Search

Around a third of the total population of Europe suffers from mental disorders. The use of electroencephalography (EEG) together with Machine Learning (ML) algorithms to diagnose mental disorders has recently been shown to be a prominent research area, as exposed by several reviews focused on the field. Nevertheless, previous to the application of ML algorithms, EEG data should be correctly preprocessed and prepared via Feature Engineering (FE). In fact, the choice of FE techniques can make the difference between an unusable ML model and a simple, effective model. In other words, it can be said that FE is crucial, especially when using complex, non-stationary data such as EEG. To this aim, in this paper we present a Systematic Mapping Study (SMS) focused on FE from EEG data used to identify mental disorders. Our SMS covers more than 900 papers, making it one of the most comprehensive to date, to the best of our knowledge. We gathered the mental disorder addressed, all the FE techniques used, and the Artificial Intelligence (AI) algorithm applied for classification from each paper. Our main contributions are: (i) we offer a starting point for new researchers on these topics, (ii) we extract the most used FE techniques to classify mental disorders, (iii) we show several graphical distributions of all used techniques, and (iv) we provide critical conclusions for detecting mental disorders. To provide a better overview of existing techniques, the FE process is divided into three parts: (i) signal transformation, (ii) feature extraction, and (iii) feature selection. Moreover, we classify and analyze the distribution of existing papers according to the mental disorder they treat, the FE processes used, and the ML techniques applied. As a result, we provide a valuable reference for the scientific community to identify which techniques have been proven and tested and where the gaps are located in the current state of the art. [ABSTRACT FROM AUTHOR]

Mining user data and constructing web user profiles of older adults from the perspective of elderly services is conducive to understanding their behavioral habits, needs, and usage preferences on the web, which provides more targeted elderly care services. In this paper, IGA-SOMK + + , which is a novel clustering method for constructing web user profiles of older adults, is proposed based on the China Family Panel Studies (CFPS) survey data, which include 6596 older adults aged greater than 60 years. The selected data aspects include basic information, work situation, health situation, living habits, and web use services. To describe the web user profiles of older adults, a hybrid method based on improved genetic algorithm (IGA) feature selection, self-organizing feature maps (SOM), and K-means + + is proposed. Data on older adults' web use behaviors are first processed, and IGA is used for feature selection based on the adaptive crossover and mutation probabilities. SOM is then used to determine the initial center vectors of K-means + + for further clustering, which is referred to as SOMK + + (SOM-K-means + +). The results of IGA-SOMK + + are compared with those of the state-of-the-art methods, including the K-means, mini batch K-means, Agnes, K-modes, FCM, K-means + + , SOMK + + , and IHPSO-KM. In addition, the significance and robustness of IGA-SOMK + + are analyzed. The experimental results show that the IGA feature selection reduces the influence of the redundant feature factors and improves the performance of the clustering algorithm. SOMK + + overcomes the sensitivity of K-means to initial cluster centers. Moreover, IGA-SOMK + + has the best clustering effect among the compared algorithms in terms of silhouette coefficient (SC), calinski-harabaz (CH) index, and davies-bouldin (DB) metrics. For example, it increases the SC from 0.280 to 0.629. Finally, by analyzing the results, the user group of older adults is segmented to perform the deep mining of CFPS data, which verifies the feasibility of the user profile model. This paper summarizes the basic situation of the current web access of older adults in China in terms of web use services, as well as the importance of the web in their lives and in the information channels. It also provides suggestions for the current problems of older adults in accessing the web. [ABSTRACT FROM AUTHOR]

Label distribution learning, where deal with label ambiguity by describing the degree of relevance of each label to a specific instance. As a novel machine learning paradigm, the curse of dimensionality is one of the prominent problems. Feature selection is a vital preprocessing step to reduce the high dimensionality of data. However, most existing label distribution feature selection methods focus on selecting a feature subset that has relevant capabilities for all labels, ignoring label-specific features with the maximum discriminatory power for each label. To tackle this issue, a label distribution feature selection algorithm based on label-specific features is proposed in this paper. Initially, we introduce a feature selection optimization model for label distribution data that simultaneously considers common and label-specific features, leveraging sparse learning to further investigate the intricate relationships between features and labels. Subsequently, the label correlation coefficient is employed to enhance the collaborative learning effect of labels. Finally, the relevance between features and labels is taken into account to guide the feature selection process, which can effectively eliminate the redundant features. Comprehensive experiments demonstrate the advantage of our proposed method over other well-established feature selection algorithms for selecting label-specific features to label distribution data. [ABSTRACT FROM AUTHOR]

This paper deals with a major challenge in cyber-security: the need to respond to ever renewed techniques used by attackers in order to avoid detection based on analysing static features of malware. These constantly renewed techniques consist of various changes in file geometry, entropy a.s.o. As a consequence, static malware features sets describe less and less accurately the malicious files; hence, the performance of machine learning models in detecting new variants of the same malware family may be severely impaired. The paper focuses on a promising approach to this detection challenge: defining file features based on OS (operating system) API (Application Program Interface) calls sequences. We explore in detail the detection potential of such features, since, in order to act maliciously, these features are highly unlikely to be hidden. We studied several tens of thousands of such features, a modest-sized subset of which were subsequently fed to several machine learning models. The database used for training and testing consists of 1.5 million files, including malicious files from the polymorphic families Emotet and Trickbot. Using this database, nearly 4,000 pairings (classifier, feature selection algorithm) were trained / tested. Our experimental results show that the API (Application Program Interface) calls-oriented feature mining process is well suited for detecting polymorphic malware. A comparative discussion of the detection results of the various models is presented; depending on the target optimisation criterion (detection rate / false positive rate / saving resources), three of the 4,000 classification models turn out to be best suited for real-world applications: Random Forrest, Legacy Neural Networks and Decision Tree. [ABSTRACT FROM AUTHOR]

Speech emotion recognition (SER) has gained significant attention due to its several application fields, such as mental health, education, and human-computer interaction. However, the accuracy of SER systems is hindered by high-dimensional feature sets that may contain irrelevant and redundant information. To overcome this challenge, this study proposes an iterative feature boosting approach for SER that emphasizes feature relevance and explainability to enhance machine learning model performance. Our approach involves meticulous feature selection and analysis to build efficient SER systems. In addressing our main problem through model explainability, we employ a feature evaluation loop with Shapley values to iteratively refine feature sets. This process strikes a balance between model performance and transparency, which enables a comprehensive understanding of the model's predictions. The proposed approach offers several advantages, including the identification and removal of irrelevant and redundant features, leading to a more effective model. Additionally, it promotes explainability, facilitating comprehension of the model's predictions and the identification of crucial features for emotion determination. The effectiveness of the proposed method is validated on the SER benchmarks Toronto emotional speech set (TESS), Berlin Database of Emotional Speech (EMO-DB), Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS), and Surrey Audio-Visual Expressed Emotion (SAVEE) dataset, outperforming state-of-the-art methods. These results highlight the potential of the proposed technique in developing accurate and explainable SER systems. To the best of our knowledge, this is the first work to incorporate model explainability into an SER framework. The source code of this paper is publicly available via this https://github.com/alaaNfissi/Unveiling-Hidden-Factors-Explainable-AI-for-Feature-Boosting-in-Speech-Emotion-Recognition. [ABSTRACT FROM AUTHOR]

Good prognostic health management (PHM) plays a crucial role in industrial production and other fields. The accurate prediction of remaining useful life (RUL) can ensure good working condition of machines, and the selection of health indicators (HI) is the key to prediction. Due to the length and noise of the data, the selection of features requires a lot of prior knowledge. Therefore, a novel convolution-based attention mechanism bidirectional long and short-term memory (CABLSTM) network is proposed to achieve the end-to-end lifetime prediction of rotating machinery in this paper. Unlike concatenating two networks, the model in this paper is used to convolute the cell states of Bi-LSTM. Firstly, the input signal is performed through CNN to obtain feature information. Secondly, the obtained features are fed into the Bi-LSTM network with attention mechanism for convolution operation to obtain time-frequency information to construct HI. Finally, the training data are normalized to predict RUL. Bi-LSTM can capture features in longer time-frequency information, and attention mechanism can give input influence weight, and highlight its effective characteristics to obtain better prediction accuracy. The complex process of feature extraction, HI construction, and RUL prediction is combined in one algorithm by deep learning. To verify the performance of the method in this paper, experiments were conducted on the bearing dataset of PRONOSTIA, and compared with other methods. The results showed that the method outperforms other methods due to its better accuracy and prediction precision. [ABSTRACT FROM AUTHOR]

Rough set theory is a mathematical tool widely employed in various fields to handle uncertainty. Feature selection, as an essential and independent research area within rough set theory, aims to identify a small subset of important features by eliminating irrelevant, redundant, or noisy ones. In human life, data characteristics constantly change over time and other factors, resulting in ordered datasets with varying features. However, existing feature extraction methods are not suitable for handling such datasets since they do not consider previous reduction results when features change and need to be recomputed, leading to significant time consumption. To address this issue, the incremental attribute reduction algorithm utilizes prior reduction results effectively reducing computation time. Motivated by this approach, this paper investigates incremental feature selection algorithms for ordered datasets with changing features. Firstly, we discuss the dominant matrix and the dominance conditional entropy while introducing update principles for the new dominant matrix and dominance diagonal matrix when features change. Subsequently, we propose two incremental feature selection algorithms for adding (IFS-A) or deleting (IFS-D) features in ordered data set. Additionally, nine UCI datasets are utilized to evaluate the performance of our proposed algorithm. The experimental results validate that the average classification accuracy of IFS-A and IFS-D under four classifiers on twelve datasets is 82.05% and 80.75%, which increases by 5.48% and 3.68% respectively compared with the original data. [ABSTRACT FROM AUTHOR]

In the real world, data often exhibits high-dimensional and complex characteristics. In addition, an object may correspond to multiple class labels. Therefore, filtering and processing such data has become a hot research topic. Multi-label feature selection is one of the key preprocessing techniques for effectively solving redundant information. However, due to the high complexity of multi-label learning feature selection and involving optimization of multiple objectives, careful consideration is required to achieve effective optimization. To address these issues, this paper proposes a multiple reference points-based multi-objective feature selection for multi-label learning. Firstly, the neighborhood crossover strategy utilizes individuals in the same neighborhood for crossover, thereby enhancing the search effect within that decision space. The distribution-based mutation strategy adjusts the mutation probability of different features based on their distribution in selected candidate solutions to increase the likelihood of potentially useful features being selected. Secondly, the multiple reference points-based selection strategy uses a scalar function to uniformly evaluate offspring individuals and control their distance from each other through multiple reference points. This approach helps in selecting convergent solutions with balanced diversity. Finally, experimental results demonstrate that the proposed algorithm attains an optimal performance of 83% across fourteen widely used datasets, outperforming ten state-of-the-art algorithms in terms of effectiveness and performance in multi-label feature selection. [ABSTRACT FROM AUTHOR]

Currently, learning physiological vital signs such as blood pressure (BP), hemoglobin levels, and oxygen saturation, from Photoplethysmography (PPG) signal, is receiving more attention. Despite successive progress that has been made so far, continuously revealing new aspects characterizes that field as a rich research topic. It includes a diverse number of critical points represented in signal denoising, data cleaning, employed features, feature format, feature selection, feature domain, model structure, problem formulation (regression or classification), and model combinations. It is worth noting that extensive research efforts are devoted to utilizing different variants of machine learning and deep learning models while transfer learning is not fully explored yet. So, in this paper, we are introducing a per-beat rPPG-to-BP mapping scheme based on transfer learning. An interesting representation of a 1-D PPG signal as a 2-D image is proposed for enabling powerful off-the-shelf image-based models through transfer learning. It resolves limitations about training data size due to strict data cleaning. Also, it enhances model generalization by exploiting underlying excellent feature extraction. Moreover, non-uniform data distribution (data skewness) is partially resolved by introducing logarithmic transformation. Furthermore, double cleaning is applied for training contact PPG data and testing rPPG beats as well. The quality of the segmented beats is tested by checking some of the related quality metrics. Hence, the prediction reliability is enhanced by excluding deformed beats. Varying rPPG quality is relaxed by selecting beats during intervals of the highest signal strength. Based on the experimental results, the proposed system outperforms the state-of-the-art systems in the sense of mean absolute error (MAE) and standard deviation (STD). STD for the test data is decreased to 5.4782 and 3.8539 for SBP and DBP, respectively. Also, MAE decreased to 2.3453 and 1.6854 for SBP and DBP, respectively. Moreover, the results for BP estimation from real video reveal that the STD reaches 8.027882 and 6.013052 for SBP and DBP, respectively. Also, MAE for the estimated BP from real videos reaches 7.052803 and 5.616028 for SBP and DBP, respectively. Proposed camera-based blood pressure monitoring system [ABSTRACT FROM AUTHOR]

In the era of big data, the amount of class labels is growing rapidly, which poses a great challenge to classification tasks. The hierarchical classification was thus introduced to address this issue by considering the structural information between different class labels. In this paper, we propose an incremental feature selection algorithm for handling the arrival of new samples by using the theory of fuzzy rough sets. As a preliminary step, we propose a non-incremental hierarchical feature selection algorithm, which is an improved version of the existing method. Then utilizing the sibling strategy, the incremental calculation of the dependency degree at the arrival of samples is discussed. Based on the analysis of dependency degree change, we design feature addition and deletion strategies, as well as the incremental feature selection algorithm. In the experimental section, two versions of algorithms are designed. The experimental results show that our improvement of the existing method is highly effective and can significantly accelerate the process of feature selection. In addition, version 2 of the incremental algorithm exhibits much higher efficiency than the improved non-incremental algorithm on several datasets, as well as the existing method. Compared to six hierarchical feature selection algorithms, our algorithm achieves better results on the classification accuracy and three hierarchical evaluation metrics. The effectiveness and efficiency of version 1 are also verified by the comparison of version 2 and other results. [ABSTRACT FROM AUTHOR]

As a reliable and valid tool for analyzing uncertain information, fuzzy rough set theory has attracted widespread concern in feature selection. However, the performance of fuzzy rough set model is generally affected by various factors, for instance, large data distribution differences, unreasonable settings for fuzzy information granules and feature evaluation functions with single perspective. Considering these problems, a fitting fuzzy rough set model with relative dependency complement mutual information is proposed in this paper. First, the relative distance is introduced to eliminate the influence of data distribution on the fuzzy similarity relation. Then, by analyzing the similarity distributions of samples with regard to decisions, a fitting fuzzy neighborhood radius is proposed to improve the fuzzy information granules, and a fitting fuzzy rough set model is proposed based on the relative distance and the fitting fuzzy neighborhood radius. Moreover, considering the complementary characteristics between fuzzy information granularity and fuzzy information entropy, the related definitions of relative complement information entropy in the fitting fuzzy rough set model are offered, and a multiview uncertainty measure based on relative dependency complement mutual information is constructed to comprehensively analyze the uncertainty of information. Finally, a heuristic feature selection algorithm is designed. A series of experiments designed in this paper prove the superiority of the proposed method. [ABSTRACT FROM AUTHOR]

Improving classification performance is an essential goal for various practical applications. Feature selection has become an important data preprocessing step in machine learning systems. However, many effective methods based on heuristic search strategies have the problem of high running costs. This paper proposes an efficient multiobjective feature selection method based on artificial immune algorithm optimization. It introduces a clone selection algorithm to explore the search space of optimal feature subsets. According to the target requirements of feature selection, combined with biological research results, this method introduces genome shuffling technology and a conditional lethal mutation mechanism to improve the search performance of the algorithm. Experimental comparisons are conducted on 21 benchmark datasets with 17 advanced feature selection methods in terms of classification accuracy, the number of feature subsets, and computational cost. The results show that the algorithm achieves the smallest number of selected features (only 3.26% compared to the lowest) and better average classification accuracy with a much lower average computational cost than others (only 3.62% compared to the lowest). [ABSTRACT FROM AUTHOR]

In Machine Learning (ML), Feature Selection (FS) plays a crucial part in reducing data's dimensionality and enhancing any proposed framework's performance. However, in real-world applications, FS work suffers from high dimensionality, computational and storage complexity, noisy or ambiguous nature, high performance, etc. The area of FS is very vast and challenging in its nature. There are lots of work that have been reported on FS over the various area of applications. This paper has discussed FS's framework and the multiple models of FS with detailed descriptions. We have also classified the various FS algorithms with respect to the data, i.e., structured or labeled data and unstructured data for the different applications of ML. We have also discussed what essential features are, the commonly used FS methods, the widely used datasets, and the widely used work done in the various ML fields for the FS task. Here we try to view the multiple comparison experimental results of FS work in different result discussions. This paper draws a descriptive survey on FS with the associated area of real-world problem domains. This paper's main objective is to understand the main idea of FS work and identify the core idea of how FS will be applicable in various problem domains. [ABSTRACT FROM AUTHOR]

A decision table consists of samples, categorical features, and a decision feature. Uncertainty measurement (UM) can supply new points of view for analyzing data. Thus, it is vital to study the uncertainty of the decision table. Some UMs, such as classification precision, rough membership degree, dependence degree, and attribute importance, cannot accurately measure the uncertainty of a decision table. For example, the dependence degree only considers the information provided by the lower approximation of the decision and ignores the upper approximation, which may lead to some information loss. This paper proposes new UMs in a decision table and gives an application for feature selection. First, new UMs such as conditional information entropy, conditional information quantity, and conditional discriminant index in a decision table are proposed. Then, statistical analysis is used to identify the strengths and weaknesses of the proposed UMs. Next, the UM with the best performance is applied to create a heuristic select feature algorithm in a decision table. Finally, the created algorithm is compared to five other feature selection algorithms, and numerical experiments demonstrate its superior performance. [ABSTRACT FROM AUTHOR]

The classification of cancer using established biological knowledge has become increasingly prevalent, primarily due to the improved accuracy and enhanced biological interpretability this method offers for classification outcomes. Despite these advances, current cancer classification methods encounter challenges in maintaining the intricate structure of gene networks and leveraging the statistical information embedded within gene data. In this paper, we introduce an adaptive hypergraph regularized logistic regression model that capitalizes on established biological knowledge and statistical information within gene data. Specifically, our model integrates a hypergraph into the objective function, an innovation that preserves the complex gene network structure more effectively. Additionally, we implement adaptive penalties in the penalty term, which facilitates the targeted selection of disease-related genes based on gene weights. To further refine our model, we incorporate constraints on gene pairs with high statistical correlations within the penalty term, thereby minimizing the inclusion of redundant genes. We adopt the block coordinate descent algorithm to address the nonconvexity of our model. Through comparative experimentation with established methodologies on real datasets, our proposed model demonstrates marked improvement in classification accuracy and adept selection of genes pertinent to specific diseases. [ABSTRACT FROM AUTHOR]

The artificial immune system and network anomaly detection system are developed with common goals and principles considered. Moreover, artificial immune-based network anomaly detection can adaptively learn and dynamically detect threats. However, existing immune recognition algorithms suffer from the curse of dimensionality, hole problems, and detector inefficiency tolerance. In this paper, we proposed a novel immune detector training mechanism for network anomaly detection. First, a hybrid filter embedded feature selection algorithm is designed to comprehensively evaluate features and select the optimal subset. Then, candidate detectors are generated based on self antigens, and the nonself region is represented using complementary space to circumvent the hole problem. Finally, considering the training efficiency during the evolution of the candidate detectors, an antigen clustering feature tree is constructed to rapidly index the tolerance objects. Furthermore, the algorithm considers the effect of the collaboration of multiple mature detectors on candidate detectors, and a Monte Carlo-based coverage estimation algorithm is designed to achieve more accurate and fine-grained maturation tolerance of candidate detectors. The theoretical analysis shows that the time complexity of our algorithm is significantly reduced. The experimental results show that our algorithm not only improves the detection accuracy but also reduces the time cost of detector training. [ABSTRACT FROM AUTHOR]

In the era of big data, the demand for multivariate time series prediction has surged, drawing increased attention to feature selection and neural networks in machine learning. However, certain feature selection methods neglect the alignment between actual data sample differences and clustering results, while neural networks lack automatic parameter adjustment in response to changing target features. This paper presents the MBSSA-Bi-AESN model, a Bi-directional Adaptive Echo State Network that utilizes the modified salp swarm algorithm (MBSSA) and feature selection to address the limitations of manually set parameters. Initial feature subset selection involves assigning weights based on the consistency of clustering results with differences. Subsequently, the four critical parameters in the Bi-AESN model are optimized using MBSSA. The optimized Bi-AESN model and selected feature subset are then integrated for simultaneous model learning and optimal feature subset selection. Experimental analysis on eight datasets demonstrates the superior prediction accuracy of the MBSSA-Bi-AESN model compared to benchmark models, underscoring its feasibility, validity, and universality. [ABSTRACT FROM AUTHOR]

Mutual information is fundamental for feature selection, and relevant conditional and joint mutual fuzzy granularities (MFGs) characterize feature correlation and redundancy in fuzzy decision systems, respectively. Recently, conditional MFGs have been utilized to generate three-stage heuristic measures, and thus, two feature selection algorithms (called GFMRI-FS and NGFMRI-FS) have emerged. However, these methods emphasize correlation connections but neglect redundancy deletion, thus the approach can be improved. In this paper, a joint MFG with redundant feature information is supplemented to reasonably modify existing heuristic measures, and new three-stage heuristic measures are constructed to obtain two improved selection algorithms (called C-GFMRI-FS and C-NGFMRI-FS). First, the measurement mechanisms of the MFGs are analyzed, and three-stage correctional heuristic measures are established to acquire the dependency semantics, bound combinations, size relationships, theoretical extensions, applied optimization, parametric monotonicity, and systematic algorithms. Then, the three-stage current and correctional heuristic measures obtain granulation monotonicity/certainty and nonmonotonicity/uncertainty, respectively, as well as uncertainty fluctuations on feature-dynamic subset chains. Furthermore, feature selection is investigated by correctional heuristic measures, and two selection algorithms (i.e., C-GFMRI-FS and C-NGFMRI-FS) are designed. Finally, table examples and data experiments validate the uncertainty heuristic measures and feature selection algorithms; accordingly, C-GFMRI-FS and C-NGFMRI-FS outperform GFMRI-FS and NGFMRI-FS and achieve better classification results. This study fuses conditional and joint MFGs to make systematic improvements in heuristic measurements and feature selection, thus facilitating data learning. [ABSTRACT FROM AUTHOR]

In recent years, Deep Neural Networks (DNN) have emerged as a practical method for image recognition. The raw data, which contain sensitive information, are generally exploited within the training process. However, when the training process is outsourced to a third-party organization, the raw data should be desensitized before being transferred to protect sensitive information. Although masks are widely applied to hide important sensitive information, preventing inpainting masked images is critical, which may restore the sensitive information. The corresponding models should be adjusted for the masked images to reduce the degradation of the performance for recognition or classification tasks due to the desensitization of images. In this paper, we propose a mask-based image desensitization approach while supporting recognition. This approach consists of a mask generation algorithm and a model adjustment method. We propose exploiting an interpretation algorithm to maintain critical information for the recognition task in the mask generation algorithm. In addition, we propose a feature selection masknet as the model adjustment method to improve the performance based on the masked images. Extensive experimentation results based on multiple image datasets reveal significant advantages (up to 9.34% in terms of accuracy) of our approach for image desensitization while supporting recognition. [ABSTRACT FROM AUTHOR]

Meta-heuristic algorithms have been extensively utilized in feature selection tasks because they can obtain the global optimal solution. However, the meta-heuristic algorithm will take too much time in the face of a large number of samples. Although most of the studies compromise to approximate optimal solutions for avoiding time-consuming problems, a new problem with reduced classification performance, especially classification stability, is then generated. Aiming to above problems, this paper proposes a new feature selection framework. First, this framework exploits a voting ensemble strategy to improve classification stability by reducing the impact of misclassified labels on the overall classification results. Second, the framework uses a data perturbation strategy to enhance classification accuracy. In particular, the data perturbation strategy is able to generate more neighborhood relationships in the dataset, which could reveal the distribution of various features of the samples. A voting ensemble of different feature distributions is capable of extracting more information from the dataset, then the initially misclassified samples are more likely to be returned to the correct classification. Third, the framework takes a random sampling accelerator into account to solve the problem of excessive time consumption by reducing the size of the search sample space. Finally, for the sake of verifying the effectiveness of the proposed framework, four meta-heuristic feature selection methods based on a neighborhood rough set are compared on 20 datasets. The experimental results indicate that our framework could improve classification performance and accelerate feature selection, particularly in confronting large sample sizes. [ABSTRACT FROM AUTHOR]

In general, fully exploiting the local structure of the original data space can effectively improve the clustering performance of nonnegative matrix factorization (NMF). Therefore, graph-based NMF algorithms have been widely studied and applied. However, traditional graph-based NMF methods generally employ predefined models to construct similarity graphs, so that the clustering results depend heavily on the quality of the similarity graph. Furthermore, most of these methods follow the ideal assumption that the importance of different features is equal in the process of learning the similarity matrix, which results in irrelevant features being valued. To alleviate the above issues, this paper develops an adaptive graph regularized nonnegative matrix factorization with self-weighted learning (SWAGNMF) method. Firstly, the proposed method learns the similarity matrix flexibly and adaptively to explore the local structure of samples based on the assumption that data points with smaller distances should have a higher probability of adjacency. Furthermore, the self-weight matrix assigns different weights automatically according to the importance of features in the process of constructing similarity graph, i.e., discriminative features are assigned more significant weights than redundant features, which can effectively suppress irrelevant features and enhance the robustness of our model. Finally, considering the duality between samples and features, the proposed method is capable of exploring the local structures of both the data space and the feature space. An effective alternative optimization algorithm is proposed, and convergence is theoretically guaranteed. Extensive experiments on benchmark and synthetic datasets show that the proposed method outperforms compared state-of-the-art clustering methods. [ABSTRACT FROM AUTHOR]

The citation count is an important factor to estimate the relevance and significance of academic publications. However, it is not possible to use this measure for papers which are too new. A solution to this problem is to estimate the future citation counts. There are existing works, which point out that graph mining techniques lead to the best results. We aim at improving the prediction of future citation counts by introducing a new feature. This feature is based on frequent graph pattern mining in the so-called citation network constructed on the basis of a dataset of scientific publications. Our new feature improves the accuracy of citation count prediction, and outperforms the state-of-the-art features in many cases which we show with experiments on two real datasets. [ABSTRACT FROM AUTHOR]

Generalized low-rank approximation of matrix (GLRAM) is a multi-linear learning method and has been widely concerned due to its outstanding performance. It can make full use of the spatial information of features and save space and time complexity by directly dealing with matrix data. However, the existing methods may have trivial solutions and cannot consider the local structure of sample data. In this paper, we propose a sparse low-rank approximation of matrix and local preserving model for unsupervised image feature selection. This model mainly employ the L2,1-norm sparse regularization of Kronecker product of two transformation matrices in GLRAM to avoid trivial solution and implement feature selection for images. Moreover, the weighted distance between the simplified representations of the original matrix data is used to preserve the local structure of image data into the transformation space. An iterative optimization algorithm is designed to efficiently solve the proposed model and the convergence and complexity of algorithm are analyzed. Comprehensive experiments on several image datasets demonstrate the effectiveness and superiority of the proposed method. [ABSTRACT FROM AUTHOR]

Log anomaly is a manifestation of a software system error or security threat. Detecting such unusual behaviours across logs in real-time is the driving force behind large-scale autonomous monitoring technology that can rapidly alert zero-day attacks. Increasingly, AI methods are being used to process voluminous log datasets and reveal patterns of correlated anomaly. In this paper, we propose an enhanced approach to learning semantic-aware embeddings for logs called the Subword Encoder Neural network (SEN). Solving upon a key limitation of previous semantic log parsing works, the proposed work introduces the concept of learning word vectors from subword-level granularity using an attention encoder strategy. The learnt embeddings reflect the contextual/lexical relationships at the word level. As a result, the learnt word representations precisely capture new log messages previously not seen by the model. Furthermore, we develop a novel feature distillation algorithm termed Naive Bayes Feature Selector (NBFS) to extract useful log events. This probabilistic technique examines the occurrence pattern of events to only select the salient ones that can aid anomaly detection. To our best knowledge, this is the first attempt to associate affinity to log events based on the target task. Since the predictions can be traced to the log messages, the AI is inherently explainable too. The model outperforms state-of-the-art methods by a fair margin. It achieves a 0.99 detection F1-score on the benchmarked BGL, HDFS and OpenStack log datasets. [ABSTRACT FROM AUTHOR]

As for single-label learning, feature selection has become a crucial data pre-processing tool for multi-label learning due to its ability to reduce the feature dimensionality to mitigate the influence of "the curse of dimensionality". In recent years, the multi-label feature selection (MFS) methods based on manifold learning and sparse regression have confirmed their superiority and gained more and more attention than other methods. However, most of these methods only consider exploiting the geometric structure distributed in the data manifold but ignore the geometric structure distributed in the feature manifold, resulting in incomplete information about the learned manifold. Aiming to simultaneously exploit the geometric information of data and feature spaces for feature selection, this paper proposes a novel MFS method named dual-graph based multi-label feature selection (DGMFS). In the framework of DGMFS, two nearest neighbor graphs are constructed to form a dual-graph regularization to explore the geometric structures of both the data manifold and the feature manifold simultaneously. Then, we combined the dual-graph regularization with a non-convex sparse regularization l2,1 − 2-norm to obtain a more sparse solution for the weight matrix, which can better deal with the redundancy features. Finally, we adopt the EM strategy to design an iterative updating optimization algorithm for DGMFS and provide the convergence analysis of the optimization algorithm. Extensive experimental results on ten benchmark multi-label data sets have shown that DGMFS is more effective than several state-of-the-art MFS methods. [ABSTRACT FROM AUTHOR]

In conventional multi-label learning, each training instance is associated with multiple available labels. Nevertheless, real-world objects usually exhibit more sophisticated properties such as abundant irrelevant features, incomplete labels, noisy labels, as well as class imbalance. Unfortunately, most existing multi-label learning algorithms only discussed one of them and failed to consider the confounding effects of these factors, which will degrade the accuracy of multi-label classification. In this paper, we propose an integrated multi-label learning framework ML-INC that trains the multi-label model while addressing the aforementioned issues simultaneously. Specifically, we first decompose the observed label matrix into an incomplete ground-truth label matrix and a noisy label matrix by employing the low-rank and sparse decomposition scheme. Secondly, a label confidence matrix is learned to supplement the incomplete label matrix by utilizing the high-order label correlation and the label consistency. Additionally, the low-rank structure is adopted to capture the label correlation. Thirdly, a label regularization matrix is introduced to alleviate the effects of class imbalance in the label matrix, and a sparse constraint is imposed on the feature mapping matrix to select relevant discriminative features. Finally, the Alternating Direction Multiplier Method (ADMM) is employed to handle the optimization problem and comprehensive experiments are conducted to certify the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Classification is the most basic and representative problem in data mining and machine learning. In recent years, more and more evolutionary algorithms (EAs) have been used to handle classification problems due to their good global search capabilities. EA-based methods construct an optimization classification model by introducing a linear equations set. Thus, EAs can be used to directly solve classification problems. However, the structure of the existing optimization classification model is closely related to the dimension of datasets, and there are often a lot of redundant, irrelevant, or noise features in the datasets. These features may lead to complex calculation and low learning efficiency, thus affecting the classification performance. Feature selection has superior performance on solving this kind of problems. Therefore, for the optimization classification model, this paper introduces the concept of feature selection and proposes an improved optimization classification model based on both structure search and weight search. Brain storm optimization (BSO) algorithm is used to solve the optimization classification model. On one hand, BSO algorithm is used to search for the optimal classification model structure which is determined by the optimal feature subset. On the other hand, BSO is used to search for the optimal weight vectors under different structures. Thus, BSO is used to search for optimal structure and weights alternately. Experimental results show that feature selection can effectively simplify the structure of the classification model and improve the classification performance. In addition, the experimental results also verify the effectiveness of BSO on searching the corresponding weights under different structures, which can further improve the classification performance. [ABSTRACT FROM AUTHOR]

Feature selection is an important preprocessing process in machine learning. It selects the crucial features by removing irrelevant features or redundant features from the original feature set. Most of feature selection algorithms focus on maximizing relevant information and minimizing redundant information. In order to remove more redundant information in the evaluation criteria, we propose a feature selection based on mutual information with correlation coefficient (CCMI) in this paper. We introduce the correlation coefficient in the paper, and combine the correlation coefficient and mutual information to measure the relationship between different features. We use the absolute value of the correlation coefficient between two different features as the weight of the redundant item denoted by the mutual information in the evaluation standard. In order to select low redundancy features effectively, we also use the principle of minimization in the evaluation criteria. By comparing with 7 popular contrast algorithms in 12 data sets, CCMI has achieved the highest average classification accuracy for two classifiers of SVM and KNN. Experimental results show that our proposed CCMI has better feature classification capability. [ABSTRACT FROM AUTHOR]

With unlabeled and high-dimensional data explosion, unsupervised feature selection has become an essential step in many machine learning and data mining tasks. Many dictionary learning based models have been successfully developed for unsupervised feature selection in recent years. These models learn an over-complete dictionary to investigate more data distribution information. However, over-complete dictionary learning will generate redundancy in the latent representations for data. Moreover, if data contain noise, dictionary learning will also yield noise in the latent representations. In this paper, we propose a novel unsupervised feature selection framework, named dictionary learning for unsupervised feature selection via dual sparse regression. In this model, dictionary learning is first embedded into a sparse regression to learn an over-complete dictionary with sparse representations for data, in which the redundancy and noise are eliminated. The data are then projected to the representations to evaluate the significance of features using the other sparse regression. We also offer an efficient algorithm to solve this problem and theoretically analyze its convergence and computational complexity, which is proportional to the data dimensionality. Finally, the evaluation results with the k-means task utilizing the selected features on 9 benchmark datasets demonstrate the superiority of our approach in terms of effectiveness and efficiency. [ABSTRACT FROM AUTHOR]

Software Fault Prediction (SFP) is an important process to detect the faulty components of the software to detect faulty classes or faulty modules early in the software development life cycle. In this paper, a machine learning framework is proposed for SFP. Initially, pre-processing and re-sampling techniques are applied to make the SFP datasets ready to be used by ML techniques. Thereafter seven classifiers are compared, namely K-Nearest Neighbors (KNN), Naive Bayes (NB), Linear Discriminant Analysis (LDA), Linear Regression (LR), Decision Tree (DT), Support Vector Machine (SVM), and Random Forest (RF). The RF classifier outperforms all other classifiers in terms of eliminating irrelevant/redundant features. The performance of RF is improved further using a dimensionality reduction method called binary whale optimization algorithm (BWOA) to eliminate the irrelevant/redundant features. Finally, the performance of BWOA is enhanced by hybridizing the exploration strategies of the grey wolf optimizer (GWO) and harris hawks optimization (HHO) algorithms. The proposed method is called SBEWOA. The SFP datasets utilized are selected from the PROMISE repository using sixteen datasets for software projects with different sizes and complexity. The comparative evaluation against nine well-established feature selection methods proves that the proposed SBEWOA is able to significantly produce competitively superior results for several instances of the evaluated dataset. The algorithms' performance is compared in terms of accuracy, the number of features, and fitness function. This is also proved by the 2-tailed P-values of the Wilcoxon signed ranks statistical test used. In conclusion, the proposed method is an efficient alternative ML method for SFP that can be used for similar problems in the software engineering domain. [ABSTRACT FROM AUTHOR]

Feature selection based on neighborhood rough sets (NRSs) has become a popular area of research in data mining. However, the limitation that NRSs inherently ignore the differences between different classes hinders their performance and development, and feature evaluation functions for NRSs cannot reflect the relevance between features and decisions effectively. The above two points restrict the effect of NRSs on feature selection. Consequently, a feature selection algorithm using the rough mutual information in adaptive neighborhood rough sets (ANRSs) is presented in this paper. First, we propose the boundaries of samples to granulate all samples and build ANRSs model by combining the boundaries of all samples in the same class, and the model discovers the characteristics of each class from the given data and overcomes the inherent limitation of NRSs. Second, inspired by combining information and algebraic views, we naturally extend the mutual information that represents the information view to ANRSs and combine it with the roughness that represents the algebraic view to propose the rough mutual information to quantitatively compute the relevance between features and decisions. Third, a maximum relevance minimum redundancy-based feature selection algorithm using the rough mutual information is studied. Additionally, to decrease the time consumption of the algorithm in processing high-dimensional datasets and improve the classification accuracy of the algorithm, the Fisher score dimensionality reduction method is introduced into the designed algorithm. Finally, the experimental results of the designed algorithm based on ANRSs and various algorithms based on NRSs are compared on eleven datasets to demonstrate the efficiency of our algorithm. [ABSTRACT FROM AUTHOR]

One of the main problems with classifier training for imbalanced data is defining the correct learning criterion. On the one hand, we want the minority class to be correctly recognized, and on the other hand, we do not want to make too many mistakes in the majority class. Commonly used metrics focus either on the predictive quality of the distinguished class or propose an aggregation of simple metrics. The aggregate metrics, such as Gmean or AUC, are primarily ambiguous, i.e., they do not indicate the specific values of errors made on the minority or majority class. Additionally, improper use of aggregate metrics results in solutions selected with their help that may favor the majority class. The authors realize that a solution to this problem is using overall risk. However, this requires knowledge of the costs associated with errors made between classes, which is often unavailable. Hence, this paper will propose the semoos algorithm - an approach based on multi-objective optimization that optimizes criteria related to the prediction quality of both minority and majority classes. semoos returns a pool of non-dominated solutions from which the user can choose the model that best suits him. Automatic solution selection formulas with a so-called Pareto front have also been proposed to compare state-of-the-art methods. The proposed approach will train a svm classifier ensemble dedicated to the imbalanced data classification task. The experimental evaluations carried out on a large number of benchmark datasets confirm its usefulness. [ABSTRACT FROM AUTHOR]

Online streaming feature selection plays an important role in dealing with multi-dimensional data problems. Many online streaming feature selection algorithms have been combined with evolutionary algorithms (EA) and play an important role, however, most of them use single-objective optimization which has some limitations. Meanwhile, they ignore the interaction between features. The combination of features with each other may generates higher relevance. Therefore, this paper proposes a new online group feature selection algorithm PSO-NRS by fusing particle swarm optimization (PSO) algorithm and neighborhood rough set theory (NRS). PSO-NRS is able to select the set of features that are highly correlated with labels by combining features randomly. Using NRS for online feature selection does not require any domain knowledge, which makes PSO-NRS generalize better and can handle different types of data. PSO-NRS applies two layers of filtering for online feature selection. In the first filtering layer, two objective functions are designed and multi-objective optimization by particle swarm is used to select the set of features with the highest relevance. In the second filtering layer, a search strategy is defined using a rough set-based evaluation method to complete the final feature selection. The interactions between features are considered and redundant features are removed during the two filtering layers. Finally, PSO-NRS is experimented on 14 different types of datasets and compared with six state-of-the-art online feature selection algorithms to strongly validate the effectiveness and generalization of this algorithm. [ABSTRACT FROM AUTHOR]

This paper proposes new improved binary versions of the Sine Cosine Algorithm (SCA) for the Feature Selection (FS) problem. FS is an essential machine learning and data mining task of choosing a subset of highly discriminating features from noisy, irrelevant, high-dimensional, and redundant features to best represent a dataset. SCA is a recent metaheuristic algorithm established to emulate a model based on sine and cosine trigonometric functions. It was initially proposed to tackle problems in the continuous domain. The SCA has been modified to Binary SCA (BSCA) to deal with the binary domain of the FS problem. To improve the performance of BSCA, three accumulative improved variations are proposed (i.e., IBSCA1, IBSCA2, and IBSCA3) where the last version has the best performance. IBSCA1 employs Opposition Based Learning (OBL) to help ensure a diverse population of candidate solutions. IBSCA2 improves IBSCA1 by adding Variable Neighborhood Search (VNS) and Laplace distribution to support several mutation methods. IBSCA3 improves IBSCA2 by optimizing the best candidate solution using Refraction Learning (RL), a novel OBL approach based on light refraction. For performance evaluation, 19 real-wold datasets, including a COVID-19 dataset, were selected with different numbers of features, classes, and instances. Three performance measurements have been used to test the IBSCA versions: classification accuracy, number of features, and fitness values. Furthermore, the performance of the last variation of IBSCA3 is compared against 28 existing popular algorithms. Interestingly, IBCSA3 outperformed almost all comparative methods in terms of classification accuracy and fitness values. At the same time, it was ranked 15 out of 19 in terms of number of features. The overall simulation and statistical results indicate that IBSCA3 performs better than the other algorithms. [ABSTRACT FROM AUTHOR]

Feature selection aims at selecting representative features from the original high-dimensional feature set, and it has drawn much attention in most real-world applications like data mining and pattern recognition. This paper studies feature selection problem from the viewpoint of feature self-representation. Traditionally, feature self-representation is only performed on the whole-level reconstruction, whereas the feature selection ability is insufficient owing to the intra-class variations. To address this problem, we propose a new feature selection method, i.e., class-level regularized self-representation (CLRSR). In the proposed method, a class-level reconstruction term is designed to reduce intra-class variations of the samples from different categories. By jointly optimizing the whole-level reconstruction and the class-level reconstruction, CLRSR is able to select more discriminative and informative features. Moreover, an iterative algorithm is proposed to minimize cost function of CLRSR, and its convergence is proven in theory. By comparing with several state-of-the-art feature selection methods, experimental evaluations on six benchmark datasets have verified effectiveness and superiority of CLRSR. [ABSTRACT FROM AUTHOR]

The recommendation system has become the primary tool used by many Internet application platforms to solve the problem of information overload, and it faces issues such as data sparsity, cold start, and scalability. At present, most social recommendation algorithms only consider the influence of the trust relationship on the user's feature vector, which indirectly affects the predicted rating, or consider directly trusting friends as neighbor users, which directly affects the predicted rating, but does not consider the direct influence and indirect combining influences to make rating predictions. Therefore, this paper proposes a collaborative filtering recommendation algorithm (ETBRec), which not only considers the trust difference between users but also proposes the definition of experts and considers the direct impact of expert users on prediction ratings and the trustees' indirect impact of ratings. Among them, the trust difference is realized through trust metrics, including direct trust metrics and indirect trust metrics; the selection of expert users takes into account the user's degree of trust and user rating attitude; experimental comparisons with various social recommendation algorithms and related recommendation algorithms on the Ciao and Douban datasets. The experimental results show that the ETBRec algorithm performs better on some evaluation indexes such as mean absolute error (MAE) and root mean squared error (RMSE). [ABSTRACT FROM AUTHOR]

Writer identification (WI) is a typical pattern recognition problem with the goal of recognizing the writer of a text from images of his or her handwriting. For handwriting-based applications, a new approach is required that can confirm the writer based on a very small amount of available text. However, due to the limited number of characters and the diverse contents of the available samples, it is still challenging to achieve a high recognition rate on large-scale databases. To solve this problem, this paper proposes an efficient method for offline text-independent WI using redundant writing patterns and dual-factor analysis of variance (DF-ANOVA). In this method, a search for writing patterns in a limited number of handwritten texts is first conducted to generate meaningful codebooks. Then, these writing patterns are described by means of the simplified Wigner distribution function (WDF) and improved directional index histogram (DIH) descriptors to extract both structural and texture features, and the two corresponding feature distances are integrated by means of a fuzzy integral rule to make the final decision. The Fisher discriminant ratio (FDR) is used for feature selection, and DF-ANOVA is used to eliminate the influence of factors associated with the labels of samples on the feature distance. The results of evaluations on the IAM (96.92%), Firemaker (96.4%), and Uyghur2016 (100%) databases illustrate that this system shows strong robustness to an increasing number of writers and a decreasing number of words in the sample texts. [ABSTRACT FROM AUTHOR]

The Pigeon-Inspired Optimization (PIO) algorithm is an intelligent algorithm inspired by the behavior of pigeons returned to the nest. The binary pigeon-inspired optimization (BPIO) algorithm is a binary version of the PIO algorithm, it can be used to optimize binary application problems. The transfer function plays a very important part in the BPIO algorithm. To improve the solution quality of the BPIO algorithm, this paper proposes four new transfer function, an improved speed update scheme, and a second-stage position update method. The original BPIO algorithm is easier to fall into the local optimal, so a new speed update equation is proposed. In the simulation experiment, the improved BPIO is compared with binary particle swarm optimization (BPSO) and binary grey wolf optimizer (BGWO). In addition, the benchmark test function, statistical analysis, Friedman's test and Wilcoxon rank-sum test are used to prove that the improved algorithm is quite effective, and it also verifies how to set the speed of dynamic movement. Finally, feature selection was successfully implemented in the UCI data set, and higher classification results were obtained with fewer feature numbers. [ABSTRACT FROM AUTHOR]

Multi-label learning has widely applied in machine learning and data mining. The purpose of feature selection is to select an approximately optimal feature subset to characterize the original feature space. Similar to single-label data, feature selection is an import preprocessing step to enhance the performance of multi-label classification model. In this paper, we propose a multi-label feature selection approach with Pareto optimality for continuous data, called MLFSPO. It maps multi-label features to high-dimensional space to evaluate the correlation between features and labels by utilizing the Hilbert-Schmidt Independence Criterion (HSIC). Then, the feature subset obtains by combining the Pareto optimization with feature ordering criteria and label weighting. Eventually, extensive experimental results on publicly available data sets show the effectiveness of the proposed algorithm in multi-label tasks. [ABSTRACT FROM AUTHOR]

Missing data is a common problem in credit evaluation practice and can obstruct the development and application of an evaluation model. Block-wise missing data is a particularly troublesome issue. Based on multi-task feature selection approach, this paper proposes a method called MMPFS to build a model for credit evaluation that primarily includes two steps: (1) dividing the dataset into several nonoverlapping subsets based on missing patterns, and (2) integrating the multi-task feature selection approach using logistic regression to perform joint feature learning on all subsets. The proposed method has the following advantages: (1) missing data do not need to be managed in advance, (2) available data can be fully used for model learning, (3) information loss or bias caused by general missing data processing methods can be avoided, and (4) overfitting risk caused by redundant features can be reduced. The implementation framework and algorithm principle of the proposed method are described, and three credit datasets from UCI are investigated to compare the proposed method with other commonly used missing data treatments. The results show that MMPFS can produce a better credit evaluation model than data preprocessing methods, such as sample deletion and data imputation. [ABSTRACT FROM AUTHOR]

Fitness landscape analysis is a tool that can help us gain insight into a problem, determine how hard it is to solve a problem using a given algorithm, choose an algorithm for solving a given problem, or choose good algorithm parameters for solving the problem. In this paper, fitness landscape analysis of hyperheuristics is used for clustering instances of three scheduling problems. After that, good parameters for tree-based genetic programming that can solve a given scheduling problem are calculated automatically for every cluster. Additionally, we introduce tree editing operators which help in the calculation of fitness landscape features in tree based genetic programming. A heuristic is proposed based on introduced operators, and it calculates the distance between any two trees. The results show that the proposed approach can obtain parameters that offer better performance compared to manual parameter selection. [ABSTRACT FROM AUTHOR]

Multi-view observations provide complementary clues for 3D object recognition, but also include redundant information that appears different across views due to view-dependent projection, light reflection and self-occlusions. This paper presents a view-relation constrained global representation network (VCGR-Net) for 3D object recognition that can mitigate the view interference problem at all phases, from view-level source feature generation to multi-view feature aggregation. Specifically, we determine inter-view relations via LSTM implicitly. Based on the relations, we construct a two-stage feature selection module to filter features at each view according to their importance to the global representation and their reliability as observations at specific views. The selected features are then aggregated by referring to intra- and inter-view spatial context to generate global representation for 3D object recognition. Experiments on the ModelNet40 and ModelNet10 datasets demonstrate that the proposed method can suppress view interference and therefore outperform state-of-the-art methods in 3D object recognition. [ABSTRACT FROM AUTHOR]

Intelligent detection of abnormal behaviors meets the need of engineering applications for identifying anomalies and alerting operators. However, most existing methods tackle the high-dimensional sequential video data with key frame extraction, which ignore the redundancy effect of inter- and intra- video frames. In this paper, a novel Abnormal Detection method based on double sparseness LSSVMoc (AD_LSSVMoc) is proposed, which combine both sample (i.e. frame) selection and feature selection simultaneously in a uniform sparse model. For the feature extraction, both handcrafted features and learned features are aggregated into effective descriptors. To achieve feature selection and sample selection, a improved LSSVMoc is proposed with sparse primal and dual optimization strategy, and alternating direction method of multipliers is used to solve the constrained linear equations problem raised in AD_LSSVMoc. Experiments show that the proposed AD_LSSVMoc method achieves a competitive detection performance and high detecting speed compared to state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Constraint-based relevant feature selection using the Markov blanket (MB) discovery in Bayesian network (BN) has attracted widespread attention in diverse data mining applications. However, several MB discovery methods have been presented to manage low- or high-dimensional data by focusing on either improving computation efficiency or boosting learning accuracy instead of considering both. This paper presents a new constraint-based algorithm for feature selection that considers the improvement and balancing of both computational efficiency and prediction accuracy, called F eature S election via Mining M arkov B lanket (FSMB). The FSMB mines the MB containing parents-children (PC) and spouses (SP) using a forward approach to induce the true positive parents-children (PC) of a given target T. The FSMB removes false-positive PC from the PC set and never considers them again. Concurrently, the FSMB finds SP of a target T through an exhaustive search from the non-parents-children set using the V-structure strategy to differentiate both true-positive PC and SP in the MB set and then use them to remove the false-positive SP. Also, the FSMB removes the non-MB descendants using the updated PC and SP set. Extensive experiments are conducted and validated on benchmark datasets for performance evaluation. The results are compared with existing algorithms, including the Incremental Association Markov Blanket (IAMB), the Max-Min Markov Blanket (MMMB), the HITON-MB, the Simultaneous Markov Blanket (STMB), the Iterative Parents-and-Children-Based Markov Blanket (IPCMB), the Balanced Markov Blanket (BAMB), the Efficient and Effective Markov Blanket (EEMB), and the Markov Blanket discovery by Feature Selection (MBFS). Experimental results show that the FSMB outperforms the existing algorithms with higher accuracy and shorter running time. [ABSTRACT FROM AUTHOR]

Information gain, as an important feature measure, plays a vital role in the process of feature selection. Most of existing information gain-based feature selection algorithms are developed on data with single type features. However, in practical applications, there are few studies on hybrid data including symbolic, numerical and missing features. In addition, because obtaining all the decision labels of data is often costly, only a small number of objects are expected to be labeled with ground truth. Therefore, this paper proposes an information gain-based semi-supervised feature selection algorithm for hybrid data. To acquire the decision labels of unlabeled objects to improve the classification accuracy of the selected feature subset, we firstly propose an extended decision annotation algorithm based on the enlarged neighborhood granule by using the information of labeled objects. On this basis, we develop an information gain-based feature selection algorithm by exploiting the information of both labeled and unlabeled objects. Finally, the experimental results on ten UCI datasets demonstrate the superiority of the proposed algorithm, in comparison with other feature selection algorithms. [ABSTRACT FROM AUTHOR]

Since the Markov blanket (MB) of a class variable captures the causal relationship between the class variable and selected features, employing the MB of a class variable for feature selection improves the interpretability and robustness of the predictive model. Online MB learning aims to identify the MB with streaming features. However, the only existing online MB learning algorithm needs to enumerate the subsets of selected PC (i.e., parents and children) and spouses and may include false-positives in the found MB, thus affecting the efficiency and accuracy on high-dimensional data. To address this issue, in this paper, we propose two online MB learning algorithms, called Online SimulTaneous MB learning (O-ST) algorithm and Online Divide-and-Conquer MB learning (O-DC) algorithm. When a new feature arrived, O-ST simultaneously learns the PC and spouses (i.e., the MB) conditioned on the currently selected MB, and O-DC learns the PC and spouses separately by sequentially comparing the mutual information in the currently selected PC. The comprehensive experimental results validate that the proposed algorithms achieve higher efficiency and better accuracy than the state-of-the-art online MB learning algorithms. [ABSTRACT FROM AUTHOR]

With the development of data mining, artificial intelligence, neural network, expert system and machine learning, information system (i-system) becomes more and more important. If the objects, attributes and information values in an i-system are replaced by cells, genes and gene expression values, respectively, then the i-system is said to be a gene space. Because gene expression data is characterized by small samples, high dimension and noise, there is considerable uncertainty in a gene space. Traditional machine learning and statistical methods are often powerless to a gene space. Granular computing (GrC) can effectively deal with various uncertainties. This paper studies the uncertainty measurement of gene space based on the class-consistent technology and discusses its application in gene selection from the perspective of GrC. A class-consistent relation between cells in a gene space is first established by the gene expression values of cells on the basis of class-consistent technology. Then, the information granules (i-granules) are obtained from a gene space by using the class-consistent relation. Next, two metrics (information granularity and information entropy) to measure the uncertainty of gene space are defined and their properties are also investigated. The results of numerical experiments and statistical tests verify their effectiveness. Furthermore, as their application to gene space, two gene selection algorithms are proposed. Finally, the clustering experiments and statistical tests on 16 gene spaces show that the designed gene selection algorithms outperform some state-of-the-art feature selection algorithms in terms of three clustering performance indicators. [ABSTRACT FROM AUTHOR]

Selecting the most discriminative features is a challenging problem in many applications. Bio-inspired optimization algorithms have been widely applied to solve many optimization problems including the feature selection problem. In this paper, the most discriminating features were selected by a new Chaotic Dragonfly Algorithm (CDA) where chaotic maps embedded with searching iterations of the Dragonfly Algorithm (DA). Ten chaotic maps were employed to adjust the main parameters of dragonflies' movements through the optimization process to accelerate the convergence rate and improve the efficiency of DA. The proposed algorithm is employed for selecting features from the dataset that were extracted from the Drug bank database, which contained 6712 drugs. In this paper, 553 drugs that were bio-transformed into liver are used. This data have four toxic effects, namely, irritant, mutagenic, reproductive, and tumorigenic effect, where each drug is represented by 31 chemical descriptors. The proposed model is mainly comprised of three phases; data pre-processing, features selection, and the classification phase. In the data pre-processing phase, Synthetic Minority Over-sampling Technique (SMOTE) was used to solve the problem of the imbalanced dataset. At the features selection phase, the most discriminating features were selected using CDA. Finally, the selected features from CDA were used to feed Support Vector Machine (SVM) classifier at the classification phase. Experimental results proved the capability of CDA to find the optimal feature subset, which maximizing the classification performance and minimizing the number of selected features compared with DA and the other meta-heuristic optimization algorithms. Moreover, the experiments showed that Gauss chaotic map was the appropriate map to significantly boost the performance of DA. Additionally, the high obtained value of accuracy (81.82-96.08%), recall (80.84-96.11%), precision (81.45-96.08%) and F-Score (81.14-96.1%) for all toxic effects proved the robustness of the proposed model. [ABSTRACT FROM AUTHOR]

The neighborhood rough set model (NRS) has been widely applied to study feature selection. Nevertheless, the dependency, as a significant feature evaluation function in NRS, only focuses on the classification information in the lower approximation and ignores the classification information in the upper approximation, which affects the evaluation effect of this function. Consequently, this paper first defines the fuzziness using the upper approximation and proposes two self-information uncertainty measures based on the dependency and fuzziness. Second, combining the above two self-information uncertainty measures, a more comprehensive approximate self-information is proposed for evaluating the uncertainty of the classification information of feature subsets. Furthermore, a heuristic feature selection algorithm is constructed based on the approximate self-information. Third, to reduce the time cost of the constructed algorithm in processing high-dimensional datasets, we propose a two-stage selection strategy, in which the first stage adopts the Fisher score dimensionality reduction method (FS) with low time cost and stable performance to retain important features in the high-dimensional dataset as a candidate feature subset. Then, the second stage employs our algorithm to further reduce the candidate feature subset. Finally, the results of various feature selection algorithms on eleven datasets are presented, and the comparison results confirm that our algorithm is efficient. [ABSTRACT FROM AUTHOR]

Prediction as a new frontier has begun to come into focus which has been widely used in traffic management and statistics. Because of traffic time series data possesses both spatial and temporal characteristics, existing models have their limitations in modeling time-space. Most methods utilize two different parts to capture the characteristics of time and space respectively and ignore the potential relationship between them, which does not meet actual traffic conditions in real world. Even if a few models take this correlation into account, they capture a tiny part of temporal characteristics to describe this correlation. These models just extract local features based on time dimensions and cannot capture full-scale temporal characteristics and spatial-temporal heterogeneity. Besides, most methods overlook the over-fitting of feature in aggregation operation by using graph neural networks to model time-space and predict. To settle these problems, this paper proposes a novel spatial-temporal corelated graph neural network (STCGNN) of data forecasting. This model can extract the full-scale characteristics of intricate spatial-temporal graph and can generate spatial-temporal adjacent correlated matrix based on whole characteristics of time series. Furthermore, this model can determine the aggregation degree of feature nodes by Neighborhood Feature Selection Network (NFSN) to avoid the over-fitting of feature. Various experiments are conducted on three real-world datasets and evaluation metrics are 15.81 of MAE, 14.59% of MAPE and 27.23 of RMSE on PEMS03 dataset. These results have demonstrated that our model STCGNN obtains the lowest errors in predicting than previous models. [ABSTRACT FROM AUTHOR]