Search

Sentiment analysis allows for the subjective information contained within a piece of media to be classified. This subjective information is particularly relevant in the finance domain as opinions and speculation may influence investment decisions and, consequently, affect asset prices. State-of-the-art sentiment classification methods in machine learning use contextual word representations from pre-trained BERT language models and several papers have fine-tuned BERT for the sentiment classification of financial texts. However, these preceding studies have not considered extensive real-world data sets or provided a robust assessment of whether the sentiment scores derived from BERT predict asset prices. This paper addresses these limitations by fine-tuning BERT to analyse the sentiment across an extensive set of news articles published by the Wall Street Journal. The analysis also extends beyond financial news stories, assessing the sentiment in economic and national news articles. An econometric evaluation of the sentiment scores shows that financial news sentiment is statistically significant in predicting daily market returns, while longer-term economic news sentiment predicts monthly returns.

To create a more inclusive workplace, enterprises are actively investing in identifying and eliminating unconscious bias (e.g., gender, race, age, disability, elitism and religion) across their various functions. We propose a deep learning model with a transfer learning based language model to learn from manually tagged documents for automatically identifying bias in enterprise content. We first pretrain a deep learning-based language-model using Wikipedia, then fine tune the model with a large unlabelled data set related with various types of enterprise content. Finally, a linear layer followed by softmax layer is added at the end of the language model and the model is trained on a labelled bias dataset consisting of enterprise content. The trained model is thoroughly evaluated on independent datasets to ensure a general application. We present the proposed method and its deployment detail in a real-world application.

Classification is a fundamental task in machine learning, and the principled design and evaluation of classifiers is vital to create effective classification systems and to characterise their strengths and limitations in different contexts. Binary classifiers have a range of well-known measures to summarise performance, but characterising the performance of multinomial classifiers (systems that classify instances into one of many classes) is an open problem. While confusion matrices can summarise the empirical performance of multinomial classifiers, they are challenging to interpret at a glance—challenges compounded when classes are imbalanced. We present a way to decompose multinomial confusion matrices into components that represent the prior and posterior probabilities of correctly classifying each class, and the intrinsic ability of the classifier to discriminate each class: the Bayes factor or likelihood ratio of a positive (or negative) outcome. This approach uses the odds formulation of Bayes’ rule and leads to compact, informative visualisations of confusion matrices, able to accommodate far more classes than existing methods. We call this method confusR and demonstrate its utility on 2-, 17-, and 379-class confusion matrices. We describe how confusR could be used in the formative assessment of classification systems, investigation of algorithmic fairness, and algorithmic auditing.

With the recent developments in digitisation, there are increasing number of documents available online. There are several information extraction tools that are available to extract information from digitised documents. However, identifying precise answers to a given query is often a challenging task especially if the data source where the relevant information resides is unknown. This situation becomes more complex when the data source is available in multiple formats such as PDF, table and html. In this paper, we propose a novel data extraction system to discover relevant and focused information from diverse unstructured data sources based on text mining approaches. We perform a qualitative analysis to evaluate the proposed system and its suitability and adaptability using cotton industry.

In the mining industry, many reports are generated in the project management process. These past documents are a great resource of knowledge for future success. However, it would be a tedious and challenging task to retrieve the necessary information if the documents are unorganized and unstructured. Document clustering is a powerful approach to cope with the problem, and many methods have been introduced in past studies. Nonetheless, there is no silver bullet that can perform the best for any types of documents. Thus, exploratory studies are required to apply the clustering methods for new datasets. In this study, we will investigate multiple topic modelling (TM) methods. The objectives are finding the appropriate approach for the mining project reports using the dataset of the Geological Survey of Queensland, Department of Resources, Queensland Government, and understanding the contents to get the idea of how to organise them. Three TM methods, Latent Dirichlet Allocation (LDA), Nonnegative Matrix Factorization (NMF), and Nonnegative Tensor Factorization (NTF) are compared statistically and qualitatively. After the evaluation, we conclude that the LDA performs the best for the dataset; however, the possibility remains that the other methods could be adopted with some improvements.

Due to the rapid developments in Intelligent Transportation System (ITS) and increasing trend in the number of vehicles on road, abundant of road traffic data is generated and available. Understanding spatio-temporal traffic patterns from this data is crucial and has been effectively helping in traffic plannings, road constructions, etc. However, understanding traffic patterns during COVID-19 pandemic is quite challenging and important as there is a huge difference in-terms of people's and vehicle's travel behavioural patterns. In this paper, a case study is conducted to understand the variations in spatio-temporal traffic patterns during COVID-19. We apply nonnegative matrix factorization (NMF) to elicit patterns. The NMF model outputs are analysed based on the spatio-temporal pattern behaviours observed during the year 2019 and 2020, which is before pandemic and during pandemic situations respectively, in Great Britain. The outputs of the analysed spatio-temporal traffic pattern variation behaviours will be useful in the fields of traffic management in Intelligent Transportation System and management in various stages of pandemic or unavoidable scenarios in-relation to road traffic.

Spam detection in online social networks (OSNs) have become an immensely challenging task with the nature and the use of online social networks. The spammers tend to change their behaviors over time which is commonly known as the spam drift in Online Social networks. The most popular spam combatting approaches, such as classification will grew mere unsuccessful when the drift is present since old labels cannot be used to train the new classifiers. This paper presents the comprehensive results of an approach developed to identify the drift by using a set of datasets. It introduces a set of real time user interest-based features which can be used for spam drift detection in OSNs. The paper presents a hierarchical test-based drift detection approach for spam drift learning overtime, where it learns features in an unsupervised manner. The feature similarity differences, KL divergence and the Peer Acceptability of a user is then used to detect and validate the drifted content of spam users in the real time. The system automatically updates the learning model and relabeled the users with new predicted label for classification with the new knowledge acquired through drift detection. The results are encouraging. The error rate of the classifiers with drift detection is below 1%.

Matching lists of addresses is an increasingly common task executed by business and governments alike. However, due to security issues, this task cannot always be performed using cloud computing. Moreover, addresses can arrive with spelling errors that can cause non-matches or ‘false negatives’ to occur. Our proposed framework, PostMatch, provides a locally-executed method for address-matching that combines the open-source ‘Libpostal’ address-parsing library with our ‘postparse’ post-processor code and machine-learning. PostMatch provides improved parsing accuracy compared with Libpostal alone, approaching 96.9%. The matching process features the Jaro-Winkler edit distance algorithm together with XGBoost machine-learning to achieve very high accuracy on public data. PostMatch is open-source (GPL3 licensed) and available as R script code on Github.

Unexpected hospital readmissions are problematic to both hospitals and patients. Prediction of patients’ readmission becomes an important task. Recurrent neural networks (RNN) and the attention mechanisms have been proposed to learn temporal relationships between patient’ admissions for readmission prediction. Existing works demonstrate that incorporating medical ontologies can be beneficial to prediction tasks. However, it ignores the importance of semantic information of medical codes which can be found in the codes’ descriptions. Therefore, we propose a model called Code Description Attention Model (CDAM), which adopts codes’ descriptions into readmission prediction model via RNN and the attention mechanisms to explore the semantic information about medical codes. Experimental results show that CDAM improves not only the performance of readmission prediction but also the quality of codes’ embeddings.

Currently, the majority of diagnoses of malaria rely on a combination of the patient's clinical presentation and the visualization of parasites on a stained blood film. Breath offers an attractive alternative to blood as the basis for simple, noninvasive

Development of reliable methodologies for determination of optimal sensor placements is an important requirement for the development of sentient structures. An optimal sensor layout is attained when a limited number of sensors are placed in an area such that the cost of the placement is minimised while the value of the obtained information is maximised. In this report, we first introduce a criterion that maximizes the value, or expected benefit, of using a sensor subset for a given sensor model relative to the environment. Defining the value in terms of the information obtained allows the sensor layout problem to be represented as an entropy optimization problem. This criterion is compared with other well-known criteria, both theoretically and experimentally, the latter by comparing the various criteria for optimal sensor layout using data from an existing wireless sensor network. This is achieved by firstly learning a spatial model of the environment using a Bayesian Network architecture, then predicting the expected sensor data in the rest of the space, and lastly verifying the predicted results with actual measurements (ground truth).