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....

2024 Vol.37 Issue.3, Published 2024-03-25 Papers and Reports    Graph Neural Network Based Recommender System        Graph Neural Network Based Recommender System 191 Recurrent Neural Network and Attention Enhanced Gated Graph Neural Network for Session-Based Recommendation LI Weiyue, ZHU Zhiguo, DONG Hao, JIANG Pan, GAO Ming Most of existing session-based recommender systems with graph neural networks are capable of capturing the adjacent contextual relation of products effectively in the session graph. However, few of them focus on the sequential relation. Both relations are important for precise recommendations in e-commerce scenarios. To solve the problem, a recurrent neural network and attention enhanced gated graph neural network for session-based recommender system is proposed based on bidirectional long short-term memory. The model is designed to complement the advantages of different network structures and learn the user's interest preferences expressed during the current session more fully. Specifically, a parallel framework is adopted in the proposed model to learn the neighborhood contextual features and temporal relation among products respectively within user session clickstreams in e-commerce scenarios. Attention mechanisms are applied to denoise the features. Finally, the adaptive fusion method of both features is employed based on gating mechanism. Experiments on three real-world datasets show the superiority of the proposed model. The model code in the paper is available at https://github.com/usernameAI/RAGGNN. 2024 Vol. 37 (3): 191-206 [Abstract] ( 416 ) [HTML 1KB] [ PDF 1266KB] ( 861 ) 207 Social Recommendation Model Based on Self-Supervised Tri-Training and Consistent Neighbor Aggregation LIU Shudong, LI Liying, CHEN Xu Integrating user social relationships into user-item rating data to construct a heterogeneous user-item graph can alleviate data sparsity and cold start in traditional recommender systems. However, due to the complexity of user social relationships, aggregating inconsistent neighbors may degrade the recommendation performance. To address this issue, a social recommendation model based on self-supervised tri-training and consistent neighbor aggregation(SR-STCNA) is proposed. Firstly, on the basis of user-item rating data, social relationships among users are introduced and diverse relations within the heterogeneous user-item graph are established. The relationships between users as well as between users and items are presented by a hypergraph. Self-supervised tri-training is employed to learn users' representations from unlabeled data and uncover the complex connectivity between user-user and user-item interactions. Then, the consistent neighbors of users and items are aggregated in the process of their representation learning by the node consistency score and relationship self-attention on the user-item heterogeneous graph. Consequently, the representation ability of users and items is enhanced, thereby improving the recommendation performance. Finally, the experimental results on CiaoDVD, FilmTrust, Last.fm and Yelp datasets validate the superiority of SR-STCNA. 2024 Vol. 37 (3): 207-220 [Abstract] ( 218 ) [HTML 1KB] [ PDF 742KB] ( 508 ) 221 Graph Neural Network Recommendation Based on Enhanced Social Influence DAI Xingyue, YE Hailiang, CAO Feilong With the rapid development of online social platforms, social recommendation becomes a critical task in recommender systems. However, the performance of recommendation systems is limited to some extent due to the sparsity of social relationships between users. Therefore, a graph neural network recommendation method based on enhanced social influence is proposed in the paper, aiming to utilize implicit social relationships between users to enhance social recommendation. The implicit social relationships are revealed, and the social graph among users is reconstructed by analyzing interaction information between users and items. On this basis, global features of the social graph are integrated with local features of users effectively via the mutual information maximization method. A learnable mechanism is integrated into the graph attention network to fully capture the interaction information between users and items. An improved Bayesian personalized ranking loss is designed to provide more accurate user and item feature representations for the rating prediction task. Extensive experiments on three public social recommendation datasets demonstrate the superiority of the proposed method. 2024 Vol. 37 (3): 221-230 [Abstract] ( 252 ) [HTML 1KB] [ PDF 705KB] ( 712 ) 231 Multimodal Recommendation Method Integrating Latent Structures and Semantic Information ZHANG Xiaoming, LIANG Zhengguang, YAO Changyu, LI Zhaoxing Multimodal recommender systems aim to improve recommendation performance via multimodal information such as text and visual information. However, existing systems usually integrate multimodal semantic information into item representations or utilize multimodal features to search the latent structure without fully exploiting the correlation between them. Therefore, a multimodal recommendation method integrating latent structures and semantic information is proposed. Based on user's historical behavior and multimodal features, user-user and item-item graphs are constructed to search the latent structure, and user-item bipartite graphs are built to learn the user's historical behavior. The graph convolutional neural network is utilized to learn the topological structure of different graphs. To better integrate latent structures and semantic information, contrastive learning is employed to align the learned latent structure representations of item with their multimodal original features. Finally, evaluation experiments on three datasets demonstrate the effectiveness of the proposed method. 2024 Vol. 37 (3): 231-241 [Abstract] ( 306 ) [HTML 1KB] [ PDF 1236KB] ( 465 ) 242 Session-Based Recommendation Model with Self Contrastive Graph Neural Network and Dual Predictor ZHANG Yusong, XIA Hongbin, LIU Yuan Session-based recommendation aims to predict user behavior based on short-term anonymous sessions. In most of the existing session-based recommendation models using graph neural network and contrastive learning, joint optimization of cross-entropy loss and contrastive learning loss is typically adopted. However, these two methods play similar roles and require the construction of a large number of complex positive and negative samples, bringing a burden to the model. Moreover, simple linear predictor struggles to predict the data with random behaviors of users. To solve the problems, a session-based recommendation model with self contrastive graph neural network and dual predictor is proposed(SCGNN). Firstly, the original session is built into two views, an improved graph neural network is employed to learn item and session embeddings, and item representation is optimized by self-contrastive learning. Then, a user behavior-aware factor is introduced to mitigate the impact of user random behaviors. Finally, the decision forest predictor and linear predictor are both utilized to predict the items, and soft label generation strategy is proposed for assist prediction by collaboratively filtering the historical sessions similar to the current session. Experiments on three benchmark datasets, Tmall, Diginetica and Nowplaying, validate the effectiveness of SCGNN. 2024 Vol. 37 (3): 242-252 [Abstract] ( 227 ) [HTML 1KB] [ PDF 753KB] ( 700 ) Papers and Reports 253 Self-Supervised Non-Isometric 3D Shape Collection Correspondence Calculation Method WU Yan, YANG Jun, ZHANG Siyang Aiming at the problem of low accuracy and poor generalization ability in existing non-isometric 3D shape collection correspondence calculation methods, a self-supervised non-isometric 3D shape collection correspondence calculation method using deep intrinsic-extrinsic feature alignment algorithm is proposed. Firstly, discriminative feature descriptors are obtained by directly learning the original 3D shape features through DiffusionNet. Then, the deep intrinsic-extrinsic feature alignment algorithm is employed to compute correspondences between non-isometric shapes. Consistency between internal and external information is realized by utilizing local manifold harmonic bases as intrinsic information of the shapes and integrating external information such as Cartesian coordinates. Consequently, correspondence results are generated automatically in an unsupervised manner. Finally, a weighted undirected graph of non-isometric shape collections is constructed. Based on the principle of inherent correlation among similar geometric shapes, a self-supervised multi-shape matching algorithm is designed to continuously enhance the cycle-consistency of the shortest path in the shape graph, and thus optimal correspondences for non-isometric 3D shape collections are obtained. Experimental results demonstrate that the proposed method achieves small geodesic errors in correspondences with accurate results, and effectively deals with the symmetric ambiguity problem with good generalization ability. 2024 Vol. 37 (3): 253-266 [Abstract] ( 195 ) [HTML 1KB] [ PDF 3461KB] ( 427 ) 267 Byzantine Robot Recognition Scheme via Blockchain-Based Reputation Management HUANG Jie, ZENG Jiazhou A reputation management system with identity authentication and task supervisor (RMS-IATS) for swarm robotics via blockchain technology is proposed to identify Byzantine robots within the swarm robotics and avoid the security threat caused by Byzantine robots to swarm robotics. Firstly, a classical blockchain-based swarm robotics reputation management system(RMS) is improved by introducing penalty factors, and a severer reputation value penalty is imposed on the robotics with long-term Byzantine behavior. Secondly, to speed up the identification of Byzantine robots within swarm robotics, an identity authentication protocol is devised, and thus lower initial reputation scores are assigned to the robots with unauthorized identities. Next, a dual-layer communication network for communication between robots is designed to solve the communication latency issue caused by blockchain in the swarm robotics system. Finally, the feasibility of the proposed blockchain-based RMS-IATS and dual-layer communication network is proved through simulations. The identification time for different types of Byzantine robots is shortened by RMS-IATS compared with the classical RMS for swarm robotics, and the maximum communication latency of the system is reduced by the proposed dual-layer communication network compared with the blockchain. 2024 Vol. 37 (3): 267-281 [Abstract] ( 223 ) [HTML 1KB] [ PDF 1323KB] ( 645 )

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