Search

The dynamicity of real-world systems poses a significant challenge to deployed predictive machine learning (ML) models. Changes in the system on which the ML model has been trained may lead to performance degradation during the system's life cycle. Recent advances that study non-stationary environments have mainly focused on identifying and addressing such changes caused by a phenomenon called concept drift. Different terms have been used in the literature to refer to the same type of concept drift and the same term for various types. This lack of unified terminology is set out to create confusion on distinguishing between different concept drift variants. In this paper, we start by grouping concept drift types by their mathematical definitions and survey the different terms used in the literature to build a consolidated taxonomy of the field. We also review and classify performance-based concept drift detection methods proposed in the last decade. These methods utilize the predictive model's performance degradation to signal substantial changes in the systems. The classification is outlined in a hierarchical diagram to provide an orderly navigation between the methods. We present a comprehensive analysis of the main attributes and strategies for tracking and evaluating the model's performance in the predictive system. The paper concludes by discussing open research challenges and possible research directions. • A new taxonomical classification of concept drift types. • Providing a classification hierarchy of performance-based detection methods. • Identifying research gaps and trends in performance-based detection methods. • Suggesting future research directions in concept drift detection based on the findings. [ABSTRACT FROM AUTHOR]

Forecasting harmful algal blooms (HABs) is an important part of marine environmental monitoring. Algae bloom observation channels includes satellite remote sensing and maritime station monitoring (MSM). Compared with satellite remote sensing image, MSM can collect more accurate data, which includes various seawater inorganic salt content related to the HABs. However, the measured data of MSM is easily affected by regional sediment content and seawater dynamic field, which leads to lack the accurate forecasting models. Therefore, this paper proposed a local spatio-temporal HABs forecasted model (STHFM) based on few impact factors in MSM. The advantage of this model is to first use principal component analysis to select main environment factors (MEFs) related to HABs. Then, this model distinguishes multiple warning levels of HABs in spatio-temporal according to the algae growth rate. Finally, this paper generates continuous MEFs time series information based on the Autoregressive Integrated Moving Average (ARIMA) model in the high-level warning area. And an improved LSTM network with MEFs time series as input is established to forecast HABs in future. The proposed model is tested on the NOAA website public dataset, which contains historical harmful algae Alexandrium data on the East Coast of US. The experimental shows that our model has good HABs monitoring performance. Under the public NOAA Alexandrium dataset, the proposed model can achieve the highest prediction accuracy of 82.1%, and has a small prediction error. • The coupling relationship between HABs and major environmental factors were investigated. • The changes of algae were successfully predicted by using the few major influencing factors. • The use of space–time clustering in combination with forecasting is recommended. • Better prediction of HABs by combining ARIMA with deep LSTM network. [ABSTRACT FROM AUTHOR]

In the last decade, machine learning-based approaches became capable of performing a wide range of complex tasks sometimes better than humans, demanding a fraction of the time. Such an advance is partially due to the exponential growth in the amount of data available, which makes it possible to extract trustworthy real-world information from them. However, such data is generally imbalanced since some phenomena are more likely than others. Such a behavior yields considerable influence on the machine learning model's performance since it becomes biased on the more frequent data it receives. Despite the considerable amount of machine learning methods, a graph-based approach has attracted considerable notoriety due to the outstanding performance over many applications, i.e., the Optimum-Path Forest (OPF). In this paper, we propose three OPF-based strategies to deal with the imbalance problem: the O 2 PF and the OPF-US, which are novel approaches for oversampling and undersampling, respectively, as well as a hybrid strategy combining both approaches. The paper also introduces a set of variants concerning the strategies mentioned above. Results compared against several state-of-the-art techniques over public and private datasets confirm the robustness of the proposed approaches. • Four variations of the oversampling algorithm O2PF for data oversampling. • Four OPF-based methods for data undersampling. • Three hybrid approaches for undersampling and oversampling. [ABSTRACT FROM AUTHOR]

In the last decade, several studies have shown that facial attractiveness can be learned by machines. In this paper, we address Facial Beauty Prediction from static images. The paper contains three main contributions. First, we propose a two-branch architecture (REX-INCEP) based on merging the architecture of two already trained networks to deal with the complicated high-level features associated with the FBP problem. Second, we introduce the use of a dynamic law to control the behaviour of the following robust loss functions during training: ParamSmoothL1, Huber and Tukey. Third, we propose an ensemble regression based on Convolutional Neural Networks (CNNs). In this ensemble, we use both the basic networks and our proposed network (REX-INCEP). The proposed individual CNN regressors are trained with different loss functions, namely MSE, dynamic ParamSmoothL1, dynamic Huber and dynamic Tukey. Our approach is evaluated on the SCUT-FBP5500 database using the two evaluation scenarios provided by the database creators: 60%–40% split and five-fold cross-validation. In both evaluation scenarios, our approach outperforms the state of the art on several metrics. These comparisons highlight the effectiveness of the proposed solutions for FBP. They also show that the proposed dynamic robust losses lead to more flexible and accurate estimators. [ABSTRACT FROM AUTHOR]

Teaching plays a fundamental role in human learning. Typically, a human teaching strategy involves assessing a student's knowledge progress for tailoring the teaching materials to enhance the learning progress. A human teacher can achieve this by tracing a student's knowledge over essential learning concepts in a task. Albeit, such a teaching strategy is not well exploited yet in machine learning as current machine teaching methods tend to directly assess the progress of individual training samples without paying attention to the underlying learning concepts in a learning task. In this paper, we propose a novel method, called Knowledge Augmented Data Teaching (KADT), which can optimize a data teaching strategy for a student model by tracing its knowledge progress over multiple learning concepts in a learning task. Specifically, the KADT method incorporates a knowledge tracing model to dynamically capture the knowledge progress of a student model in terms of latent learning concepts. We further develop an attention-pooling mechanism to distill knowledge representations of a student model with respect to class labels, which enables to develop a data teaching strategy on critical training samples. We have evaluated the performance of the KADT method on four different machine learning tasks, including knowledge tracing, sentiment analysis, movie recommendation, and image classification. The KADT method consistently outperforms the state-of-the-art methods on all these tasks. [ABSTRACT FROM AUTHOR]

Automated machine learning has the ability to improve the efficiency of time series classification due to the ability to combine multiple feature extraction methods and classification models. In the paper, we propose a flexible AutoML approach that combines multiple feature generation strategies (spectral, wavelet, topological, quantile) and classifiers as parts of the modeling pipeline. It allows obtaining a more robust and lightweight solution for the time series classification problem. Generation of the pipeline is based on an evolutionary algorithm. Comparison with approaches with the highest results was conducted on the UEA/UCR archive for modeling quality analysis. The proposed approach allows, on the one hand, to solve the task of time series classification automatically and could be used as part of an industrial data processing pipeline. On the other hand, it could be used as a tool for modeling and studying natural process properties described as time series. [ABSTRACT FROM AUTHOR]

The Successor Feature framework for Reinforcement Learning algorithms improves task transfer by decomposing the learned state–action value function. The decomposition involves two components, one that captures future-expected state features and the other that models the task-related reward structure. However, successful transfer between tasks depends heavily on how the reward function changes, possibly leading to failure of the original Successor Feature formulation. This paper proposes the Dynamic Successor Feature framework, DynSF, by extending the mathematical formulation of the original Successor Feature framework to center around a learned state-transition model. Under this formulation, the state-transition model dynamically induces the acting policy. The flexibility of DynSF also extends to the architecture, requiring only a state-transition model and a small vector of parameters. This architecture provides immense flexibility in the choice of the model used to learn the state-transition model. The DynSF framework is evaluated and compared to other baseline algorithms through several experiments in a continuous grid world environment, a robotic Reacher, and pixels in the Doom environment. • DynSF enables a state-transition model to be used in Reinforcement Learning. • Policy and discount factors can be parameterized dynamically. • Different supervised ML algorithms can be used to learn the state transition model. • DynSF can be used for guided exploration and transfer learning. • DynSF performs better during task transfer than the Successor Feature framework. [ABSTRACT FROM AUTHOR]

Medical experts are often skeptical of data-driven models due to the lack of their explainability. Several experimental studies commence with wide-ranging unsupervised learning and precisely with clustering to obtain existing patterns without prior knowledge of newly acquired data. Explainable Artificial Intelligence (XAI) increases the trust between virtual assistance by Machine Learning models and medical experts. Awareness about how data is analyzed and what factors are considered during the decision-making process can be confidently answered with the help of XAI. In this paper, we introduce an improved hybrid classical-quantum clustering (improved qk-means algorithm) approach with the additional explainable method. The proposed model uses learning strategies such as the Local Interpretable Model-agnostic Explanations (LIME) method and improved quantum k-means (q k -means) algorithm to diagnose abnormal activities based on breast cancer images and Knee Magnetic Resonance Imaging (MRI) datasets to generate an explanation of the predictions. Compared with existing algorithms, the clustering accuracy of the generated clusters increases trust in the model-generated explanations. In practice, the experiment uses 600 breast cancer (BC) patient records with seven features and 510 knee MRI records with five features. The result shows that the improved hybrid approach outperforms the classical one in the BC and Knee MRI datasets. [ABSTRACT FROM AUTHOR]

Credit scoring systems have seen revolutionary development in recent decades, with many classification algorithms being proposed. However, with the increase in the data volume, the performance of traditional algorithms tends to encounter bottlenecks. Although deep learning methods have advantages in handling big data, they are not commonly applied in credit scoring. As one of the most frequently used methods in deep learning, convolutional neural network (CNN) use convolutional kernels as feature extraction tools and has been very successful in tasks related to images or text. This is because image and text data naturally have a structural characteristic called spatial local correlation, which means that the pixels or tokens covered by the same convolutional kernel are highly correlated, and they can be jointly processed to extract meaningful feature representations. However, the tabular data used for credit scoring do not naturally have such a characteristic. The main contribution of this paper is to propose a novel end-to-end soft reordering one-dimensional CNN (SR-1D-CNN), which can adaptively reorganize the original tabular data and make them more conducive to CNN learning. Several real-world credit scoring datasets of different sizes are used for a comprehensive comparison with traditional machine learning classifiers and other deep learning methods. The experimental results demonstrate that the soft reordering mechanism can effectively improve the classification effect of the CNN for tabular data. With the increase in the data scale, the proposed approach obtains superior results to those of other benchmark models. [ABSTRACT FROM AUTHOR]

Adversarial attacks seriously threaten the security of machine learning models. Thus, detecting adversarial examples has become an important and interesting research topic facing various adversarial attacks. However, the majority of existing adversarial example detection algorithms cannot perform well in detecting adversarial examples with slight perturbations. In this paper, we propose a novel attention-based dual stream detector (ADS-Detector) that can address the detection of adversarial examples with both slight and large perturbations. Specifically, we first design a data process module to generate pixel and prediction confidence stream data from the raw image. Then, we propose an N -layer attention module to extract the channel and spatial feature weights between the pixel and prediction confidence stream data. Eventually, we feed the dual-stream data into the same subdetection model with a convolutional block attention module; then, the output results are combined to determine whether the input image is an adversarial example or not. To validate the performance, we conduct extensive experiments on three public datasets: CIFAR10, Dogs vs. Cats and ImageNet. After sufficient analysis of the simulation results, we find that our proposed method outperforms the others for the detection of adversarial attacks generated by the considered attack methods. • Adversarial example detection based on prediction confidence and attention mechanism. • A dual-stream framework is designed to detect both slight and large perturbations. • Spatial filters and attention mechanisms are utilized to enhance detection performance. • Extensive experiments have demonstrated the effectiveness of our method. [ABSTRACT FROM AUTHOR]

The objective of this paper is to present MLDA, a tool for the exploration and analysis of multi-label datasets with both simple and multiple views. MLDA comprises a GUI and a Java API, providing the user with a wide set of charts, metrics, methods for transforming and preprocessing data, as well as comparison of several datasets. The paper introduces the main features of the framework, and introduces its use toward some illustrative examples. [ABSTRACT FROM AUTHOR]

In this paper, a novel optimization technique based on artificial bee colony algorithm (ABC), which is called as PS-ABCII, is presented. In PS-ABCII, there are three major differences from other ABC-based techniques: (1) the opposition-based learning is applied to the population initialization; (2) the greedy selection mechanism is not adopted; (3) the mode that employed bees become scouts is modified. In order to illustrate the superiority of the proposed modified technique over other ABC-based techniques, ten classical benchmark functions are employed to test. In addition, a hybrid model called PS-ABCII-ELM is also proposed in this paper, which is combined of the PS-ABCII and Extreme Learning Machine (ELM). In PS-ABCII-ELM, the PS-ABCII is applied to tune input weights and biases of ELM in order to improve the generalization performance of ELM. And then it is applied to model and optimize the thermal efficiency of a 300 MW coal-fired boiler. The experimental results show that the proposed model is very convenient, direct and accurate, and it can give a general and suitable way to predict and improve the boiler efficiency of a coal-fired boiler under various operating conditions. [ABSTRACT FROM AUTHOR]

In real application domains, acquiring fine-grained data has a higher cost than coarse-grained data. To achieve the best results at the lowest cost, it is necessary to select an optimal granularization for different data precisions. However, existing work in cost-sensitive learning focuses mainly on a fixed granularization, while related test costs and misclassification costs are also fixed. In this paper, we propose an optimal cost-sensitive granularization based on rough sets for variable costs. The major contributions of this paper are threefold. First, different granularizations are used with confidence levels correlated to the precision of the data. In this context, we build a confidence-level-based covering rough set model. The test cost is then represented by a function of each feature and a confidence level, and misclassification costs are computed according to test costs. The variable cost settings are more reasonable for representing the relationship between the granularizations and costs than previous approaches. Finally, a granularization approach is proposed to obtain a trade-off between the different data precisions and variable costs. The experimental results show that our approach satisfactorily handles data with different precisions under different cost settings. In addition, the optimal granularization is adaptive to the data involved rather than being fixed, and is thus more efficient and more versatile than the existing fixed granularization. [ABSTRACT FROM AUTHOR]

Abstract: Business process engineering and mining is a technique that allows discovery, analysis and modeling of possible Business Processes based on information gathered from enterprise information systems. Most of currently available business process engineering and mining techniques either focus on machine learning techniques to mine, discover and model any possible Business Processes from raw data, or use semantically-enabled process models and service descriptions to construct and represent complex Business Processes. However, in real-life scenario, all the required services are not always available and hence exact matching of the services in order to construct Business Process is not possible. In this paper, we present our approach of using fuzzy Web Service discovery to construct and represent Business Processes. It helps in relaxing the matching criteria of Web Services, and allows service consumers to specify business requirements in a more fuzzy way, and hence increases the possibility of finding required Web Services that could construct Business Processes. The paper presents the proposed solution then reports and discusses the evaluation. [Copyright &y& Elsevier]

Abstract: In this paper, we propose a VECtor DIScovery approach, called VECDIS, which enhances the learning performance of existing classifiers directly from various data types and is able to discover features made of multiple feature types for explanatory purposes. The data types could be combinations of multivariate, short time-series or short sequential data. The features in the dataset could have single item or/and a list of ordered items of different sizes. The present approach allows handling raw vector data without prior manipulation (i.e., preprocessing). The discovered features are made of vector and non-vector mathematical relations. The algorithm generates new vector features and mathematical expression features that are transmitted or exchanged with previously generated features, to the next iterative step. The approach is able to search and automatically discover thousands of different features (sequence manipulation), performed on the sequence features. We performed large number of experiments with various synthetic and simulated datasets and with a wide range of classifiers. The results show that VECDIS enhanced significantly the classification performance of existing classifiers to handle datasets having multiple feature types with short sequence features. Nevertheless, there is no guarantee that the mathematical library as presented in this paper is suitable to all sequence datasets and would lead to discovering a valuable feature set. Therefore, VECDIS enables expanding or exchanging the mathematical library as desire. [Copyright &y& Elsevier]

Abstract: This paper proposes demographic estimation algorithms for profiling Twitter users, based on their tweets and community relationships. Many people post their opinions via social media services such as Twitter. This huge volume of opinions, expressed in real time, has great appeal as a novel marketing application. When automatically extracting these opinions, it is desirable to be able to discriminate discrimination based on user demographics, because the ratio of positive and negative opinions differs depending on demographics such as age, gender, and residence area, all of which are essential for market analysis. In this paper, we propose a hybrid text-based and community-based method for the demographic estimation of Twitter users, where these demographics are estimated by tracking the tweet history and clustering of followers/followees. Our experimental results from 100,000 Twitter users show that the proposed hybrid method improves the accuracy of the text-based method. The proposed method is applicable to various user demographics and is suitable even for users who only tweet infrequently. [Copyright &y& Elsevier]

Activity recognition plays a key role in many fields, such as health monitoring and elderly care. Handling changes in user habits is a significant technical challenge in activity recognition. Ideally, a model should adapt to newly emerging classes and concept drift dynamically. This paper proposes a novel semi-supervised class incremental learning method, namely, disagreement-based class incremental random forest (Di-CIRF). The proposed model can detect newly emerging classes and update a previously established activity recognition model through streaming data. First, it is necessary to identify novel candidates by employing the disagreement-based confidence voting mechanism and minimum bounding box (MBB)-based separation detection to annotate newly emerging data accurately. Then, the coarse coding-based cohesion detection strategy is adopted to filter out the true novelty instances. This paper also proposes the iterative MBB-based splitting strategy and the pseudo-instance generation mechanism in Di-CIRF for updating the activity model without retaining the trained data. According to experimental results on four public activity recognition datasets, Di-CIRF outperforms the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

This paper focuses on the numerical solutions of generalized Black–Scholes partial differential equations and proposes a novel structure automatic-determined Fourier extreme learning machine (SAD-FELM) to solve the partial differential equations. In the SAD-FELM, the Fourier extreme learning machine (FELM) is designed. The product of two Fourier polynomials is used to serve as basis functions in the hidden layer. The optimal sampling frequencies of two Fourier polynomials depend on the distribution of data samples. Then, the improved extreme learning machine algorithm is used to solve the linear equation systems derived from the generalized Black–Scholes partial differential equations. Furthermore, to determine the optimal sampling frequencies and the number of neurons, this paper designs a structure automatic-determined algorithm based on metaheuristic algorithms. The research conducts comparative experiments among the SAD-FELMs based on different metaheuristic algorithms. The experimental results show that the numerical solutions obtained by the SAD-FELM based on the hunger games search algorithm are the most accurate. Finally, compared with other methods, the superiority and feasibility of the SAD-FELM are proved. [ABSTRACT FROM AUTHOR]

This study demonstrates that a time-frequency (TF) image pattern recognition approach offers significant advantages over standard signal classification methods that use t -domain only or f -domain only features. Two approaches are considered and compared. The paper describes the significance of the standard TF approach for non-stationary signals; TF signal (TFS) features are defined by extending t -domain or f -domain features to a joint ( t, f ) domain resulting in e.g. TF flatness and TF flux. The performance of the extended TFS features is comparatively assessed using Receiver Operating Characteristic (ROC) analysis Area Under the Curve (AUC) measure. Experimental results confirm that the extended TFS features generally yield improved performance (up to 19%) when compared to the corresponding t -domain and f -domain features. The study also explores a second approach based on novel TF image (TFI) features that further improves TF-based classification of non-stationary signals. New TFI features are defined and extracted from the ( t, f ) domain; they include TF Hu invariant moments, TF Haralick features, and TF Local Binary Patterns (LBP). Using a state-of-the-art classifier, different metrics based on confusion matrix performance are compared for all extended TFS features and TFI features. Experimental results show the improved performance of TFI features over both TFS features and t -domain only or f -domain only features, for all TF representations and for all the considered performance metrics. The experiment is validated by comparing this new proposed methodology with a recent study, utilizing the same large and complex data set of EEG signals, and the same experimental setup. The resulting classification results confirm the superior performance of the proposed TFI features with accuracy improvement up to 5.52%. [ABSTRACT FROM AUTHOR]

In this paper, a novel decremental subspace-based learning method called Decremental Generalized Discriminative Common Vectors method (DGDCV) is presented. The method makes use of the concept of decremental learning, which we introduce in the field of supervised feature extraction and classification. By efficiently removing unnecessary data and/or classes for a knowledge base, our methodology is able to update the model without recalculating the full projection or accessing to the previously processed training data, while retaining the previously acquired knowledge. The proposed method has been validated in 6 standard face recognition datasets, showing a considerable computational gain without compromising the accuracy of the model. [ABSTRACT FROM AUTHOR]

Multigranulation rough sets have attracted more and more attentions in recent years. In real-life applications, with the development of information technology, the attribute values often dynamically evolve over time. How to update useful knowledge is of great importance for dynamic information systems. Approximations of a concept are fundamental concepts of multigranulation rough sets, which need to be updated incrementally while refining or coarsening attribute values. Motivated by the requirements of dynamic knowledge acquisition due to refining or coarsening attribute values, in this paper, we present the dynamic mechanisms for updating approximations in multigranulation rough sets while refining or coarsening attribute values. Then, the corresponding dynamic algorithms for updating multigranulation approximations are designed on the basis of the proposed mechanisms. Extensive experiments on six data sets from UCI demonstrate that the proposed dynamic algorithms for updating approximations in multigranulation rough sets are more effective in comparison with the static algorithm. [ABSTRACT FROM AUTHOR]

Several code smells detection tools have been developed providing different results, because smells can be subjectively interpreted and hence detected in different ways. Machine learning techniques have been used for different topics in software engineering, e.g., design pattern detection, code smell detection, bug prediction, recommending systems. In this paper, we focus our attention on the classification of code smell severity through the use of machine learning techniques in different experiments. The severity of code smells is an important factor to take into consideration when reporting code smell detection results, since it allows the prioritization of refactoring efforts. In fact, code smells with high severity can be particularly large and complex, and create larger issues to the maintainability of software a system. In our experiments, we apply several machine learning models, spanning from multinomial classification to regression, plus a method to apply binary classifiers for ordinal classification. In fact, we model code smell severity as an ordinal variable. We take the baseline models from previous work, where we applied binary classification models for code smell detection with good results. We report and compare the performance of the models according to their accuracy and four different performance measures used for the evaluation of ordinal classification techniques. From our results, while the accuracy of the classification of severity is not high as in the binary classification of absence or presence of code smells, the ranking correlation of the actual and predicted severity for the best models reaches 0.88–0.96, measured through Spearman’s ρ . [ABSTRACT FROM AUTHOR]

There are situations in which lexicon-based methods for Sentiment Analysis (SA) are not able to generate a classification output for specific instances of a given dataset. Most often, the reason for this situation is the absence of specific terms in the sentiment lexicon required in the classification effort. In such cases, there were only two possible paths to follow: (1) add terms to the lexicon (off-line process) by human intervention to guarantee no noise is introduced into the lexicon, which prevents the classification system to provide an immediate answer; or (2) use the services of a word-frequency dictionary (on-line process), which is computationally costly to build. This paper investigates an alternative approach to compensate for the lack of ability of a lexicon-based method to produce a classification output. The method is based on the combination of the classification outputs of non lexicon-based tools. Specifically, firstly the outcome values of applying two or more non-lexicon classification methods are obtained. Secondly, these non-lexicon outcomes are fused using a uninorm based approach, which has been proved to have desirable compensation properties as required in the SA context, to generate the classification output the lexicon based approach is unable to achieve. Experimental results based on the execution of two well-known supervised machine learning algorithms, namely Naïve Bayes and Maximum Entropy, and the application of a cross-ratio uninorm operator are presented. Performance indices associated to options (1) and (2) above are compared against the results obtained using the proposed approach for two different datasets. Additionally, the performance of the proposed cross-ratio uninorm operator based approach is also compared when the aggregation operator used is the arithmetic mean instead. It is shown that the combination of non lexicon-based classification methods with specific uninorm operators improves the classification performance of lexicon-based methods, and it enables the offering of an alternative solution to the SA classification problem when needed. The proposed aggregation method could be used as well as a replacement of ensemble averaging techniques commonly applied when combining the results of several machine learning classifiers’ outputs. [ABSTRACT FROM AUTHOR]

Feature selection is an important research topic in machine learning and computer vision in that it can reduce the dimensionality of input data and improve the performance of learning algorithms. Low-rank approximation techniques can well exploit the low-rank property of input data, which coincides with the internal consistency of dimensionality reduction. In this paper, we propose an efficient unsupervised feature selection algorithm, which incorporates low-rank approximation as well as structure learning. First, using the self-representation of data matrix, we formalize the feature selection problem as a matrix factorization with low-rank constraints. This matrix factorization formulation also embeds structure learning regularization as well as a sparse regularized term. Second, we present an effective technique to approximate low-rank constraints and propose a convergent algorithm in a batch mode. This technique can serve as an algorithmic framework for general low-rank recovery problems as well. Finally, the proposed algorithm is validated in twelve publicly available datasets from machine learning repository. Extensive experimental results demonstrate that the proposed method is capable to achieve competitive performance compared to existing state-of-the-art feature selection methods in terms of clustering performance. [ABSTRACT FROM AUTHOR]

In reality, the number of labeled time series data is often small and there is a huge number of unlabeled data. Manually labeling these unlabeled examples is time-consuming and expensive, and sometimes it is even impossible. In this paper, we combine active learning and semi-supervised learning to obtain a confident and sufficient labeled training data for multivariate time series classification. We first propose a sampling strategy by ranking the informativeness of unlabeled examples based on its uncertainty and its local data density. Next, an active semi-supervised learning framework is introduced to make best use of the advantage of active learning and semi-supervised learning for data annotation. Finally, we advance a valid stopping criterion of active learning to provide a sufficient and reliable labeled training dataset by costing human resources as less as possible. Our experimental results show that our approach can manually annotate examples as small as possible and simultaneously obtain a confident and informative labeled dataset, which is sufficient to learn an efficient classification. [ABSTRACT FROM AUTHOR]

There have been increasing interests in natural language processing to explore effective methods in learning better representations of text for sentiment classification in product reviews. However, most existing methods do not consider subtle interplays among words appeared in review text, authors of reviews and products the reviews are associated with. In this paper, we make use of a heterogeneous network to model the shared polarity in product reviews and learn representations of users, products they commented on and words they used simultaneously. The basic idea is to first construct a heterogeneous network which links users, products, words appeared in product reviews, as well as the polarities of the words. Based on the constructed network, representations of nodes are learned using a network embedding method, which are subsequently incorporated into a convolutional neural network for sentiment analysis. Evaluations on the product reviews, including IMDB, Yelp 2013 and Yelp 2014 datasets, show that the proposed approach achieves the state-of-the-art performance. [ABSTRACT FROM AUTHOR]

Computer aided diagnosis systems based on brain imaging are a powerful tool to assist in the diagnosis of Alzheimer's Disease (AD). The goal is the automatic recognizing of neurodegenerative patterns that characterize the disease. In this regard, determining regions related to the disease results crucial to select the most discriminative voxels and to optimize the number of features to be used in the learning algorithm. In this paper, we propose a method based on the robust principal component analysis (Robust PCA) algorithm that allows to automatically compute Regions Of Interest (ROIs) over a training set of images and rank them according to their diagnostic relevance. Robust PCA is used to compute the sparse error matrix, which is, in turn, employed to determine the brain areas related to the Alzheimer's disease. These areas are further used as a mask to select and weight the most discriminative voxels to construct a classification model. We then describe a method to fuse the features computed from different image modalities based on the weights assigned by the individual Support Vector Classifiers during the training process. The method presented here has been applied to multimodal image containing both functional (18F-FDG PET) and structural (Magnetic Resonance) data. Experiments, conducted using 68 control subjects and 70 CE patients, show the effectiveness of the proposed approach for the exploratory analysis. At the same time, classification experiments using the features computed by the proposed method and assessed by cross-validation showed accuracy values up to 92% and AUC (Area Under the Curve) of 0.95. Thus, the proposal seems as an effective technique to reveal ROIs in differential diagnosis applications and to combine multimodal image data, outperforming other classification methods, including the voxel-as-features (VAF) baseline. [ABSTRACT FROM AUTHOR]

Liver transplantation is a promising and widely-accepted treatment for patients with terminal liver disease. However, transplantation is restricted by the lack of suitable donors, resulting in significant waiting list deaths. This paper proposes a novel donor-recipient allocation system that uses machine learning to predict graft survival after transplantation using a dataset comprised of donor-recipient pairs from the King's College Hospital (United Kingdom). The main novelty of the system is that it tackles the imbalanced nature of the dataset by considering semi-supervised learning, analysing its potential for obtaining more robust and equitable models in liver transplantation. We propose two different sources of unsupervised data for this specific problem (recent transplants and virtual donor-recipient pairs) and two methods for using these data during model construction (a semi-supervised algorithm and a label propagation scheme). The virtual pairs and the label propagation method are shown to alleviate the imbalanced distribution. The results of our experiments show that the use of synthetic and real unsupervised information helps to improve and stabilise the performance of the model and leads to fairer decisions with respect to the use of only supervised data. Moreover, the best model is combined with the Model for End-stage Liver Disease score (MELD), which is at the moment the most popular assignation methodology worldwide. By doing this, our decision-support system considers both the compatibility of the donor and the recipient (by our prediction system) and the recipient severity (via the MELD score), supporting then the principles of fairness and benefit. [ABSTRACT FROM AUTHOR]

Abstract: One key issue for most classification algorithms is that they need large amounts of labeled samples to train the classifier. Since manual labeling is time consuming, researchers have proposed technologies of active learning and semi-supervised learning to reduce manual labeling workload. There is a certain degree of complementarity between active learning and semi-supervised learning, and therefore some researches combine them to further reduce manual labeling workload. However, researches on combining active learning and semi-supervised learning for SVM classifier are rare. Of numerous SVM active learning algorithms, the most popular is the one that queries the sample closest to the current classification hyperplane in each iteration, which is denoted as SVMAL in this paper. Realizing that SVMAL is only interested in samples that are more likely to be on the class boundary, while ignoring the usage of the rest large amounts of unlabeled samples, this paper designs a semi-supervised learning algorithm to make full use of the rest non-queried samples, and further forms a new active semi-supervised SVM algorithm. The proposed active semi-supervised SVM algorithm uses active learning to select class boundary samples, and semi-supervised learning to select class central samples, for class central samples are believed to better describe the class distribution, and to help SVMAL finding the boundary samples more precisely. In order not to introduce too many labeling errors when exploring class central samples, the label changing rate is used to ensure the reliability of the predicted labels. Experimental results show that the proposed active semi-supervised SVM algorithm performs much better than the pure SVM active learning algorithm, and thus can further reduce manual labeling workload. [Copyright &y& Elsevier]

Noise addition is a data distortion technique widely used in data intensive applications. For example, in machine learning tasks it helps to reduce overfitting, whereas in data privacy protection it adds uncertainty to personally identifiable information. Yet, due to its mathematical operating principle, noise addition is a method mainly intended for continuous numerical data. In fact, despite the large amount of nominal data that are being currently compiled and used in data analysis, only a few alternative techniques have been proposed to distort nominal data in a similar way as standard noise addition does for numerical data. Furthermore, all these alternative methods rely on the distribution of the data rather than on the semantics of nominal values, which negatively affects the utility of the distorted outcomes. To tackle this issue, in this paper we present a semantically-grounded alternative to numerical noise suitable for nominal data, which we name semantic noise . By means of semantic noise, and by exploiting structured knowledge sources such as ontologies, we are able to distort nominal data while preserving better their semantics and thus, their analytical utility. To that end, we provide semantically and mathematically coherent versions of the statistical operators required in the noise addition process, which include the difference, the mean, the variance and the covariance. Then, we propose semantic noise addition algorithms that cope with the finite, discrete and non-ordinal nature of nominal data. The proposed algorithms cover both uncorrelated noise addition, which is suited to independent attributes, and correlated noise addition, which can cope with multivariate datasets with dependent attributes. Empirical results show that our proposals offer general and configurable mechanisms to distort nominal data while preserving data semantics better than baseline methods based only on the distribution of the data. [ABSTRACT FROM AUTHOR]

To learn an effective Sum-Product Network (SPN) for probabilistic inference, one needs to have a substantial amount of data. In the case when the training dataset is small, SPN performance can be degraded. In this paper, we investigate how transfer learning can improve a SPN when the number of training examples is limited. In particular, we consider a structural transfer setting where (i) one does not have a source dataset but a source SPN, and (ii) there is some kind of similarity between the source SPN and the target domain. We propose a transfer learning approach called TopTrSPN, utilizing the information of the first layer clusters in the source SPN to learn the first layer of the target SPN. Our approach is motivated by transfer learning characteristics of Convolution Neural Network (CNN) as SPN can be viewed as a probabilistic, general-purpose convolution network. Moreover, since the source SPN may have some distribution differences from the target domain, we perform matching between the two by filtering out inconsistent variables. Empirical results on twenty benchmark datasets show the feasibility of our proposed transfer learning approach for SPN structure learning a target domain. Moreover, our proposed transfer learning approach shows encouraging performance when it is applied to text datasets with different set of variables. [ABSTRACT FROM AUTHOR]

Constructing a graph to represent the structure among data objects plays a fundamental role in various data mining tasks with graph-based learning. Since traditional pairwise distance-based graph construction is sensitive to noise and outliers, sparse representation based graphs (e.g., ℓ 1 -graphs) have been proposed in the literature. Although ℓ 1 -graphs prove powerful and robust for many graph-based learning tasks, it suffers from weak locality and high computation costs. In this paper, we propose a locality weighted sparse representation (LWSR), which aims for good preservation of the locality structure among data objects and a significant reduction of the computation time. LWSR approximates each object as a sparse linear combination of its nearest neighbors, and weights their corresponding coefficients by their distances to the target object. Experimental results show that LWSR-graph based learning methods outperform state-of-the-art methods in both effectiveness and efficiency for graph-based learning. [ABSTRACT FROM AUTHOR]

Topic models, such as probabilistic latent semantic analysis (PLSA) and latent Dirichlet allocation (LDA), have shown impressive success in many fields. Recently, multi-view learning via probabilistic latent semantic analysis (MVPLSA), is also designed for multi-view topic modeling. These approaches are instances of generative model, whereas they all ignore the manifold structure of data distribution, which is generally useful for preserving the nonlinear information. In this paper, we propose a novel multiple graph regularized generative model to exploit the manifold structure in multiple views. Specifically, we construct a nearest neighbor graph for each view to encode its corresponding manifold information. A multiple graph ensemble regularization framework is proposed to learn the optimal intrinsic manifold. Then, the manifold regularization term is incorporated into a multi-view topic model, resulting in a unified objective function. The solutions are derived based on the Expectation Maximization optimization framework. Experimental results on real-world multi-view data sets demonstrate the effectiveness of our approach. [ABSTRACT FROM AUTHOR]

Document embedding is a technology that captures informative representations from high-dimensional observations by some structure-preserving maps over corpus and has been intensively explored in machine learning. Recently, some manifold-inspired embedding methods become a hot topic, mainly due to their ability in capturing discriminative embedding. However, the existing methods capture the embeddings based on the geometrical information of nearest neighbors without considering the intrinsic documents-generating structure on a subspace, thus leads to a limitation to uncover intrinsic semantic information. In this paper, we propose a semi-supervised local-invariant method, called Discriminative Locally Document Embedding (Disc-LDE), aiming to build a smooth affine map for document embedding by preserving documents-generating structure on a subspace. Disc-LDE models the documents-generating structure as a pseudo-document by a generative probabilistic model of subspace, where the subspace is acquired by a transductive learning of multi-agent random walk on neighborhood graph, and regularizes the training of Auto-Encoders (AEs) to jointly recover the input document and its pseudo-document. Under a general regularized function learning framework, the regularized training can impact the parameterized encoder network become smooth to variations along the documents-generating structure of the local field on manifold. The experimental results on three widely-used corpora demonstrate Disc-LDE could efficient capture the intrinsic semantic structure to improve the clustering and classification performance to the state-of-the-arts methods. [ABSTRACT FROM AUTHOR]

Most existing algorithms for learning Bayesian networks (BNs) can be categorized as constraint-based or score-based methods. Constraint-based algorithms use conditional independence (CI) tests to judge the presence or absence of an edge. Though effective and applicable to (high-dimensional data) large-scale networks, CI tests require a large number of samples to determine the independencies. Thus they can be unreliable especially when the sample size is small. On the other hand, score-based methods employ a score metric to evaluate each candidate network structure, but they are inefficient in learning large-scale networks due to the extremely large search space. In this paper, we propose a new hybrid Bayesian network learning method, SAR (the abbreviation of S eparation A nd R eunion), which maintains the merits of both types of learning techniques while avoiding their drawbacks. Extensive experiments show that our method generally achieves better performance than state-of-the-art methods. [ABSTRACT FROM AUTHOR]

The task of document categorization is to classify documents from a stream as relevant or non-relevant to a particular user interest so as to reduce information overload. Existing solutions typically perform classification at the document level, i.e., a document is returned as relevant if at least a part of the document is of interest of the user. In this paper, we propose a novel multiple-instance stream learning framework for adaptive document categorization, named MIS-DC. Our proposed approach has the ability of making accuracy prediction at both the document level and the block level, while only requires labeling the training documents at the document level. In addition, our proposed approach can also provide adaptive document categorization by detecting and handling concept drift at a finer granularity when data streams evolve over time, thereby yielding higher prediction accuracy than existing data stream algorithms. Experiments on benchmark and real-world datasets have demonstrated the effectiveness of our proposed approach. [ABSTRACT FROM AUTHOR]

Uncertainty measures are critical evaluating tools in machine learning fields, which can measure the dependence and similarity between two feature subsets and can be used to judge the significance of features in classifying and clustering algorithms. In the classical rough sets, there are some uncertainty tools to measure a feature subset, including accuracy, roughness, information entropy, rough entropy, etc. These measures are applicable to discrete-valued information systems, but not suitable to real-valued data sets. In this paper, by introducing the neighborhood rough set model, each object is associated with a neighborhood subset, named a neighborhood granule. Several uncertainty measures of neighborhood granules are proposed, which are neighborhood accuracy, information quantity, neighborhood entropy and information granularity in the neighborhood systems. Furthermore, we prove that these uncertainty measures satisfy non-negativity, invariance and monotonicity. The maximum and minimum of these measures are also given. Theoretical analysis and experimental results show that information quantity, neighborhood entropy and information granularity measures are better than the neighborhood accuracy measure in the neighborhood systems. [ABSTRACT FROM AUTHOR]

Distance metric plays a significant role in machine learning methods(classification, clustering, etc.), especially in k -nearest neighbor classification( k NN), where the Euclidean distances are computed to decide the labels of unknown points. But Euclidean distance ignores the statistical structure which may help to measure the similarity of different inputs better. In this paper, we construct an unified framework, including two eigenvalue related methods, to learn data-dependent metric. Both methods aim to maximize the difference of intra-class distance and inter-class distance, but the optimization is considered in global view and local view respectively. Different from previous work in metric learning, our methods straight seek for equilibrium between inter-class distance and intra-class distance, and the linear transformation decomposed from the metric is to be optimized directly instead of the metric. Then we can effectively adjust the data distribution in transformed space and construct favorable regions for k NN classification. The problems can be solved simply by eigenvalue-decomposition, much faster than semi-definite programming. After selecting the top eigenvalues, the original data can be projected into low dimensional space, and then insignificant information will be mitigated or eliminated to make the classification more efficiently. This makes it possible that our novel methods make metric learning and dimension reduction simultaneously. The numerical experiments from different points of view verify that our methods can improve the accuracy of k NN classification and make dimension reduction with competitive performance. [ABSTRACT FROM AUTHOR]

Support vector machine is a classification model which has been widely used in many nonlinear and high dimensional pattern recognition problems. However, it is inefficient or impracticable to implement support vector machine in dealing with large scale training set due to its computational difficulties as well as the model complexity. In this paper, we study the support vector recognition problem mainly in the context of the reduction methods to reconstruct training set for support vector machine. We focus on the fact of uneven distribution of instances in the vector space to propose an efficient self-adaption instance selection algorithm from the viewpoint of geometry-based method. Also, we conduct an experimental study involving eleven different sizes of datasets from UCI repository for measuring the performance of the proposed algorithm as well as six competitive instance selection algorithms in terms of accuracy, reduction capabilities, and runtime. The extensive experimental results show that the proposed algorithm outperforms most of competitive algorithms due to its high efficiency and efficacy. [ABSTRACT FROM AUTHOR]

Contrast pattern-based classifiers are an important family of both understandable and accurate classifiers. Nevertheless, these classifiers do not achieve good performance on class imbalance problems. In this paper, we introduce a new contrast pattern-based classifier for class imbalance problems. Our proposal for solving the class imbalance problem combines the support of the patterns with the class imbalance level at the classification stage of the classifier. From our experimental results, using highly imbalanced databases, we can conclude that our proposed classifier significantly outperforms the current contrast pattern-based classifiers designed for class imbalance problems. Additionally, we show that our classifier significantly outperforms other state-of-the-art classifiers not directly based on contrast patterns, which are also designed to deal with class imbalance problems. [ABSTRACT FROM AUTHOR]

With the development of machine learning and data science, data sharing is very common between companies and research institutes to avoid data scarcity. However, sharing original datasets that contain private information can cause privacy leakage. A reliable solution is to utilize private synthetic datasets which preserve statistical information from original datasets. In this paper, we propose MC-GEN, a privacy-preserving synthetic data generation method under differential privacy guarantee for machine learning classification tasks. MC-GEN applies multi-level clustering and differential private generative model to improve the utility of synthetic data. In the experimental evaluation, we evaluated the effects of parameters and the effectiveness of MC-GEN. The results showed that MC-GEN can achieve significant effectiveness under certain privacy guarantees on multiple classification tasks. Moreover, we compare MC-GEN with three existing methods. The results showed that MC-GEN outperforms other methods in terms of utility. [ABSTRACT FROM AUTHOR]

The gradient-based meta learning and its approximation algorithms have been widely used in the few-shot scenarios. In practice, it is common for the trained meta-model to employ uniform settings for gradient descent steps across different tasks. However, the meta-model may be biased toward some tasks. The convergence issue occurs that some tasks may see convergence in a few steps while others fail to approach the optimum in the whole inner loop. The bias problem may cause the trained meta-model works well in some tasks but has unexpected bad performance in other tasks, which hurts the generality of the meta-model. To address this issue, in this paper, we formally establish the approximation between the metric-based strategy and gradient descent in meta-test. By directly calculating similarity to classify data, the trained meta-model avoids the convergence issue. We point out that the metric-based methods can closely approximate the gradient descent in meta-test if the representation capability of the derived features and the convergence of the inner loop during meta-training are guaranteed. Based on such observation, we propose a new meta-learning model GMT2 (Gradient-based Meta-Train with Metric-based meta-Test) by combining gradient descent in meta-training with metric-based methods in meta-test. GMT2 employs a new first-order approximation scheme using the adversarial update strategy which not only enhances the feature representation of inner layers, but also allows enough inner gradient steps without calculating second-order derivatives. Experiments show that GMT2 achieves better efficiency and competitive accuracy comparing with popular meta-learning models. [ABSTRACT FROM AUTHOR]

Feature selection (FS) is extensively applied in many machine learning applications for the selection of relevant features from data sets. A lot of unlabeled data are available in a variety of applications that can be exploited for semi-supervised FS to address the lack of labeled data and improve learning performance. Recently, semi-supervised sparse FS based on graph Laplacian has obtained considerable research interest, which uses the correlation between features in the process of FS. However, the Laplacian regularization has a weak extrapolating power and a bias towards the constant geodesic function, and cannot retain the local topology well. In this paper, a spline regression-based framework for semi-supervised sparse FS (SRS3FS) is proposed, which uses the mixed convex and non-convex ℓ 2 , p -norm (0 < p ≤ 1) regularization to select the relevant features and consider the correlation between features. The framework exploits local spline regression to retain the geometry structure of labeled and unlabeled data and encodes the data distribution. A unified iterative algorithm is presented to solve the proposed framework for the convex and non-convex cases, and its convergence is theoretically and experimentally proved. Experiments on several data sets illustrate the effectiveness of our framework in the selection of the most relevant and discriminative features. [ABSTRACT FROM AUTHOR]

Machine and deep learning survival models demonstrate similar or even improved time-to-event prediction capabilities compared to classical statistical learning methods yet are too complex to be interpreted by humans. Several model-agnostic explanations are available to overcome this issue; however, none directly explain the survival function prediction. In this paper, we introduce SurvSHAP(t), the first time-dependent explanation that allows for interpreting survival black-box models. It is based on SHapley Additive exPlanations with solid theoretical foundations and a broad adoption among machine learning practitioners. The proposed methods aim to enhance precision diagnostics and support domain experts in making decisions. Experiments on synthetic and medical data confirm that SurvSHAP(t) can detect variables with a time-dependent effect, and its aggregation is a better determinant of the importance of variables for a prediction than SurvLIME. SurvSHAP(t) is model-agnostic and can be applied to all models with functional output. We provide an accessible implementation of time-dependent explanations in Python at https://github.com/MI2DataLab/survshap. • We introduce the first time-dependent explanation for interpreting machine learning survival models. • The proposed SurvSHAP(t) method accurately explains the predicted survival function. • SurvSHAP(t) is able to detect time-dependent variable effects and its aggregation determines the local variable importance. • An open-source implementation of SurvSHAP(t) is available on GitHub. [ABSTRACT FROM AUTHOR]

Federated learning (FL) enables learning a global machine learning model from data distributed among a set of participating workers. This makes it possible (i) to train more accurate models due to learning from rich, joint training data and (ii) to improve privacy by not sharing the workers' local private data with others. However, the distributed nature of FL makes it vulnerable to targeted poisoning attacks that negatively impact on the integrity of the learned model while, unfortunately, being difficult to detect. Existing defenses against those attacks are limited by assumptions on the workers' data distribution and/or are ill-suited to high-dimensional models. In this paper, we analyze targeted attacks against FL, specifically label-flipping and backdoor attacks, and find that the neurons in the last layer of a deep learning (DL) model that are related to these attacks exhibit a different behavior from the unrelated neurons. This makes the last-layer gradients valuable features for attack detection. Accordingly, we propose FL-Defender to combat FL targeted attacks. It consists of (i) engineering robust discriminative features by calculating the worker-wise angle similarity for the workers' last-layer gradients, (ii) compressing the resulting similarity vectors using PCA to reduce redundant information, and (iii) re-weighting the workers' updates based on their deviation from the centroid of the compressed similarity vectors. Experiments on three data sets show the effectiveness of our method in defending against label-flipping and backdoor attacks. Compared to several state-of-the-art defenses, FL-Defender achieves the lowest attack success rates while maintaining the main task accuracy. [ABSTRACT FROM AUTHOR]

For addressing the "One-Class Collaborative Filtering" (OCCF) problem in recommendation systems, in which the obtained user information is all single-type positive feedback, the current mainstream methods are all based on the idea of pairwise preference learning. The Bayesian Personalized Ranking (BPR) method is a classical representative of such an idea. However, the assumption in BPR, that "users tend to prefer items that they have once interacted", may not always hold in reality. This is because for non-interacted items, a user may have different perspectives, such as potentially favorite, dislike, or something in between. For mitigating the above-mentioned issue, this paper proposes a M ulti-pairwise preference and S imilarity based BPR method, termed as MSBPR for brevity. MSBPR utilizes additional auxiliary feedback information to excavate and infer deeper-level user preferences. Subsequently, MSBPR borrows ideas from traditional item/user-based collaborative filtering methods to further divide non-interacted items from the angle of item/user, respectively. Afterwards, MSBPR constructs four preferences on top of the divided items, and accordingly builds up the multiple pairwise preference assumption. To optimize MSBPR, we derive an efficient learning algorithm based on the stochastic gradient descent algorithm. The computational complexity of MSBPR is also theoretically analyzed. Comprehensive experimental results demonstrate the effectiveness and efficiency of MSBPR over ten state-of-the-art methods on six benchmark real-world datasets. • We propose a M ulti-pairwise preference and Similarity based BPR method, MSBPR. • We propose three item preference classification strategies for MSBPR. • We propose a novel multi-pairwise preference assumption for MSBPR. • We derive an SGD-based learning algorithm for optimizing MSBPR. • MSBPR outperforms nine state-of-the-art models on four real datasets. [ABSTRACT FROM AUTHOR]

Path planning is a complicated optimization problem that is crucial for the safe flight of unmanned aerial vehicles (UAVs). Especially in the scenarios involving multiple UAVs, this problem is highly challenging due to the constraints of complex environments, various tasks and inherent UAV maneuverability. In this paper, a cooperative path planning model for multiple UAVs is presented. In addition to common limitations such as the path length minimization, UAV maneuverability limitation and collision avoidance, the communication requirements between UAVs and the impact of obstacles in the flight environment on the quality of communication are also taken into account in the presented path planning model. On this basis, the corresponding objective function is designed. Then, an improved particle swarm optimization (PSO) algorithm is proposed to solve the above path planning problem. Utilizing the ideas of the dynamic multi-swarm PSO (DMSPSO) algorithm and the comprehensive learning PSO (CLPSO) algorithm, the proposed algorithm, denoted as CL-DMSPSO, further improves the performance of both algorithms. The effectiveness and superiority of the novel CL-DMSPSO algorithm is verified on benchmark functions, especially for complex multimodal functions. Finally, we present an effective path planning method using CL-DMSPSO to generate optimized flyable paths for multiple UAVs. And simulation and comparison results on the designed scenario indicate the proposed UAV path planning method can efficiently plan high-quality paths for UAVs and demonstrate the advantages of the proposed CL-DMSPSO algorithm compared with other PSO algorithms in UAV path planning. • We put forward a cooperative path planning model for multiple UAVs. • The communication constraints are taken into account in the presented path planning model. • We propose an improved particle swarm optimization algorithm denoted as CL-DMSPSO. • Extensive simulation experiments are conducted on benchmark functions and the designed path planning scenario. [ABSTRACT FROM AUTHOR]

In this paper, we combine context-based Meta-Reinforcement Learning with task-aware representation to efficiently overcome data-inefficiency and limited generalization in the hyperparameter optimization problem. First, we propose a new context-based meta-RL model that disentangles task inference and control, which improves the meta-training efficiency and accelerates the learning process for unseen tasks. Second, the task properties are inferred on-line, which includes not only the dataset representation but also the task-solving experience, thus encouraging the agent to explore in a much smarter fashion. Third, we employ amortized meta-learning to meta-train the agent, which is simple and runs faster than the gradient-based meta-training method. Experimental results suggest that our method can search for the optimal hyperparameter configuration with limited computational cost in a reasonable time. [Display omitted] • A task encoder is proposed to achieve the task inference. • The rich task-specific information allows the agent to make better decision. • An amortized meta-learning approach is proposed to accelerate the meta-training. [ABSTRACT FROM AUTHOR]

Feature selection is an important preprocessing technique in the fields of data mining and machine learning. With the promotion of privacy protection awareness, recently it becomes a very practical and challenging issue to select high-quality feature subsets while ensuring the privacy of all participants. However, there is a lack of research results on this issue, i.e., feature selection under privacy protection. Aiming at the issue, this paper proposes a federated feature selection framework for the first time. In the framework, inspiring by the idea of federated learning, a credible third participant is introduced to process and integrate optimal feature subsets from multiple participants. On the basis of the framework, a federated evolutionary feature selection algorithm based on particle swarm optimization is proposed to effectively solve feature selection problems with multiple participants under privacy protection. Two new operators satisfying the requirement of privacy protection, i.e., the feature assembling strategy with multi-participant cooperation and the swarm initialization strategy guided by assembling solution, are designed to improve the ability of the proposed algorithm. Compared with several typical assembling feature selection algorithms on 15 data sets, experimental results show that the proposed algorithm can significantly improve the classification accuracy of the feature subset selected by each participant, while protecting the privacy of data. [ABSTRACT FROM AUTHOR]

The prediction of missing links of open knowledge graphs (OpenKGs) poses unique challenges compared with well-studied curated knowledge graphs (CuratedKGs). Unlike CuratedKGs whose entities are fully disambiguated against a fixed vocabulary, OpenKGs consist of entities represented by non-canonicalized free-form noun phrases and do not require an ontology specification, which drives the synonymity (multiple entities with different surface forms have the same meaning) and sparsity (a large portion of entities with few links). How to capture synonymous features in such sparse situations and how to evaluate the multiple answers pose challenges to existing models and evaluation protocols. In this paper, we propose VGAT, a variational autoencoder densified graph attention model to automatically mine synonymity features, and propose CR, a cluster ranking protocol to evaluate multiple answers in OpenKGs. For the model, VGAT investigates the following key ideas: (1) phrasal synonymity encoder attempts to capture phrasal features, which can automatically make entities with synonymous texts have closer representations; (2) neighbor synonymity encoder mines structural features with a graph attention network, which can recursively make entities with synonymous neighbors closer in representations. (3) densification attempts to densify the OpenKGs by generating similar embeddings and negative samples. For the protocol, CR is designed from the significance and compactness perspectives to comprehensively evaluate multiple answers. Extensive experiments and analysis show the effectiveness of the VGAT model and rationality of the CR protocol. • Propose a new model to automatically mine synonymous features. • Design a novel evaluation protocol to evaluate multiple answers. • Densify OpenKGs by variational autoencoder and negative samples. • Improve the performance of link prediction. [ABSTRACT FROM AUTHOR]