Enhancing graph transformer encoding with graph heterogeneous memory for improved recommendation performance

Recommender systems are crucial for mitigating information overload, yet their effectiveness is often constrained by data sparsity. While graph neural networks (GNNs) have shown promise, existing models face a critical limitation: they typically handle either complex heterogeneous relationships or capture historical user dependencies, but struggle to address both simultaneously. To bridge this gap, we propose the Heterogeneous Graph Memory Transformer (HMT), a novel architecture that synergistically integrates a heterogeneous graph transformer with dedicated graph memory modules. HMT is designed to concurrently learn rich semantic representations from heterogeneous graphs and preserve long-range contextual information from user interaction histories, thereby creating more robust and accurate user and item embeddings. Extensive experiments on three benchmark datasets-Amazon, iFashion, and Yelp2018-demonstrate HMT's superior performance. Notably, our model achieves state-of-the-art N@5 scores of 0.3295, 0.4273, and 0.2748, respectively, significantly outperforming strong baselines. Further analyses confirm the model's robustness against data imbalance and noise. The HMT framework offers a promising direction for developing next-generation recommender systems capable of contending with the complex and dynamic nature of real-world user-item ecosystems.