ANGCN: Adaptive Neighborhood-Awareness for Recommendation

In the field of recommender systems, GNN-based collaborative filtering models, through the configuration of multiple convolutional layers, enable the utilization of subgraph structures from high-hop neighbors to enhance the learning of embeddings for users and items. However, the increase of convolutional layers not only leads to long-tail effects, over-smoothing during training, and the introduction of noise in recommended projects, but also limits the propagation of high-order information solely among interacted users and items due to the dependence on the adjacency matrix in GNN convolution. In response to these challenges, we propose a novel recommendation model, ANGCN, which incorporates a collaborative perception network on top of the GNN-based collaborative filtering model. This network dynamically adjusts information propagation between nodes through loss learning, ensuring that high-order information is no longer solely reliant on the multi-layer stacking of graph convolution. Experimental evaluations on two real-world datasets demonstrate the effectiveness of ANGCN, exhibiting significant performance improvements compared to various state-of-the-art baseline algorithms.