KSGAN: Knowledge-aware subgraph attention network for scholarly community recommendation

Abstract

On online scholarly platforms, recommending suitable communities to researchers matters for researchers’ communication and collaboration. Previous studies on community recommendation either treat a community as a single item or simply aggregate its member features while ignoring rich user interactions and side information in scholarly communities. Existing knowledge-aware recommenders fail to capture the complicated knowledge graph structures and profile the rich information in scholarly communities, and thus not suitable for the scenario of scholarly community recommendation. In this paper, we propose a knowledge-aware subgraph attention network (KSGAN) for scholarly community recommendation. Specifically, by using a scholarly KG to profile rich information of scholarly communities, we design a biquaternion-based embedding method to capture its multiple relational patterns and hierarchical structures. Then, by profiling a scholarly community as a subgraph, we design a scalable subgraph representation learning module to learn enhanced community representation. Last, we design an attention-based historical community fusion module that captures both global dependencies and target dependencies for recommendation. Extensive experiments on two real-world scholarly datasets show that KSGAN significantly outperforms state-of-the-art baselines for scholarly community recommendation. The proposed KSGAN can find potential practical implementations on scholarly platforms to recommend scholarly communities.