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....

2015 Vol.28 Issue.11, Published 2015-11-30 Papers and Reports    Researches and Applications        Papers and Reports 961 Active Semi-supervised Affinity Propagation Clustering Algorithm LEI Qi, YU Hui-Ping, WU Min Affinity propagation clustering algorithm is completely unsupervised and it has the problem of ignoring the internal structure and the hierarchy of clusters. An active learning and pair-wise constrains based semi-supervised affinity propagation (AP) clustering algorithm is proposed. An active learning strategy is designed and the data with the biggest uncertainty are queried to gain the valuable constraint information as much as possible. Based on the constraint information, the similarity matrix is adjusted to guide the clustering process in AP algorithm. To verify the effectiveness of the proposed algorithm, experiment is carried out on the UCI machine learning repository and face image datasets and the results indicate that the proposed clustering algorithm improves the clustering performance significantly. 2015 Vol. 28 (11): 961-968 [Abstract] ( 472 ) [HTML 1KB] [ PDF 721KB] ( 417 ) 969 Algorithm of Detecting Community in Bipartite Network with Autonomous Determination of the Number of Communities GUO Gai-Gai, QIAN Yu-Hua, ZHANG Xiao-Qin, LI Ye-Bin The existing algorithms can find the community structure in bipartite network. However, they can not predict the number of communities and the relevant information and discover the real community structure accurately due to the variety and the complexity of the real network. In this paper, an algorithm of detecting community structure in bipartite network-cluster assign algorithm (CAA) is proposed and it determines the number of communities autonomously. In this algorithm, the interaction information between two types of nodes is used effectively and the problem of determining the number of communities is solved. The T-type nodes of the network are clustered, then the B-type nodes are assigned to the existing classes according to the allocation mechanism. Experiments show CAA obtains a higher quality community and has a higher accuracy than the algorithms based on resource distribution matrix and edge cluster coefficient. 2015 Vol. 28 (11): 969-975 [Abstract] ( 515 ) [HTML 1KB] [ PDF 625KB] ( 736 ) 976 Deep Learning Expert Ranking Method Based on Listwise LI Xian-Hui, YU Zheng-Tao, WEI Si-Chao, GAO Sheng-Xiang, WANG Li-Ren The traditional expert list ranking method is easy to fall into local minimum, its training time is long, and the ranking function can not be approximated well. Combining listwise expert ranking with deep neural network, a deep learning expert ranking method based on listwise is proposed. Firstly, a deep learning expert ranking model is presented. Through unsupervised self-training, better parameters are obtained to initialize weights layer by layer. Then, the training instances formed by the expert documents corresponding to the queries are inputted into the restricted Boltzmann machines for the training. Finally, cosine value is used instead of matrix subtraction to compute weight. Thus, the whole replacement of weights is finished and the deep learning expert ranking model is constructed. The comparative experiments of expert ranking show that the proposed method is efficient and it improves the accuracy of ranking effectively. 2015 Vol. 28 (11): 976-982 [Abstract] ( 629 ) [HTML 1KB] [ PDF 597KB] ( 1615 ) 983 Overlapping Community Detection Algorithm Based on Two-Stage Clustering JIANG Sheng-Yi, YANG Bo-Hong, LI Min-Min, WU Mei-Ling, WANG Lian-Xi Aiming at the complex network overlapping community detection, an overlapping community detection algorithm based on two-stage clustering is proposed. Eigen decomposition is applied to network adjacency matrix. The nodes are projected into k-dimensional Euclidean space, and then they are clustered by hard and soft clustering algorithm to detect the structure of overlapping community efficiently and adaptively. At the stage of hard clustering, a clustering algorithm based on the principle of minimum distance is introduced to divide nodes autonomously, and the number of communities and cluster centers for the soft clustering stage are determined. At the stage of soft clustering, fuzzy C-means algorithm is introduced and the fuzzy modularity is considered as objective function for the algorithm. Through iterative optimization of the fuzzy modularity, a soft partition is realized and overlapping community structures in network can be figured out. Experiments are carried out on a number of real network datasets, and the results indicate that the proposed algorithm can mine overlapping community structure in complex network with high efficiency and effectiveness. 2015 Vol. 28 (11): 983-991 [Abstract] ( 406 ) [HTML 1KB] [ PDF 572KB] ( 688 ) 992 n-grams Features Weighting Algorithm Based on Relevance and Semantic QIU Yun-Fei, LIU Shi-Xing, LIN Ming-Ming, SHAO Liang-Shan When n-grams are considered as text classification features, the classification accuracy is decreased. The redundancy and relevance between words are ignored while n-grams are weighted. Thus, n-grams features weighting algorithm based on relevance and semantic is proposed. To decrease the internal redundancy, feature reduction is conducted to n-grams during text preprocessing. Then, n-grams are weighted according to the relevance of words and classes in n-grams and the semantic similarity of n-grams and testing dataset. The experimental results on Sougo Chinese news corpse and NetEase text corpse show that the proposed algorithm can select n-grams features of high relevance and low redundancy, and reduce the sparse data while quantifying the testing dataset. 2015 Vol. 28 (11): 992-1001 [Abstract] ( 426 ) [HTML 1KB] [ PDF 901KB] ( 410 ) Researches and Applications 1002 Cross-Language Sentiment Classification Algorithm Based on Dependency Analysis Parser and Weight on Property Probability ZHANG Ling-Ling, JI Jun-Zhong, BEI Fei, WU Chen-Sheng In the document-level sentiment classification methods,only the distribution information of emotion is taken into account, while the semantic emotion knowledge is ignored. To solve these problems,a cross-language sentiment classification algorithm based on the dependency analysis and property probability weight is proposed. Firstly, dependency relations are got by dependency relation parsing before translation. Then, based on the correlation between the distribution of dictionary polar and the document-level sentiment classification, the weight feature of property probability is merged into Naive bayesian classification to improve the classification effect. Finally, extensive experiments are performed on English datasets for training and standard Chinese datasets for testing. The results show that the proposed algorithm is superior to other existing algorithms in performance. 2015 Vol. 28 (11): 1002-1012 [Abstract] ( 488 ) [HTML 1KB] [ PDF 793KB] ( 515 ) 1013 Variable Granularity and Simulated Feedback Mechanism Based Burning State Intelligent Cognitive Method of Rotary Kiln Sintering Process CHEN Ke-Qiong, WANG Jian-Ping, LI Wei-Tao, ZHAO Li-Xin An improved compressed Gabor filter bank is used for flame image pre-processing, and the scale-invariant feature transform descriptor is combined with bag of visual words and latent semantic analysis to extract the local configuration features of the flame image region of interests. A simple feature space is constructed based on the definition of feature resolution, cognitive granular entropy, and feature weight in the given level of cognitive information granularity. The multi-dimensional reverse normal particle cloud model of training samples is generated and the pattern classifier is constructed based on cloud-membership to obtain the burning state classification rules of rotary kiln sintering process. Variable granularity and simulated feedback mechanism based burning state intelligent cognitive method of rotary kiln sintering process is presented based on the definition of cognitive error. Experiments show that the proposed method is superior in cognizing the burning state to other methods. 2015 Vol. 28 (11): 1013-1022 [Abstract] ( 368 ) [HTML 1KB] [ PDF 657KB] ( 333 ) 1023 Image Encoding Algorithm Based on Contour and Curvature-Driven Inpainting WANG Xiang-Hai , JI Hai-Wei , SONG Chuan-Ming A perception-oriented image coding algorithm based on contour and curvature-driven diffusion is presented. Firstly, the gradient variance and binary edge map are used to separate the image into non-overlapped structural blocks, gradient blocks and common blocks. Subsequently, all common blocks, necessary structural blocks and gradient blocks, and their assistant information are compressed by JPEG and arithmetic coding. Finally, common blocks are decoded by the decoder with JPEG. Moreover, the gradient weighted linear interpolation is employed to reconstruct the gradient blocks, and the structural blocks are decoded by combining contour and curvature-driven diffusion methods to obtain a better preservation of integrity and strength of principal edges. Experimental results show that the proposed algorithm achieves higher compression ratio, better visual quality and lower computational complexity than JPEG and parameter-assistant inpainting. 2015 Vol. 28 (11): 1023-1032 [Abstract] ( 414 ) [HTML 1KB] [ PDF 596KB] ( 517 ) 1033 Fast Keyword Spotting in Handwritten Chinese Documents Using Index YU Geng, YIN Fei, CHEN You-Bin, LIU Cheng-Lin In document retrieval, high retrieval precision and speed can hardly be achieved simultaneously. A fast keyword spotting method for handwritten Chinese documents is proposed. By this method, keyword spotting is accelerated with accuracy preserved. Firstly, compressed index files are generated from the candidate segmentation recognition lattice of text lines recognition, then keywords are retrieved from the index files. Experimental results on the handwritten Chinese documents database CASIA-HWDB demonstrate the effectiveness of the proposed method. Moreover, it reduces the size of index and the retrieval time. 2015 Vol. 28 (11): 1033-1040 [Abstract] ( 400 ) [HTML 1KB] [ PDF 653KB] ( 566 ) 1041 The Application of Improved Random Forest in the Telecom Customer Churn Prediction DING Jun-Mei, LIU Gui-Quan, LI Hui An improved random forest algorithm (IRFA) is proposed to handle imbalanced classification and improve the prediction accuracy of high-value customers in telecom customer churn prediction. The node partition method for generating each tree is improved. Nodes are divided based on the life value of customers. Thus the problem of imbalanced data distribution is solved, and the accuracy of churn prediction of high-value customers is raised. IRFA is applied to customer churn prediction for a telecom company. Experimental results show that compared with other methods, the proposed algorithm has a better performance in classification and it improves the accuracy of churn prediction of high-value customers. 2015 Vol. 28 (11): 1041-1049 [Abstract] ( 566 ) [HTML 1KB] [ PDF 664KB] ( 937 ) 1050 Nodes Deployment Strategy for Underwater Sensor Network Based on Fuzzy Data Fusion ZHANG Ju-Wei, LIU Ya-Chuang, YANG Ting For nodes deployment of underwater sensor networks, based on the probability sensing model of passive sonar, a fuzzy sensing model and a fuzzy data fusion model are established with the consideration of underwater environment. The application of fuzzy data fusion algorithm for the nodes deployment of underwater sensor networks is studied, and nodes-deployment strategy based on fuzzy data fusion for underwater sensor networks(NAFC) is proposed. Thus, the energy consumption and the number of deployment nodes are reduced. Experimental results show that NAFC can improve the detection performance of the sensor network. 2015 Vol. 28 (11): 1050-1056 [Abstract] ( 367 ) [HTML 1KB] [ PDF 470KB] ( 497 )

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