Search

Machine learning models, especially deep neural network (DNN) models, have widespread and valuable applications in business activities. Training a deep learning model for commercial use requires plenty of private data, expert knowledge, and computing resources. The huge commercial value of such trained models has attracted the attention of attackers. Attackers can construct a dataset by repeatedly querying the target model for the output of the requested samples and then train a substitute model on this dataset that functions similarly to the target model. In this paper, we propose a defense scheme based on physical unclonable function (PUF) against such black-box model extraction attacks. We deploy a PUF on the user side and the corresponding PUF model on the service provider side to ensure that only legitimate users can obtain the correct model predictions. Our experimental results show that by choosing a suitable fuzzy extractor threshold d , legitimate users can recover more than 99.5% of the prediction results with a little additional computational overhead to the service provider. We perform a model extraction attack in the most favorable case for the attacker, and the prediction accuracy of the obtained substitute model is only about 10%, which demonstrates the effectiveness of our proposed scheme. Compared to existing defenses, our scheme not only effectively prevents black-box model extraction attacks but also ensures that the accuracy of the prediction service for legitimate users is not affected. • Defense based on PUF obfuscation against black-box model extraction attacks. • A PUF and the corresponding PUF model ensure that legitimate users can obtain correct model predictions. • Legitimate users can correctly recover more than 99.5% of the predictions with a little additional computational overhead. • The prediction accuracy of the model extracted by the attacker is only about 10%. [ABSTRACT FROM AUTHOR]

Feature importance and interaction are among the main issues in explainable artificial intelligence or interpretable machine learning. To measure feature importance and interaction, several methods, such as H-statistic and partial dependency, have been proposed. However, it is difficult to understand the practical implications of importance and interaction. In this paper, a new method for measuring feature importance and interaction is proposed. For the classification model, we observed correctly predicted cases in a predictive model and grouped them according to the characteristics of the cases. We derived a method for feature importance and interaction from group information. For the regression model, we grouped cases according to the change in the size of the prediction error. The proposed method supports the same rationale for feature importance and interaction. It also supports the decomposition of feature importance to feature power and feature interactions. To implement the proposed method, three visualization tools, including a feature interaction graph, are implemented. Through the proposed work, we can better understand the working mechanism of a predictive model. [ABSTRACT FROM AUTHOR]

• Emerging Mobile Health systems are examples of novel technologies. • Data are collected from sensor nodes and forwarded to local databases. • From cloud computing services, the data are collected for further analysis. • This paper presents a detailed overview of M-Health systems. • We propose a secure Android-based architecture to collect patient data. Nowadays, new highly-developed technologies are changing traditional processes related to medical and healthcare systems. Emerging Mobile Health (M-Health) systems are examples of novel technologies based on advanced data communication, deep learning, artificial intelligence, cloud computing, big data, and other machine learning methods. Data are collected from sensor nodes and forwarded to local databases through new technologies that enable cellular networks and then store the information in cloud storage systems. From cloud computing services or medical centres, the data are collected for further analysis. Furthermore, machine learning techniques are being used for accurate prediction of disease analysis and for purposes of classification. This paper presents a detailed overview of M-Health systems, their model and architecture, technologies and applications and also discusses statistical and machine learning approaches. We also propose a secure Android-based architecture to collect patient data, a reliable cloud-based model for data storage. Finally, a predictive model able to classify cardiovascular diseases according to their seriousness will be discussed. Moreover, the proposed prediction model has been compared with existing models in terms of accuracy, sensitivity, and specificity. The experimental results show encouraging results in terms of the proposed predictive model for an M-Health system. Keywords: Machine Learning, Predictive, Models, M-Health, Classification, SVM, Decision Tree, Accuracy [ABSTRACT FROM AUTHOR]

This paper presents an investigation into the short-term versus long-term performance of various particle swarm optimization (PSO) control parameter configurations. While evidence suggests that the best PSO parameter values to employ are time-dependent, this paper provides an in-depth examination of a small set of parameter values to provide a more concrete quantification of the performance degradation observed with specific control parameter configurations over time. Given that the short-term performance is not necessarily indicative of long-term performance, this paper proposes that machine learning techniques be used to build predictive models based on two easily-observable landscape characteristics. Finally, using the predictive models as a basis, this paper also proposes a parameter-free PSO algorithm, which performs on par with other top-performing PSO variants, namely the three best performing static PSO configurations, particle swarm optimization with time-varying acceleration coefficients (PSO-TVAC), and particle swarm optimization with improved random constants (PSO-iRC). [ABSTRACT FROM AUTHOR]

Abstract In machine learning tasks, it is essential for a data set to be partitioned into a training set and a test set in a specific ratio. In this context, the training set is used for learning a model for making predictions on new instances, whereas the test set is used for evaluating the prediction accuracy of a model on new instances. In the context of human learning, a training set can be viewed as learning material that covers knowledge, whereas a test set can be viewed as an exam paper that provides questions for students to answer. In practice, data partitioning has typically been done by randomly selecting 70% instances for training and the rest for testing. In this paper, we argue that random data partitioning is likely to result in the sample representativeness issue, i.e., training and test instances show very dissimilar characteristics leading to the case similar to testing students on material that was not taught. To address the above issue, we propose a subclass-based semi-random data partitioning approach. The experimental results show that the proposed data partitioning approach leads to significant advances in learning performance due to the improvement of sample representativeness. [ABSTRACT FROM AUTHOR]

The increasing presence of renewable energy plants has created new challenges such as grid integration, load balancing and energy trading, making it fundamental to provide effective prediction models. Recent approaches in the literature have shown that exploiting spatio-temporal autocorrelation in data coming from multiple plants can lead to better predictions. Although tensor models and techniques are suitable to deal with spatio-temporal data, they have received little attention in the energy domain. In this paper, we propose a new method based on the Tucker tensor decomposition, capable of extracting a new feature space for the learning task. For evaluation purposes, we have investigated the performance of predictive clustering trees with the new feature space, compared to the original feature space, in three renewable energy datasets. The results are favorable for the proposed method, also when compared with state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]

Reinforcement learning has shown to be profoundly successful at learning optimal policies for simulated environments using distributed training with extensive compute capacity. Model-free reinforcement learning uses the notion of trial and error, where the error is a vital part of learning the agent to behave optimally. In mission-critical, real-world environments, there is little tolerance for failure and can cause damaging effects on humans and equipment. In these environments, current state-of-the-art reinforcement learning approaches are not sufficient to learn optimal control policies safely. On the other hand, model-based reinforcement learning tries to encode environment transition dynamics into a predictive model. The transition dynamics describes the mapping from one state to another, conditioned on an action. If this model is accurate enough, the predictive model is sufficient to train agents for optimal behavior in real environments. This paper presents the Dreaming Variational Autoencoder (DVAE) for safely learning good policies with a significantly lower risk of catastrophes occurring during training. The algorithm combines variational autoencoders, risk-directed exploration, and curiosity to train deep-q networks inside "dream" states. We introduce a novel environment, ASRS-Lab, for research in the safe learning of autonomous vehicles in grid-based warehousing. The work shows that the proposed algorithm has better sample efficiency with similar performance to novel model-free deep reinforcement learning algorithms while maintaining safety during training. [ABSTRACT FROM AUTHOR]

• We propose a deep rule ensemble creation approach. • It is driven by iterative expansion of the feature space. • It involves multiple ways of heuristic creation of diversity. • We compare the proposed approach with existing rule learning and ensemble methods. • The proposed approach achieves significant advances in classification accuracy. Rule learning approaches, which essentially aim to gerenate a decision tree or a set of "if-then" rules, have been popularly used in practice for automatically building rule-based models for prediction tasks, e.g., classification and regression. The key strength of rule-based models is their ability to interpret how an output is obtained given an input, in comparison with models trained by other machine learning approaches, e.g., neural networks. Moreover, ensemble learning approaches have been adopted as a popular way for advancing the performance of rule-based prediction through producing multiple rule-based models with diversity. Traditional approaches of ensemble learning are typically designed to train a single ensemble. In recent years, there have been some studies on creation of multiple ensembles towards increasing the diversity among rule-based models and the depth of ensemble learning. In this paper, we propose a feature expansion driven approach for automatic creation of deep rule ensembles, i.e., the dimensionality of the feature space is increased at each iteration by adding features newly created at the previous iteration. The proposed approach is compared with more recent approaches of rule learning and ensemble creation. The experimental results show that the proposed approach achieves improved performance on various data sets. [ABSTRACT FROM AUTHOR]

Machine learning (ML) classification has been one of the most important techniques of popular Internet services that aim to provide accurate predictions of data by means of a classifier model. In a machine-learning-as-a-service (MLaaS) system, the service provider allows classifier owners to upload their classifier models and charge other users for access on a pay-per-query basis, so that a user can query a classifier with data instances and then obtain their classification results. However, in a traditional way, either stored classifiers or queried data, which are potentially sensitive and economical, will be exposed by the service. Due to privacy concerns, both the classifiers and the data should remain confidential. In this paper, we propose a novel scheme to enable a classifier owner to outsourcely store his/her classifier model on a cloud server for users' queries, while protecting the confidentiality of classifier and data. We adopt a trusted processor to design efficient classification protocols for two concrete classifiers respectively. For the communicational efficiency, users only need to interact with the server no more than twice for each query in our scheme. We implement the prototype of the scheme and conduct experiments in an Intel SGX enclave. The experimental results show that the scheme is practical. [ABSTRACT FROM AUTHOR]

Drug–drug interactions are one of the major concerns of drug discovery, and the accurate prediction of drug–drug interactions is important for drug safety surveillance. However, most existing methods are suitable for predicting unobserved interactions between drugs with known interactions, but not suitable for those without known interactions. In this paper, we propose a sparse feature learning ensemble method with linear neighborhood regularization, abbreviated as SFLLN, to predict drug–drug interactions. First, we integrate four drug features, chemical substructures, targets, enzymes and pathways, by mapping drugs in different feature spaces into the common interaction space through sparse feature learning. Then, we introduce the linear neighborhood regularization to describe drug–drug relations in the interaction space by using known drug–drug interactions. In computational experiments, SFLLN produces high accuracy and outperforms benchmark methods. We present an iterative algorithm to solve the objective function of SFLLN, and analyze the convergence of the algorithm. The experiment shows that SFLLN has high efficiency and costs a reasonable amount of running time. Moreover, the case studies demonstrate that the proposed method can find novel interactions, which are validated by public evidence. In conclusion, SFLLN has great potential for predicting drug–drug interactions. [ABSTRACT FROM AUTHOR]

Abstract In regression, a predictive model which is able to anticipate the output of a new case is learnt from a set of previous examples. The output or response value of these examples used for model training is known. When learning with aggregated outputs, the examples available for model training are individually unlabeled. Collectively, the aggregated outputs of different subsets of training examples are provided. In this paper, we propose an iterative methodology to learn linear models from this type of data. In spite of being simple, its competitive performance is shown in comparison with a straightforward solution and state-of-the-art techniques. A real world problem is also illustrated which naturally fits the aggregated outputs framework: the estimation of marine litter beaching along the south-east coastline of the Bay of Biscay. [ABSTRACT FROM AUTHOR]

Recommender systems have the ability to filter unseen information for predicting whether a particular user would prefer a given item when making a choice. Over the years, this process has been dependent on robust applications of data mining and machine learning techniques, which are known to have scalability issues when being applied for recommender systems. In this paper, we propose a k-means clustering-based recommendation algorithm, which addresses the scalability issues associated with traditional recommender systems. An issue with traditional k-means clustering algorithms is that they choose the initial k centroid randomly, which leads to inaccurate recommendations and increased cost for offline training of clusters. The work in this paper highlights how centroid selection in k-means based recommender systems can improve performance as well as being cost saving. The proposed centroid selection method has the ability to exploit underlying data correlation structures, which has been proven to exhibit superior accuracy and performance in comparison to the traditional centroid selection strategies, which choose centroids randomly. The proposed approach has been validated with an extensive set of experiments based on five different datasets (from movies, books, and music domain). These experiments prove that the proposed approach provides a better quality cluster and converges quicker than existing approaches, which in turn improves accuracy of the recommendation provided. [ABSTRACT FROM AUTHOR]

The predictive performance of traditional supervised methods heavily depends on the amount of labeled data. However, obtaining labels is a difficult process in many real-life tasks, and only a small amount of labeled data is typically available for model learning. As an answer to this problem, the concept of semi-supervised learning has emerged. Semi-supervised methods use unlabeled data in addition to labeled data to improve the performance of supervised methods. It is even more difficult to get labeled data for data mining problems with structured outputs since several labels need to be determined for each example. Multi-target regression (MTR) is one type of a structured output prediction problem, where we need to simultaneously predict multiple continuous variables. Despite the apparent need for semi-supervised methods able to deal with MTR, only a few such methods are available and even those are difficult to use in practice and/or their advantages over supervised methods for MTR are not clear. This paper presents an extension of predictive clustering trees for MTR and ensembles thereof towards semi-supervised learning. The proposed method preserves the appealing characteristic of decision trees while enabling the use of unlabeled examples. In particular, the proposed semi-supervised trees for MTR are interpretable, easy to understand, fast to learn, and can handle both numeric and nominal descriptive features. We perform an extensive empirical evaluation in both an inductive and a transductive semi-supervised setting. The results show that the proposed method improves the performance of supervised predictive clustering trees and enhances their interpretability (due to reduced tree size), whereas, in the ensemble learning scenario, it outperforms its supervised counterpart in the transductive setting. The proposed methods have a mechanism for controlling the influence of unlabeled examples, which makes them highly useful in practice: This mechanism can protect them against a degradation of performance of their supervised counterparts – an inherent risk of semi-supervised learning. The proposed methods also outperform two existing semi-supervised methods for MTR. [ABSTRACT FROM AUTHOR]

Identifying Drug-Target Interactions (DTIs) is an important process in drug discovery. Traditional experimental methods are expensive and time-consuming for detecting DTIs. Therefore, computational approaches provide many effective strategies to deal with this issue. In recent years, most of computational methods only use the information of drug-drug similarity or target-target similarity, which cannot perfectly capture all characteristics to identify DTIs. In this paper, we propose a novel computational model of DTIs prediction, based on machine learning methods. To improve the performance of prediction, we further use molecular substructure fingerprints, Multivariate Mutual Information (MMI) of proteins and network topology to represent drugs, targets and relationship between them. Moreover, we employ Support Vector Machine (SVM) and Feature Selection (FS) to construct model for predicting DTIs. Experiments of evaluation show that proposed approach achieves better results than other outstanding methods for feature-based DTIs prediction. The proposed approach achieves AUPRs of 0.899, 0.929, 0.821 and 0.655 on Enzyme, Ion Channel (IC), GPCR and Nuclear Receptor datasets, respectively. Compared with existing best methods, AUPRs are increased by 0.016 on Ion Channel datasets. In addition, our method obtains the second best performance on GPCR and Enzyme datasets. The source code and all datasets are available at https://figshare.com/s/53bf5a6065f3911d46f6 . [ABSTRACT FROM AUTHOR]

Supervised classification consists in learning a predictive model using a set of labeled samples. It is accepted that predictive models accuracy usually increases as more labeled samples are available. Labeled samples are generally difficult to obtain as the labeling step if often performed manually. On the contrary, unlabeled samples are easily available. As the labeling task is tedious and time consuming, users generally provide a very limited number of labeled objects. However, designing approaches able to work efficiently with a very limited number of labeled samples is highly challenging. In this context, semi-supervised approaches have been proposed to leverage from both labeled and unlabeled data. In this paper, we focus on cases where the number of labeled samples is very limited. We review and formalize eight semi-supervised learning algorithms and introduce a new method that combine supervised and unsupervised learning in order to use both labeled and unlabeled data. The main idea of this method is to produce new features derived from a first step of data clustering. These features are then used to enrich the description of the input data leading to a better use of the data distribution. The efficiency of all the methods is compared on various artificial, UCI datasets, and on the classification of a very high resolution remote sensing image. The experiments reveal that our method shows good results, especially when the number of labeled sample is very limited. It also confirms that combining labeled and unlabeled data is very useful in pattern recognition. [ABSTRACT FROM AUTHOR]

In this paper, we propose an unsupervised learning framework based on Genetic Programming (GP) to predict the position of any particular target event (defined by the user) in a time-series. GP is used to automatically build a library of candidate temporal features. The proposed framework receives a training set S = { ( V a ) | a = 0 ⋯ n } , where each V a is a time-series vector such that ∀ V a ∈ S , V a = { ( x t ) | t = 0 ⋯ t max } where t max is the size of the time-series. All V a ∈ S are assumed to be generated from the same environment. The proposed framework uses a divide-and-conquer strategy for the training phase. The training process of the proposed framework works as follow. The user specifies the target event that needs to be predicted (e.g., Highest value, Second Highest value , …, etc.). Then, the framework classifies the training samples into different Bins , where Bins = { ( b i ) | i = 0 ⋯ t max } , based on the time-slot t of the target event in each V a training sample. Each b i ∈ Bins will contain a subset of S . For each b i , the proposed framework further classifies its samples into statistically independent clusters. To achieve this, each b i is treated as an independent problem where GP is used to evolve programs to extract statistical features from each b i ’s members and classify them into different clusters using the K-Means algorithm. At the end of the training process, GP is used to build an ‘event detector’ that receives an unseen time-series and predicts the time-slot where the target event is expected to occur. Empirical evidence on artificially generated data and real-world data shows that the proposed framework significantly outperforms standard Radial Basis Function Networks, standard GP system, Gaussian Process regression, Linear regression, and Polynomial Regression. [ABSTRACT FROM AUTHOR]

Abstract: This paper focuses on the problem of defect prediction, a problem of major importance during software maintenance and evolution. It is essential for software developers to identify defective software modules in order to continuously improve the quality of a software system. As the conditions for a software module to have defects are hard to identify, machine learning based classification models are still developed to approach the problem of defect prediction. We propose a novel classification model based on relational association rules mining. Relational association rules are an extension of ordinal association rules, which are a particular type of association rules that describe numerical orderings between attributes that commonly occur over a dataset. Our classifier is based on the discovery of relational association rules for predicting whether a software module is or it is not defective. An experimental evaluation of the proposed model on the open source NASA datasets, as well as a comparison to similar existing approaches is provided. The obtained results show that our classifier overperforms, for most of the considered evaluation measures, the existing machine learning based techniques for defect prediction. This confirms the potential of our proposal. [Copyright &y& Elsevier]

Abstract: A proper representation of the uncertainty involved in a prediction is an important prerequisite for the acceptance of machine learning and decision support technology in safety–critical application domains such as medical diagnosis. Despite the existence of various probabilistic approaches in these fields, there is arguably no method that is able to distinguish between two very different sources of uncertainty: aleatoric uncertainty, which is due to statistical variability and effects that are inherently random, and epistemic uncertainty which is caused by a lack of knowledge. In this paper, we propose a method for binary classification that does not only produce a prediction of the class of a query instance but also a quantification of the two aforementioned sources of uncertainty. Despite being grounded in probability and statistics, the method is formalized within the framework of fuzzy preference relations. The usefulness and reasonableness of our approach is confirmed on a suitable data set with information about patients suffering from chest pain. [Copyright &y& Elsevier]

Abstract: Noise detection for software measurement datasets is a topic of growing interest. The presence of class and attribute noise in software measurement datasets degrades the performance of machine learning-based classifiers, and the identification of these noisy modules improves the overall performance. In this study, we propose a noise detection algorithm based on software metrics threshold values. The threshold values are obtained from the Receiver Operating Characteristic (ROC) analysis. This paper focuses on case studies of five public NASA datasets and details the construction of Naive Bayes-based software fault prediction models both before and after applying the proposed noise detection algorithm. Experimental results show that this noise detection approach is very effective for detecting the class noise and that the performance of fault predictors using a Naive Bayes algorithm with a logNum filter improves if the class labels of identified noisy modules are corrected. [Copyright &y& Elsevier]