Dual-view hypergraph attention network for news recommendation
Wenxuan Liu,Zizhuo Zhang,Bang Wang
DOI: https://doi.org/10.1016/j.engappai.2024.108256
IF: 8
2024-03-18
Engineering Applications of Artificial Intelligence
Abstract:News Recommendation (NR) helps users to quickly find their mostly interested information. Recently, some NR systems based on graph neural networks has achieved significant performance improvement in that they use a graph structure to pair-wisely link two nodes (e.g., user, news, topic node) for representation learning. However, such kind of pairwise edges may be not enough to describe multifaceted relations involving more than two nodes; While a hypergraph structure with one hyperedge connecting two and more nodes might be a better choice. Considering this, we propose a dual-view hypergraph attention network for news recommendations (Hyper4NR) in this paper. In particular, we design a dual-view hypergraph structure to model users' click history, containing both topic-view hyperedges and semantic-view hyperedges. On the constructed hypergraph, we use hyperedge-specific attention network (HSAN) to pass messages in between hyperedges and nodes to encode their representations based on a self-supervised learning approach. Furthermore, we construct another kind of candidate hypergraph on which we apply the HyperGAT to obtain enhanced candidate news encoding. Extensive experiments on the widely used MIND news dataset and Adressa dataset show that our Hyper4NR outperforms these state-of-the-art NR methods.
automation & control systems,computer science, artificial intelligence,engineering, electrical & electronic, multidisciplinary