Iterative heterogeneous graph learning for knowledge graph-based recommendation

Tieyuan Liu, Hongjie Shen, Liang Chang, Long Li, Jingjing LiScientific 2023 · DOI

Abstract

Incorporating knowledge graphs (KGs) into recommendation systems has gained significant traction due to their ability to provide rich, multi-type relational data. However, existing methods often struggle to fully exploit the heterogeneous nature of KGs, treating different relation types with insufficient distinction. This paper introduces HGKR (Iterative Heterogeneous Graph Learning on Knowledge Graphs), a novel framework that models KGs as fine-grained heterogeneous graphs. By employing Graph Neural Networks (GNNs) for iterative message passing at both graph and semantic levels, and incorporating an attention-based knowledge-perceiving filter, HGKR captures nuanced user interests. Experimental results on MovieLens and MOOCCube datasets demonstrate that HGKR consistently outperforms state-of-the-art baselines in recommendation accuracy.

Introduction

Modern recommendation systems face persistent challenges, notably the "cold start" problem and the difficulty of capturing complex, high-order relationships between users and items. Knowledge Graphs (KGs) offer a solution by providing a structured web of auxiliary information. For instance, knowing a movie's director or genre can bridge the gap when user interaction data is sparse. Yet, most current KG-based models fail to adequately handle heterogeneity—the fact that a "directed by" link carries different semantic weight and information than a "genre" link. Treating these relations uniformly leads to suboptimal performance.

The HGKR Framework

To address these limitations, the authors propose the HGKR framework. The core innovation lies in treating the KG not as a monolithic graph, but as a collection of interacting heterogeneous layers.

The HGKR framework architecture, illustrating the flow from initial embeddings through heterogeneous graph learning and knowledge perceiving to final prediction.

The framework consists of four primary modules:

  1. Bipartite Extraction: The complex KG is decomposed into multiple bipartite graphs, each corresponding to a specific relation type (e.g., User-Item, Movie-Director).
  2. Hierarchical Message Passing: This occurs at two levels. At the graph level, GNNs (specifically GraphSAGE and GAT) aggregate local neighbor information. At the semantic level, information is iteratively propagated across different bipartite graphs, allowing the model to learn cross-relational dependencies.
  3. Knowledge-Perceiving Item Filter: An attention mechanism evaluates a user's historical interactions against target items, refining the user's preference representation based on KG-derived insights.
  4. Predicting Module: Refined user and item embeddings are combined to calculate the final interaction probability.

Modeling Heterogeneity

The transition from a standard KG to a layered heterogeneous representation is central to HGKR's success. By isolating relations into distinct layers, the model can apply specialized GNN architectures—such as Graph Attention Networks (GAT) for item-related entities to reduce noise, and GraphSAGE for others.

Comparison between a standard Knowledge Graph (a) and the layered Knowledge Heterogeneous Graph (b) used in the HGKR framework.

Experimental Results

HGKR was evaluated on two real-world datasets: MovieLens Latest (enriched with Wikidata) and MOOCCube (a dense online course dataset). The model demonstrated superior performance over state-of-the-art baselines like RippleNet and KGAT.

  • CTR Prediction: HGKR achieved a 2.21% improvement in AUC on the MOOCCube dataset, suggesting it excels in semantically dense environments where multiple relational layers overlap.
  • Top-K Recommendation: The model showed significant gains in Precision@K and nDCG@K (normalized discounted cumulative gain), particularly as the number of recommendations increased.

Conclusion

HGKR represents a significant step forward in KG-based recommendation by explicitly modeling the semantic heterogeneity of relational data. While the current manual selection of GNN types for specific relations is a limitation, the iterative refinement process proves highly effective at capturing potential user interests. Future work will focus on automating model configuration and integrating KG reasoning to handle data sparsity more robustly.

Summary updated

Abstract

Incorporating knowledge graphs into recommendation systems has attracted wide attention in various fields recently. A Knowledge graph contains abundant information with multi-type relations among multi-type nodes. The heterogeneous structure reveals not only the connectivity but also the complementarity between the nodes within a KG, which helps to capture the signal of potential interest of the user. However, existing research works have limited abilities in dealing with the heterogeneous nature of knowledge graphs, resulting in suboptimal recommendation results. In this paper, we propose a new recommendation method based on iterative heterogeneous graph learning on knowledge graphs (HGKR). By treating a knowledge graph as a heterogeneous graph, HGKR achieves more fine-grained modeling of knowledge graphs for recommendation. Specifically, we incorporate the graph neural networks into the message passing and aggregating of entities within a knowledge graph both at the graph and the semantic level. Furthermore, we designed a knowledge-perceiving item filter based on an attention mechanism to capture the user's potential interest in their historical preferences for the enhancement of recommendation. Extensive experiments conducted on two datasets in the context of two recommendations reveal the excellence of our proposed method, which outperforms other benchmark models.
BibTeX
@inproceedings{liu2023iterative,
  author = {Tieyuan Liu and Hongjie Shen and Liang Chang and Long Li and Jingjing Li},
  title = {{Iterative heterogeneous graph learning for knowledge graph-based recommendation}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-33984-5},
}
Figure