Adaptive Long-term Embedding with Denoising and Augmentation for Recommendation
New GNN-based recommender outperforms baselines on 4 datasets, improving robustness.
The rapid growth of online content has made personalized recommendation systems essential, but graph-based sequential recommenders often suffer from noise and static representations. To address this, researchers Zahra Akhlaghi and Mostafa Haghir Chehreghani propose Adaptive Long-term Embedding with Denoising and Augmentation for Recommendation (ALDA4Rec). The method first builds an item-item graph and applies community detection to filter out noisy interactions. Graph Convolutional Networks (GCNs) then learn short-term user-item embeddings, while averaging, GRUs, and attention mechanisms capture long-term dynamics. A key innovation is an MLP-based adaptive weighting strategy that dynamically optimizes long-term user preferences, allowing the model to flexibly balance recent and historical behavior.
ALDA4Rec was evaluated on four real-world datasets spanning e-commerce and streaming domains. The results show consistent improvements over state-of-the-art baselines in both accuracy (e.g., Recall, NDCG) and robustness to sparse or noisy data. The open-source release enables practitioners to integrate these techniques directly into production recommendation pipelines. By reducing noise and better modeling long-term preferences, ALDA4Rec offers a practical upgrade for platforms seeking more relevant and resilient recommendations.
- Constructs an item-item graph and filters noise via community detection
- Uses GCNs for short-term embeddings, GRU/attention for long-term, with MLP adaptive weighting
- Outperforms state-of-the-art baselines on 4 real-world datasets in accuracy and robustness
Why It Matters
Delivers more accurate and robust recommendations for e-commerce and content platforms by adaptively modeling long-term user preferences.