Knowledge Retrieval-Augmented Interest Learning for Recommendation

Existing recommendation methods rely on user-item interactions for interest modeling but struggling to the data sparsity issue. Knowledge graphs (KGs) provide rich semantic signals for mitigating this issue, yet existing KG-based methods often introduce redundant/irrelevant signals causing semantic noise that degrades recommendation performance. A possible solution is to filter task-relevant knowledge from KG to aid learning in the task domain. In this work, we propose a knowledge retrieval-augmented interest learning for recommendation (KRAR) to transfer semantic knowledge from KG to enhance interest learning on behavioral interaction graph. KRAR elaborates a dual retrieval from both users and items to filter interest-relevant items for augmenting the neighborhood in interest learning. The retrieved items participate in encoding users’ interests implementing a cross-graph knowledge transfer, which enhances interest learning for recommendation. Experiments show that KRAR outperforms the baselines, demonstrating the superiority of knowledge retrieval and embedding augmentation in handling data sparsity for promoting recommendation performance.