KGIF: Optimizing Relation-Aware Recommendations with Knowledge Graph Information Fusion

TL;DR

The paper tackles sparsity and opacity in knowledge-graph augmented recommender systems by proposing KGIF, a Knowledge Graph Attention Network with Information Fusion. KGIF explicitly fuses entity and relation embeddings using dynamic projection vectors within a TransD-based CK G embedding, followed by an attentive propagation layer that captures high-order interactions. Its contributions include an explicit fusion mechanism, robustness in sparse knowledge graphs, and explainable recommendations through interpretable path visualizations, with experimental evidence showing state-of-the-art gains on Amazon-book, Last-FM, and Yelp2018. The method enhances recommendation accuracy while increasing transparency, offering practical value for real-world systems that rely on rich side information and multi-hop relational reasoning.

Abstract

While deep-learning-enabled recommender systems demonstrate strong performance benchmarks, many struggle to adapt effectively in real-world environments due to limited use of user-item relationship data and insufficient transparency in recommendation generation. Traditional collaborative filtering approaches fail to integrate multifaceted item attributes, and although Factorization Machines account for item-specific details, they overlook broader relational patterns. Collaborative knowledge graph-based models have progressed by embedding user-item interactions with item-attribute relationships, offering a holistic perspective on interconnected entities. However, these models frequently aggregate attribute and interaction data in an implicit manner, leaving valuable relational nuances underutilized. This study introduces the Knowledge Graph Attention Network with Information Fusion (KGIF), a specialized framework designed to merge entity and relation embeddings explicitly through a tailored self-attention mechanism. The KGIF framework integrates reparameterization via dynamic projection vectors, enabling embeddings to adaptively represent intricate relationships within knowledge graphs. This explicit fusion enhances the interplay between user-item interactions and item-attribute relationships, providing a nuanced balance between user-centric and item-centric representations. An attentive propagation mechanism further optimizes knowledge graph embeddings, capturing multi-layered interaction patterns. The contributions of this work include an innovative method for explicit information fusion, improved robustness for sparse knowledge graphs, and the ability to generate explainable recommendations through interpretable path visualization.

Figures (7)

• Figure 1: The SOTA recommender models work without fusion. KGIF utilizes not only attribute but also relation information to update both user and item representation for fused CKG embedding.
• Figure 2: Overall process of KGIF. (a) Learn a set of initial CKG embedding with TransD. (b) Information fusion by using projection vectors from the previous step. (c) Attentive embedding propagation with self-attention of triplets for updating CKG embedding. (d) Recommendation-making based on user-item information aggregation.
• Figure 3: Comparison between CF and CKG. CF recommends items solely based on user-item interactions, while CKG incorporates item attributes in addition to user-item interactions when making recommendations to users.
• Figure 4: First-order ego-network of a node $h$. The head entity is in the middle denoted as $\mathbf{h}_{Ego}$ and tails are denoted as $t_{0-4}$.
• Figure 5: Comparative Analysis of Loss and Recall: Fusion vs. Without Fusion (Original) Approach