Dynamic Graph with Similarity-Aware Attention Graph Neural Network for Recommender Systems
New GNN framework beats LightGCN by adapting user graphs during training
Modern recommender systems struggle with static user-item graphs and limited similarity measures, failing to capture evolving user preferences. The new Dynamic Graph with Similarity-Aware Attention Graph Neural Network (DG-SA-GNN) from Senapati et al. tackles this head-on. The framework constructs four parallel user similarity graphs using Cosine, Jaccard, Discounted Pearson Correlation Coefficient (Discount PCC), and IPIJ similarity functions. Each graph is processed by a dedicated UserGNN module, then a Graph Transformer fuses the four views. A Cross-Attention module refines user embeddings through interaction with item embeddings. Crucially, the graphs are reconstructed at scheduled epochs during training, allowing the model to adapt to the learned embedding space dynamically.
Benchmarked on the MovieLens100K dataset, DG-SA-GNN achieves a Recall@20 of 0.162 and NDCG@20 of 0.065, surpassing the LightGCN baseline in recall. Mini-batch training with hard negative sampling enhances both scalability and convergence. This work validates that dynamic multi-similarity graph construction, combined with attention-based fusion, can capture nuanced user preferences and deliver superior recommendation performance. The paper is available on arXiv (2605.05238) and presents a promising direction for more adaptive, real-time recommender systems.
- Uses four parallel user similarity graphs (Cosine, Jaccard, Discounted PCC, IPIJ) processed by separate UserGNN modules.
- Graphs are dynamically reconstructed at scheduled training epochs to adapt to evolving user embeddings.
- Achieves Recall@20 of 0.162 and NDCG@20 of 0.065 on MovieLens100K, outperforming LightGCN baseline.
- Employs mini-batch training with hard negative sampling for scalability and faster convergence.
Why It Matters
Dynamic multi-similarity graphs enable recommenders to capture shifting user tastes, improving personalization for platforms handling millions of users.