Search

Multi-label classification has been successfully applied to image annotation, information retrieval, text categorization, etc. When the number of classes increases significantly, the traditional multi-label learning models will become computationally impractical. Label space dimension reduction (LSDR) is then developed to alleviate the effect of the high dimensionality of labels. However, almost all the existing LSDR methods focus on single-view learning. In this paper, we develop a multi-view label embedding (MVLE) model by exploiting the multi-view correlations. The label space and feature space of each view are bridged by a latent space. To exploit the consensus among different views, multi-view latent spaces are correlated by Hilbert–Schmidt independence criterion(HSIC). For a test sample, it is firstly embedded to the latent space of each view and then projected to the label space. The prediction is conducted by combining the multi-view outputs. Experiments on benchmark databases show that MVLE outperforms the state-of-the-art LSDR algorithms in both multi-view settings and different multi-view learning strategies.

Hierarchical classification is significant for big data, where the original task is divided into several sub-tasks to provide multi-granularity predictions based on a tree-shape label structure. Obviously, these sub-tasks are highly correlated: results of the coarser-grained sub-tasks can reduce the candidates for the fine-grained sub-tasks, while results of the fine-grained sub-tasks provide attributes describing the coarser-grained classes. A human can integrate feedbacks from all the related sub-tasks instead of considering each sub-task independently. Therefore, we propose a deep collaborative multi-task network for hierarchical image classification. Specifically, we first extract the relationship matrix between every two sub-tasks defined by the hierarchical label structure. Then, the information of each sub-task is broadcasted to all the related sub-tasks through the relationship matrix. Finally, to combine this information, a novel fusion function based on the task evaluation and the decision uncertainty is designed. Extensive experimental results demonstrate that our model can achieve state-of-the-art performance.

Face photo-sketch synthesis refers transforming a face image between photo domain and sketch domain. It plays a crucial role in law enforcement and digital entertainment. A great deal of efforts have been devoted on face photo-sketch synthesis. However, limited by the weak identity supervision, existing methods mostly yield indistinct details or great deformation, resulting in poor perceptual appearance or low recognition accuracy. In the past several years, face identification achieved great progress, which represents the face images much more precisely than before. Considerring the face image translation is also a type of face image re-representation, we attempt to introduce face recognition models to improve the synthesis performance. First, we applied existing synthesis models to augment the training set. Then, we proposed a full-scale identity supervision method to reduce redundant information introduced by these pseudo samples and take the valid information to enhance the intra-class variations. The proposed framework consists of two sub-networks: cross-domain translation (CT) network and intra-domain adaptation (IA) network. The CT network translates the input image from source domain to latent image of target domain, which overcomes the great gap between two domains with less structural deformation. The IA network adapts the perceptual appearance of latent image to target image by adversarial learning. Experimental results on CUHK Face Sketch Database and CUHK Face Sketch FERET Database demonstrate the proposed method preserved best perceptual appearance and more distinct details with less deformation.

To ensure the operational safety and reliability, fault recognition of complex systems is becoming an essential process in industrial systems. However, the existing recognition methods mainly focus on common faults with enough data, which ignore that many faults are lack of samples in engineering practice. Transfer learning can be helpful, but irrelevant knowledge transfer can cause performance degradation, especially in complex systems. To address the above problem, a hierarchy guided transfer learning framework (HGTL) is proposed in this paper for fault recognition with few-shot samples. Firstly, we fuse domain knowledge, label semantics and inter-class distance to calculate the affinity between categories, based on which a category hierarchical tree is constructed by hierarchical clustering. Then, guided by the hierarchical structure, the samples in most similar majority classes are selected from the source domain to pre-train the hierarchical feature learning network (HFN) and extract the transferable fault information. For the fault knowledge extracted from the child nodes of one parent node are similar and can be transferred with each other, so the trained HFN can extract better features of few samples classes with the help of the information from similar faults, and used to address few-shot fault recognition problems. Finally, a dataset of a nuclear power system with 65 categories and the widely used Tennessee Eastman dataset are analyzed respectively via the proposed method, as well as state-of-the-art recognition methods for comparison. The experimental results demonstrate the effectiveness and superiority of the proposed method in fault recognition with few-shot problem.

Long-tail distribution is widespread in many practical applications, where most categories contain only a small number of samples. As sufficient instances cannot be obtained for describing the intra-class diversity of the minority classes, the separating hyperplanes learned by traditional machine learning methods are usually heavily skewed. Resampling techniques and cost-sensitive algorithms have been introduced to enhance the statistical power of the minority classes, but they cannot infer more reliable class boundaries beyond the description of samples in the training set. To address this issue, we cluster the original categories into super-class to produce a relatively balanced distribution in the super-class space. Moreover, the knowledge shared among categories belonging to a certain super-class can facilitate the generalization of the minority classes. However, existing super-class construction methods have some inherent disadvantages. Specifically, taxonomy-based methods suffer a gap between the semantic space and the feature space, and the performance of learning-based algorithms strongly depends on the features and data distribution. In this paper, we propose a deep super-class learning (DSCL) model to tackle the problem of long-tail distributed image classification. Motivated by the observation that classes belonging to the same super-class usually have more similar evaluations on the features than those belonging to different super-classes, we design a block-structured sparse constraint and attach it on the top of a convolutional neural network. Thus, the proposed DSCL model can accomplish representation learning, classifier training, and super-class construction in a unified end-to-end learning procedure. We compared the proposed model with several super-class construction methods on two public image datasets. Experimental results show that the super-class construction strategy is effective for the long-tail distributed classification, and the DSCL model can achieve better results than the other methods.

The consistently increasing of the feature dimension brings about great time complexity and storage burden for multi-label learning. Numerous multi-label feature selection techniques are developed to alleviate the effect of high-dimensionality. The existing multi-label feature selection algorithms assume that the labels of the training data are complete. However, this assumption does not always hold true for labeling data is costly and there is ambiguity among classes. Hence, in real-world applications, the data available usually have an incomplete set of labels. In this paper, we present a novel multi-label feature selection model under the circumstance of missing labels. With the proposed algorithm, the most discriminative features are selected and missing labels are recovered simultaneously. To remove the irrelevant and noisy features, the effective l2, p-norm (0

Unsupervised feature selection (UFS) aims to reduce the time complexity and storage burden, improve the generalization ability of learning machines by removing the redundant, irrelevant and noisy features. Due to the lack of training labels, most existing UFS methods generate the pseudo labels by spectral clustering, matrix factorization or dictionary learning, and convert UFS to a supervised problem. The learned clustering labels reflect the data distribution with respect to classes and therefore are vital to the UFS performance. In this paper, we proposed a novel subspace clustering guided unsupervised feature selection (SCUFS) method. The clustering labels of the training samples are learned by representation based subspace clustering, and features that can well preserve the cluster labels are selected. SCUFS can well learn the data distribution in that it uncovers the underlying multi-subspace structure of the data and iteratively learns the similarity matrix and clustering labels. Experimental results on benchmark datasets for unsupervised feature selection show that SCUFS outperforms the state-of-the-art UFS methods. HighlightsA novel subspace clustering guided unsupervised feature selection (SCUFS) model is proposed.SCUFS learns a similarity graph by self-representation of samples and can uncover the underlying multi-subspace structure of data.The iterative updating of similarity graph and pseudo label matrix can learn a more accurate data distribution.

Most of manifold learning based feature extraction methods are two-step methods, which first construct a weighted neighborhood graph and then use the pre-constructed graph to perform subspace learning. As a result, these methods fail to use the underlying correlation structure of data to learn an adaptive graph to preciously characterize the similarity relationship between samples. To address this problem, we propose a novel unsupervised feature extraction method called low-rank adaptive graph embedding (LRAGE), which can perform subspace learning and adaptive probabilistic neighborhood graph embedding simultaneously based on reconstruction error minimization. The proposed LRAGE is imposed with low-rank constraint for the sake of exploring the underlying correlation structure of data and learning more informative projection. Moreover, the L 2 , 1 -norm penalty is imposed on the regularization to further enhance the robustness of LRAGE. Since the resulting objective function has no closed-form solutions, an iterative optimization algorithm is elaborately designed. The convergence of the proposed algorithm is proved and the corresponding computational complexity analysis is also presented. In addition, we explore the potential properties of the proposed LRAGE by comparing it with several similar models on both synthetic and real-world data sets. Extensive experiments on five well-known face data sets and three non-face data sets demonstrate the superiority of the proposed LRAGE.

A deep discriminative structured model, convolutional neural random fields (CNRFs), is proposed for action recognition problem. In the proposed model, a spatio-temporal convolutional neural network (CNN) is developed for invariant feature learning from raw input frames, and the CNN is combined with conditional random fields (CRFs) for capturing the interdependencies between outputs. The parameters from both CRF and CNN are learned in a joint fashion which enables structured prediction and feature learning as well. We also explore different combinations of observation and transition feature functions based on the learned high level features from convolution part. The approach enjoys the advantages of both CNN and CRF, it has the invariant feature learning ability possessed by the former and structured prediction ability of the latter. The experimental results on both segmented and unsegmented human action recognition datasets show that CNRF boosts the performance over the comparison methods by a large margin. HighlightsWe develop a novel spatio-temporal CNN architecture for feature learning from video frames.CRF is coupled with CNN to achieve feature learning and structured prediction simultaneously.We explore different combinations of feature functions for sequence labeling.We validate our framework on segmented and unsegmented action datasets, respectively.

A support vector machine (SVM) is a popular algorithm for classification learning. The classical SVM effectively manages classification tasks defined by means of numerical attributes. However, both numerical and nominal attributes are used in practical tasks and the classical SVM does not fully consider the difference between them. Nominal attributes are usually regarded as numerical after coding. This may deteriorate the performance of learning algorithms. In this study, we propose a novel SVM algorithm for learning with heterogeneous data, known as a heterogeneous SVM (HSVM). The proposed algorithm learns an mapping to embed nominal attributes into a real space by minimizing an estimated generalization error, instead of by direct coding. Extensive experiments are conducted, and some interesting results are obtained. The experiments show that HSVM improves classification performance for both nominal and heterogeneous data. HighlightsWe propose an algorithm to map nominal features to a numerical space via minimizing estimated generalization errors.We integrate the mapping algorithm with support vector machines and result in an improved learning algorithm from heterogeneous data.Experiments show the proposed technique is effective for learning with heterogeneous data and also help deal with imbalanced tasks.

By removing the irrelevant and redundant features, feature selection aims to find a compact representation of the original feature with good generalization ability. With the prevalence of unlabeled data, unsupervised feature selection has shown to be effective in alleviating the curse of dimensionality, and is essential for comprehensive analysis and understanding of myriads of unlabeled high dimensional data. Motivated by the success of low-rank representation in subspace clustering, we propose a regularized self-representation (RSR) model for unsupervised feature selection, where each feature can be represented as the linear combination of its relevant features. By using L 2 , 1 -norm to characterize the representation coefficient matrix and the representation residual matrix, RSR is effective to select representative features and ensure the robustness to outliers. If a feature is important, then it will participate in the representation of most of other features, leading to a significant row of representation coefficients, and vice versa. Experimental analysis on synthetic and real-world data demonstrates that the proposed method can effectively identify the representative features, outperforming many state-of-the-art unsupervised feature selection methods in terms of clustering accuracy, redundancy reduction and classification accuracy. HighlightsA regularized self-representation (RSR) model is proposed for unsupervised feature selection.An iterative reweighted least-squares algorithm is proposed to solve the RSR model.The proposed method shows superior performance to state-of-the-art.

Ensemble learning has attracted considerable attention owing to its good generalization performance. The main issues in constructing a powerful ensemble include training a set of diverse and accurate base classifiers, and effectively combining them. Ensemble margin, computed as the difference of the vote numbers received by the correct class and the another class received with the most votes, is widely used to explain the success of ensemble learning. This definition of the ensemble margin does not consider the classification confidence of base classifiers. In this work, we explore the influence of the classification confidence of the base classifiers in ensemble learning and obtain some interesting conclusions. First, we extend the definition of ensemble margin based on the classification confidence of the base classifiers. Then, an optimization objective is designed to compute the weights of the base classifiers by minimizing the margin induced classification loss. Several strategies are tried to utilize the classification confidences and the weights. It is observed that weighted voting based on classification confidence is better than simple voting if all the base classifiers are used. In addition, ensemble pruning can further improve the performance of a weighted voting ensemble. We also compare the proposed fusion technique with some classical algorithms. The experimental results also show the effectiveness of weighted voting with classification confidence.

Feature subset selection has become an important challenge in areas of pattern recognition, machine learning and data mining. As different semantics are hidden in numerical and categorical features, there are two strategies for selecting hybrid attributes: discretizing numerical variables or numericalize categorical features. In this paper, we introduce a simple and efficient hybrid attribute reduction algorithm based on a generalized fuzzy-rough model. A theoretic framework of fuzzy-rough model based on fuzzy relations is presented, which underlies a foundation for algorithm construction. We derive several attribute significance measures based on the proposed fuzzy-rough model and construct a forward greedy algorithm for hybrid attribute reduction. The experiments show that the technique of variable precision fuzzy inclusion in computing decision positive region can get the optimal classification performance. Number of the selected features is the least but accuracy is the best.