Search

Multi-task feature selection (MTFS) has been proven effective for reducing the curse of dimensionality in large-scale classification. Many existing MTFS methods assume that all tasks learn concurrently in a static environment without considering the dynamism of tasks in real-world scenarios. However, new tasks emerge dynamically in practical applications, meaning that the aforementioned static assumption is insufficient. In this paper, we construct a dynamic multi-task feature selection framework to achieve feature reduction for constantly arriving new tasks. First, we modify the traditional mapping by shifting from hard labels to soft labels. Unlike the conventional rigid mapping, the new flexible loss function changes the direct mapping strategy to an indirect one. Second, we use the orthogonal regularization term to constrain the independent relationship between new and old tasks. This ensures that the selected relevant features for new tasks differ from prior tasks. Finally, we integrate the flexible loss and the orthogonal regularization term in the dynamic multi-task feature selection framework. Our method outperforms nine other advanced feature selection methods in terms of effectiveness and efficiency across six datasets. For example, the ACC value of our method is almost 1% higher than the next-best method on the large-scale SUN dataset. • A dynamic feature selection is proposed to select features for new tasks. • A flexible loss is presented to incorporate potential knowledge. • An orthogonal constraint is used for selecting task-specific features. [ABSTRACT FROM AUTHOR]

The class imbalance problem is one of the main challenges that hinders classifiers from identifying unknown instances. When class distribution imbalance and class overlap coexist in the same data set, the recognition of minority instances by traditional classifiers will be seriously affected. In the binary classification problem, the relationship between the majority class and the minority class is clear, and the overlap problem between classes can be well solved by removing the majority instances in the class-overlap area. However, for multi-class imbalance problems, the relationships between classes will become complex. Clustering techniques are effective methods for identifying overlapping instances; however, a single clustering method cannot effectively identify clusters of different shapes. Therefore, in this paper, a H eterogeneous C lustering E nsemble learning method for M ultiple C lass-overlap D etection (HCE-MCD) is proposed for multi-class imbalance problems. This method uses a genetic algorithm to select a combination of heterogeneous clustering techniques with the best fitness function from a pool of clustering algorithms and uses majority voting to fuse the results of the clustering algorithm. Furthermore, instances located near the decision boundary are more likely to be overlapping instances. Therefore, according to the clustering results of the heterogeneous clustering ensemble method, we can identify the overlapping instances existing in the data set, thereby eliminating the class overlap problem in the multi-class dataset. Experimental results on 19 open-access datasets show that our proposed method outperforms or partially outperforms state-of-the-art research protocols. [ABSTRACT FROM AUTHOR]

Feature selection is a fundamental step in machine learning, serving to reduce dataset redundancy, accelerate training speed, and improve model quality. This is particularly crucial in high-dimensional datasets, where the excess of features presents challenges for pattern recognition and data analysis. Recent methods proposed for high-dimensional data are often tailored for specific domains, leaving a lack of consensus on a universally recommended solution for general use cases. This paper proposes a hybrid feature selection approach using a multi-objective genetic algorithm to enhance classification performance and reduce dimensionality across diverse classification tasks. The proposed approach narrows the search space of possible relevant features by exploring the combined outputs of classical feature selection methods through novel genetic algorithm operators. This enables the evolution of combined solutions potentially not explored by the original methods, generating optimized feature sets in a process that adapts to different data conditions. Experimental results demonstrate the effectiveness of the proposed method in high-dimensional use cases, offering improved classification performance with reduced feature sets. In summary, our hybrid method offers a promising solution for addressing the challenges of high-dimensional datasets by enhancing classification performance in varying domains and data conditions. • Hybrid method optimizes feature selection with multi-objective genetic algorithm. • Achieves improved classification metrics with optimized and smaller feature sets. • Combines multiple feature selection mechanisms for superior results. • Efficiently represents information in complex datasets from multiple domains. [ABSTRACT FROM AUTHOR]

Empirical Risk Minimization (ERM) with the squared loss has become one of the most popular principles for designing learning algorithms. However, the existing mean regression models under ERM usually suffer from poor generalization performance due to their sensitivity to atypical observations (e.g., outliers). For alleviating this sensitivity, some strategies have been proposed by utilizing the quantitative relationship of error values with different observations to form robust learning objectives. Instead of focusing on error values, this paper considers the error density to uncover the structure information of observations and proposes a new learning objective, called Error Density-dependent Empirical Risk Minimization (EDERM), for robust regression under complex data environments. Property characterizations and experimental analysis validate the robustness and competitiveness of the proposed EDERM-based learning models. The implemented codes can be found at https://github.com/zhangxuelincode/EDERM. • Error density-dependent learning paradigm beyond the error value modeling. • Robust estimation for both regression and classification. • Property characterization and empirical validations. [ABSTRACT FROM AUTHOR]

The accurate classification of Typhoon Intensity (TI) based on cloud patterns in satellite images is crucial for effective disaster warning and management. However, this task is particularly challenging due to the striking visual similarities among different sub-classes of typhoons, coupled with a long-tailed distribution of class occurrences. Existing deep learning methods often struggle with biased classification stemming from imbalanced datasets and face difficulty discerning subtle differences between categories. This paper proposes a novel solution to these challenges through a hybrid framework that integrates contrastive learning and classifier learning. Contrastive learning is employed to increase the separation of similar sub-classes within the feature space, while classifier learning aims to train a discriminative and unbiased classifier. The proposed approach is evaluated extensively on the publicly available DeepTI dataset, demonstrating enhanced performance for both prominent and less frequent classes. The model achieves the 70.88% accuracy for seven categories typhoon intensity classification. The code implementation is at https://github.com/chen-huanxin/contrastive-hybrid-network. • The proposed network combines contrastive learning with classifier learning. • The proposed network can learn the relation between samples. • Contrastive learning provides the network with more supervision. • The proposed network can learn an unbiased classifier from imbalanced dataset. [ABSTRACT FROM AUTHOR]

Images have become a significant medium for information transmission, while image forensics has garnered widespread attention from researchers. Due to the scarcity of finely annotated images in the field of image manipulation detection and localization, existing methods for locating manipulated images can only utilize a limited amount of data for model training. However, deep learning models typically require a large number of finely annotated images to fully leverage their powerful fitting capabilities. In this paper, we propose a semi-supervised image manipulation localization framework, employing semi-supervised learning to use unannotated images for deep learning model training. To achieve this goal, we first design a residual enhancement module that contains an encoding-decoding structure. The regression target of this branch is the pristine regions in the images to generate the reconstructed images. Next, we perform a residual operation between the reconstructed and original images to roughly locate the anomalous regions and refine the features extracted by the encoder. Secondly, we employ weakly-supervised learning by adding an image-level classification head to the final layer of the encoder. The classification head generates a class active map based on image-level classification results, which further guides the model's feature augmentation. Finally, we formulate a specialized semi-supervised framework tailored for image manipulation detection, enabling the utilization of a large volume of unannotated data for model training. Extensive experimental results highlight the notable efficacy of the proposed approach, and outperforms state-of-the-art approaches. • Unannotated images are utilized for image manipulation detection. • Semi-supervised framework tailored for image manipulation detection. • Weakly-supervised framework for features augmentation. • Residual enhancement module refines the features extracted by encoder. • Approach shows state-of-the-art performance. [ABSTRACT FROM AUTHOR]

With rapid growth of online healthcare systems, chief complaint classification plays an important role in areas such as triage or doctor recommendation. Existing medical text classification techniques such as rule-based or learning-based methods fail to effectively utilize the inherent hierarchical structure of label descriptions that contain strong domain knowledge. In this paper, we propose a novel text classification framework for chief complaint by embedding both input text and hierarchical structure of label descriptions based on deep neural networks. The proposed framework makes use of not only three branches (i.e. chief complaint branch, main-category branch, and sub-category branch) with a Sequence Information Encoder to encode semantics from chief complaint and hierarchical structure of label descriptions but also a Hierarchical Relational Network with Attention module to capture complex relationships among them focusing on informative words with attentional scores. We evaluate our framework on two public medical datasets with label descriptions extracted from medical books and websites. Experimental results show that the proposed method outperforms other baseline techniques by a significant margin. The source code of our framework is available at ANONYMISED. [Display omitted] • Novel chief complaint classification model using hierarchical label descriptions. • SIE effectively embeds input text to capture sequential information with context. • HRNA captures relationships among chief complaint and category descriptions. • The framework greatly outperforms SoTA methods on two real world datasets. [ABSTRACT FROM AUTHOR]

• We analyze the limitations in the Enhanced Fuzzy Min-Max (EFMM) model. • Proposing a new FEFMM model to overcome the EFMM limitations. • The performance of FEFMM is evaluated using benchmark data sets. • Results show FEFMM's superior classification accuracy. In this paper, the existing enhanced fuzzy min–max (EFMM) neural network is improved with a flexible learning procedure for undertaking pattern classification tasks. Four new contributions are introduced. Firstly, a new training strategy is proposed for avoiding the generation of unnecessary overlapped regions between hyperboxes of different classes. The learning phase is simplified by eliminating the contraction procedure. Secondly, a new flexible expansion procedure is introduced. It eliminates the use of a user-defined parameter (expansion coefficient) to determine the hyperbox sizes. Thirdly, a new overlap test rule is applied during the test phase to identify the containment cases and activate the contraction procedure (if necessary). Fourthly, a new contraction procedure is formulated to overcome the containment cases and avoid the data distortion problem. Both the third and fourth contributions are important for preventing the catastrophic forgetting issue and supporting the stability-plasticity principle pertaining to online learning. The performance of the proposed model is evaluated with benchmark data sets. The results demonstrate its efficiency in handling pattern classification tasks, outperforming other related models in online learning environments. [ABSTRACT FROM AUTHOR]

Flow field segmentation and classification help researchers to understand vortex structure and thus turbulent flow. Existing deep learning methods are mainly based on global vorticity information and focused on 2D circumstance. In this paper, based on flow field theory, criteria for flow field segmentation and classification in 3D space using local velocity information and the physical relationship between local vorticity and vortex wake are conducted, and then novel physical local information-embedded flow field segmentation and classification methods in 3D space using Multilayer Perceptron (MLP) are proposed. Our methods consist of three modules: a preprocessing module, a flow field segmentation module and a flow field classification module. The preprocessing module calculates input for succeeding modules based on the above criteria. The segmentation and classification modules process corresponding input to identify vortex structure, and further classify the type of vortex wakes in a 3D flow field. Experimental results show that our method can identify vortex area with an accuracy rate that reaches 90% while reducing time consumption by more than 90% compared with existing segmentation methods; our classification method reaches an accuracy rate of 99.7% while reducing time consumption by more than 90% compared with existing classification methods. Our proposed methods proved to achieve good performance both on accuracy and efficiency. • A novel classification criterion based on Re-vorticity relationship is constructed. • 3D flow field classification learning method using above criterion is proposed. • Explore new 3D flow field segmentation criterion using local velocity information. • A deep learning method for efficient 3D flow field segmentation is introduced. [ABSTRACT FROM AUTHOR]

With the rapid advancement of the internet, there has been a dramatic increase in short-text data. Due to the brevity of short texts, sparse features, and limited contextual information, short-text classification has become a challenging task in natural language processing. However, current methods primarily capture semantic information from locally-sequenced words in short text, which ignores the intricate feature relationships that pervade both the intra-text and inter-text. Therefore, this paper proposes a novel Edge-Enhanced Minimum-Margin Graph Attention Network (EMGAN) for short text classification to address this issue. Specifically, we construct a Heterogeneous Information Graph (HIG) to represent complex relationships among short text features. HIG mainly considers the relationship between document features and three attribute features, such as entities, topics, and keywords, and can represent short text features from multiple dimensions and levels. Then, to enhance the connectivity and expressiveness of the HIG for more effective propagation of feature information within it, we present a novel X-shaped structure edge-enhancement method. It enriches their relationships by reconstructing the edge structures. Furthermore, we design a Minimum Margin Graph Attention Network (MMGAN) for short text classification. Specifically, this method aims to explore the minimum margin between high-order neighbors and central nodes at the minimum cost, efficiently extracting and aggregating feature information. Extensive experimental results demonstrate that our proposed EMGAN model outperforms existing methods on five datasets, validating its effectiveness in short-text classification. Our code is submitted at https://github.com/w123yy/EMGAN. • A new method for efficient classification of short texts. • Heterogeneous infographics represent complex relationships among short texts. • The edge enhancement method enriches edge relations and enhances graph structure. • Minimum margin attention network aggregates feature to solve feature sparsity. [ABSTRACT FROM AUTHOR]

Many deep graph neural networks perform well in tackling semi-supervised node classification tasks. However, they often focus on learning a holistic representation for nodes, which lacks interpretability and compactness for classification purposes. In this paper, we propose a Multi-Relation Enhanced Graph Disentangled Network (MREGDN), which can effectively disentangle the latent factors that affect graph topology, while enhancing the interpretability and low-redundancy of the learned node representations for classification. There are three types of relation constraints designed elaborately to optimize the disentangled representation learning process: (a) constraints of intra-factor compactness to encourage a factor to represent a distribution, (b) constraints of inter-factor irrelevance to ensure diversity among factors, and (c) constraints of keeping inter-node similarity in different spaces to retain necessary information related to category. Extensive experiments verify that MREGDN significantly outperforms its best-known competitors on various semi-supervised node classification tasks. • MREGDN improves the interpretability and discrimination of node representations. • MREGDN learns more compact and category-relevant node representations. • MREGDN greatly outperforms existing competitors on various classification tasks. [ABSTRACT FROM AUTHOR]

Feature selection refers to the process of identifying the most relevant and valuable features or characteristics in a dataset, and it is an important step that can improve the performance and accuracy of predictive models for machine learning and data mining. Feature selection is crucial in efficiently reducing the number of features, thereby enhancing classification accuracy. This, in turn, alleviates computational burdens and elevates the overall performance of machine learning algorithms. This paper extended the conventional manta ray foraging optimization (MRFO) algorithm to an improved binary MRFO (IBMFO) algorithm. IBMRFO serves as a search strategy for designing a wrapper-based feature selection approach. To commence, we introduce a novel mechanism, termed the chaotic tent map (CTM), aimed at initializing the population to fortify both the algorithm's exploitation capabilities and population diversity. Second, an adaptive somersault foraging (ASF) factor is tuned to avoid premature convergence. Finally, a binary mechanism is introduced to adapt the algorithm for binary feature selection dilemmas. Experiments are performed on 30 renowned UC Irvine machine learning repository datasets. The results of the student's t-test show that the classification algorithm has statistical significance. The findings unequivocally exhibit the superior performance of the proposed IBMRFO compared to other algorithms under scrutiny. Notably, the efficiency of the enhanced factors is assessed through rigorous testing. For post-publications supporting this work, readers can refer to https://github.com/ELF-CHEUNG/IBMRFO. • Novel chaotic tent map for population initialization from evolutionary computation. • Optimization methods based on enhanced evolutionary computation. • Adaptive factor enhances convergence rate and search efficiency of algorithm. • Extensive numerical and scientific simulations for verifications with UCI dataset. [ABSTRACT FROM AUTHOR]

Online social networks (OSNs) are receiving great attention from the research community for different purposes, such as event detection, crisis management, and forecasting, among others. The increasing amount of research conducted with social networks opens the need for a classification methodology regarding trends in the field. This work does not cover all types of social networks; it focuses on the analysis of microblogs as a data source in the context of recommender systems (RSs). The main goal of this work is to provide authors with insights on the trends of academic literature reviews in the proposed context and to provide a comparison of different research approaches. The authors searched for up-to-date research papers related to RS methods using microblogs within a time period of five years, from 2012 to January 2018. Starting from 2012, a significant amount of research related to the subject field of RSs was conducted and identified by the authors of this work. After the filtering process, 39 papers were finally selected from journals and conferences in four different databases related to Internet technologies (i.e., IEEE, ACM, Science Direct, and Springer). A general classification presented in this work is then adopted and used to describe state-of-the-art social network recommendation approaches for microblogging. This work can be extended in the future to include novel methodologies and trends of RSs for microblogs. [ABSTRACT FROM AUTHOR]

In this paper, motivate by quantile in the field statistics and bounded linex loss function a novel robust bounded quantile loss is proposed for improving the performance of support vector machine (SVM) and support vector regression (SVR). The bounded quantile loss has some important properties such as asymmetry, non-convexity, which make SVM and SVR based on bounded quantile (BQ-SVM and BQ-SVR) robust to noise. The Fisher consistency and generalization error bound of BQ-SVM guarantees the generalization capability of BQ-SVM. In addition, we derive the influence functions of BQ-SVM and BQ-SVR to show that they are less sensitive to outliers. However, the non-convexity of the proposed bounded quantile makes it difficult to optimize. We utilize the concave-convex procedure (CCCP) technique to solve the proposed models. A host of numerical studies on artificial and benchmark datasets are conducted, and various evaluation criteria were considered to demonstrate the effectiveness and the efficiency of the proposed method. Experimental results indicate that our proposed methods is more robust compare to classical and advanced methods when the dataset contains outliers in both classification and regression applications. • A novel bounded quantile loss (L b q loss) is proposed for robust learning problems. • The support vector machines based methods with L b q loss (BQ-SVM and BQ- SVR) are robust. • The designed ClipDCD-based CCCP algorithm is efficient to solve the BQ-SVM and BQ-SVR. • Numerical and theoretical results demonstrate the effectiveness of BQ-SVM and BQ-SVR. [ABSTRACT FROM AUTHOR]

The high dimensionality of data hinders the learning ability of machine learning algorithms. Feature selection techniques can be used to reduce dimensionality, which is an important step for processing high-dimensional data. Feature selection solves this problem by removing irrelevant and redundant information, which can improve learning models, reduce calculation time, and improve learning accuracy. In this paper, a novel filter in mixed-attribute datasets for feature selection is proposed. The independent attributes are mixed or heterogeneous in the sense that both numerical and categorical attribute types may appear together in the same dataset. Based on the preordonnances theory, we use a new concept to quantify the relevance and redundancy of features even if there are heterogeneous (mixed-type) data. The technique for order preference by similarity to the ideal solution is one of the well-known multicriteria decision-making methods; it is utilized as a weighting and informative feature selection filter. To assess the effectiveness of the proposed method, several experiments, both simulated and real, are performed, including a comparison to other well-known filter methods. The experimental results show that, in most cases, the method yielded competitive results in comparison to other methods. • Introducing new criterions for relevance and complementarity. • Dealing with mixed-attribute datasets without requiring a preprocessing step. • The multi-criteria decision method, namely TOPSIS, is used to score the explanatory features. • Investigate the performance of the proposed filter in high-dimensional data by using simulated and real data sets. [ABSTRACT FROM AUTHOR]

Local binary pattern (LBP) with a fixed sampling template is sensitive to scale changes. Furthermore, under rotation changes or noise corruptions, one uniform LBP pattern can be corrupted to fall into a non-uniform pattern which loses its discrimination power to describe the corresponding texture feature. To overcome these two main drawbacks, we propose a scale and pattern adaptive local binary pattern (SPALBP). Firstly, in the gradient-based sampling radius adaptive scheme, eight directional adaptive sampling radius of each center pixel can be obtained by using its eight Kirsch gradient values. Secondly, in the noise and rotation robust neighborhood sampling scheme, three neighborhood sampling templates are used to extract three kinds of averaging neighborhood pixels. Thirdly, for each center pixel, three kinds of L B P r i u 2 patterns can be extracted by sampling these three kinds of averaging neighborhood pixels along eight directional adaptive sampling radius. Finally, an optimal SPALBP uniform pattern can be adaptively selected from these three L B P r i u 2 patterns. Hence, all SPALBP patterns show more robustness against scale changes, rotation changes and noise corruptions. Extensive experiments are conducted on four standard texture databases: Outex, UIUC, CUReT and XU_HR. Comparing with state-of-the-art LBP-based variants, the proposed SPALBP method consistently shows superior performance both in dramatic environment changes and high-levels of noise conditions, meanwhile it maintains a lower texture feature dimension. • The paper presents a scale and pattern adaptive Local Binary Pattern. • Using the novel gradient-based sampling radius adaptive scheme. • Using the novel noise and rotation robust neighborhood sampling scheme. • The classification results of the proposed SPALBP are significant. [ABSTRACT FROM AUTHOR]

Point cloud classification has emerged as a vital research area in several emerging applications, including robotics and autonomous driving. However, discriminative feature learning has long been a challenging issue owing to the irregularity and disorder of point clouds. Recently, although Transformer-based methods can achieve high accuracy in point cloud learning, plenty of Transformer layers bring huge computation and memory consumption. This paper presents a novel hierarchical local–global framework based on Transformer network for point clouds, named TNPC. TNPC contains two serial Stages implementing downsampling operation, and each Stage is composed of a local feature extracting (LFE) block and a global feature extracting (GFE) block, which can reduce computation and memory consumption obviously. LFE block consists of two parallel branches, namely Transformer branch and shared multilayer perceptron (SMLP) branch, which are designed to learn relevant feature of any two points and local high-dimensional semantic features of each point between sampling centroid and its neighborhoods, respectively. The proposed two parallel branches not only improve the feature extraction effect, but also reduce the computation and memory consumption. GFE block consists of a center point contact (CPC) module and a global point cloud transformer layer (PCTL) module, which can improve the effect of global features extracting without adding the number of parameters and computation. The performance of our method is validated experimentally on ModelNet40 and ScanObjectNN datasets. Our method improves the mean accuracy on each category (m A c c) to 91.6% and 79.8% on the ModelNet40 dataset and ScanObjectNN dataset, respectively. In terms of efficiency, our method leads to a significant reduction, with only 4.73MB parameters and only 1.91GB floating-point operations (FLOPs). Experimental results demonstrate that the proposed method achieves state-of-the-art performance on classification accuracy and efficiency. • A novel hierarchical local–global framework is proposed for 3D point clouds. • LFE block contains parallel Transformer branch and SMLP branch. • LFE block can learn local high-dimensional semantic features of each point. • GFE block is designed to extract the global feature of point cloud effectively. [ABSTRACT FROM AUTHOR]

The classification of huge complex datasets to extract the information needed for making informed decisions has gained a lot of importance in today's scenario due to the outspurt of data. With an increase in dataset size, the means of obtaining the desired information via data clustering techniques also became complex. This paper illustrates a framework for selecting the best combination of the Kernel function and the Lattice structure for better classification based on the size of the data. The performance of the SOM majorly depends on the choice of Kernel Function, Lattice Structures, and the size of the datasets. The types of kernel functions deployed in this study are Gaussian, Exponential, and Laplacian Kernel along with the lattice Structures such as hexagonal, Rectangular, and Square which result in nine combinations respectively. These nine combinations of the Kernel Function and the Lattice Structure have been tested on datasets ranging from small to large size using the performance metrics. A frequency table has been arrived at based on the classification of the dataset size, combinations of Kernel functions, and Lattice Structures. The results of this study infer that the combination of Gaussian Kernel -Square Lattice performs better for small-size datasets wherein the Exponential and Laplacian Kernel functions along with Hexagonal Lattice perform better for mid-size datasets, while the Laplacian Kernel – Square Lattice perform better for large size datasets. [ABSTRACT FROM AUTHOR]

This paper introduces a three-step framework for classifying multiclass radiography images. The first step utilizes a de-noising technique based on wavelet transform (WT) and the statistical Kolmogorov Smirnov (KS) test to remove noise and insignificant features of the images. An unsupervised deep belief network (DBN) is designed for learning the unlabelled features in the second step. Although small-scale DBNs have demonstrated significant potential, the computational cost of training the restricted Boltzmann machine is a major issue when scaling to large networks. Moreover, noise in radiography images can cause a significant corruption of information that hinders the performance of DBNs. The combination of WT and KS test in the first step helps improve performance of DBNs. Discriminative feature subsets obtained in the first two steps serve as inputs into classifiers in the third step for evaluations. Five frequently used classifiers including naive Bayes, radial basis function network, random forest, sequential minimal optimization, and support vector machine and four different case studies are implemented for experiments using the Image Retrieval in Medical Application data set. The experimental results show that the three-step framework has significantly reduced computational cost and yielded a great performance for multiclass radiography image classification. Along with effective applications in image processing in other fields published in the literature, deep learning network in this paper has again demonstrated its robustness in handling a complex set of medical images. This implies that the proposed approach can be implemented in real practice for analysing noisy radiography images, which have many useful medical applications such as diagnosis of diseases related to lung, breast, musculoskeletal or pediatric studies. [ABSTRACT FROM AUTHOR]

Contrast patterns, which lie in the core of most understandable classifiers, are frequently evaluated by quality measures. Since many different quality measures are available, they should be compared to select the most appropriate for each applications. This paper introduces a method to compare quality measures, using a set of mined patterns and a collection of objects not used for mining. The comparison is performed by correlating quality values with a quality estimation of the patterns. Additionally, a meta-learning study is performed to show that combining quality measures could be better than using the best single measures in isolation. The results of this paper can help researchers to create new quality measures or to find new combinations of quality measures to create better understandable classification systems. [ABSTRACT FROM AUTHOR]

This paper proposes a novel hybrid classifier method using a mixed-integer linear programming (MILP) model and an improved online learning algorithm for general fuzzy min–max neural networks (IOL_GFMM). IOL_GFMM is used to generate initial hyper-boxes for the MILP model in order to improve its efficiency. This paper is the first attempt that integrates the fuzzy min–max neural networks and the MILP model. The applicability of the proposed hybrid model is proven on both real and artificial datasets. Performance evaluation parameters of the novel hybrid classifier are classification accuracy, training time, number of iterations, and number of misclassified data. Results show that the proposed hybrid method provides satisfying improvements in performance evaluation parameters compared with the pure MILP model. Our proposed approach achieves to speed the training time up to 76 times and decreases the number of misclassified data by up to 17 times. Results of artificial datasets show another contribution that the proposed hybrid approach is very efficient in order to classify data samples not only when they belong to separate regions but also when there exists significant overlap. [ABSTRACT FROM AUTHOR]

Abstract: Even though rapid advances in intelligent firearm identification have been made in recently years, the major practical and theoretical problems are still unsolved. From the practical point of view, capturing high quality images from ballistics specimen is a difficult task. From the theoretical point of view, extracting the descriptive features from projectile and cartridge images is an open research question in firearm identification. The aim of this paper is to address the research issues with respect to feature extraction and intelligent ballistics recognition. In this paper, different image processing techniques are employed for digitizing the ballistics images. Due to some segments in an image systematically distributed by the image’s geometrical circular center, the existing moment invariants however cannot extract the required pattern features for intelligent recognition. This paper presents the novel feature set called circle moment invariants to overcome the shortcoming of existing moment invariants. In addition, an intelligent system is designed for classifying and evaluating the extracted features of ballistics images. The experimental results indicate that the proposed approach and feature criteria are capable of classifying the cartridge images very efficiently and effectively. Consequently, the circle moment invariants are proved to be the adequate descriptors for describing the pattern features in cartridge images. [Copyright &y& Elsevier]

Automatic detection of brain tumors in single-spectral magnetic resonance images is a challenging task. Existing techniques suffer from inadequate performance, dependence on initial assumptions, and, sometimes, the need for manual interference. The research reported in this paper seeks to reduce some of these shortcomings, and to remove others, achieving satisfactory performance at reasonable computational costs. The success of the system described here is explained by the synergy of the following aspects: (1) a broad choice of high-level features to characterize the image’s texture, (2) an efficient mechanism to eliminate less useful features (3) a machine-learning technique to induce a classifier that signals the presence of a tumor-affected tissue, and (4) an improved version of the skippy greedy snake algorithm to outline the tumor’s contours. The paper describes the system and reports experiments with synthetic as well as real data. [ABSTRACT FROM AUTHOR]

Smuggling has long played an important role in the inefficiency of economies. To secure the borders against this illegal act, X-Ray Inspection Systems are often deployed at the borders and customs. In this paper, we present a new method for classification of shipping containers X-Ray images, produced in the inspection lines. The aim is to improve the matching accuracy of the presented manifest, which lists the claimed contents of the shipping containers, with the real contents of the container. The proposed method is based on utilizing Scale Invariant Feature Transforms (SIFT) feature vectors, Bag of visual words (BOVW) and tree augmented naive Bayes (TAN) approach for classifying containers X-Ray images. The prior research on classification of X-Ray images of shipping containers has focused mostly on working with greedy algorithms such as sliding windows for task of classification. More recent studies introduced filter banks and visual words for extraction of features. The proposed method for the first time considers the salient points and keypoints for the task of feature extraction. In addition, this paper presents a framework using the tree augmented naive Bayes based on the theory of learning Bayesian networks, which is proved to have a significant improvements upon the prior designed systems by considering the correlations among the extracted features. For experimental evaluations, our method is compared with two recently proposed methods on containers X-Ray images categorization. The results show that the proposed method is more accurate and time-efficient in categorization of X-Ray images. [ABSTRACT FROM AUTHOR]

The human liver is one of the major organs in the body and liver disease can cause many problems in human life. Fast and accurate prediction of liver disease allows early and effective treatments. In this regard, various data mining techniques help in better prediction of this disease. Because of the importance of liver disease and increase the number of people who suffer from this disease, we studied on liver disease through using two well-known methods in data mining area. In this paper, novel decision tree based algorithms is used which leads to considering more factors in general and predictions with high accuracy compared to other studies in liver disease. In this application, 583 UCI instances of liver disease dataset from the UCI repository are considered. This dataset consists of 416 records of liver disease and 167 records of healthy liver. This dataset is analyzed by two algorithms named Boosted C5.0 and CHAID algorithms. Until now there is no work in the literature that uses boosted C5.0 and CHAID for creating the rules in liver disease. Our results show that in both algorithms, the DB, ALB, SGPT, TB and A/G factors have a significant impact on predicting liver disease which according to the rules generated by both algorithms important ranges are DB = [10.900–1.200], ALB [4.00–4.300], SGPT = [34–37], TB = [0.600–1.200] (by boosted C5.0), A/G = [1.180–1.390], as well as in the Boosted C5.0 algorithm, Alkphos, SGOT and Age have significant impact in prediction of liver disease. By comparing the performance of these algorithms, it becomes clear that C5.0 algorithm via Boosting technique has an accuracy of 93.75% and this result reveals that it has a better performance than the CHAID algorithm which is 65.00%. Another important achievement of this paper is about the ability of both algorithms to produce rules in one class for liver disease. The results of our assessment show that Boosted C5.0 and CHAID algorithms are capable to produce rules for liver disease. Our results also show that boosted C5.0 considers the gender in liver disease, a factor which is missing in many other studies. Meanwhile, using the rules generated in boosted C5.0 algorithm, we obtained the important result about low susceptibility of female to liver disease than male. This factor is missing in other studies of liver disease. Therefore, our proposed computer-aided diagnostic methods as an expert and intelligent system have impressive impact on liver disease detection. Based on obtained results, we observed that our model had better performance compared to existing methods in the literature. [ABSTRACT FROM AUTHOR]

Applying effective supervised learning methods is difficult when dealing with complex industrial equipment such as Electrical Submersible Pump as the labeling work is costly and requires the dedication of an expert in the field. Moreover, the appearance of a new, unknown fault may require more time from the expert, not only for labeling the data but also for an arduous search for samples of this new type of fault. The solution developed in this paper uses uncertainty-based active learning for aiding experts through intelligent fault diagnosis and the search for potential samples of a new type of fault. The proposed approach was tested with a classical classification algorithm (random forest) applied to features extracted from the vibration signals. Two different feature extraction methods were investigated and contrasted: one using features defined by the expert and another using automatic feature extraction by a deep learning procedure. The algorithms were trained and tested on a dataset composed of 5616 vibration signals labeled in the frequency domain. The results show enough evidence to claim, at a significance value of 5%, that the active learning procedure proposed in this paper outperforms active learning based on random search. A new acquisition strategy for active learning is proposed to aid experts to find potential samples of a new fault. The recovered samples of the new fault aid improving performance of feature extractors based on deep learning since this type of method is highly affected by imbalanced data. • Active learning to enable class sample recovery of electrical submersible pump. • Active learning increases performance even with a restricted number of samples. • The proposed acquisition strategy better acquires samples of a certain class. • Proposed acquisition strategy reduces human effort on data for deep leaning models. [ABSTRACT FROM AUTHOR]

In this paper, we introduce a new adaptive rule-based classifier for multi-class classification of biological data, where several problems of classifying biological data are addressed: overfitting, noisy instances and class-imbalance data. It is well known that rules are interesting way for representing data in a human interpretable way. The proposed rule-based classifier combines the random subspace and boosting approaches with ensemble of decision trees to construct a set of classification rules without involving global optimisation. The classifier considers random subspace approach to avoid overfitting, boosting approach for classifying noisy instances and ensemble of decision trees to deal with class-imbalance problem. The classifier uses two popular classification techniques: decision tree and k-nearest-neighbor algorithms. Decision trees are used for evolving classification rules from the training data, while k-nearest-neighbor is used for analysing the misclassified instances and removing vagueness between the contradictory rules. It considers a series of k iterations to develop a set of classification rules from the training data and pays more attention to the misclassified instances in the next iteration by giving it a boosting flavour. This paper particularly focuses to come up with an optimal ensemble classifier that will help for improving the prediction accuracy of DNA variant identification and classification task. The performance of proposed classifier is tested with compared to well-approved existing machine learning and data mining algorithms on genomic data (148 Exome data sets) of Brugada syndrome and 10 real benchmark life sciences data sets from the UCI (University of California, Irvine) machine learning repository. The experimental results indicate that the proposed classifier has exemplary classification accuracy on different types of biological data. Overall, the proposed classifier offers good prediction accuracy to new DNA variants classification where noisy and misclassified variants are optimised to increase test performance. [ABSTRACT FROM AUTHOR]

Identifying standard fetal ultrasound (US) planes with key anatomical structures during mid-pregnancy prenatal screening is crucial for measuring fetal growth parameters and early detection of abnormalities. However, obtaining these standard planes is laborious and time-consuming and depends on the clinical experience of sonographers. Automatic detection of these planes can aid sonographers in identifying the correct standard planes. In recent times, various deep learning techniques have developed to automate the detection of standard fetal US planes. However, a common limitation among these approaches is their dependence on a single model prediction to make the final decision, which introduces the possibility of inaccuracies. Therefore, we propose an automated identification of commonly used standard fetal US planes based on the stacking ensemble of deep convolutional neural networks (CNN). The stacking ensemble method employs three pre-trained deep CNNs: AlexNet, VGG-19, and DarkNet-19. Softmax and random forest classifiers are used to get predictions from deep CNNs. The final prediction is made using the absolute majority voting technique. A publicly available fetal US dataset is employed to evaluate the performance of the stacking ensemble approach. The proposed ensemble model classifies fetal US planes into six distinct classes: abdomen, brain, femur, thorax, maternal cervix, and other (less commonly employed planes, such as kidney, and limbs) fetal planes. Experimental findings demonstrate that the stacking ensemble approach achieved high performance with an accuracy of 95.69%, precision of 94.02%, recall of 96.28%, F1-score of 95.08%, specificity of 99.12%, and Matthews correlation coefficient of 94.19% compared to individual deep CNN models and other competing methods. • This paper presents a stacking ensemble of CNNs to classify fetal US images. • The stacking ensemble learning model includes AlexNet, VGG-19, modified VGG-19 and DarkNet-19. • The final classification is made by using the absolute majority voting method. • A publicly available fetal US dataset is used to validate the proposed method. • Proposed method outperformed the state-of-the-art and competing algorithms. [ABSTRACT FROM AUTHOR]

In texture classification, local binary pattern (LBP) is currently one of the most widely-concerned feature encoding models. Most existing LBP-based texture classification methods are usually limited to single-kind texture features. In fact, an across-domain fusion of LBP features with other features, such as image contours, could be another potential path to promote texture classification. To enhance the feature modeling ability of LBP-based methods, this paper firstly designs a Cellular Neural Network (CellNN) with recurrent convolutions, initially trained by a simplified simulated-annealing algorithm, to extract informative image contours. For better reliability, a new three-channel contour extractor of deep CellNNs (i. e. , dCellNNs) is proposed. This extractor contains the initially-trained CellNNs of more than three layers, and it is further optimized by fine-tuning parameters. Moreover, a new weighted-base algorithm is designed to fulfill the fusion of the multi-scale texture features by LBPs and the contour features by dCellNNs to enhance feature representation. Finally, these enhanced features are concatenated together to generate the final multi-scale features of given texture image. On texture datasets KTH, Brodatz, OTC12 and UIUC, experiment results verify that the across-domain fusion of multi-scale LBPs and dCellNNs is efficient in capturing & enhancing texture features. With moderate feature dimensionality and computational costs, it could improve texture classification, acquiring an obvious accuracy increase on previous state-of-the-art ones, e.g. , a rise of 2.58% on KTH-TIPS2b, a rise of 3.11% on Brodatz, a rise of 0.71% on OTC12 and a rise of 0.42% on UIUC. • A multi-scale across-domain feature fusion of textures and contours is proposed. • The first deep model of CellNN (dCellNNs) with two-stage training is proposed. • The three-channel texture contour extractor of dCellNNs is designed. • A weighted-base fusing is designed for the across-domain fusion of features. [ABSTRACT FROM AUTHOR]

Due to the enormous amount of user-generated content being generated on the web, labeling such data is a time-consuming and expensive endeavor. As a result, we have limited annotated data and the vast majority of data are unlabeled. Analysis reveals that extracting (external) knowledge from unlabeled data and integrating it with knowledge extracted from labeled data is a beneficial task for text information processing, in particular text classification. In this paper, we present a hybrid approach for classifying sentiments that employs external knowledge, which is categorized as either general-purpose sentiment knowledge or domain-related knowledge. General-purpose sentiment knowledge is extracted from sentiment lexicons, whereas domain-related knowledge is extracted from unlabeled data from the same or related domains. Similar domains for a given domain are identified based on their similarity score in terms of overlapping features. The proposed approach utilizes both forms of external knowledge and combines them with logistic regression to train an improved classification model. The classification model uses the conventional gradient descent algorithm for optimization, and its convergence analysis indicates that it is convex and converges to the global optimum. The proposed approach is empirically evaluated and compared to three baselines and one state-of-the-art method using standard performance evaluation metrics on a multi-domain sentiment dataset. The experiment results are encouraging, demonstrating that the proposed approach considerably outperforms the baseline approaches and outperforms the state-of-the-art approach by up to 2% in terms of both f-score and accuracy. • A hybrid approach for document-level sentiment classification that makes use of external knowledge. • A domain-specific knowledge extraction approach from unlabeled text documents. • A detailed convergence analysis of the proposed optimization model used in the proposed approach. [ABSTRACT FROM AUTHOR]

Recently meta-learning-based few-shot learning methods have been widely used for relation classification. Previous work reveals that meta-learning performs poorly in scenarios where the edge probability distribution of the target domain dataset appears to be significantly different from the source domain. In this paper, we enhance the meta-learning framework with high-dimensional semantic feature extraction and hyperplane projection metrics for meta-tasks. First, we enhance the focus of BERT on entity words by adding entity markers and vector pooling. After that, the high-dimensional semantic features of the support set are extracted and transformed into hyperplanes. Finally, we obtain the classification results by calculating the projection distance between the query sample and the hyperplane. In addition, we design a auxiliary function with a plane correction factor, which can better amplify the plane spacing and reduce the degree of category confusion, which is important for solving the problem of metric spatial loss. Experiments on two real-world few-shot datasets show that our model HPN is more effective in classifying few-shot relations in the same domain and domain-adapted scenarios. And HPN is more stable on NOTA tasks. • A hyperplane projection network is proposed for few shot relation classification. • Propose an effective entity encoder that extracts better features. • First apply a novel hyperplane module to generate class-level feature. [ABSTRACT FROM AUTHOR]

Neighborhood rough set theories are commonly used in global feature selection to achieve high performance in continuous data classification. However, selecting a single feature subset to represent the entire dataset may degrade the performance when there are intra-class dissimilarities among objects. Therefore, this paper proposes a novel feature-selection method, Granule-specific Feature Selection (GFS) to select local feature subsets for continuous data classification. The feature selection approach constitutes a novel feature selection algorithm and a novel feature evaluation function and uses existing approaches for granule identification and classification with some adjustments. The neighborhood rough set theories are used in granule (subclass) identification within each class when there are no subclass label information available in the training data, while an improved k-Nearest Neighbors approach is used in classification with granule-specific feature subsets. Experimental results show GFS outperforms most of the global, class-specific, and local feature selection baselines in terms of classification performance. • Granule-specific feature selection using neighborhood rough set theories. • Feature selection while managing intra-class dissimilarities of objects. • Handle data uncertainty of continuous data when selecting features. • Improve classification performance using granule-specific feature subsets. [ABSTRACT FROM AUTHOR]

• Propose an improved random forest based on the improvement of decision trees. • Improve the evaluation mechanism for the classification effect of decision trees. • Propose a method for quantifying the diversity between decision trees. • Multiple tests verify the superiority of the proposed improved random forest. Random forest is one of the most widely used machine learning algorithms. Decision trees used to construct the random forest may have low classification accuracies or high correlations, which affects the comprehensive performance of the random forest. Aiming at these problems, the authors proposed an improved random forest based on the classification accuracy and correlation measurement of decision trees in this paper. Its main idea includes two parts, one is retaining the classification and regression trees (CARTs) with better classification effects, the other is reducing the correlations between the CARTs. Specifically, in the classification effect evaluation part, each CART was applied to make predictions on three reserved data sets, then the average classification accuracies were achieved, respectively. Thus, all the CARTs were sorted in descending order according to their achieved average classification accuracies. In the correlation measurement part, the improved dot product method was proposed to calculate the cosine similarity, i.e., the correlation, between CARTs in the feature space. By using the achieved average classification accuracy as reference, the grid search method was used to find the inner product threshold. On this basis, the CARTs with low average classification accuracy among CART pairs whose inner product values are higher than the inner product threshold were marked as deletable. The achieved average classification accuracies and correlations of CARTs were comprehensively considered, those with high correlation and weak classification effect were deleted, and those with better quality were retained to construct the random forest. Multiple experiments show that, the proposed improved random forest achieved higher average classification accuracy than the five random forests used for comparison, and the lead was stable. The G-means and out-of-bag data (OBD) score obtained by the proposed improved random forest were also higher than the five random forests, and the lead was more obvious. In addition, the test results of three non-parametric tests show that, there were significant diversities between the proposed improved random forest and the other five random forests. This effectively proves the superiority and practicability of the proposed improved random forest. [ABSTRACT FROM AUTHOR]

• A reconstruction framework is developed to classify mass news titles. • This framework can identify out-of-domain titles with in-domain words. • A cosine-based classification layer is designed to use domain knowledge. • This framework is trained on in-domain titles without out-of-domain titles. • This framework is an explainable classification approach. News title classification is a widely-used task for people to organize mass news. But traditional classification methods usually depend on a pre-defined label set, while many news articles may belong to unseen labels instead of pre-defined labels. To address this challenge, this paper develops a domain-knowledge based reconstruction framework to classify both out-of-domain (titles with unseen labels) and in-domain (titles with pre-defined labels) titles. Specifically, out-of-domain titles are identified through the reconstructed difference from autoencoder, while in-domain titles are classified to the pre-defined label with the highest cosine similarity between title and each category explanation text. Those category explanation texts are regarded as domain knowledge for in-domain aspect classification, aiming to make reconstruction framework more stable and meanwhile provide extra information for short titles. Furthermore, our method adopts BERT model to generate pre-trained embeddings for input texts. We conducted experiments on BRI (Belt and Road Initiative) news dataset and demonstrated that our method has achieved F1-score of 0.735 and 0.838 for out-of-domain title identification and in-domain aspect classification respectively. Moreover, we used titles from four countries to indicate our method's application value in cross-border business analysis. [ABSTRACT FROM AUTHOR]

The accumulation of large amounts of historical data has led to the wide application of ensemble learning over the past few decades, but the balance between the individual accuracy of base classifiers (BCs) and the diversity among these BCs is rarely considered in the construction of ensemble models. Since such a balance is crucial to the success of ensemble models, this paper proposes a Genetic Algorithm-based sequential instance selection framework to address this research gap. The novelties of the proposed framework include: transforming the balance between the individual accuracy of BCs and the diversity among BCs into a general combinatorial optimization model and designing a Genetic Algorithm-based evolutionary instance selection method to solve this model. The proposed framework not only overcomes the inherent limitations of the Genetic Algorithm in some high-dimensional tasks but also provides an explicit and automatic way to balance the accuracy and diversity by searching appropriate training data subsets for different component BCs. Based on obtained training data subsets, the component BCs of the ensemble model are generated sequentially, and their predictions are further combined with the weighted majority voting rule. Using 30 real datasets collected from various practical applications, such as medicine, business, and industry, the effectiveness of the proposed framework in constructing powerful ensemble models is examined and compared with six benchmark ensemble learning methods. In addition, the capability of the proposed framework to improve convergence performances is also examined by the comparison with the traditional Genetic Algorithm. [ABSTRACT FROM AUTHOR]

Deciding the signal length is an important challenge for one-class time-series classification (OCTSC). This paper aims to develop an OCTSC algorithm that does not require model retraining for different signal lengths. For this purpose, a distance-based one-class time-series classification approach using local cluster balance (OCLCB) is proposed. OCLCB extracts feature vectors, namely, local cluster balance (LCB), from the clustering results of sliding windows. K-means clustering is applied to the sliding windows extracted from the training signal. Then, the local prototype (LP) is calculated as the average of the local cluster balance (LCB) in the training data. Unseen scores are computed as the distance metrics between LP and LCBs in the testing data. Since the sliding window size is independent of the entire signal size, OCLCB does not need to retrain the model. This aspect gives the benefit of reducing the parameter tuning costs. The source code is uploaded at https://github.com/ToshiHayashi/OCLCB. [ABSTRACT FROM AUTHOR]

Inductive Logic Programming (ILP) and Relational Data Mining (RDM) address the task of inducing models or patterns from multi-relational data. One of the established approaches to RDM is propositionalization, characterized by transforming a relational database into a single-table representation. This paper presents a propositionalization technique called wordification which can be seen as a transformation of a relational database into a corpus of text documents. Wordification constructs simple, easy to understand features, acting as words in the transformed Bag-Of-Words representation. This paper presents the wordification methodology, together with an experimental comparison of several propositionalization approaches on seven relational datasets. The main advantages of the approach are: simple implementation, accuracy comparable to competitive methods, and greater scalability, as it performs several times faster on all experimental databases. Furthermore, the wordification methodology and the evaluation procedure are implemented as executable workflows in the web-based data mining platform ClowdFlows. The implemented workflows include also several other ILP and RDM algorithms, as well as the utility components that were added to the platform to enable access to these techniques to a wider research audience. [ABSTRACT FROM AUTHOR]

Distance metric learning is the task that aims to automate this process of learning task-specific distance functions in a supervised manner. In this paper, we study how to learn a Mahalanobis distance metric that can improve nearest neighbor classification. Our paper makes two contributions. First, we propose a novel framework named CLE_LMNN for Mahalanobis distance learning. CLE_LMNN builds on a recently proposed framework known as large margin nearest neighbor (LMNN) classification. Compared with LMNN, CLE_LMNN learns a Mahalanobis distance in a fine-grained way by first partitioning the labeled examples into subsets. As shown by our experiments, this fine-grained learning way is inclined to obtain a Mahalanobis distance more suitable for classification. Second, we present a novel algorithm named CLE for clustering labeled examples. Different from traditional unsupervised clustering algorithms, CLE fully employ the class information of labeled examples to effectively partition the set of examples with the same class into different subsets. To evaluate our proposed framework, we conduct extensive experiments on three real datasets. The experimental results show the effectiveness of CLE_LMNN when applied to classification. [ABSTRACT FROM AUTHOR]

The paper presents application of data mining methods for recognizing the most significant genes and gene sequences (treated as features) stored in a dataset of gene expression microarray. The investigations are performed for autism data. Few chosen methods of feature selection have been applied and their results integrated in the final outcome. In this way we find the contents of small set of the most important genes associated with autism. They have been applied in the classification procedure aimed on recognition of autism from reference group members. The results of numerical experiments concerning selection of the most important genes and classification of the cases on the basis of the selected genes will be discussed. The main contribution of the paper is in developing the fusion system of the results of many selection approaches into the final set, most closely associated with autism. We have also proposed special procedure of estimating the number of highest rank genes used in classification procedure. [ABSTRACT FROM AUTHOR]

The work presented in this papers is directed at mechanisms where by 3D surfaces can be represented to support the generation and application of classification techniques. Three different mechanisms are presented to allow for the representation of 3D surfaces in such a way that key features are retained while at the same time ensuring compatibility with prediction (classification) techniques. The three representation techniques are: (i) Local Geometry Matrices (LGMs) founded on the concept of local binary patterns, (ii) Local Distance Measure (LDM) founded on the idea that distances from edges (critical points) may be significant, and (iii) Point Series (PS) whereby local geometries are represented in terms of a linearisation of space. The representations are designed to capture the nature of 3D surfaces in terms of their local geometry and predict class labels associated with such local geometries. To act as a focus for the work the prediction of “springback” within the context of sheet metal forming is considered, where springback is a form of deformation that occurs (in a non-uniform manner) across a manufactured 3D surface as a result of the application of some sheet metal forming process. The evaluation of each of the techniques, and variations thereof, using sheet metal parts that have been manufactured especially for the purpose, is fully described. The paper also reports on a statistical significance test concerning the results. [ABSTRACT FROM AUTHOR]

• A theoretical framework of set oriented dimensionality reduction is proposed. • A new criterion for set oriented feature extraction based on CR is proposed. • An optimization framework based on race ratio maximization is developed. • Extensive experiments are conducted to show the effectiveness of our method. Given that collaborative representation (CR) methods have achieved great success in traditional single image based classification, recently, researchers have exploited the mechanism of collaborative representation to handle the case of image set based classification problem. However, without considering a proper criterion for feature extraction, this extension of collaborative representation mechanism suffers from the misleading coefficients of the incorrect classes on the uncontrolled datasets with small class separability. To address this limitation, inspired by large margin principle in discriminative analysis that aims to separately exploit the inter-class and intra-class variability, this paper proposes a novel theoretical framework of set oriented multiple kernel learning for dimensionality reduction based on collaborative representation classification. To achieve this framework, we integrate the learning of an optimal kernel from the multiple base kernels and a discriminative projection into a unified formulation. Moreover, robust feature information can be effectively extracted by minimizing the intra-class reconstruction residual and maximizing the inter-class reconstruction residual of the regularized hull modeled for the image sets. Since the criterion of feature extraction conforms to the mechanism of the collaborative representation classifier, the collaborative representation coefficients in our model can be much discriminative across classes. Notably, this research has important theoretical significance in improving the classification performance for collaborative representation classifier from the perspective of large margin discriminative learning. By employing the method of trace ratio maximization, we also develop a framework to solve the resulting nonconvex optimization problem efficiently. Extensive experiments on benchmark datasets well demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

• A model based on deep learning for environmental sound classification is proposed. • Model provides the trade-off between the audio length, accuracy and parameter space. • Model omits the signal processing modules and learns representations from audio. • Model has a small parameter space compared to other architectures in the literature. • The frequency and magnitude responses of some of the learned filters are analyzed. In this paper, we present an end-to-end approach for environmental sound classification based on a 1D Convolution Neural Network (CNN) that learns a representation directly from the audio signal. Several convolutional layers are used to capture the signal's fine time structure and learn diverse filters that are relevant to the classification task. The proposed approach can deal with audio signals of any length as it splits the signal into overlapped frames using a sliding window. Different architectures considering several input sizes are evaluated, including the initialization of the first convolutional layer with a Gammatone filterbank that models the human auditory filter response in the cochlea. The performance of the proposed end-to-end approach in classifying environmental sounds was assessed on the UrbanSound8k dataset and the experimental results have shown that it achieves 89% of mean accuracy. Therefore, the proposed approach outperforms most of the state-of-the-art approaches that use handcrafted features or 2D representations as input. Moreover, the proposed approach outperforms all approaches that use raw audio signal as input to the classifier. Furthermore, the proposed approach has a small number of parameters compared to other architectures found in the literature, which reduces the amount of data required for training. [ABSTRACT FROM AUTHOR]

• Our approach outperforms previous ones in identifying different query types. • Our approach has practical advantages related to the reduced number of features. • RandomForest outperformed other algorithms with 99.6% accuracy. In real-life classification problems, prior information about the problem and expert knowledge about the domain are often used to obtain reliable and consistent solutions. This is especially true in fields where the data is ambiguous, such as text, in which the same words can be used in seemingly similar texts, but have a different meaning. A promising avenue for text classification is machine learning, which has been shown to perform well in a variety of applications including query classification and sentiment analysis. Many of the proposed approaches rely on the bag-of-words representation, which loses the information about the structure of the text. In this paper, we propose a Customised Grammar Framework for text classification, which exploits domain-related information and a new way to represent text as a series of syntactic categories forming syntactic patterns. The framework employs a formal grammar approach for transforming the text into the syntactic patterns representation. We applied the framework for the query classification problem and our results show that our approach outperforms previous ones in terms of classification performance. [ABSTRACT FROM AUTHOR]

• In this study, we model a dynamic texture as a randomized neural network. • A randomized neural network is a feed forward neural network with a single-hidden layer. • Slices of the dynamic textures are divided into windows to build a training dataset. • Each window provides an input feature vector and corresponding label. • The weights of the output neuron layer are the dynamic texture descriptors. Dynamic texture analysis has been the focus of intensive research in recent years. Thus, this paper presents an innovative and highly discriminative dynamic texture analysis method, whose signature is composed of the weights of the output layer of a randomized neural network after a training procedure. This training is performed by using the pixels of slices of each orthogonal plane of the video (XY, YT , and XT) as input feature vectors and corresponding output labels. The obtained video signature provided an accuracy of 97.05%, 98.54%, 97.74% and 96.51% on the UCLA-50 classes, UCLA-9 classes, UCLA-8 classes and Dyntex++, respectively. These results, when compared to other dynamic texture analysis methods, demonstrate that our descriptors are very effective and that our proposed approach can contribute significantly to the field of dynamic texture analysis. [ABSTRACT FROM AUTHOR]

• A simultaneous prototypes and local metric learning algorithm is proposed. • The approach is used in biometrics and object recognition based on image set. • Learned prototypes dramatically reduces the storage and time costs. • Learned local metric significantly improves the discrimination ability of a set. • Experimental results of our method superior/comparable to the literatures. Classification based on image set is recently a competitive technique, where each set contains multiple images of a person or an object. As a widely used model, affine hull has shown its power in modeling image set. However, due to the existence of noise and outliers, the over-large affine hull usually matches fails when two hulls overlapped. Aiming at alleviating this handicap, this paper proposes a novel method for image set classification, namely Learning of Reduced Prototypes and Local Metric (LRPLM). Specifically, for each gallery image set, a reduced set of prototypes and an optimal local feature-wise metric are simultaneously learned, which jointly minimize the loss function involved the estimation of classification error probability. In doing so, LRPLM inherits the merits of affine hull with better representation to account for the unseen appearances and makes use of the powerful discriminative ability improved by the local metric. It looks like that LRPLM pulls similar image sets with the same class label "closer" to each other, while pushing dissimilar ones "far away". Extensive experiments illustrate the considerable effectiveness of LRPLM on three widely used datasets. As we know, classification is a research hotspot in expert and intelligent systems. Different from the previous classification methods, LRPLM focuses on image set-based classification technology, while most of them are single-shot classification technology. Thus, the proposed method can be considered as an expert system technology for medical diagnosis, security monitoring, object categorization, and biometrics recognition applications. [ABSTRACT FROM AUTHOR]

• A novel text classifier having the ability to visually explain its rationale. • Domain-independent classification that does not require feature engineering. • Support for incremental learning and text classification over streams. • Efficient framework for addressing early risk detection problems. • State-of-the-art performance on early depression detection task. With the rise of the Internet, there is a growing need to build intelligent systems that are capable of efficiently dealing with early risk detection (ERD) problems on social media, such as early depression detection, early rumor detection or identification of sexual predators. These systems, nowadays mostly based on machine learning techniques, must be able to deal with data streams since users provide their data over time. In addition, these systems must be able to decide when the processed data is sufficient to actually classify users. Moreover, since ERD tasks involve risky decisions by which people's lives could be affected, such systems must also be able to justify their decisions. However, most standard and state-of-the-art supervised machine learning models (such as SVM, MNB, Neural Networks, etc.) are not well suited to deal with this scenario. This is due to the fact that they either act as black boxes or do not support incremental classification/learning. In this paper we introduce SS3, a novel supervised learning model for text classification that naturally supports these aspects. SS3 was designed to be used as a general framework to deal with ERD problems. We evaluated our model on the CLEF's eRisk2017 pilot task on early depression detection. Most of the 30 contributions submitted to this competition used state-of-the-art methods. Experimental results show that our classifier was able to outperform these models and standard classifiers, despite being less computationally expensive and having the ability to explain its rationale. [ABSTRACT FROM AUTHOR]

• An efficient and effective UAV classification system using FMCW surveillance radar echo signals. • Micro-Doppler signature analysis on FMCW radar echoes from UAV. • UAV classification results on physically measured FMCW radar echo signals. Due to its small size, slow flying speed, and low flying altitude, classification of mini-sized unmanned aerial vehicles (UAVs) using a frequency-modulated continuous wave (FMCW) surveillance radar is a challenging task. This is because the FMCW radar echo signals are acquired at a short dwell time and thus contain limited information about targets. In this paper, we first analyze FMCW radar returns from various types of UAVs and non-UAV objects in terms of the micro-Doppler signature (m-DS) pattern. Based on the analysis results, we propose an effective and efficient UAV classification system using FMCW radar echo signals. The proposed system consists of five main parts namely, (i) burst selection, (ii) rule-based scan pruning, (iii) the empirical mode decomposition based m-DS analysis and features extraction, (iv) error counting minimization based class label estimation, and (v) scan-to-scan filtering. Our experimental results on physically measured FMCW radar echo signals from several types of UAVs and non-UAV objects show that the proposed system consistently outperforms a commercial-off-the-shelf UAV classification system in terms of the classification accuracy. [ABSTRACT FROM AUTHOR]

• A new method is proposed to detect key points in the time series. • Extracting shapelets based on key points and pruning redundant shapelets. • Experiments on real datasets verify effectiveness and efficiency of our proposal. Shapelet is a time series subsequence that can best represent the time series of one class. Shapelet can improve the accuracy and efficiency of classification, as well as the interpretability of classification results. Although shapelet has good classification performance, how to efficiently find the optimal shapelet is still an important challenge due to the large number of shapelet candidates contained in a time series. In this paper, a new shapelet discovery method, referred to as Pruning Shapelets with Key Points (PSKP), is proposed. PSKP first finds the key points in time series according to the standard deviation of each time tick of time series, and then extracts shapelet candidates with these key points. Finally, PSKP classifies the time series through a decision tree constructed based on the optimal shapelet. We make experiments on various data sets and evaluate the performance of the proposed method with compared candidates. The experimental results demonstrate that the proposed method is feasible and effective. [ABSTRACT FROM AUTHOR]

• We proposed a new ensemble learning framework in the setting of granular computing. • We conducted data transformation by information granulation for in-depth training. • We developed multi-level fusion strategies for in-depth classifiers fusion. • The proposed framework outperforms popular ensemble learning methods. • The proposed framework outperforms popular standard learning methods. Classification is a special type of machine learning tasks, which is essentially achieved by training a classifier that can be used to classify new instances. In order to train a high performance classifier, it is crucial to extract representative features from raw data, such as text and images. In reality, instances could be highly diverse even if they belong to the same class, which indicates different instances of the same class could represent very different characteristics. For example, in a facial expression recognition task, some instances may be better described by Histogram of Oriented Gradients features, while others may be better presented by Local Binary Patterns features. From this point of view, it is necessary to adopt ensemble learning to train different classifiers on different feature sets and to fuse these classifiers towards more accurate classification of each instance. On the other hand, different algorithms are likely to show different suitability for training classifiers on different feature sets. It shows again the necessity to adopt ensemble learning towards advances in the classification performance. Furthermore, a multi-class classification task would become increasingly more complex when the number of classes is increased, i.e. it would lead to the increased difficulty in terms of discriminating different classes. In this paper, we propose an ensemble learning framework that involves transforming a multi-class classification task into a number of binary classification tasks and fusion of classifiers trained on different feature sets by using different learning algorithms. We report experimental studies on a UCI data set on Sonar and the CK+ data set on facial expression recognition. The results show that our proposed ensemble learning approach leads to considerable advances in classification performance, in comparison with popular learning approaches including decision tree ensembles and deep neural networks. In practice, the proposed approach can be used effectively to build an ensemble of ensembles acting as a group of expert systems, which show the capability to achieve more stable performance of pattern recognition, in comparison with building a single classifier that acts as a single expert system. [ABSTRACT FROM AUTHOR]

• Variations can provide more valuable information for MI classification. • Sensitivity of the representative bispectral features extraction method. • Reduces the impacts of non-linear and non-Gaussian noises. • Enhances the separability of bispectral features. Motor imagery (MI)-related brain activities can be effectively described by frequency analysis. Bispectrum is developed to overcome the drawback of power spectrum that the estimation of power spectrum discards the phase relationship among frequency components. However, the widely used bispectral features extraction method adds up all bispectral values as one feature, which could lead to the loss of effective information and increase of the sensitivity to non-linear and non-Gaussian noises. Thus, the representative bispectral features extraction method may be inefficient for MI classification. In addition, recent research suggests that the variations of EEG signals could provide more useful underlying information of event-related brain responses. This paper presents an advanced variations based bispectral feature extraction method to improve the performance of MI classification. The proposed method calculates the variations of MI-related EEG signals as input to bispectrum estimation. Besides, a new segmented bispectral sum features are developed to reduce the influence of non-linear and non-Gaussian noises and emphasize the valuable information for MI classification. The dataset collected in our laboratory and BCI Competition IV dataset 2b were adopted to validate the proposed method. The results indicate that the proposed method outperforms the power spectrum based methods and the representative bispectral features based methods. Moreover, compared to other state-of-the-art works, our approach also achieves the greater performance for MI classification. [ABSTRACT FROM AUTHOR]