Unifying Task-Oriented Knowledge Graph Learning and Recommendation

Qianyu Li, Xiaoli Tang, Tengyun Wang, Haizhi Yang, Hengjie SongIEEE Access 2019 · DOI

Abstract

Incorporating knowledge graphs (KGs) into recommender systems (knowledge-aware recommendation) to improve the recommendation accuracy and explainability has attracted considerable research efforts. However, existing methods largely assume that KGs are complete when transferring knowledge from them, which may lead to suboptimal performance for those KGs, can be hardly complete in real-life scenarios. In this paper, we present a robustly co-learning model (RCoLM) that takes the incompleteness nature of KGs into consideration when incorporating them into recommendation. The RCoLM aims at transferring knowledge between recommendation task and knowledge graph completion (KG completion) task by utilizing a transfer learning model. An earlier version of this paper appeared in KDD 2019. This version is an extension of the previous submission and two major innovations are presented here. At first, distinct from previous knowledge-aware recommendation methods, which mainly focus on transferring knowledge from KGs to item recommendations, the RCoLM attempts to exploit user-item interactions from recommendations for KG completion, and unifies the two tasks in a joint model for mutual enhancements. Second, the RCoLM provides a general task-oriented negative sampling strategy on KG completion task, which further improves the adaptive ability of the proposed algorithm and plays an essential role for obtaining superior performance in various sub-tasks of the KG completion. The extensive experiments on two real-world public datasets demonstrate that RCoLM outperforms not only state-of-the-art knowledge-aware recommendation methods but also existing KG completion methods.

Introduction

Modern recommendation systems increasingly rely on Knowledge Graphs (KGs) to provide context and improve explainability. By linking items to a web of entities and relations—such as a movie linked to its director, genre, or lead actors—recommenders can better understand latent user preferences. However, a significant hurdle remains: real-world KGs are notoriously incomplete. Missing links can lead to "blind spots" in recommendation logic, where a system fails to suggest a relevant item simply because a relationship was never explicitly recorded in the graph.

The Challenge of Incomplete Knowledge

Traditional knowledge-aware recommendation models often operate under the "Completeness Assumption," treating the KG as a static, finished product. This one-way flow of information—from the KG to the recommender—ignores the potential for user-item interactions to inform the KG itself.

Illustration of the overlap between user-movie interactions and a knowledge graph.

As illustrated above, user interactions (like Alice and Bob watching movies) and the structured Knowledge Graph (linking movies to entities like Charles Chaplin) are deeply intertwined. The Robustly Co-Learning Model (RCoLM) addresses this by treating recommendation and KG completion as two sides of the same coin, creating a "mutual enhancement" loop.

The RCoLM Framework

RCoLM introduces a bi-directional knowledge transfer mechanism. It doesn't just use the KG to help the recommender; it uses recommendation data to help fill in the gaps in the KG. The architecture is built on several sophisticated components designed to capture deep user interests.

The architecture of the Robustly Co-Learning Model (RCoLM).

Key Components:

  • User Preference Tree Construction: The model builds a multi-layer tree starting from a user's clicked items, traversing the KG to find related entities. This allows the model to capture deep, multi-hop interests (e.g., a user likes a movie because they like the director's other work).
  • Dual-Layer Attention: A combination of self-attention (to weigh entity relevance within a layer) and target-specific attention (to weigh the importance of different tree levels for a specific candidate item).
  • KG Completion via TransR: RCoLM employs TransR to model entities and relations in separate spaces, effectively handling complex one-to-many relationships.
  • Task-Oriented Negative Sampling: Unlike standard methods that only swap entities during training, RCoLM introduces a strategy that also swaps relations. This forces the model to learn the specific nature of connections, significantly improving KG completion accuracy.

Experimental Evaluation

The researchers evaluated RCoLM on two widely used real-world benchmarks: MovieLens-1M (movie ratings) and Book-Crossing (book recommendations).

  • Superior Accuracy: RCoLM consistently outperformed state-of-the-art baselines like RippleNet and CKE. On MovieLens-1M, it achieved a 10.28% improvement in AUC (Area Under Curve) over the best baseline.
  • Mutual Enhancement: The authors observed a strong positive correlation ($\rho \approx 0.95$) between the learning curves of the recommendation and KG completion tasks, empirically validating that the two tasks are mutually beneficial during training.
  • Robustness to Sparsity: The model showed significant gains on the sparse Book-Crossing dataset, demonstrating that the KG effectively "fills the gaps" when interaction data is thin.

Conclusion

RCoLM represents a significant shift toward more robust, bi-directional knowledge-aware recommendation. By acknowledging KG incompleteness and leveraging user behavior as a source of truth for KG completion, it achieves state-of-the-art performance in both domains. While the current model focuses on entity-centric transfer, future work will explore relation-specific knowledge transfer and optimizations for large-scale industrial applications.

Summary updated

BibTeX
@article{li2019unifying,
  author = {Qianyu Li and Xiaoli Tang and Tengyun Wang and Haizhi Yang and Hengjie Song},
  title = {{Unifying Task-Oriented Knowledge Graph Learning and Recommendation}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2932466},
}
Figure