Adaptive Memory-Augmented Graph Attention Network for Social Recommendation

To address the key challenges in social recommendation systems inadequate modeling of heterogeneous relationships, difficulties in coordinating long-term and short-term user preferences, and fixed multi-relationship weights, this paper proposes an Adaptive Memory-enhanced Network (Hierarchical Attention Memory Network, HAMNet). The model achieves breakthroughs through four key aspects: First, it dynamically perceives association weights between users and neighboring nodes based on graph attention mechanisms, accurately capturing local interaction patterns; Second, it designs a memory enhancement module that preserves users' long-term behavioral characteristics through gating mechanisms and dynamically integrates them with real-time interests to alleviate data sparsity interference; Third, it introduces an adaptive weighting strategy that automatically adjusts the fusion ratios of multi-source heterogeneous information (including social relationships and user-item interactions) according to scenario requirements; Finally, a cross-layer memory update strategy is adopted to propagate historical features and mitigate the over-smoothing phenomenon in deep networks. Experiments on Ciao, Epinions and Yelp datasets show that HAMNet achieves maximum improvements of 7.55% in recommendation accuracy (HR@10) and 8.51% in ranking quality (NDCG@10), with particularly significant performance advantages in sparse interaction scenarios, verifying its effectiveness in dynamic relationship modeling, long-short-term preference coordination, and multi-source information adaptation in complex social recommendation scenarios.