Reproduce “InBox: Recommendation with Knowledge Graph using Interest Box Embedding” on OpenResearch

About this paper

Zezhong Xu, Yincen Qu, Wen Zhang, Lei Liang, Huajun Chen

InBox introduces a recommender system that utilizes novel box embeddings for tags, relations, and user interests to model complex user preferences as compositional regions within knowledge graphs. The model consistently outperforms state-of-the-art baselines across four real-world datasets, achieving recall@20 gains up to 13.10%.

Key insights

• Geometric box embeddings provide a more suitable and expressive representation for modeling one-to-many relationships and complex concept combinations in knowledge-aware recommendations than traditional point embeddings.
• Item-Relation-Tag (IRT) triplets within knowledge graphs are particularly crucial for recommendation performance, allowing InBox to be effective even with less comprehensive KG data compared to multi-hop GNNs.
• Explicitly enforcing item points to reside within the intersection region of their associated concept boxes (concept combination) significantly enhances recommendation accuracy and interpretability.

Summary by alphaXiv.