Knowledge Graph-based Recommendation System: Contextualized Graph Attention Network

In recent years, with the explosive growth of online information and content, recommendation systems have become increasingly important in various scenarios such as e-commerce websites and social media platforms. These systems typically aim to provide users with a list of items she might be interested in. However, traditional methods based on user behavior data (e.g., collaborative filtering, deep learning) face issues such as data sparsity and cold start problems. To address these issues, researchers have attempted to incorporate various auxiliary information (side information) into recommendation systems, which includes users’ social networks, review texts, etc.

Research Background

Among these auxiliary information, item knowledge graphs (Knowledge Graph, KG) contain rich relationships between items and have been proven to significantly improve the performance of recommendation systems. Essentially, a knowledge graph is a heterogeneous network where nodes represent entities and edges represent relationships. However, effectively integrating these graph information into recommendation systems remains a challenge. Existing methods can be roughly divided into three categories: regularization-based methods, path-based methods, and graph neural network (GNN)-based methods. However, these methods have certain shortcomings, making it difficult to solve the following challenges:

1. Challenge 1: Most GNN-based methods do not consider users’ personalized preferences for entities when aggregating local graph context, causing the recommendation system to fail in effectively capturing users’ personalized needs for neighboring information.
2. Challenge 2: Existing methods perform poorly in capturing non-local graph context (i.e., higher-order neighbors) of entities because knowledge graphs are often incomplete, and some important connections between entities may be missing.
3. Challenge 3: Users’ historical items usually play an important role in predicting her preferences for candidate items. Most previous methods directly aggregate users’ historical item information when learning user representations, ignoring the importance of different historical items to target items.

Paper Source

This paper, titled “Contextualized Graph Attention Network for Recommendation with Item Knowledge Graph,” was authored by Yong Liu, Susen Yang, Yonghui Xu, Chunyan Miao, Min Wu, and Juyong Zhang, from Alibaba Group, the National University of Singapore, Nanyang Technological University, Shandong University, A*STAR, and other institutions. The dates are as follows: received on June 20, 2020, revised on April 28, 2021, and accepted on May 10, 2021. The paper was published in IEEE Transactions on Knowledge and Data Engineering (TKDE), with the current version dated December 7, 2022.

Detailed Research Process

Research Methods

The paper proposes a novel recommendation framework called Contextualized Graph Attention Network (CGAT), which combines local and non-local graph context information and uses users’ historical behavior data for recommendations.

1. Local Graph Context:

   • It uses a user-specific graph attention mechanism to aggregate neighbors’ information, where different users’ dependencies on the same neighbor entity may vary.
   • The representations of entities and their neighbors are combined through linear transformation, and relationship vectors are introduced to understand the importance of neighbor entities.

2. Non-Local Graph Context:

   • A biased random walk sampling strategy (Biased Random Walk Sampling, BRWS) is designed to extract non-local context. Multiple paths are generated starting from the target entity through random walks, extracting important higher-order neighbors.
   • Gated Recurrent Unit (GRU) is used to model the dependency on non-local context, thereby retaining frequently occurring contextual entities.

3. User Interaction Graph Context:

   • An item-specific attention mechanism is proposed to weight users’ historical items relative to candidate items, modeling user preferences.

Experiment and Results

Extensive experiments were conducted on real datasets, comparing CGAT with existing state-of-the-art KG-based recommendation methods, validating the effectiveness of CGAT.

• CGAT significantly outperforms benchmark methods such as GraphSAGE, FMG, MCRec, CFKG, RippleNet, MKR, KGNN-LS, and KGAT in terms of precision and recall on FM, ML, and DF datasets.
• Evaluating different components of the framework revealed that both local and non-local context information significantly improved recommendation accuracy. Especially, the biased random walk method for capturing non-local context and the user-specific graph attention mechanism for processing local context proved effective.
• Furthermore, exploring data sparsity demonstrated CGAT’s excellent performance in sparse interaction data scenarios, mainly because the framework can effectively integrate contextual information from KGs.

Conclusion and Significance

The CGAT model offers new insights into recommendation systems by explicitly capturing local and non-local context information in knowledge graphs and integrating users’ historical behavior data, significantly enhancing the accuracy and personalization capabilities of recommendation systems. Future research can further extend CGAT’s applications in more KG recommendation scenarios and explore more types of aggregation strategies to improve recommendation effects.

This study is significant both scientifically and practically, providing effective technical means to help e-commerce platforms and social media precisely recommend items of interest to users, enhancing user experience and business profits.