Live streaming channel recommendation based on viewers' interaction behavior: A hypergraph approach

• Published in: Decision Support Systems 2024-09, Vol.184, p.114272, Article 114272
• Main Authors: Yu, Li, Gong, Wei, Zhang, Dongsong
• Format: Article
• Language: English
• Subjects: Graph neural networks; Hypergraph; Interaction behavior; Live streaming; Recommender systems
• ISSN: 0167-9236
• DOI: 10.1016/j.dss.2024.114272

Live streaming has become increasingly popular in recent years. Viewers of live streaming channels can interact with live streamers through various behaviors, such as sending virtual gifts and Danmaku. It is very critical to accurately model such viewers' behaviors, which reflect their interest, for recommending live streaming channels. However, existing studies on live streaming channel recommendation usually model viewers' interaction behaviors through traditional graphs where an edge only connects two nodes, which cannot capture interaction relationships between multi-viewers and multi-channels. In this study, we propose a novel approach to live streaming recommendation based on Viewers' Interaction Behavior Modeled by Hypergraphs (VIBM-Hyper). Specifically, VIBM-Hyper first constructs two hypergraphs to model viewers' interaction behaviors, including a channel-oriented behavior hypergraph and a viewer-oriented behavior hypergraph. Then, it employs a hypergraph convolution technique to learn the representations of viewers and live streaming channels, respectively, which are finally used to predict a viewer's preference for a certain live streaming channel. We analyzed viewers' multiple types of behaviors in live streaming channels and conducted empirical evaluation to investigate the effectiveness of VIBM-Hyper with two real-world datasets. The evaluation results demonstrate its superior performance in live streaming channel recommendation in comparison to the state-of-the-art methods. •We proposed a novel approach to live streaming recommendation based on Viewers' Interaction Behavior Modeled by Hypergraphs.•We constructed two hypergraphs to model viewers' interaction behaviors.•We employed a hypergraph convolution technique to learn the representations of viewers and live streaming channels.•We conducted empirical evaluation to investigate the effectiveness of VIBM-Hyper with two real-world datasets.