Knowledge-aware Multi-view Cross Learning for Edge-based Collaborative Recommendation

• Published in: 2023 IEEE 29th International Conference on Parallel and Distributed Systems (ICPADS) pp. 2483 - 2490
• Main Authors: Dai, Yang, Meng, Shunmei, Gu, Huanhuan, Liu, Nan, Tu, Longchuan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.12.2023
• Subjects: Collaboration; Computational modeling; Data preprocessing; Edge Computing; Graph Attention Network; Knowledge engineering; Knowledge Graph; Knowledge graphs; Multi-view Cross Learning; Predictive models; Recommender System; Spread spectrum communication
• ISSN: 2690-5965
• DOI: 10.1109/ICPADS60453.2023.00331

In order to effectively overcome the sparsity and cold start challenges of collaborative filtering in actual scenarios, and provide more accurate suggestions, knowledge graphs are commonly utilized in recommender systems. Graph neural networks are increasingly used in knowledge graph-based recommender systems. Nevertheless, conventional recommender systems overlook the relationship information between items and entities as well as the sparse supervision signal, potentially leading to a decline in the actual recommendation performance. In this investigation, we propose KMCL, a Knowledge-aware Multi-view Cross-learning framework for edge-based collaborative recommendation. In contrast to the typical manner of dealing with knowledge graphs, our approach investigates three different graph views to improve users' and items' embedding representations, respectively. Firstly, the merged user-item KG graph is utilized to integrate the multi-hop neighbor information of each user, with the Graph Attention Network assigning various weights to different neighbors. Secondly, Graph Neural Network collects the top-x neighbor nodes for learning in the collaborative view of user-item graph and item-entity graph. Finally, a cross module is developed to share high-level information between embeddings with the same features. In addition, we also introduce the idea of edge computing, which preprocesses and calculates the locally obtained user preferences first, and then inputs them into different views to improve the recommendation efficiency. Experimental results demonstrate that KMCL delivers better suggestions than state-of-the-art baseline models in delivering superior recommendations. Furthermore, our approach maintains higher accuracy in predicting outcomes in less familiar interaction scenarios.