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....

2021 Vol.34 Issue.2, Published 2021-02-25 Papers and Reports    Researches and Applications    Network Science and Information Recommendation        Network Science and Information Recommendation 95 Link Prediction Model Based on Adversarial Graph Convolutional Network TANG Chen, ZHAO Jieyu, YE Xulun, YU Shushi Most link prediction models rely too much on the known link information while mining node similarity. However, the number of the known observed links is small in the real world. To improve the robustness of the model, it is crucial to decouple the dependence of the model on the link information and mine the underlying features of nodes. In this paper, a link prediction model based on adversarial graph convolutional network is proposed with the consideration of the potential relationship between node features and links. Firstly, the similarity metric between nodes is utilized to fill in some unknown links in the adjacency matrix to alleviate the influence of link sparsity on the graph convolution model. Then, the adversarial network is employed to deeply mine the underlying connections between node features and links to reduce the dependence of the model on links. Experiments on real datasets show that the proposed model achieves better performance on link prediction problem and the performance remains relatively stable under link sparsity. Moreover, the proposed model is applicable to large-scale datasets. 2021 Vol. 34 (2): 95-105 [Abstract] ( 977 ) [HTML 1KB] [ PDF 807KB] ( 832 ) 106 Point of Interest Recommendation Algorithm Integrating Geo-Category Information and Implicit Social Relationship DONG Chanjuan, LI Sheng, HE Xiongxiong, MA Yue Aiming at the problems in the existing point of interest recommendation algorithms, such as check-in data sparsity, difficulties in obtaining social relation and lack of consideration of user individuality, a point of interest recommendation algorithm integrating geo-category information and implicit social relationship is proposed. Firstly, user check-in category information is considered, and user check-in location matrix and category matrix are decomposed simultaneously to reduce the impact of data sparsity. On the basis of explicit social relations, the method of information entropy is employed to measure user implicit social relations to alleviate the sparse problem of social networks, and then the user implicit social relations are added to the matrix factorization model by regularization method. Finally, the adaptive kernel density estimation method is adopted to personalize the impact of geographic information on user check-in behavior to improve the accuracy of recommendation. Experiments on Foursquare and Yelp datasets verify the effectiveness of the proposed algorithm. 2021 Vol. 34 (2): 106-116 [Abstract] ( 505 ) [HTML 1KB] [ PDF 831KB] ( 547 ) 117 Dynamic Network Link Prediction Based on Node Representation and Subgraph Structure HAO Xiaorong, WANG Li, LIAN Tao The key to dynamic link prediction is modeling network dynamics and extracting local structural features. Therefore, a method for dynamic network link prediction based on node representation and subgraph structure is proposed. To model node evolution dynamics, the node2vec model is introduced, and the node representations in historical snapshots are concatenated in temporal order. To model the local subgraph structure information, a graph isomorphism algorithm is employed to encode the topology structure of the local subgraph. In each historical snapshot, the node vectors of the target node pair and the topology structure of the local subgraph are fused by the ultimate feature representation of the target link. Extensive experiments demonstrate that the proposed method achieves better performance. 2021 Vol. 34 (2): 117-126 [Abstract] ( 547 ) [HTML 1KB] [ PDF 781KB] ( 442 ) 127 Friend Recommendation Feedback Algorithm Combining Cognition and Interest YIN Yunfei, SUN Jingqin, HUANG Faliang, BAI Xiangyu In the existing friend recommendation algorithm, important information is lost in the portrayal of the friend relationship. Inspired by the user's cognitive behavior of the item, a friend recommendation feedback algorithm based on cognition and interest is proposed in this paper. Hybrid similarity is utilized to conduct online friend relationship research and explore friendship issues in online social networks. Aiming at the open loop problem of the friend recommendation process, a positive and negative feedback optimization adjustment strategy based on historical recommendation information is proposed. The user similarity correction formula is employed for friend feedback dynamic recommendation, and it is proved that friend recommendation is a complex process of gradual correction. The psychological and cognitive problems portrayed by friend relationships in online social networks and the dynamic changes of recommendations are presented. The experiments show that the proposed algorithm improves the recommendation quality and realizes the dynamic adjustment of the user similarity matrix and it is superior in accuracy, recall, robustness and scalability. 2021 Vol. 34 (2): 127-136 [Abstract] ( 543 ) [HTML 1KB] [ PDF 767KB] ( 461 ) Papers and Reports 137 Optimal Individual Convergence Rate of Heavy-Ball Based Momentum Methods Based on Adaptive Step-Size Strategy HUANG Jianzhi, LONG Sheng, TAO Qing Optimization techniques, including adaptive step-size and momentum, are both utilized in AMSGrad. Compared with the adaptive step-size algorithms, there is one more logarithm factor in AMSGrad for convergence analysis. To solve the problem, the non-smooth convex problems are studied in this paper. By selecting the time-varying step-size and momentum parameters correctly, it is proved that the Heavy-ball-based momentum methods with adaptive step-size obtain the optimal individual convergence rate. It is indicated that the Heavy-ball-based momentum methods with adaptive step-size hold the advantages in both adaptation and acceleration. Hinge loss problem under L1-norm constraint is solved to verify the correctness of theoretical analysis. 2021 Vol. 34 (2): 137-145 [Abstract] ( 429 ) [HTML 1KB] [ PDF 1276KB] ( 287 ) 146 Label Distribution Learning Method Based on Low-Rank Representation LIU Ruixin, LIU Xinyuan, LI Chen Label correlations, noises and corruptions are ignored in label distribution learning algorithms. Aiming at this problem, a label distribution learning method based on low-rank representation(LDL-LRR)is proposed. The base of the feature space is leveraged to represent the sample information, and consequently dimensionality reduction of the data in the original feature space is achieved. To capture the global structure of the data, low-rank representation is transferred to the label space to impose low-rank constraint to the model. Augmented Lagrange method and quasi-Newton method are employed to solve the LRR and objective function, respectively. Finally, the label distribution is predicted by the maximum entropy model. Experiments on 10 datasets show that LDL-LRR produces good performance and stable effect. 2021 Vol. 34 (2): 146-156 [Abstract] ( 501 ) [HTML 1KB] [ PDF 2491KB] ( 346 ) Researches and Applications 157 Aspect-Level Sentiment Classification Model Based on Context-Preserving Capability HE Li, FANG Wanlin, ZHANG Hongyan Hidden emotional characteristics of the statement in various aspects can be discovered by aspect-level sentiment classification. Based on the framework of aspect-specific graph convolutional network, an aspect-level sentiment classification model based on context-preserving capability is proposed. A context gating unit is introduced into the graph convolution layer to reintegrate the useful information in the output of the previous layer. A multi-grained attention computing module is added to the proposed model to describe the interrelation in emotional expression between aspect words and their context. Experimental results on five public datasets show the advantages of the proposed model in classification accuracy and macro-F1. 2021 Vol. 34 (2): 157-166 [Abstract] ( 577 ) [HTML 1KB] [ PDF 721KB] ( 455 ) 167 Bus Passenger Flow Forecast Based on Attention and Time-Sharing Graph Convolutional Network ZHANG Wei, ZHU Fenghua, CHEN Yuanyuan, LÜ Yisheng 1cvarying system. Therefore, the spatiotemporal correlation between different bus lines can hardly be built effectively. To solve this problem, an attention and time-sharing graph convolution based long short-term memory network for bus passenger flow forecast is proposed. Firstly, temporal features of historical data are extracted by long short-term memory network(LSTM), and then they are weighted by a channel-wise attention module. A time-sharing graph convolution approach is utilized to analyze the spatial dependencies among bus lines. Different adjacent matrices are selected according to time intervals, and non-Euclidean pair-wise correlations are modeled via graph convolution. Finally, the final prediction result is obtained by integrating the extracted spatiotemporal features and vector representations of external factors, like weather and holiday information. Experiments on real bus passenger flow datasets indicate that the proposed model improves the prediction accuracy and learning speed evidently. 2021 Vol. 34 (2): 167-175 [Abstract] ( 715 ) [HTML 1KB] [ PDF 956KB] ( 550 ) 176 Query Expansion Combining Copulas Theory and Association Rules Mining HUANG Mingxuan, Hu Xiaochun The Copulas theory is introduced into the association pattern mining of text feature terms, and a query expansion algorithm combining Copulas theory and association rules mining is proposed. Firstly, top n documents of the document set returned by the query are extracted to construct the pseudo-relevance feedback document set (PRFDS) or user relevance feedback document set(URFDS). Then, the support and the confidence based on Copulas theory are applied to mine the feature term frequent itemsets and association rule patterns with the original query terms in PRFDS or URFDS, and the expansion terms are obtained from the patterns to realize query expansion. The experimental results on NTCIR-5 CLIR Chinese and English corpus show that the proposed expansion algorithm effectively restrains the problems of query topic drift and word mismatch, and enhances the performance of information retrieval with the quality of expansion terms improved and the invalid expansion terms reduced. 2021 Vol. 34 (2): 176-188 [Abstract] ( 404 ) [HTML 1KB] [ PDF 779KB] ( 723 )

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