Wednesday, Apr. 23, 2025 Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文

Judgement and Disposal of Academic Misconduct Article Copyright Transfer Agreement Proof of Confidentiality Requirements for Electronic Version More.... Chinese Association of Automation National ResearchCenter for Intelligent Computing System Institute of Intelligent Machines,Chinese Academy of Sciences More....

2014 Vol.27 Issue.2, Published 2014-02-28 Papers and Reports    Researches and Applications        Papers and Reports 97 Multi-View Classification Method Based on Cross-View Constraints XUE Hui, CHEN Song-Can, LIU Jie, HUANG Ji-Jian A multi-view paired model, cross-view constraint, is taken into account and thus the pairwise constraints are extended in single-view learning. Instead of the strict paired constraints, the weaker constraint information is used, i.e. whether the data pairs between different views belong to the same class or not. Therefore, the cross-view constraints can not only include the totally paired constraints, but also be generalized to the case that the data are unpaired completely. Based on the cross-view constraints, a multi-view classification method is proposed. The proposed method can deeply mine the potential discriminative information in cross-view constraints and utilize the structural information of the data pairs as well. Experimental results demonstrate the effectiveness of the proposed method. 2014 Vol. 27 (2): 97-102 [Abstract] ( 503 ) [HTML 1KB] [ PDF 440KB] ( 892 ) 103 Kernel-Based Continuous-Action Actor-Critic Learning CHEN Xing-Guo, GAO Yang, FAN Shun-Guo, YU Ya-Jun In reinforcement learning, the learning algorithms frequently have to deal with both continuous state and continuous action spaces to control accurately. In this paper, the great capacity of kernel method for handling continuous state space problems and the advantage of actor-critic method in dealing with continuous action space problems are combined. Kernel-based continuous-action actor-critic learning(KCACL) is proposed grounded on the combination. In KCACL, the actor updates each action probability based on reward-inaction, and the critic updates the state value function according to online selective kernel-based temporal difference(OSKTD) learning. The experimental results demonstrate the effectiveness of the proposed algorithm. 2014 Vol. 27 (2): 103-110 [Abstract] ( 602 ) [HTML 1KB] [ PDF 488KB] ( 1058 ) 111 Supervised Multi-Manifold Learning Algorithm Based on ISOMAP SHAO Chao, WAN Chun-Hong The existing supervised multi-manifold learning algorithms adjust the distances between data points according to their class labels, and hence the multiple manifolds can be classified successfully. However,the poor generalization ability of these algorithms results in unfaithful display of the intrinsic geometric structure of some manifolds. A supervised multi-manifold learning algorithm based on Isometric mapping (ISOMAP) is proposed. The shortest path algorithm suitable for the multi-manifold structure is used to compute the shortest path distances which can effectively approximate the corresponding geodesic distances even in the multi-manifold structure. Then, Sammon mapping is used to further preserve shorter distances in the low-dimensional embedding space. Consequently, the intrinsic geometric structure of each manifold can be faithfully displayed. Moreover, the manifolds of new data points can be precisely judged based on the similarities between neighboring local tangent spaces according to the local Euclidean nature of the manifold, and thus the proposed algorithm obtains a good generalization ability. The effectiveness of the proposed algorithm is verified by experimental results. 2014 Vol. 27 (2): 111-119 [Abstract] ( 506 ) [HTML 1KB] [ PDF 1532KB] ( 926 ) 120 An Image Classification Method Based on Hyperedge Correlation XU Jie, JING Li-Ping, YU Jian Traditional hypergraph-based image classification methods overlook the correlation among hyperedges in hypergraph construction, which results in poor classification performance. A method based on hyperedge correlation is proposed in this paper. The correlation among hyperedses is quantified by combining the image vision and its corresponding tags information. The tags corresponding to the image are introduced into the image classification as indicator information and thus better classification performance is obtained. The effectiveness of the proposed method is verified by experiments conducted on datasets such as LabelMe and UIUC. 2014 Vol. 27 (2): 120-126 [Abstract] ( 416 ) [HTML 1KB] [ PDF 916KB] ( 954 ) 127 Multi-Sample Incremental Manifold Learning Algorithm Based on Isogonal Mapping TAN Chao, GUAN Ji-Hong, ZHOU Shui-Geng In the classical dimension reducing manifold learning algorithms, the distance is used to measure the similarity between data, and the problem of subspace deviation caused by noise can not be solved.A multi-sample incremental manifold learning algorithm based on Isogonal mapping is proposed. The covariance matrix of the high dimensional samples with sample mean as the center is turned into the covariance matrix with neighborhood mean as the center. Thus, the error of the subspace caused by distance measurement is eliminated, the covariance matrix is weighted, and the effect of noise or irregular new samples on dimension reduction is reduced. Experimental results show an improvement of the proposed algorithm compared with other algorithms. Moreover, the proposed algorithm can be well applied to image recognition. 2014 Vol. 27 (2): 127-133 [Abstract] ( 572 ) [HTML 1KB] [ PDF 441KB] ( 915 ) 134 Improved Ordinal Decisions Trees Algorithms Based on Rank Entropy CHEN Jian-Kai, WANG Xi-Zhao, GAO Xiang-Hui When the expanded attributes are selected for decision tree learning based on rank entropy, computing the rank mutual information of every single cut for each of the continuous-valued attributes is required to get the expanded attribute by comparing the values of rank mutual information. Therefore, the computational complexity is high. Aiming at this problem, cut-points are divided into stable and unstable cut-points and a mathematical model is established in this paper. The proposed model theoretically proves that the rank mutual information function achieves its maximum not at stable cut-points, but at unstable cut-points. The result means that in the algorithm only traversing the unstable cut-points is required instead of computing the values of the stable cut-points. Thus, the computational efficiency of building decision trees is greatly improved, which is confirmed by the numerical experimental results. 2014 Vol. 27 (2): 134-140 [Abstract] ( 433 ) [HTML 1KB] [ PDF 444KB] ( 1320 ) 141 New Words Discovery in Microblog Content HUO Shuai, ZHANG Min, LIU Yi-Qun, MA Shao-Ping New words discovery is of great significance in the field of natural language processing. It is more difficult to find new words in microblog than in other corpus. In this paper, an algorithm based on context entropy is proposed, and the new word candidates are filtered based on the context. To improve the precision, lexical features are introduced and an algorithm combining them with term frequency is put forward. Thus, the precision rate and the recall rate are greatly improved, and the F-measure value is up to 89.6%. 2014 Vol. 27 (2): 141-145 [Abstract] ( 530 ) [HTML 1KB] [ PDF 388KB] ( 2769 ) Researches and Applications 146 Dynamic Probabilistic Particle Swarm Optimization Based on Heterogeneous Multiple Population Strategy NI Qing-Jian, DENG Jian-Ming, XING Han-Cheng Aiming at premature convergence and the slow search speed of the traditional particle swarm optimization, a heterogeneous multiple population strategy is combined with the characteristics of dynamic probabilistic particle swarm optimization (DPPSO). In the evolutionary process of DPPSO with the strategy, multiple sub-populations are maintained and each sub-population evolves with different DPPSO variants. According to certain rules, communication between the sub-populations are executed to maintain the information exchange inside the entire population and coordinate exploration and exploitation. DPPSO algorithms with the strategy are tested on four benchmark functions which are commonly used in the evolutionary computation. Experimental results demonstrate that the DPPSO with the strategy significantly improves the convergence speed and stability with strong global search ability. 2014 Vol. 27 (2): 146-152 [Abstract] ( 451 ) [HTML 1KB] [ PDF 563KB] ( 963 ) 153 Integration of Random Subspace and Min-Max Modular SVM YU Yi, WU Jiao-Gao, LI Yun The min-max modular support vector machine (M3-SVM) is a powerful tool for dealing with large-scale data. To improve the classification performance of M3-SVM for unblanced data with high dimension, several random subspace strategies are analyzed and combined with M3-SVM to reduce the dimensionality and add the ensemble mechanism on feature level. Thus, the min-max modular support vector machine with random subspace is proposed. The experimental results on real-world datasets including unbalanced data indicate that the proposed random subspace strategy enhances the classification of M3-SVM. Moreover, the diversity between sub-modules (base learner) in M3-SVM is discussed. 2014 Vol. 27 (2): 153-159 [Abstract] ( 518 ) [HTML 1KB] [ PDF 537KB] ( 846 ) 160 Density-Punished Support Vector Data Description ZHANG Li, ZHANG Hai-Fei, ZHOU Wei-Da, LIN Ying, LI Fan-Zhang Based on the concept of relative density degrees, a density-punished support vector data description method is presented. The relative density degrees are associated with punishing misclassifications. If the relative density degree of the sample is large, it is likely to be a target sample. Thus, a large penalty should be put on its misclassification. Similarly, if the relative densitydegree of the sample is small, it might be a boundary or noise point so that the corresponding penalty for its misclassification should be small as well. The experimental results on UCI datasets show that the proposed method has better performance compared with support vector data description and density-induced support vector data description. 2014 Vol. 27 (2): 160-165 [Abstract] ( 428 ) [HTML 1KB] [ PDF 407KB] ( 963 ) 166 Linked Social Media Data Based Semi-Supervised Feature Selection Method WANG Yi-Bing, PAN Zhi-Song, WU Jun-Qing, JIA Bo, HU Gu-Yu Mountains of high-dimensional, unlabeled data are produced by the social media network, which brings tremendous challenges to the data processing. Meanwhile, the linked graph information between data samples can not be effectively used in the existing pattern recognition algorithms. A semi-supervised feature selection method (SSLFS) based on linked relations is proposed combined with a little supervised information after mining the linked graph of social media network. Through spectral analysis and sparsity constraint, SSLFS selects feature subsets which maintain the characteristics of local manifold and sparsity. The experimental results on the Flickr dataset show that the subset obtained by SSLFS is more effective when applied to classification compared with those by other methods. 2014 Vol. 27 (2): 166-172 [Abstract] ( 439 ) [HTML 1KB] [ PDF 437KB] ( 976 ) 173 Improved Covariance Feature Based Lie-KNN Classification Algorithm WANG Bang-Jun, LI Fan-Zhang, ZHANG Li, YU Jian, HE Shu-Ping K-nearest neighbor(KNN) classification is simple, efficient and widely used for classification problems or as a base of comparison. However, the data, especially those with complex high-dimensional structures, do not always belong to the Euclidean space in practical application. How to select the features of samples and calculate the distances between them is a hard problem in KNN. With full consideration of various factors, a multi-covariance Lie-KNN classification method is put forward based on the image region covariance. In this method, the simplicity and the validity of KNN and the abilities of Lie group structure to represent complex data and calculate distances are fully used. It effectively solves the classification problems of complex high-dimensional data. Experimental results on handwritten numerals verify its effectiveness. 2014 Vol. 27 (2): 173-178 [Abstract] ( 424 ) [HTML 1KB] [ PDF 402KB] ( 771 ) 179 Sparsity Preserving Canonical Correlation Analysis with Missing Samples ZU Chen, ZHANG Dao-Qiang Based on the canonical correlation analysis (CCA), a supervised learning algorithm, sparsity preserving CCA with missing samples (SPCCAM), is proposed. The class information of samples is introduced by sparsity preserving and cross correlation is used to overcome the limitations of the CCA and its extensions that the paired samples of different views are required. SPCCAM can combine features from different views with unpaired training samples. The experimental results on the artificial dataset, multiple feature database and PIE facial database show that the proposed SPCCAM effectively enhances the classification performance by using class information. 2014 Vol. 27 (2): 179-186 [Abstract] ( 457 ) [HTML 1KB] [ PDF 585KB] ( 774 ) 187 EasyEnsemble.M for Multiclass Imbalance Problem LI Qian-Qian, LIU Xu-Ying The potential useful information in the majority class is ignored by stochastic under-sampling. When under-sampling is applied to multi-class imbalance problem, this situation becomes even worse. In this paper, EasyEnsemble.M for multi-class imbalance problem is proposed. The potential useful information contained in the majority classes which is ignored is explored by stochastic sampling the majority classes for multiple times. Then, sub-classifiers are learned and a strong classifier is obtained by using hybrid ensemble techniques. Experimental results show that EasyEnsemble.M is superior to other frequently used multi-class imbalance learning methods when G-mean is used as performance measure. 2014 Vol. 27 (2): 187-192 [Abstract] ( 742 ) [HTML 1KB] [ PDF 378KB] ( 2201 )

Supervised by China Association for Science and Technology Sponsored by Chinese Association of Automation NationalResearchCenter for Intelligent Computing System Institute of Intelligent Machines, Chinese Academy of Sciences Published by Science Press

Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn Supported by Beijing Magtech 　Email:support@magtech.com.cn