Graph-Based Sampling-Free Knowledge Graph Enhanced Recommendation

In recent years, knowledge graph (KG) enhanced recommendation systems, aiming to address cold start problems and the interpretability of recommendation systems, have garnered substantial research interest. Existing recommendation systems typically focus on implicit feedback such as purchase history but lack negative feedback. Most systems use negative sampling strategies to handle implicit feedback data, which may overlook potential positive user-item interactions. Meanwhile, some other works adopt a sampling-free strategy, viewing all unobserved interactions as negative samples and assigning weights to each negative sample to represent the probability that the sample is a positive one. However, these methods employ simple and intuitive weight assignment strategies, failing to capture potential relationships in all interaction data.

Research Background and Motivation

With the rapid development of the internet, information overload has become increasingly severe. To enhance user search experiences and increase revenue for product suppliers, recommendation systems have emerged and achieved great success in various applications such as e-commerce and social networks. Recently, introducing knowledge graphs as content information into recommendation systems has addressed cold start and interpretability issues. For instance, merely based on the watch history of four users, the movie “Avatar” cannot be recommended to users 1 and 2. However, through related knowledge information (e.g., shared director James Cameron), more accurate and reasonable recommendations can be generated. Most existing methods focus on exploring new architectures that combine knowledge information with user-item interaction data in different recommendation systems, such as convolutional neural networks (CNN), attention mechanisms, and graph neural networks (GNN).

Research Origin and Author Information

This paper was written by Shuang Liang, Jie Shao, Jiasheng Zhang, and Bin Cui, affiliated with research institutions including the Future Media Center, School of Computer Science and Engineering, University of Electronic Science and Technology of China, and the Sichuan Institute of Artificial Intelligence, University of Electronic Science and Technology of China. The paper was published in September 2023 in the IEEE Transactions on Knowledge and Data Engineering journal.

Research Content and Process

This paper proposes a graph-based sampling-free strategy to achieve efficient performance in knowledge graph-enhanced recommendations. The proposed method significantly improves recommendation performance by leveraging node centrality while combining knowledge graph embeddings and recommendation tasks. It efficiently captures high-order connection information in knowledge graph-enhanced recommendations through a local aggregation block. Experiments on three datasets demonstrate that the proposed method is competitive in efficiency, reaching the latest research level.

Research Process

1. Processing User-Item Interaction Data: The study first converts user-item interaction data into graph-structured data.
2. Node Centrality Calculation: Node centrality is used to determine the weight assignment for each node, especially nodes in the knowledge graph.
3. Knowledge Graph Embedding: Knowledge graph structure information is utilized to train embeddings.
4. Multi-Hop Top-k Neighbor Aggregation: High-order connection information is captured by sampling the most important neighboring nodes for the target node based on node centrality and updating these nodes’ embeddings.
5. Model Optimization: Finally, a loss function for the sampling-free strategy is used to optimize the model parameters.

Experiment and Result Analysis

The paper conducted extensive experiments on three public datasets: Amazon-Book, Yelp2018, and Last-FM, showing that the proposed method outperforms existing state-of-the-art methods in both recommendation performance and efficiency. Specific experimental results include:

1. Recommendation Performance Comparison: The proposed method achieved the best performance on all datasets, with significant improvements especially in scenarios with low user-item interaction density.
2. Effectiveness of Weight Assignment Strategies: Different weight assignment strategies (such as uniform assignment, random assignment, frequency assignment, degree centrality, and PageRank centrality) were compared, confirming the effectiveness of PageRank centrality in recommendations.
3. Effectiveness of Multi-Hop Top-k Neighbor Aggregation: Multi-level neighbor sampling and information aggregation better capture high-order connection information, significantly improving recommendation accuracy.

Conclusion and Value

The graph-based sampling-free strategy proposed in this paper not only has significant advantages in computational efficiency but also markedly improves the recommendation performance of knowledge graph-enhanced recommendation systems. By introducing graph structure information, the method can more reasonably assign weights to negative samples and capture high-order connection information. This has important application value in recommendation scenarios with sparse data, such as e-commerce and social networks.

Highlights and Innovations

1. Establishing a New Research Paradigm: By introducing graph structure information into the sampling-free strategy, the proposed method offers a new approach for knowledge graph-enhanced recommendation systems.
2. Efficient Computation: The method not only improves recommendation accuracy but also maintains computational efficiency, offering a noticeable speed improvement compared to traditional methods.
3. Multi-Hop Top-k Neighbor Sampling: By selecting the most important neighboring nodes during aggregation, the method efficiently captures high-order connection information, enhancing recommendation quality.

Future work will further explore advanced graph neural networks and different node centrality algorithms to optimize the effectiveness of the graph-based sampling-free strategy.