HinMAD3R: Representation learning on heterogeneous information networks via multiple attentions with dual dropout and dual residual

Abstract

Heterogeneous information networks (HINs) contain rich semantic information, and effectively utilizing this information can enhance the quality of representation learning. Existing models based on the message-passing paradigm typically focus on either node-type or edge-type information, neglecting the synergistic effect of various heterogeneous information. Moreover, these models are prone to over-smoothing as network depth increases, which degrades both performance and generalization ability. To address these issues, we propose a novel multiple attention mechanism that simultaneously considers node-features, node-types, and edge-types, aiming to maximize the utilization of diverse semantic information. Additionally, we introduce dual dropout and dual residual strategies to mitigate the over-smoothing problem and enhance the model’s generalization capability. Extensive experiments conducted on seven datasets demonstrate that the proposed model outperforms state-of-the-art baselines.