PDA-GNN: propagation-depth-aware graph neural networks for recommendation

Embedding learning of users and items can reveal latent interaction information in recommender systems. Most existing recommendation approaches implicitly treat users and items as integral individuals and assume embeddings of users and items propagate following a holistic pattern. However, this may be inappropriate in real-world scenarios because individuals possess multiple attribute facets, which present different propagation depths. Therefore, in this paper, we propose a novel framework, PDA-GNN, for Propagation-Depth-Aware Graph Neural Networks, to distinguish fine-grained attributes of users and items in recommender systems and distribute different propagation depths on the graph. In PDA-GNN, we first divide individual attributes into different embedding patterns to model the fine-grained attribute propagation process, with each attribute embedding possessing a distinct propagation depth. Accordingly, we devise an attention-based attribute aggregation mechanism to highlight specific attribute aspects and integrate different attribute embeddings with different attention weights. Moreover, we design a novel attribute distance normalization approach to constrain the distances between individual attribute embeddings. Extensive experiments conducted on three real-world datasets demonstrate that our model consistently outperforms the state-of-the-art recommendation methods.