Reproduce “Learning Intents behind Interactions with Knowledge Graph for Recommendation” on OpenResearch

About this paper

Xiang Wang, Tinglin Huang, Dingxian Wang, Yancheng Yuan, Zhenguang Liu, Xiangnan He, Tat-Seng Chua

KGIN, a Graph Neural Network (GNN)-based model, refines knowledge-aware recommender systems by learning fine-grained user intents from Knowledge Graphs (KGs) and employing a relational path-aware aggregation scheme. The model demonstrated consistent performance improvements, achieving up to 14.51% higher ndcg@20 compared to strong baselines, while also providing interpretable explanations for recommendations.

Key insights

• User interactions are driven by multiple, distinct underlying intents that can be effectively decomposed from a monolithic 'interact-with' relation by leveraging the rich information within Knowledge Graphs.
• Explicitly modeling relational paths and the dependencies between relations within a GNN's aggregation scheme is crucial for fully leveraging KG structural information and learning richer, more discriminative user and item representations.
• Ensuring independence among learned user intents leads to more distinct and non-redundant behavioral patterns, which enhances both the model's capacity to capture complex preferences and the clarity of generated explanations.

Summary by alphaXiv.