Knowledge Graph Convolutional Networks for Recommender Systems (1904.12575v1)

Abstract: To alleviate sparsity and cold start problem of collaborative filtering based recommender systems, researchers and engineers usually collect attributes of users and items, and design delicate algorithms to exploit these additional information. In general, the attributes are not isolated but connected with each other, which forms a knowledge graph (KG). In this paper, we propose Knowledge Graph Convolutional Networks (KGCN), an end-to-end framework that captures inter-item relatedness effectively by mining their associated attributes on the KG. To automatically discover both high-order structure information and semantic information of the KG, we sample from the neighbors for each entity in the KG as their receptive field, then combine neighborhood information with bias when calculating the representation of a given entity. The receptive field can be extended to multiple hops away to model high-order proximity information and capture users' potential long-distance interests. Moreover, we implement the proposed KGCN in a minibatch fashion, which enables our model to operate on large datasets and KGs. We apply the proposed model to three datasets about movie, book, and music recommendation, and experiment results demonstrate that our approach outperforms strong recommender baselines.

Knowledge Graph Convolutional Networks for Recommender Systems

The paper "Knowledge Graph Convolutional Networks for Recommender Systems," presented at WWW 2019, introduces an innovative approach to addressing the sparsity and cold start problems inherent in collaborative filtering (CF)-based recommender systems. The authors propose Knowledge Graph Convolutional Networks (KGCN), an end-to-end framework designed to capture and utilize the inter-item relatedness by leveraging the attributes present in knowledge graphs (KGs).

Problem Statement

Traditional collaborative filtering methods rely heavily on user-item interaction data, which often suffers from sparsity and cold start issues, where new users or items have little to no interaction history. To mitigate these problems, researchers have incorporated additional attributes into the recommendation algorithms. These attributes, however, are not isolated; they are interconnected, forming a knowledge graph. This paper aims to harness the power of knowledge graphs to better address the limitations of traditional recommender systems by capturing both high-order structure and semantic information inherent in KGs.

Proposed Method

KGCN extends the principles of graph convolutional networks (GCNs) to the domain of knowledge graphs, offering a novel way to utilize these rich sources of structured information. The key components of KGCN include:

1. Receptive Fields and Neighborhood Sampling: For each entity (item or item attribute) in the KG, a receptive field is defined that extends to its neighbors up to a specified number of hops. This allows KGCN to capture high-order proximity information. A fixed-size neighborhood is sampled for computational efficiency.
2. Attention Mechanism: KGCN employs an attention mechanism to assign weights to the neighbors based on their relevance to the user and the type of relation connecting them. This personalized aggregation process ensures that the entity representations are tailored to individual users' interests.
3. Multiple Aggregation Methods: The paper investigates several aggregation strategies such as sum, concatenation, and neighbor-only aggregation, determining the best method empirically.
4. Minibatch Implementation: To ensure scalability, KGCN is implemented in a minibatch fashion, making it suitable for large datasets and extensive KGs.

Experimental Evaluation

The efficacy of KGCN was validated using three real-world datasets across different recommendation scenarios: MovieLens-20M (movies), Book-Crossing (books), and Last.FM (music). The performance was evaluated based on metrics like AUC, F1-score for click-through rate (CTR) prediction, and Recall@K for top-K recommendation.

Empirical Results:

• KGCN showed substantial improvements over state-of-the-art baselines, including SVD, LibFM, and RippleNet, especially in sparse datasets like Book-Crossing and Last.FM.
• The influence of various hyperparameters such as neighbor sampling size $K$, depth of the receptive field $H$, and dimension of embeddings $d$ was thoroughly analyzed.
• Personalized attention mechanisms significantly boosted performance, underscoring the importance of accounting for user-specific interests in relations within the KG.

Implications and Future Work

The practical implications of this research are considerable for the development of more effective recommender systems capable of dealing with sparsity and cold start problems. By leveraging the rich interconnected information within KGs, KGCN provides a robust solution that enhances recommendation precision and diversity.

Future research directions include:

1. Advanced Sampling Techniques: Exploring non-uniform or importance sampling methods to further enhance the efficiency and effectiveness of the receptive field construction.
2. User-end Knowledge Graphs: Investigating the potential benefits of incorporating user-related KGs in addition to item-related KGs.
3. Unified KG Incorporation: Designing algorithms that integrate multiple KGs from both the user and item ends to fully exploit the rich relational data available.

Conclusion

KGCN stands out as a significant advancement in the field of recommender systems, providing a robust framework that effectively harnesses the structure and semantics of knowledge graphs. Its applicability to various domains such as movies, books, and music highlights its versatility and potential for widespread adoption in personalized content recommendation. The open-source release of the KGCN code further paves the way for future research and development in this area.