Heterogeneous graph attention network with motif clique

Heterogeneous information networks (HINs) have been proven to be powerful in modeling various real-world networks, such as academic networks and social media networks. By distinguishing the types of relationships and nodes, HINs obtain the capability to model multifarious data, while bringing challenges to data mining and analysis. Specifically, how to comprehensively mine the semantic information in HINs remains an open question. Although recently studied network schemas such as meta-paths or meta-graphs have achieved empirical success, they still have limitations in modeling the complex underlying relations to facilitate the graph neural networks, e.g., different schemas may overlap in semantics and lead to redundant computations. To address this issue, we leverage a recently introduced higher-order network schema known as the motif clique (abbreviated as m-clique). This schema offers greater expressiveness, enabling us to effectively construct the semantic neighborhood of nodes. We further propose a novel heterogeneous graph attention network with m cliques, named HAMC, which employs a two-level attention mechanism (node-level and semantic-level) to learn node representations. The two-level attention measures the importance of neighbors and m-cliques for each node, respectively. Extensive experiments demonstrate that the proposed HAMC outperforms the state-of-theart methods in many heterogeneous network analytic tasks such as node classification and clustering. Our code is available at https://github.com/wcx21/HAMC-Heterogeneous-Graph-Attention-Network-with-Motif-Clique.