in the transmission of government spending shocks. Output expectations contribute 60%-90% to the effect of government spending shocks on the fiscal multiplier at different impulse response horizons. We also make several methodological contributions. Firstly, we provide a novel way to identify shocks to expectations using lagged expectations as internal instruments. Secondly, we illustrate how to combine external and internal instruments to estimate a VAR's impact effects with a single reduced form. Finally, we show that constructing counterfactual IRFs with a hypothetical shock series or setting zeros in the structural matrices of a VAR is equivalent.

The dissertation consists of three essays that contribute to the literature on machine learning in empirical finance. In the first paper, I create proxies for managers' cultural fit using one of the latest machine learning technologies - the sentence embedding model - by analysing 11.5 million speeches in earnings calls. A better cultural fit is significantly and positively correlated with managerial tenure. I demonstrate that the effect of cultural fit on managerial tenure is causal using random survival forests. Firms that hire culturally disruptive managers have lower future market values and performance. The stock market reacts positively to signals that indicate low cultural dispersion within the firm. In the second paper, we document a gender-based attention effect in the sensitivity of mutual fund flows to fund performance using individual-level fund data from a fintech platform in China. Investors increase (decrease) flows to funds following positive and strong (negative and weak) prior-month performance. However, although there is no significant difference in the performance of male and female managers, the sensitivity effect significantly weakens if the fund manager is female. The effect persists after controlling for the tone of news articles on fund managers, measured using a state-of-the-art machine learning model. Simply put, investors react less to the performance of female fund managers. In the third paper, we document a significant, up to 10-fold increase in the synchronicity of intra-day, ultra-high frequency stock returns over the last decade. This surge in the intra-day synchronicity across stocks coincided with the advent of electronic, automated trading in U.S. markets. Using changes to the S&P500 index, we establish evidence of a causal relationship using a new machine learning tool - causal random forests. When firms are included in this major index, they enter the radar of high frequency arbitrageurs and market-making bots. These automated trading bots, who monitor prices in major securities closely and continuously, increase their quoting activities significantly and cause individual stocks' returns to synchronize at the microstructure level.

Machine learning is growing at an exponential rate in industry today, with contemporary machine learning models having parameter counts in the high hundred millions to trillions paired with a huge computing requirement. As the field continues to grow and the models get larger and more computationally expensive, there is a greater need for specialized hardware that can accelerate the inference and training computations that are required of these machine learning models. The design of a custom hardware device specifically made to accelerate machine learning models is very intricate, with software and hardware development that needs to work together at the cutting edge. The following paper will cover these hardware accelerators and focus specifically on the Grayskull e150, a chip optimized for machine learning inference developed by Tenstorrent. Tenstorrent is a hardware company out of Toronto, Canada. This paper will detail the hardware design choices in the Grayskull that make it optimal for machine learning inference, along with the software design that facilitates compatibility with a large number of machine learning frameworks and models. The paper will end with tests of the Grayskull card on various machine learning models, comparing its performance against a Nvidia GPU and an Intel CPU.

Primary biological aerosol particles, including bacteria, fungal spores and pollen are a crucial component of the atmosphere. These particles have a significant impact on human health, especially for those who suffer from hay fever and asthma. They can also influence the weather, acting as ice nuclei (IN) or cloud condensation nuclei (CCN) and can be impactful on agriculture through destruction of crops caused by fungal diseases. It is also vital to monitor these particles in an indoor environment: in hospitals as well as cleanrooms within the pharmaceutical manufacturing process. There are several different techniques that scientists can use to analyse bio-aerosol. Manual techniques, including tape and filter sampling, are often used by meteorological institutions to provide a pollen forecast. Alternatively, in a clean indoor environment, pharmaceutical manufacturers may use culture-based techniques to help prevent contamination. These approaches, however, can suffer from low time resolution and substantial human labour costs. As a consequence, interest in ultraviolet light-induced fluorescence (UV-LIF), which can provide real-time particle measurements, has increased in the last decade. Such technological advancements, especially instruments that produce high resolution data, bring a wealth of complex data that can be difficult to analyse. Consequently, it becomes ever more important to question the algorithmic approach used to analyse the data. One of the potential solutions to overcome this obstacle is machine learning, which has seen increased interest in recent years, with a vast number of applications. As more laboratory data have also become available, it has been possible to test the applicability of a wider variety of algorithms. In this thesis, we evaluate the applicability of these algorithms to data collected using both an earlier developed instrument called the WIBS and an updated version of the WIBS called the MBS. Hierarchical agglomerative clustering is a technique that has been used by researchers in previous studies using the WIBS. In this thesis, we present the first paper, using the MBS as well as extend the current body of research that uses the WIBS, showing some previously unseen disadvantages to this technique. We demonstrate excellent discrimination between laboratory particles for the newer instrument using supervised machine learning techniques but identify critical areas where the application of such algorithms to ambient data are limited and need improvement. Primary biological aerosol particles, including bacteria, fungal spores and pollen are a crucial component of the atmosphere. These particles have a significant impact on human health, especially for those who suffer from hay fever and asthma. They can also influence the weather, acting as ice nuclei (IN) or cloud condensation nuclei (CCN) and can be impactful on agriculture through destruction of crops caused by fungal diseases. It is also vital to monitor these particles in an indoor environment: in hospitals as well as cleanrooms within the pharmaceutical manufacturing process. There are several different techniques that scientists can use to analyse bio-aerosol. Manual techniques, including tape and filter sampling, are often used by meteorological institutions to provide a pollen forecast. Alternatively, in a clean indoor environment, pharmaceutical manufacturers may use culture-based techniques to help prevent contamination. These approaches, however, can suffer from low time resolution and substantial human labour costs. As a consequence, interest in ultraviolet light-induced fluorescence (UV-LIF), which can provide real-time particle measurements, has increased in the last decade. Such technological advancements, especially instruments that produce high resolution data, bring a wealth of complex data that can be difficult to analyse. Consequently, it becomes ever more important to question the algorithmic approach used to analyse the data. One of the potential solutions to overcome this obstacle is machine learning, which has seen increased interest in recent years, with a vast number of applications. As more laboratory data have also become available, it has been possible to test the applicability of a wider variety of algorithms. In this thesis, we evaluate the applicability of these algorithms to data collected using both an earlier developed instrument called the WIBS and an updated version of the WIBS called the MBS. Hierarchical agglomerative clustering is a technique that has been used by researchers in previous studies using the WIBS. In this thesis, we present the first paper, using the MBS as well as extend the current body of research that uses the WIBS, showing some previously unseen disadvantages to this technique. We demonstrate excellent discrimination between laboratory particles for the newer instrument using supervised machine learning techniques but identify critical areas where the application of such algorithms to ambient data are limited and need improvement.

In today’s world with ever increasing need for data collection, there is a rise in demand for privacy-preserving synthetic data generation and privacy auditing techniques to safeguard sensitive user information and data from privacy attacks. This paper explores the adversarial privacy auditing of synthetically generated data produced by Large Language Models (LLMs) using the TAPAS “Toolbox for Adversarial Privacy Auditing of Synthetic Data” framework. This paper uses a healthcare dataset with sensitive user information of Breast Cancer to evaluate the privacy of the data using adversarial techniques. The paper compares and contrasts the data quality, data distributions and privacy-preserving metrics of the real dataset with synthetically generated datasets from several sources including LLMs such as the GReaT framework and OpenAI's GPT4, Generative Adversarial Networks (GANs), and an AI-generated dataset produced using a proprietary technique from an industry startup, mostly.ai.

Much of the manufacturing industry is focused towards transitioning from Industry 3.0 towards Industry 4.0. Data science is the basis of Industry 4.0 for the bio-composites industry. The bio-composites industry consists of enterprises which produce products from renewable raw materials. This research project was applied, and predictive models were developed of the strength properties of final products for a particleboard mill. Quality control measures in bio-composite manufacturing tend toward reactivity given the time from when samples are taken from the production line and destructive test data are presented to operations personal. The implementation of a feedback system for final board strength properties will allow for a more proactive approach to quality control and process improvement. A gap in the literature, identified by Liao et al (2017) found most Industry 4.0 research done in labs (95%), leaving little applied research (5%). This research aims for a better knowledge of how useful and feasible implementing machine learning algorithms in a bio-composite manufacturing setting are. Quality control lab data was combined with process data to create fifteen predictive models. The merged data sets established a link between final board strength properties and process parameters from sensors along the process. The database starts March 3rd, 2021 and ends January 4th, 2022. The dataset was divided into seven product-grade categories by board strength property and grade. The algorithm ‘Multivariate Adaptive Regression Splines - four interaction terms’ (MARS4) performed best across all product categories in 10 validation studies. The R2 of MARS4 in ten validation studies ranged from 0.588 to 0.845 (¯ = 0.731) for all internal bond product-grade categories. The R2 of MARS4 in ten validation studies ranged from 0.745 to 0.923 (¯ = 0.826) of all modulus of rupture categories. Weighted predictions of the three top performing models (MARS4, MARS3, MARS2) in validation had promising results. Automated decision making, which is powered by predictive analytics, is a foundation of Industry 4.0, which goes beyond Industry 3.0. The real-time feedback loop had to be simulated in this study due to an unforeseen occurrence and the subsequent disconnection of the modeling servers from the process servers.

All researchers have experienced the problem of fishing out the most relevant scientific papers from an ocean of publications, and some may have wished that their text editor suggested these papers automatically. This thesis is vertebrated by this task: recommending contextually relevant citations to the author of a scientific paper, which we call Contextual Citation Recommendation (CCR). Like others before, we frame CCR as an Information Retrieval task and we evaluate our approach using existing publications. That is, an existing in-text citation to one or more documents in a corpus is replaced with a placeholder and the task is to retrieve the cited documents automatically. We carry out a cross-domain study and evaluate our approaches using two separate document collections in two different domains: computational linguistics and biomedical science. This thesis is comprised of three parts, which build cumulatively. Part I establishes a framework for the task using a standard Information Retrieval setup and explores different parameters for indexing documents and for extracting the evaluation queries in order to establish solid baselines for our two corpora in our two domains. We experiment with symmetric windows of words and sentences for both query extraction and for integrating the anchor text, that is, the text surrounding a citation, which is an important source of data for building document representations. We show for the first time that the contribution of anchor text is very domain dependent. Part II investigates a number of scientific discourse annotation schemes for academic articles. It has often been suggested that annotating discourse structure could support Information Retrieval scenarios such as this one, and this is a key hypothesis of this thesis. We focus on two of these: Argumentative Zoning (AZ, for the domain of computational linguistics) and Core Scientific Concepts (for the domain of biomedical sciences); both of these sentence-based, scientific discourse annotation schemes which define classes such as Hypothesis, Method and Result for CoreSC and Background/ Own/Contrast for AZ. By annotating each sentence in every document with AZ/CoreSC and indexing them separately by sentence class, we discover that consistent citing patterns exist in each domain, such as that sentences of type Conclusion in cited papers are consistently cited by other sentences of type Conclusion or Background in citing biomedical articles. Finally, Part III moves away from simple windows over terms or over sentences for extracting the query from a citation's context, and investigates methods for supervised query extraction using linguistic information. As part of this, we first explore how to automatically generate training data in the form of citation contexts paired with an optimal query to generate. Second, we train supervised machine learning models for automatically extracting these queries with limited prior knowledge of the document collection and show important improvements over our baselines in the domain of computational linguistics. We also investigate the contribution of stopwords to each corpus and we explore the performance of human annotators at this task.

Background: The software industry spends a lot of money on finding and fixing defects. It utilises software defect prediction models to identify code that is likely to be defective. Prediction models have, however, reached a performance bottleneck. Any improvements to prediction models would likely yield less defects-reducing costs for companies. Aim: In this dissertation I demonstrate that different families of classifiers find distinct subsets of defects. I show how this finding can be utilised to design ensemble models which outperform other state-of-the-art software defect prediction models. Method: This dissertation is supported by published work. In the first paper I explore the quality of data which is a prerequisite for building reliable software defect prediction models. The second and third papers explore the ability of different software defect prediction models to find distinct subsets of defects. The fourth paper explores how software defect prediction models can be improved by combining a collection of classifiers that predict different defective components into ensembles. An additional, non-published work, presents a visual technique for the analysis of predictions made by individual classifiers and discusses some possible constraints for classifiers used in software defect prediction. Result: Software defect prediction models created by classifiers of different families predict distinct subsets of defects. Ensembles composed of classifiers belonging to different families outperform other ensemble and standalone models. Only a few highly diverse and accurate base models are needed to compose an effective ensemble. This ensemble can consistently predict a greater number of defects compared to the increase in incorrect predictions. Conclusion: Ensembles should not use the majority-voting techniques to combine decisions of classifiers in software defect prediction as this will miss correct predictions of classifiers which uniquely identify defects. Some classifiers could be less successful for software defect prediction due to complex decision boundaries of defect data. Stacking based ensembles can outperform other ensemble and stand-alone techniques. I propose new possible avenues of research that could further improve the modelling of ensembles in software defect prediction. Data quality should be explicitly considered prior to experiments for researchers to establish reliable results.

This paper aims to present a study on developing a program that assists nonprofit organizations in determining the ideal location for building their facilities based on community needs, thus maximizing their potential for success. The study highlights the importance of location in the success of nonprofit organizations, and the challenges they face in identifying suitable areas for their operations. The paper reviews existing literature on nonprofit organizations, location analysis, and data analysis techniques, and proposes a methodology for developing the program. The methodology involves data collection and analysis, and machine learning algorithms to predict community needs. The program provides a user-friendly interface for nonprofit organizations to access and analyze the data and offers recommendations for suitable locations based on their criteria. The study concludes that the proposed program can be a valuable tool for nonprofit organizations to make informed decisions about their location and maximize their potential for success in serving their communities.

Learning view-invariant representation is essential to improving feature extraction for action recognition. Existing approaches cannot effectively capture details for human actions due to fast-paced gameplay and implicit view- dependent representation. In this paper, it goes beyond recognizing human actions from a fixed view and focusing on action recognition from an arbitrary view. This paper purposes a method to build an efficient data generating pipeline due to lack of original input data. This paper also provides a pipeline combining capturing modified I3D human actions features and use Multilayer Perception to achieve human action recognition and classification. The use of information captured from combination of virtual and real-life data, as well as different viewing angles, leads to high classification performance.

Semiconductor microelectronics integrated circuits (ICs) are increasingly integrated into modern life-critical applications, from intelligent infrastructure and consumer electronics to the Internet of Things (IoT) and advanced military and medical systems. Unfortunately, these applications are vulnerable to new hardware security attacks, including microelectronics counterfeits and hardware modification attacks. Physical Unclonable Functions (PUFs) are state-of-the-art hardware security solutions that utilize process variations of integrated circuits for device authentication, secret key generation, and microelectronics counterfeit detection. The negative impact of aging on Static Random \linebreak Access Memory Physical Unclonable Functions (SRAM PUFs) has significant consequences for microelectronics authentication, security, and reliability. This research thoroughly \linebreak examines the effect of aging on the reliability of SRAM PUFs used for secure and trusted microelectronics integrated circuit applications. It initially provides an overview of SRAM PUFs, highlighting their significance and essential features while addressing encountered challenges. The study then covers mitigation techniques, including multi-modal PUFs, that already exist to boost the resilience of SRAM PUFs against aging impacts, highlighting their advantages and the gap in the research addressed in this research. This work proposes a novel AI-enabled security for reliable SRAM PUFs. The proposed approach aims to study and countermeasure the impact of aging on SRAM PUF by analyzing data samples, including Bias Temperature Instability (BTI), Bit Flips, Accelerated aging, and Hot Carrier Injection (HCI) and to study their effects on SRAM PUF cell properties and output. Accelerated aging is a direct result of a change in the environmental temperature and voltage for a few hours. We aim to mitigate the impact of accelerated aging on the reliability authentication and encryption keys of SRAM PUFs. Further, AI-assisted approach is used to analyze SRAM PUF stability under accelerated aging operating conditions. Illegal memory chips from shady companies around the world have compromised the safety and reliability of electronic devices. Detecting these manufacturer details is crucial to detecting illegal and substandard manufacturing prior to their integration into key systems. This paper presents a novel approach designed specifically to analyze the SRAM PUF manufacturers. The suggested technique assigns a distinct process variation to every manufacturer, \linebreak facilitating the classification of the manufacturer, without the need for elaborate registration or verification protocols.The experimental results show the reliability and stability factors of SRAM PUF, including its influence on environmental conditions and the associated effects of aging. Our findings stress the importance of reliable and trustworthy PUF-based hardware security to reliably classify older microelectronic devices from new ones. Using 345 FPGA tested chips, the results show that using different ML models, we can efficiently classify these chips to correctly distinguish the manufacturer, aging impacts, and component, with F1 scores of 96\% and 98\%.

Why do firms publish science and release open-source software for free? The classical view in economics argues that firms will underproduce innovation relative to the social optimum because they cannot capture the social value available in cumulative ideas. And yet, firms do freely share their innovative knowledge, do so frequently, and do so with dramatic effect (with examples ranging in time from the transistor to the transformer architecture). Why? An influential perspective in the literature is that firms do capture value from this strategy, but they do so elsewhere – controlling the complement, whether through vertical innovation, co-invention, or the presence of unique complementary skills. Yet, while this perspective has impressive explanatory power across a wide range of industries, it fails to explain recent developments in artificial intelligence (AI) technologies. This dissertation therefore empirically examines this innovation strategy in the context of firm involvement in AI research, in search of new explanations. The first chapter of this dissertation observationally studies the impact of firm involvement on foundational scientific research, and, using a special set of “dual-affiliated” researchers located on the boundary of universities and firms, shows that firms can surprisingly have a positive effect on the scientific impact of a researcher’s papers. The second chapter uses a field experiment to estimate the impact of open source software libraries on the productivity of data scientists and illustrate the types of skills necessary for realizing those productivity gains. The final chapter uses the unexpected shift in governance of PyTorch (a leading machine learning framework) from Meta to the Linux Foundation in order to illustrate the importance of technology control rights in understanding why firms get involved in open-source software development. A key theme that emerges from these studies is that open firm innovation at scale (due to digitization) can lead to new possibilities for value capture that can align with value creation.

Image classification is a sub-field of computer vision that focuses on identifying objects within digital images. In order to improve image classification we must address the following areas of improvement: 1) Single and Multi-View data quality using data pre-processing techniques. 2) Enhancing deep feature learning to extract alternative representation of the data. 3) Improving decision or prediction of labels. This dissertation presents a series of four published papers that explore different improvements of image classification. In our first paper, we explore the Siamese network architecture to create a Convolution Neural Network based similarity metric. We learn the priority features that differentiate two given input images. The metric proposed achieves state-of-the-art Fβ measure. In our second paper, we explore multi-view data classification. We investigate the application of Generative Adversarial Networks GANs on Multi-view data image classification and few-shot learning. Experimental results show that our method outperforms state-of-the-art research. In our third paper, we take on the challenge of improving ResNet backbone model. For this task, we focus on improving channel attention mechanisms. We utilize Discrete Wavelet Transform compression to address the channel representation problem. Experimental results on ImageNet shows that our method outperforms baseline SENet-34 and SOTA FcaNet-34 at no extra computational cost. In our fourth paper, we investigate further the potential of orthogonalization of filters for extraction of diverse information for channel attention. We prove that using only random constant orthogonal filters is sufficient enough to achieve good channel attention. We test our proposed method using ImageNet, Places365, and Birds datasets for image classification, MS-COCO for object detection, and instance segmentation tasks. Our method outperforms FcaNet, and WaveNet and achieves the state-of-the-art results.

This paper presents two novel techniques for securing electronic control units (ECUs) in the controller area network (CAN) bus network of autonomous vehicles (AVs). Method 1 proposes an ECU fingerprinting technique to detect malicious ECUs in the in-vehicle CAN bus network. The technique extracts unique digital signatures based on intrinsic characteristics of the ECUs to identify the ECU responsible for broadcasting counterfeit messages received on the CAN bus. The proposed work employs three machine learning (ML) algorithms, namely k-Nearest Neighbors (k-NN), Support Vector Machine (SVM), and Logistic Regression (LR), to analyze the data from seven distinct ECUs. The performance of the proposed cybersecurity framework is evaluated and compared using these algorithms. Method 2 proposes a solution for efficient ECU classification to protect against multiple threats and attacks, including hacking and spoofing attacks, that can be perilous for the vehicle and its occupants. This technique visualizes signal loss and distortion in the ECU voltage signals caused by the ECU position in the CAN bus and utilizes Euclidean distance-based image classification on the principal components. This methodology is based on the commercially feasible eigenface-based algorithm, which has found extensive application in real-time scenarios like face recognition, fingerprint recognition, image recognition in photo-based applications, and real-time object identification. By profiling the ECU using this method, we can make the signal analyzer commercially viable. In this paper, we analyze different ECU configurations in the daisy chain network to evaluate the effectiveness of our proposed method. Our approach achieves a high accuracy rate of 97.14%. The proposed techniques address the security concerns related to the CAN bus network of AVs and provide efficient and effective ways to secure ECUs against malicious activities. The use of machine learning algorithms and visualization techniques in these methods not only enhances the accuracy of ECU detection and classification but also provides a better understanding of the underlying data. These techniques can be implemented in AVs to improve the security of the CAN bus network and ensure safe and secure operation of the vehicles.

Affective polarization - the growing dislike and distrust between supporters of opposing political parties - has become a growing concern in the United States and other Western democracies. While a number of potential explanations for this phenomena have been offered, such as elite ideological polarization or income inequality, in this dissertation I focus on the role of social identity and its relationship to polarization. I explore whether the growing inter-party hatred we witness is due less to what our opponents think, and more about who they represent. I build on social sorting theory to argue that individuals who perceive opposing parties as looking less like them will form more polarized views of the parties. In this dissertation, I offer three tests of this theory, and find several instances in which the alignment between an individual’s social identities — whether that be racial, gender, religious, age or other core identities — and those they attach to a political party, influences their feelings towards those parties.The first paper develops a theory of identity alignment, explaining how and why we might expect individuals to form views of the parties that are based on a comparison between an individual’s view of themselves and their views of those who support each party. I then provide three experimental tests of this theory, two studies using undergraduate subjects from a large public university, and a third replication study using a national sample, and find tentative evidence that supports the idea that the groups we associate with a political party influence our affective evaluations of the party, perceived ideological placement of the party, and responsiveness to party cues.The second paper studies the rise of social sorting in the United States and its relationship with affective polarization at the aggregate and individual level. I provide a new way to measure social sorting using the tools of machine learning, and validate this method by also applying it to the more well established case of ideological sorting. I then demonstrate that social sorting has risen at roughly the same rate as ideological sorting within the US, and that both phenomena ar2e independently associated with increasing levels of affective polarization. I show that individuals do indeed dislike their out-party because of what they believe, but they also form increasingly negative views because the two parties now represent wholly different social groups, comprised of few overlapping identities. This chapter provides strong evidence that the US parties now represent two groups who find it increasingly difficult to see themselves in the other side.The final chapter provides a comparative perspective by studying the relationship between social sorting and one core social identity — gender. I explore whether the gender gap between two parties — the difference in the proportion of female supporters within each party — impacts the affective evaluations that partisans make of out-parties. I argue that the gender composition of one’s own party represents an important identity marker, an indication of the identity composition of the party, and a yard stick by which individuals judge other parties. I argue that the more a party differs in its gender composition from an individual’s own party, the less warmly the individual will feel towards the party. I test this, along with competing theories regarding the gender composition of party supporters, on a dataset containing evaluations of 159 parties across 25 countries between 1996-2020, and find that an increased gender gap between parties is indeed associated with increasingly polarized evaluations of the out-party. I also find tentative evidence that individuals, particularly women, prefer parties with more women supporters, but that this relationship is attenuated by the effects of the gender gap.Taken together, these three papers present evidence that the increased partisan animosity that we see in many contemporary polities is driven, at least in part, by the social divisions between parties. As parties (particularly those in the United States) come to represent wholly different social groups who no longer share large numbers of cross-cutting identities, individuals find it increasingly difficult to find those among their political opponents who look like them, and so become increasingly negative, distrustful, and unresponsive to out-parties. Unlike other potential explanations for affective polarization such as elite ideological polarization, income inequality, or majoritarian electoral systems, this polarization driven by social sorting may be a much more difficult problem to solve. Without incentives for political parties to court supporters of social groups that are currently associated with their political opponents, it is hard to see how and why parties would take the necessary steps to decrease social-based polarization. Indeed, it is more likely that the results here are part of a vicious cycle — as groups become sorted with one party or another, this feeds into the stereotypes people have about the parties, forcing them to update their own identification and retrench into the party associated with their group, thus increasing levels of social sorting, and driving further polarization.

Developable surfaces are a special class of surfaces that are locally isometric to the plane. In a practical sense this means that they can be fabricated by bending inextensible sheets of material, such as sheet metal, wood panels, paper and up to a certain extent glass panels. Because of this property they form a set of surfaces that are relatively inexpensive to construct, which contributes to their popularity in architecture and industrial fabrication. Although the physical construction of developable surfaces is straightforward, the contrary is true for their digital design. General 3D modeling programs typically work with triangle mesh representations, which are suitable for covering a wide range of deformations and 3D shapes but are not easily transformable to fabrication instructions. The ideal representation of developables for fabrication purposes is to use a ruling-based representation. In the first part of this thesis we provide a solution for this discrepancy between desired surface representation for the modeling and fabrication parts of a developable surface construction pipeline. We approach this remeshing problem through a vector field optimization algorithm. The optimized vector field is integrated into a parameterization that we then use to extract a mesh that contains mostly quadrilateral faces. We add alignment and straightness constraints to the vector field optimization such that the parameterization that we obtain leads us to the ruling-based meshing of the surface. In the second part of this thesis we take another approach to generating developable surfaces. Rather than designing them directly, it is also possible to approximate an existing (non-developable) surface by a (piecewise) developable one. In order to solve this problem we make use of the fact that developable surfaces have a one-dimensional Gauss image. We create an iterative optimization that alternates between optimizing the surface normals such that its Gauss image becomes more developable and moving the surface triangles to match these newly prescribed surface normals as closely as possible. Our method is very intuitive to use and lets users easily transform any triangle mesh surface into a (piecewise) developable approximation of it. In the third part of this thesis we look at developables from a new perspective, namely we will consider them as they appear in images. Specifically we study images of warped documents and how we can unwarp these using a deep learning approach. We develop a convolutional neural network for the purpose of unwarping images of warped documents, as if the image came from a flat-bed scanner. Even though there is a large amount of training data already available for this task, we found that this data differs too much from real-world images of warped documents in order to achieve good unwarping performance. Therefore we produce a new dataset consisting of images of warped documents with rich ground truth annotations including the ground-truth unwarping function. Our data capturing pipeline enables us to circumvent the use of a rendering engine; instead we use an image compositing technique which ensures that the document images will keep a photo-realistic appearance. We provide a detailed description of our data capturing and generation pipeline to enable others to create customized data as necessary.

Context. Reports suggest that defects in code cost the US in excess of $50billion per year to put right. Defect Prediction is an important part of Software Engineering. It allows developers to prioritise the code that needs to be inspected when trying to reduce the number of defects in code. A small change in the number of defects found will have a significant impact on the cost of producing software. Aims. The aim of this dissertation is to investigate the factors which a ect the performance of defect prediction models. Identifying the causes of variation in the way that variables are computed should help to improve the precision of defect prediction models and hence improve the cost e ectiveness of defect prediction. Methods. This dissertation is by published work. The first three papers examine variation in the independent variables (code metrics) and the dependent variable (number/location of defects). The fourth and fifth papers investigate the e ect that di erent learners and datasets have on the predictive performance of defect prediction models. The final paper investigates the reported use of di erent machine learning approaches in studies published between 2000 and 2010. Results. The first and second papers show that independent variables are sensitive to the measurement protocol used, this suggests that the way data is collected a ects the performance of defect prediction. The third paper shows that dependent variable data may be untrustworthy as there is no reliable method for labelling a unit of code as defective or not. The fourth and fifth papers show that the dataset and learner used when producing defect prediction models have an e ect on the performance of the models. The final paper shows that the approaches used by researchers to build defect prediction models is variable, with good practices being ignored in many papers. Conclusions. The measurement protocols for independent and dependent variables used for defect prediction need to be clearly described so that results can be compared like with like. It is possible that the predictive results of one research group have a higher performance value than another research group because of the way that they calculated the metrics rather than the method of building the model used to predict the defect prone modules. The machine learning approaches used by researchers need to be clearly reported in order to be able to improve the quality of defect prediction studies and allow a larger corpus of reliable results to be gathered.

Although machine learning researchers have introduced a plethora of useful constructions for learning over Euclidean space, numerous types of data found in various applications areas benefit from, if not necessitate, a non-Euclidean treatment. In this thesis I cover the need for Riemannian geometric constructs to (1) build more principled generalizations of common Euclidean operations used in geometric machine learning models as well as to (2) enable general manifold density learning in contexts that require it. Said contexts include theoretical physics, robotics, and computational biology. I will cover two of my papers that fit into (1) above, namely the ICML 2020 paper "Differentiating through the Fréchet Mean" as well as the ICML 2022 workshop paper "Riemannian Residual Neural Networks", presented at the second annual "Topology, Algebra, and Geometry in Machine Learning" workshop. I will also cover two of my papers that fit into (2) above, namely the NeurIPS 2020 paper "Neural Manifold Ordinary Differential Equations" and the NeurIPS 2021 paper "Equivariant Manifold Flows." Finally, I will briefly discuss the implications of the work done as well as potential directions for future work.

Image compression algorithms are the basis of media transmission and compression in the field of image processing. Decades after their inception, algorithms such as the JPEG image codec continue to be the industry standard. A notable research topic gathering momentum in the field of compression is deep learning (DL). This paper explores the opti- mization of DL models for ideal image compression and object detection (OD) applications.The DL model to be optimized is based upon an existing compression framework known as the CONNECT model. This framework wraps the traditional JPEG image codec within two convolutional neural networks (CNNs). The first network, ComCNN, focuses on com- pressing an input image into a compact representation to be fed into the image codec. The second network, RecCNN, focuses on reconstructing the output image from the codec as similarly as possible to the original image. To enhance the performance of the CONNECT model, an optimization software called Optuna wraps the framework. Hyperparameters are selected from each CNN to be evaluated and optimized by Optuna. Once the CONNECT model produces ideal results, the output images are applied to the YOLOv5 OD network.This paper explores the impact of DL hyperparameters on image quality and compres- sion metrics. In addition, a detection network will provide context to the effect of image compression on computer vision applications.

There is increasing understanding that many complex systems of interest - everything from the global economy, to social group dynamics, to biochemical processes in the brain - require holistic modelling, rather than the consideration of units of the population in isolation. Network science techniques, which commence by viewing the system as a set of vertices or nodes (the units of the population, e.g. individual people) and edges (the relationships between them, e.g. friendship), are one popular approach to do so. Naturally, such complex systems express a wide array of important properties that we ought to account for when modelling them, beyond simply the presence or absence of a particular relationship. Most pertinently for this work, they evolve over time - i.e. they are 'dynamic' - and the units of the population may have distinct properties, or attributes, which further differentiate them from each other. We define any such extra information we might possess outside of the simple node/edge paradigm to be 'metadata'. Despite the potential utility of such metadata, it is only quite recently that methods have begun to jointly model both network and metadata together. In this thesis, we provide a new class of models that do so - specifically, with the purpose of finding groups in networks that change over time. We describe distinct versions of this class of models that allow the networks to be weighted and directed, as well as avoid the potential issue of placing nodes with similar degrees in the same group. In addition to elaborating such models, we derive novel requirements for the efficient detectability of groups given the presence of metadata - and in the process explain why a recent paper which claims to do the same for a similar static model is flawed. The inference method we leverage to investigate detectability is also highly scalable, and we further accelerate the process by proposing both a 'greedy' scheme, and a recursive procedure that effectively provides a top-down hierarchy of the network groups. We conclude by using our models as one component of a larger method, that provides an entirely novel means of estimating the influence of an author. We use a causal framing of the problem that to our knowledge has not previously been explored in this context, and depends upon recent ideas from the causal inference literature.

With the current rate of research activities, it is widely accepted that scientists face a challenge of keeping up-to-date with new findings, even within a sub-field of a discipline. This difficulty extends to methods that have been used in the research. Understanding reported methods gives us confidence that the findings have resulted from an appropriate, rigorous and sound scientific process. However, the modern dynamic of science is also characterised with ever-changing methods, so scientists need to be able to learn about new ones and identify the common or most appropriate methods to use in a given situation. One of the best sources of information about methods is the scientific literature. In this thesis, we developed a computational model to automatically represent the text that describes reported methods as an abstract method workflow. We focus on computational sciences, which centre on data processing. Specifically, we consider data events as a representation of processes and changes that happen to data. A data event contains the main components of each step in computational experiments, such as input/output data, processes and operations on data, databases where the data is stored and software and tools that are used in these processes. An abstract method workflow then models relationships between data events, ordering them in a way that represents the methodology as reported in the literature. This thesis introduces ODNoRFlow, a text mining method that extracts and represents an abstract method workflow from a Methods section of a publication. It relies on a hybrid text mining approach (ODNoR) that combines machine learning and a rule-based method to recognise data event components, normalise them to existing ontologies and identify the links and relations between them. Specifically, we fine-tuned a pre-trained transformer model (BioBERT) to extract mentions of data and operations, and used an existing named entity recognition system (bioNerDS) to extract software and database mentions. Mentions were normalised to the EDAM ontology. We used a combination of syntactic rules and a pre-trained attention-based BiLSTM model to identify relations and links between components, and considered whether an automated discourse analysis tool can be used to improve the outcomes. We used the microarray analysis literature as a case study to demonstrate the feasibility of the proposed approaches. At the data event level, the approach achieved F-scores for the identification and normalisation of components between 78% (for data) and 92% (for operations), whereas the relationship extraction F-scores were between 62% and 92.5%. At the workflow level, we manually analysed automatically reconstructed workflows from 25 papers, with the F-score between 61% and 93.5%. We also applied ODNoRFlow to a large corpus of the microarray analysis literature to identify and analyse the distribution of data events components, the differences in their usage and the associations between them. Overall, the thesis provides a new computational framework that contributes to the automated extraction, representation and analysis of methods used in the computational biology literature.

This portfolio of compositions uses novel biometrics from gestural interfaces called Myo armbands to compose five pieces of electroacoustic music within a continuum of spatial environments, ranging from the real to the virtual. The portfolio achieves this by processing biometric data from the Myo armbands within developed software (using Max 8), applying machine learning to such data (via Wekinator) and mapping predictive outputs to audiovisual materials across the spatial continuum. The results of this approach show how unique compositional affordances can be created when using biometrics within different spatial environments and how implementing machine learning methods can help to create such opportunities. The outcomes are presented via five composed portfolio works, three published papers within the scientific community, and numerous music performances navigating the overlapping areas of music composition and computer science. By using novel biometrics to compose music within a continuum of space, this interdisciplinary research topic contributes original knowledge to both the fields of music composition and computer science (human-computer interaction).

Bayesian methods constitute a popular approach to perform statistical inference and predict phenomena of interest. Surely, part of the popularity of the Bayesian paradigm can be linked to their intuitive core idea: to take advantage of the user's prior knowledge and integrate it in the statistical procedure. The result of this synergy is the posterior distribution, focal point of Bayesian inference and instrument to quantify the uncertainty of the estimation. This thesis collects the work done as a research student on statistical models, built with the tools of Bayesian nonparametrics, to describe power-law distributed data. The motivation of the proposed models are to be found in two different fields of application: complex networks and privacy assessment. After the introduction, in the first half of the thesis I concentrate on the proposal and analysis of models for networks, the mathematical objects that describe relations among entities by representing them as nodes connected through links. Power-laws usually appear in real-world networks as the distribution function of the degrees, namely the number of links of the nodes. The two pieces of work presented belong to the graphex process framework, a flexible generating process which allows to mimic empirically observed networks characteristics. Chapter 2 fits into this framework and extends the original proposal of Caron and Fox (2017) to provide a novel modelling approach to describe sparse networks with spatial structure or other covariates. An approximate inference strategy is provided and tested on simulated data. The paper presented in chapter 3 casts light on the asymptotic properties of networks generated under the graphex process, proving the desirable properties of sparsity, power-law degree distributions, clustering and two central limit theorems. The second half of the thesis is devoted to the development of statistical methods to quantify the risk of disclosure, which arises whenever datasets with records of individuals are published: an intruder could match the data with prior information and disclose the identity, and therefore the sensitive features, of a person. Chapter 4 develops an estimator to quantify this risk using the Pitman-Yor process, a popular prior on probability distributions that has a distinguishable power-law tail. A closed form posterior distribution of the estimator is provided in a convenient mixture representation, and experiments on both simulated and real data show the effectiveness of the method. Chapter 5 deals with the estimation of the same risk under no distributional assumptions, using a fully nonparametric method. The estimator is extremely easy to understand, fast to compute and has provable guarantees of optimality. Chapter 6 concludes the thesis with a final summary and some proposals to extend the current work towards new research directions.

This paper develops a possible explanation for a facet of visual processing inspired by the biological brain's mechanisms for information gathering. The primary focus is on how humans observe elements in their environment and reconstruct visual information within the brain. Drawing on insights from diverse studies, personal research, and biological evidence, the study posits that the human brain captures high-level feature information from objects rather than replicating exact visual details, as is the case in digital systems. Subsequently, the brain can either reconstruct the original object using its specific features or generate an entirely new object by combining features from different objects, a process referred to as "Imagination." Central to this process is the "Imagination Core," a dedicated unit housing a modified diffusion model. This model allows high-level features of an object to be employed for tasks like recreating the original object or forming entirely new objects from existing features. The experimental simulation, conducted with an Artificial Neural Network (ANN) incorporating a Convolutional Neural Network (CNN) for high-level feature extraction within the Information Processing Network and a Diffusion Network for generating new information in the Imagination Core, demonstrated the ability to create novel images based solely on high-level features extracted from previously learned images. This experimental outcome substantiates the theory that human learning and storage of visual information occur through high-level features, enabling us to recall events accurately, and these details are instrumental in our imaginative processes.

According to security.org, as of 2023, 65% of credit card (CC) users in the US have been subjected to fraud at some point in their lives, which equates to about 151 million Americans. The proliferation of advanced machine learning (ML) algorithms has also contributed to detecting credit card fraud (CCF). However, using a single or static ML-based defense model against a constantly evolving adversary takes its structural advantage, which enables the adversary to reverse engineer the defense's strategy over the rounds of an iterated game. This paper proposes an adaptive moving target defense (MTD) approach based on deep reinforcement learning (DRL), termed RESONANT to identify the optimal switching points to another ML classifier for credit card fraud detection. It identifies optimal moments to strategically switch between different ML-based defense models (i.e., classifiers) to invalidate any adversarial progress and always stay a step ahead of the adversary. We take this approach in an iterated game theoretic manner where the adversary and defender take turns to take their action in the CCF detection contexts. Via extensive simulation experiments, we investigate the performance of our proposed RESONANT against that of the existing state-of-the-art counterparts in terms of the mean and variance of detection accuracy and attack success ratio to measure the defensive performance. Our results demonstrate the superiority of RESONANT over other counterparts, including static and naïve ML and MTD selecting a defense model at random (i.e., Random-MTD). Via extensive simulation experiments, our results show that our proposed RESONANT can outperform the existing counterparts up to two times better performance in detection accuracy using AUC (i.e., Area Under the Curve of the Receiver Operating Characteristic (ROC) curve) and system security against attacks using attack success ratio (ASR).

Out-of-distribution (OOD) detection aims to identify test examples that do not belong to the training distribution and are thus unlikely to be predicted reliably. Despite a plethora of existing works, most of them focused only on the scenario where OOD examples come from semantic shift (e.g., unseen categories), ignoring other possible causes (e.g., covariate shift). In this paper, we present a novel, unifying framework to study OOD detection in a broader scope. Instead of detecting OOD examples from a particular cause, we propose to detect examples that a deployed machine learning model (e.g., an image classifier) is unable to predict correctly. That is, whether a test example should be detected and rejected or not is “model-specific”. We show that this framework unifies the detection of OOD examples caused by semantic shift and covariate shift, and closely addresses the concern of applying a machine learning model to uncontrolled environments. We provide an extensive analysis that involves a variety of models (e.g., different architectures and training strategies), sources of OOD examples, and OOD detection approaches, and reveal several insights into improving and understanding OOD detection in uncontrolled environments.

In recent years, Internet-of-Things (IoT) devices generate a large amount of personal data. However, due to the privacy concern, collecting the private data in cloud centers for training Machine Learning (ML) models becomes unrealistic. To address this problem, Federated Learning (FL) is proposed. Yet, central bottleneck has become a severe concern since the central node in traditional FL is responsible for the communication and aggregation of mil- lions of edge devices. In this paper, we propose Scalable Federated Learning via Distributed Hash Table Based Overlays for network (Scaled) to conduct multiple concurrently running FL-based applications over edge networks. Specifically, Scaled adopts a fully decentral- ized multiple-master and multiple-slave architecture by exploiting Distributed Hash Table (DHT) based overlay networks. Moreover, Scaled improves the scalability and adaptability by involving all edge nodes in training, aggregating, and forwarding. Overall, we make the following contributions in the paper. First, we investigate the existing FL frameworks and discuss their drawbacks. Second, we improve the existing FL frameworks from centralized master-slave architecture by using DHT-based Peer-to-Peer (P2P) overlay networks. Third, we implement the subscription-based application-level hierarchical forest for FL training. Finally, we demonstrate Scaled's scalability and adaptability over large scale experiments.

We routinely encounter too much information in the form of social media posts, blogs, news articles, research papers, and other formats. This represents an infeasible quantity of information to process, even for selecting a more manageable subset. The process of condensing a large amount of text data into a shorter form that still conveys the important ideas of the original document is text summarization. Text summarization is an active subfield of natural language processing. Extractive text summarization identifies and concatenates important sections of a document sections to form a shorter document that summarizes the contents of the original document. We discuss, implement, and compare several unsupervised machine learning algorithms including latent semantic analysis, latent dirichlet allocation, and k-means clustering. ROUGE-N metric was used to evaluate summaries generated by these machine learning algorithms. Summaries generated by using tf-idf as a feature extraction scheme and latent semantic analysis had the highest ROUGE-N scores. This computer-level assessment was validated using an empirical analysis survey.

Bridge inspection and infrastructure inspection are critical steps in the lifecycle of the built environment. Emerging technologies and data are driving factors which are disrupting the traditional processes for conducting these inspections. Because inspections are mainly conducted visually by human inspectors, this paper focuses on improving the visual inspection process with data-driven approaches. Data driven approaches, however, require significant data, which was sparse in the existing literature. Therefore, this research first examined the present state of the existing data in the research domain. We reviewed hundreds of image-based visual inspection papers which used machine learning to augment the inspection process and from this, we compiled a comprehensive catalog of over forty available datasets in the literature and identified promising, emerging techniques and trends in the field. Based on our findings in our review we contributed six significant datasets to target gaps in data in the field. The six datasets comprised of structural material segmentation, corrosion condition state segmentation, crack detection, structural detail detection, and bearing condition state classification. The contributed datasets used novel annotation guidelines and benefitted from a novel semi-automated annotation process for both object detection and pixel-level detection models. Using the data obtained from our collected sources, task-appropriate deep learning models were trained. From these datasets and models, we developed a change detection algorithm to monitor damage evolution between two inspection videos and trained a GAN-Inversion model which generated hyper-realistic synthetic bridge inspection image data and could forecast a future deterioration state of an existing bridge element. While the application of machine learning techniques in civil engineering is not wide-spread yet, this research provides impactful contribution which demonstrates the advantages that data driven sciences can provide to more economically and efficiently inspect structures, catalog deterioration, and forecast potential outcomes.

Diabetes is one of the highest causes of death in the world. Diabetes is caused when theblood glucose level is too high in the body. Gradually, high blood glucose leads to heartdisease, stroke, eye, and foot problems. To prevent the dreadful effects among people,early detection is required that would lead to proper medical treatment and change inlifestyle. Therefore, with the rise of machine learning we can predict if a patient hasdiabetes or not. Furthermore, we will integrate the trained model to a web applicationthat will connect the model to generate predictions in real-time considering factorsresponsible for diabetes like body mass index (BMI), age, insulin, etc. In this paper, we areusing the Pima Indian dataset that is originally from the National Institute of Diabetes,Digestive and Kidney Diseases for diabetes prediction model design using machinelearning. The proposed system in this paper is the Soft Voting ensemble classifier. Thealgorithm with the best accurate result was used in making predictions. This model wasdeployed to the web using flask (a python framework), it takes inputs from the user tomake predictions. This model is implemented using python programming language andflask (a web base framework) hosted in GCP. Soft Voting ensemble classifiers evenperform better than other classifiers with an accuracy of 91.55% which is quite promisingconsidering the other classification models in the literature for this problem.

This dissertation consists of three essays studying the cross-section of asset returns using new empirical approaches. In Essay 1, I forecast firm-level expected growth via machine learning and study the relationship between the expected growth and future stock returns. In Essay 2, I study the drivers of stocks’ migration across BE/ME portfolios and how the migration affects the value premium. In Essay 3, I study the protection effect on corporate bonds issued by state-owned enterprises.Essay 1: Expected growth is an important firm fundamental variable but is unobservable and difficult to estimate. In this study, I apply machine learning (ML) to forecast growth at the firm level. Compared with conventional linear regression, ML models produce more accurate forecasts out of sample. Among the ML models, non-linear models perform the best. In particular, the gradient boosting regression reduces the mean and median forecast errors by 9.52% and 20.95%, respectively, relative to linear regression. Moreover, the ML-based growth measure exhibits sensible properties at the firm level, in subsamples, and at the aggregate level. Consistent with theory, the firm-level expected growth predicts cross-sectional stock returns positively, controlling for past growth. Firms in the top decile of the expected change in growth beat those in the bottom decile by an average of 0.56% per month (t = 4.49), which cannot be explained by most benchmark models. The aggregate expected change in growth also forecasts stock market returns positively over a period ranging from 24 months to 60 months.Essay 2: In this paper, we revisit the migration of stocks across BE/ME portfolios contributing to the value premium. We seek to address an unanswered question in Fama and French (2007) regarding the underlying drivers of this migration. We investigate whether the movement of stocks is affected by changes in fundamental characteristics or is merely a result of random fluctuations. Our analysis reveals significant variation in stocks’ transition frequencies over time, with notable increases during recessions. Through a novel approach, we demonstrate that these transitions are not random but rather driven by changes in fundamental characteristics and firm-specific risks. Specifically, we identify operating leverage and investment as the primary drivers behind stocks’ movements across BE/ME portfolios.Essay 3: We model an implicit government guarantee as a partial protection for corporate bonds issued by state-owned enterprises (SOEs) and present pervasive evidence of its existence in the Chinese corporate bond market. The protection effect is stronger among poorly rated SOE bonds than for highly rated SOE bonds – for newly issued corporate bonds, the average guarantee premium is 72 basis points (bps) for AAA bonds while it is 166 bps for AA- (lowest issuance rating) bonds. We find the guarantee premiums are higher for bonds issued by central government-sponsored SOEs and by issuers from strategically important sectors, raising concerns of the potential too-interconnected-to-fail problem.

Machine learning has achieved unprecedented levels of success recently, in the areas of language processing and modeling, image and video classification and generation, and recommendation and dynamic pricing systems. The control of dynamic systems has also benefited from these advancements in learning, particularly in the areas of reinforcement learning for tasks such as robotic navigation and control of nuclear fusion processes. We wish to study learning in another area where it can naturally be applied: adaptive control systems. These systems must estimate and identify uncertainties in the plant inorder to apply their adaptive control laws. We study the areas of stochastic contraction and convex projection, persistency of excitation, and function approximation, with an eye towards this application. The first part of the thesis is motivated by the problem of quantitatively bounding the convergence of adaptive control methods for stochastic systems to a stationary distribution. Such bounds are useful for analyzing statistics of trajectories and determining appropriate step sizes for simulations. To this end, we extend a methodology from (unconstrained) stochastic differential equations (SDEs) which provides contractions in a specially chosen Wasserstein distance. This theory focuses on unconstrained SDEs with fairly restrictive assumptions on the drift terms. Typical adaptive control schemes place constraints on the learned parameters and their update rules violate the drift conditions. To this end, we extend the contraction theory to the case of constrained systems represented by reflected stochastic differential equations and generalize the allowable drifts. We show how the general theory can be used to derive quantitative contraction bounds on a nonlinear stochastic adaptive regulation problem. ivThe second part of the thesis defines geometric criteria which are then used to establish sufficient conditions for persistency of excitation with vector functions constructed from single hidden-layer neural networks with step or ReLU activation functions. We show that these conditions hold when employing reference system tracking, as is commonly done in adaptive control. We demonstrate the results numerically on a system with linearly parameterized activations of this type and show that the parameter estimates converge to the true values with the sufficient conditions met. The third part of the thesis studies function approximation. Classical results in neural network approximation theory show how arbitrary continuous functions can be approximated by networks with a single hidden layer, under mild assumptions on the activation function. However, the classical theory does not give a constructive means to generate the network parameters that achieve a desired accuracy. Recent results have demonstrated that for specialized activation functions, such as ReLUs, high accuracy can be achieved via linear combinations of randomly initialized activations. These recent works utilize specialized integral representations of target functions that depend on the specific activation functions used. This paper defines mollified integral representations, which provide a means to form integral representations of target functions using activations for which no direct integral representation is currently known. The new construction enables approximation guarantees for randomly initialized networks using any activation for which there exists an established base approximation which may not be constructive. We extend the results to the supremum norm and show how this enables application to an extended, approximate version of (linear) model reference adaptive control.

Software models are used to analyze and understand the properties of the system, providing stakeholders with an overview of how the system should work before actually implementing it. Such models are usually created informally, such as drawing sketches on a whiteboard or paper, especially during the early design phase, because these methods foster communication and collaboration among stakeholders. However, these informal sketches must be formalized to be useful in later applications, such as analysis, code generation, and documentation. This formalization process is often tedious, error-prone, and time-consuming. In an effort to avoid recreating formal models from scratch, this thesis presents SkeMo, a sketch-based software modeling tool. SkeMo is built on a CNN-based image classifier using 3000 input sketches of class diagram components and integrated into the functionality of an existing web-based model editor, the Instructional Modeling Language (IML), with a newly implemented touch interface. SkeMo was evaluated using a ten-fold cross-validation to assess the image classifier and through a user study involving 20 participants to collect metrics and feedback. The results demonstrate the promising potential of sketch-based modeling as an intuitive and efficient modeling practice, allowing users to quickly and easily create models to design complex software systems.

As the fields of machine learning, natural language processing, and big data become increasingly important throughout various industries, including finance, it becomes crucial to evaluate how they are being utilised and understand the motivations behind the recommendations of algorithms, in order to make sound decisions based on this information. In finance, particularly, being able to create concise and easy to comprehend tools for understanding machine learning models is extremely beneficial. These tools could serve multiple purposes, including allowing managers to better explain to and convince investors of machine learning based trading strategies. Decision-makers themselves might also be able to act according to the information provided by these tools, which could assist in, for example, making more ethically informed decisions. Thus, this type of research not only has implications in the areas of machine learning, Natural Language Processing (NLP), and finance but also in fields like AI explainability, model interpretability and Responsible Research and Innovation. The central problem this thesis addresses is whether machine learning methods can be used to mine representative sentences from financial text that are able to capture the majority of sentiment in a full document with only a few sentences. These mined sentences are referred to as 'justifications', and the process has been called 'justification mining'. The full documents used here are data taken from 10-K filings. Before examining justification mining, however, transfer learning methods suitable for training data annotated at a different level than testing data (sentence- versus document-level) are assessed. The purpose of this is to address a problem that often occurs in NLP research, where no annotated training data perfectly suited to the research is readily available. Therefore, methods must be created to make use of what is available, in this case sentence-annotated training data being employed to train classifiers for prediction on document-level testing data. These transfer learning methods are then employed in the next step, which focuses on justification mining. The process of justification mining itself is first developed using transformer models to encode embeddings and, then, using clustering algorithms and cosine similarity to extract or mine justifications. This process is evaluated by comparing sentiment from mined justifications to sentiment from full documents, in order to assess the ability of justification mining to capture or summarise sentiment. It is also evaluated by correlating predicted sentiment from mined justifications and full documents to future stock returns, to gauge whether justification mining offers any benefit in identifying signals in the data for financial purposes. Little work has been done previously that evaluates the results of financial sentiment analyses in this way. In the NLP domain, the best way of extracting aggregate justifications for sentiment is still an open question. Moreover, few research papers attempt to apply transfer learning from lower-level data to entire documents. Nor are 10-K filings widely studied in this context, as they are difficult to parse. The methods created in this thesis might offer a novel means of providing information that can assess the motivations behind sentiment analyses. In the transfer learning process, feature engineering and preprocessing steps were modified to obtain accuracies up to 0.903. For justification mining, considering statistically significant (p≤0.05) correlations of |r| > 0.2, sentiment from mined justifications more often correlated with future stock returns than did full document sentiment. Although the correlations (r) were somewhat weak overall, they could potentially be combined with traditional alpha signals to enhance these signals (see Section 11 for further discussion). Moreover, high degrees of similarity were found between aggregated sentiment from mined justifications and full documents, with similarity scores up to 0.9999, supporting the efficacy of justification mining in capturing full document sentiment. In these evaluations, transformer models performed better in numericising text for input into ML models than traditional approaches like Bag of Words. In fact, every statistically significant sentiment to stock return correlation bar one, as well as every mined justification to full document similarity and correlation score above 0.7, used transformer models for numericisation. These results imply that justification mining might be successful in eliminating sentiment noise in financial data, as well as in capturing the majority of sentiment from a full document. Moreover, transformer models might provide an advantage over traditional approaches like Bag of Words for numericisation of text. Mined justifications, themselves, provide an easily interpretable and presentable means of explaining the output of sentiment analysis and have numerous uses, including identifying the driving factors behind sentiment in a financial document, which could be helpful for making more ethically informed decisions or building investor trust in the methodology of sentiment analysis algorithms.

In this thesis we develop several state-of-the-art generative modelling-based approaches for a variety of supervised, unsupervised and private learning problems. In the (almost) supervised domain, we tackle the problems of treatment effect estimation, imputation and feature selection. For treatment effect estimation we begin by developing a GAN-based approach that generates the ``missing'' counterfactuals, which enables learning a fully supervised model. In SCIGAN, we then go on to adapt this method to the continuous-intervention setting, introducing novel generator and discriminator architectures to handle the continuous nature of the treatments. For imputation, we introduce GAIN, a GAN-based imputation approach that maximally leverages both the adversarial and generative nature of GANs to learn to impute missing data according to the ground truth distribution while simultaneously allowing for multiple imputation to improve reliability and robustness of the imputed data. Finally, for feature selection, we introduce KnockoffGAN, a GAN-based knockoff generation procedure which again leverages the adversarial nature of GANs to learn the appropriate ``knockoff'' distribution. In using a GAN, we are able to learn to generate knockoff non-parametrically and thus reliable performance on non-Gaussian data. In the domain of private learning we address the problem of private synthetic data generation. This challenge has three key components: (1) generating faithful synthetic data; (2) ensuring privacy; and (3) measuring the quality of the generated data. In PATEGAN, we tackle all 3 of these problems by proposing a GAN-based approach to generating differentially private synthetic data. In addition, we introduce a new synthetic data metric, the Synthetic Ranking Agreement, which measures how well machine learning research on the synthetic data represents the same research on the real data. In the final paper presented, we present a new differential privacy building block - a differentially private classification/prediction algorithm, that builds upon the subsample-and-aggregate paradigm. The new method affords better privacy and utility than subsample-and-aggregate and can be used in place of PATE in the PATEGAN framework.

This PhD thesis develops a typology of common Torah codices from the Cairo Genizah. As many of these codices have a non-standard form of Tiberian vocalisation (NST) the thesis also develops a typology of this vocalisation present in the corpus. These codices lack the full features associated with exemplary Bibles (having at least two columns of text, written on parchment, with Masoretic notes), and NST is the use of Tiberian Hebrew vowel signs and orthography which deviates from standard Tiberian vocalisation (ST), the main reference point for which is Codex Leningradensis. Previous research only studied a small number of NST MSS in order to contextualise them within the development of standard Tiberian, and they pay little attention to codicology. Likewise, Bibles which do not fit a certain level of codicological sophistication have never been systematically assessed (for a significant corpus). This thesis takes a representative sample of ~ 1,500 Torah fragments (~1,800 if numbers from a preliminary case study are included in the count) and subjects them to a holistic statistical, codicological, and linguistic analysis. My thesis aims to establish whether the codicological and linguistic features of these Torahs fall into meaningful, correlating patterns reflecting regional variations in codicological style and vowel sign usage. I store the MS data in a MySQL database which I built, and I employ statistical machine learning algorithms to find naturally occurring patterns in this data. These patterns are then analysed for their linguistic and codicological meaningfulness. Each chapter analyses a subsection of MSS: (a) parchment Torahs with 2 columns, but no Masoretic notes; (b) single-column parchment Torahs; (c) paper Torahs written in scribal hands; (d) Torahs written by children and non-scribe laymen. This research has uncovered some crucial insights for our knowledge of the popular writing of the Torah in the Masoretic and immediate post-Masoretic period. First, common Torahs show rich diversity of regional codicological styles. Second, their NST falls into different regional patterns; these can reflect a mixture of phonological language contact with dialectal Arabic, Aramaic influence from Targums, and an imperfect application of Standard Tiberian vocalisation rules. Third, these codicological and NST patterns tend to complement each other, so that NST and codicology generally correlate, region-by-region, throughout the world of the Genizah.

Probabilistic programming is a paradigm that enables us to efficiently write probabilistic models as program code that we can sample, infer underlying parameters and predict outcomes based on complete or incomplete observations. Naturally, stochastic simulators, a special sub-class of simulators containing random variables, internal inference procedures, and the simulation of observations, are structurally rich probabilistic models. However, most simulators are not written in the probabilistic programming paradigm, as they are written in arbitrary programming code. This means that it is challenging to automatically update the variables in these simulators to account for observations from conducted experiments, which limits the simulators' use. Furthermore, there are two components to a probabilistic programming system i) the language and compilation procedure, ii) the inference procedures. These components can limit our ability to compile particular classes of probabilistic models, such as models that contain mixtures of parameter types, due to restrictions in the expressiveness of the language. Restrictions in the expressivity of the language can also inhibit our ability to generate efficient inferences, as this naturally influences the design of the probabilistic programming system and the set of available inference backends. Creating probabilistic programming systems that are expressive enough for different probabilistic models leads to the creation of many different probabilistic programming systems, which is inefficient - it would be more efficient if we could repurpose existing probabilistic programming systems. In this thesis, we develop three pieces of original work through four papers. The first piece of work describes how to extend differentiable first-order probabilistic programming systems to perform statistically correct and computationally efficient inference on models with mixtures of continuous and non-continuous parameters, without having to modify the underlying language, or develop an entirely new probabilistic programming system. The second describes how to translate real-world stochastic simulators written in arbitrary program languages to probabilistic programming systems. And finally, in the third piece of work, we develop two new Bayesian inference schemes to make inference more computationally and statistically efficient in nested models, models that contain probabilistic programs, within probabilistic programs, which arise in many real-world stochastic simulators.

For efficient long-term storage and use of bacteria for environmental applications, understanding and identifying desiccation resistance in bacteria is key. In the past, desiccation tolerance was a common way of characterizing bacteria, so there is much data on the desiccation tolerance of a wide range of bacterial species. Since the advent of transcriptomics, multiple papers have been published on the expression level of genes during desiccation stress. Additionally, many reviews have described mechanisms and genes relevant to desiccation tolerance in bacteria, but an overarching framework for the prediction of desiccation survival in bacteria is lacking. Model building based on data collected from the literature has been used to successfully predict aerobic vs anaerobic phenotype, enzyme function and substrate specificity (Robinson et al., 2020; Jabłońska et al, 2019) Building on this wealth of previous research, machine learning was used to create a robust model that predicts desiccation tolerance given bacterial genomes. Validation and accuracy of the machine learning model was tested using a desiccation assay carried out over three months. To build the model, a literature review was conducted to find genes that were upregulated greater than two-fold during desiccation stress in bacteria. From the review, 2609 genes from 11 papers were found and condensed to 1082 non-homologous and non near-zero variance genes. A second literature search was conducted to identify bacterial species with a known desiccation response, either tolerant or sensitive, and a publicly available genome. Thirty-five desiccation tolerant and 33 desiccation sensitive genomes were chosen and then queried for the previously curated desiccation upregulated genes list. Approximately 176,800 genes were analyzed, and genes with non-zero variance were removed. The remaining 75,982 genes are included in the model (Rogozin et al., 2002). A random forest supervised machine learning approach was used to create a preliminary model for desiccation resistance. The genomes were split into 80% training data and 20% test data and the model was run 100 times with different seeds, 10-fold cross validation, and three repeats. The average accuracy for the 100 iterations of the model was 0.898 ± 0.0266, indicating the model could accurately predict the desiccation phenotype of the testing data 89.8% of the time. The experimental validation of the desiccation model looked at the viability of 28 bacteria, seven with documented desiccation phenotypes and 21 bacteria with no known desiccation phenotype. For all organisms tested the model had an accuracy of 0.75 demonstrating good model performance.

Image compression is a foundational topic in the world of image processing. Reducing an image's size allows for the image to be stored in less memory and speeds up the processing and storage time. In addition, deep learning (DL) has been a featured topic. This paper seeks to find a model that uses DL for optimal image compression.There are several image codecs that already are used for image compression. The framework that is designed in this paper does not focus on eliminating these codecs; rather, it uses a method that incorporates standard codecs. The image codec is wrapped with two convolutional neural networks (CNNs). The first network, ComCNN, has the goal of compressing an image into an optimal compact representation that can be passed into an image codec for maximum compression. The second network,, RecCNN, has the goal of reconstructing the decoded compact representation of the image into an output that is as similar to the original image as possible. By continuing to use tradition image codes such as JPEG and JPEG2000, the process is standardized while still producing optimal results.The paper gives an overview of image compression, machine learning, and different quality and compression metrics that determine the success of the network. In addition, the model is described in great detail, and results with different parameters and data types are presented.

This dissertation presents three papers demonstrating how integrating graph (network) and language (text) data in machine learning models can enhance computational social science models. These two types of data are ubiquitous across many contexts in which computational social scientists work (e.g., social media platforms, online markets, and the Web as a whole). Relatively little research has analyzed how to model network and text data together at scale, partly since models for these data are often computationally expensive but also because statistical models for them require expert-driven decisions on feature engineering and how they are related within models. The first paper in this dissertation combines node and text embeddings in a downstream classification model to study mental health dynamics on Reddit. The second paper cascades knowledge graph classifications to a text clustering model to study how demographic confounding causes extreme instances of lifestyle politics using aggregated Facebook interest data. Finally, the third paper uses graph and language data from Amazon to study the spread of political and lifestyle polarization in the large online market and tests how network and morality features explain the presence of lifestyle polarization. Together, the three studies show how integrating graph and language data in machine learning models can facilitate computational social science not only by improving such models’ power, efficiency, and ease of use but also by allowing us to test new hypotheses and explain black box models. The conclusion contextualizes findings for academia and industry.

We present prediction methodology motivated by problems arising from a large dataset containing football match information that can be applied to a wide range of areas. Much of the thesis has been driven by the prediction of goals, however we also attempt to predict observable features such as the number of shots as well as unobservable features such as the newly popular expected goals (xG) metric. This thesis can be found in three main parts. We begin by recounting a summary of pre-existing research in the area of football modelling that motivates the concepts used in the forthcoming chapters. We also describe the data available to test and evaluate our methodology, as well as outlining plans to contribute papers to the academic community. The middle part focuses on new time series methodology with extensions to football match results found in the later sections. We first introduce the reader to some basic time series concepts before proceeding to the novel material. We provide a new technique for the estimation of nonstationary parameters in SARUMA models using the nonlinear transformation obtained from the Levinson-Durbin algorithm. By defining a set of polynomial coefficients in terms of autocorrelations we provide a stable and compact parameter space on which the estimation can occur. Then we provide a new class of multivariate time series models that are used to predict the number of goals scored by each team in a football match. The model assumes that each margin follows a conditional Poisson distribution (a common assumption for the goal scoring processes in a football match) and includes an intensity feedback mechanism that allows past form to influence the expected (future) performance. We study asymptotic properties of the model and compare different model specifications to obtain the best predictive model. Finally, we combine machine learning and statistical techniques to provide new methodologies for prediction problems. We again start with a foundation chapter containing basic machine learning concepts and algorithms. We then outline our own contributions as we investigate the covariates of a goal prediction model and their subsequent effect on match predictions. We start by creating a new class of gradient boosted models for predicting goals scored based on match statistics such as shots, passes, and an xG metric. We train several machine learning algorithms for the calculation of the metric and choose what we believe to be the most appropriate. We then gather the match statistics, which we call features, and use the boosting model framework to build the models. We analyse the importance of the included features in the best model and acknowledge that real-time match data are used to predict the result. Whilst suitable for studying feature important feature, it would be inappropriate for predicting the future results. Therefore we study a new approach to feature projection, which uses a state space approach to model the features directly and produce projections for the next match ahead of time. These projections are utilised in a variety of machine learning algorithms to assess whether the more sophisticated technique produces a higher quality of match prediction.

Accurately assessing workers’ decision quality is fundamental for management, and the efficiency of expert and crowd-sourcing markets. This paper establishes novel ML and AI methods to accurately evaluate workers’ decision accuracy and bias with scarce ground truth (GT or gold standard GS) data, and to further improve accuracy assessment through costly effectively acquiring GT if given an acquisition budget. Without the proposed methods, assessing workers’ decision quality typically requires GT data to compare with workers’ noisy decisions. However, GT is often prohibitively costly to acquire for even a small fraction of each worker’s decisions. For example, physicians may determine a diagnosis and initiate a treatment, yet the correct decision, such as the one that can be established by a panel of physicians. Consequently, in practice, there is often poor transparency regarding physicians’ decision quality. In my dissertation, I collaborating with my coauthors developed the groundwork for achieving scalable and inexpensive assessments of workers’ decision accuracy and bias. The empirical results show that the decision accuracy assessment with very limited GT improves the best available approach by 60% to 93%; my bias assessment produces either comparable to or outperforms the commonly used existing approach; my cost-effective GT acquisition strategy applied in Amazon Mechanical Workers’ accuracy assessment achieves the same performance only using 1/3 of the GT or improve the assessment by 24%. All proposed methods have significant implications in many impactful domains including health care, fraud detection, fact checking, and online labor markets. The methods proposed in this dissertation address the problem of estimating workers’ decision accuracy and bias from historical data with scarcely available ground truth, and achieve the state of the art performance. This dissertation lays the groundwork towards increasing transparency in workers’ (sources’) decision quality.

Abstract: Emergency departments (ED) around the globe rely on triage to sort which patients can wait safely for care when the demand for service outstrips the resources available. The triage process is typically performed by experienced emergency nurses and can be the earliest documented assessment for patients receiving hospital-based care. Because the ED is the most common entry point for patients, triage documentation, specifically triage narratives, are nearly ubiquitous for hospitalized patients. The data collected during triage can be structured (information such as age and sex) or unstructured (in the case of free-text narratives). In Canada, ED data are reported as a national dataset, much of the reported data is collected at triage, but little of the national dataset reflects nurses own work and the triage narratives are not included. The goal of this paper-based thesis is threefold: first, to describe the role of ED nurse in data collection; second, to provide a scoping review protocol for examining triage narratives; and, to describe how triage narratives have been used in the academic literature. The first portion of this thesis reviews the systems used in Canada for ED data collection. This chapter highlights how little nursing data is collected and proposes that national organizations need to demonstrate that there is utility in collecting these triage data. It calls on the National Emergency Nurses Association to take concrete first steps to press for ED nurse data to be included in national ED data registries. In the second phase of this thesis, I provide an outline and rationale for how to review one specific component of triage nurse data: the unstructured narrative. The narratives, which have been included in other reviews of clinical documentation have show promise for a number of uses, but they have not been reviewed as a unique form of documentation. The scoping review protocol outlined each step that should be taken in conducting a review on the topic to ensure that subsequent reviews are rigorous. It offers a template upon which future reviews can use and improve upon. The third component of this thesis are the results of the scoping review, which mapped the literature around triage narratives and described the narrative structure and its uses. In this review we screened 18,074 studies published between 1990 and 2022 in CINAHL, MEDLINE, Embase, Cochrane, and ProQuest Central. 96 studies from this initial sample examined triage nurse narratives. Theses studies represent over 12 million ED visits and 2438 EDs. We grouped reviewed studies as using narratives for one of the following: case identification, as input variables for predictive modeling, and for quality improvement. Despite the breadth of research there was not much data on the structure of narratives: 27 studies reported the keywords that they used to search triage narratives, 7 offered descriptions of the narratives (typically the number of words used), but only one study described the narratives in any detail. Unfortunately, none of the studies that described the structure of narratives also described the nurses who wrote them – of the 9 studies that included a sample size for the nurses, none described the narratives. Additionally, fewer than 10% of studies (n=8) used reporting guidelines. This thesis demonstrates that there are myriad uses for triage narratives, and potential for additional synthesis of triage narrative studies. It also details how there is a striking paucity of research describing the structure of narratives or the nurses who generate them. There is a clear opportunity to describe triage narratives. In doing there is an opportunity to highlight the importance of this data type and to strengthen the validity of studies that have used them to conduct research. Canadian EDs are well positioned to do support this work given the national triage system and existing data collection structures.

Abstract: Smallholder farmers in rural regions of developing countries are often vulnerable to climate events. Climate Smart Agriculture (CSA) seeks to sustain or improve agricultural yields while mitigating climate change. The CGIAR Research Program on Climate Change, Agriculture and Food Security (CCAFS) has made substantial investments in developing and scaling CSA programs in developing countries. Using a multi-country dataset, this paper runs two complementary analyses. The first uses a double/debiased machine learning approach to estimate the impact of participating in a CCAFS CSA program on household food security. I estimate this impact for the entire sample and within three sub-samples, which categorize households according to their CSA adoption strategy (i.e., non-adoption, specialized adoption, or diversified adoption). Results indicate that the probability of a household being food secure is 6.0 percentage points higher (p

Employee attrition is a significant issue that occurs in every company today, regardless of external environment changes. According to the definition of attrition, the number of employees gradually decreases due to retirement, death, and resignations (Marais, 2022). Attrition can occur when a well-trained and well-adjusted talented person leaves the company for any reason, leaving a gap in the workplace (BasuMallick, 2021). It is extremely difficult for HR employees to fill the gap that has been created. For today's managers, minimizing turnover rates is a major concern, and modern HR managers do this in several ways, the employee's decision was motivated by several factors (Charaba, 2022). It is the company's responsibility to recognize employee turnover as it has a significant impact on processes. Retaining the employee reduces the need to find new employees, improves stability, and wastes less time on training. In this paper, a detailed interpretive model is used to identify the causes of turnover based on a study that will be implemented to better determine the correlation of several factors and understand which factors have the greatest impact on employee turnover in order to solve this major problem and predict when an employee will leave and end their work life. Various models, including Logistic Regression, PCA, Random Trees, Neural Networks, and LSVM, will undergo testing to identify the most accurate model to predict employee attrition and determine its suitability for implementation.

Parkinson’s Disease is a neurological degenerative disease affecting more than six million people worldwide. It is a progressive disease, impacting a person’s movements and thought processes. In recent years, computer vision and machine learning researchers have been developing techniques to aid in the diagnosis. This thesis is motivated by the exploration of hand tremor symptoms in Parkinson’s patients from the Archimedean Spiral test, a paper-and-pencil test used to evaluate hand tremors. This work presents a novel Fourier Domain analysis technique that transforms the pencil content of hand-drawn spiral images into frequency features. Our technique is applied to an image dataset consisting of spirals drawn by healthy individuals and people with Parkinson’s Disease. The Fourier Domain analysis technique achieves 81.5% accuracy predicting images drawn by someone with Parkinson’s, a result 6% higher than previous methods. We compared this method against the results using extracted features from the ResNet-50 and VGG16 pre-trained deep network models. The VGG16 extracted features achieve 95.4% accuracy classifying images drawn by people with Parkinson’s Disease. The extracted features of both methods were also used to develop a tremor severity rating system which scores the spiral images on a scale from 0 (no tremor) to 1 (severe tremor). The results show correlation to the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) developed by the International Parkinson and Movement Disorder Society. These results can be useful for aiding in early detection of tremors, the medical treatment process, and symptom tracking to monitor the progression of Parkinson’s Disease.

The evolution of technology is impressive. Before digital, there was analog; before software, there needed hardware. This evolution is natural as we try and optimize technology for our needs. The shift to digital was fueled by the space saved from using digital systems compared to analog. When it came to software, the ability to use generic hardware in the forms of cen- tral processing units; CPUs, Graphics Processing Units; GPUs, and Random Access Memory; RAM allowed for complex software solutions to be able to run on many different devices with- out much need for translations. With software development getting so advanced, we need to look back to go forward. This paper explores hardware acceleration and its role in optimizing software using hardware. This is done by documenting and analyzing the application of the Google Coral Dev Board with built-in EDGE TPUs.