Knowledge Graph-Based Socially Enhanced Explainable Recommendation System

Introduction

With the increasing amount of information in the Internet world, the relevant information about users and products has rapidly expanded, leading to a growing problem of information overload. Recommendation systems can effectively alleviate this problem by recommending a small number of items that match the user’s preferences, helping users quickly access content of interest and assisting businesses in achieving targeted marketing, thereby enhancing customer loyalty. Recommendation systems play an increasingly important role across various platforms, such as e-commerce, social media, and search engines.

The performance of recommendation systems highly depends on the recommendation techniques used. Early Collaborative Filtering (CF) methods make recommendations by suggesting items similar to those with which the user has previously interacted. However, CF methods perform poorly when user-item interactions are sparse. To address this limitation, various auxiliary information, such as item attributes, user information, and context, has been integrated into models. Especially, when side information is transformed into feature vectors and input into supervised learning models alongside user and item IDs, this approach performs well in some scenarios but has the drawback that each interaction is treated as an independent instance, unable to aggregate information from higher-order neighbors. Recent studies have shown that relationships between attributes are closely related and can be organized into Knowledge Graphs (KGs). In this context, many methods have been developed to model the structure of knowledge graphs for recommendations, which can be categorized into regularization methods, path-based methods, and Graph Neural Network (GNN) methods.

Most current works mainly consider item knowledge graphs and ignore the social influence on user decision-making. In fact, users are indeed influenced by social relationships when making decisions, and this influence is transmissive. For example, user A may become interested in an application under the recommendation of their friend B, and B’s interest can be passed on to their other friends. Hence, comprehensively modeling the social graph is of significant importance for improving recommendation systems.

Paper Source

This study was conducted by the following authors: Chunyu Liu, Wei Wu, Siyu Wu, Lu Yuan, Rui Ding, Fuhui Zhou, and Qihui Wu, from the School of Communications and Information Engineering at Nanjing University of Posts and Telecommunications, the College of Electronic and Information Engineering at Nanjing University of Aeronautics and Astronautics, and the Department of Learning and Performance Systems at Pennsylvania State University. This research was published in the February 2024 issue, Volume 36, No. 2 of the IEEE Transactions on Knowledge and Data Engineering.

Research Process

Research Subjects and Process

1. Data Construction: This paper primarily uses two real datasets, last-fm and yelp2018. The last-fm dataset is derived from an online music system, containing user comments on musicians and social relationships. The yelp2018 dataset originates from the Yelp Challenge, containing users’ ratings of restaurants and bars, as well as social relationships. These data are then constructed into user-item interaction graphs, item knowledge graphs, and user social graphs.

2. Embedding Layer: The TransR model is used for embedding the knowledge graph, transforming entities and relationships into vectors while capturing the graph’s topology. TransR assumes that entities and relationships are not in the same semantic space, projecting entity embeddings from entity space to relationship space via a trainable matrix. The loss function for embeddings is optimized using the negative sampling method.

3. User Aggregation Layer: A Graph Convolutional Network (GCN) framework is used to recursively aggregate information from neighboring nodes. A node-level attention mechanism is also designed to highlight nodes important to the central node. In the social graph, a new user vector is generated by integrating the influence of friends.

4. Graph-Level Attention Network: A graph-level attention network is designed to generate weights for the user, distinguishing information from CKG and USG. The final user-side information is derived from a linear combination of information from CKG and USG.

5. Item Aggregation: Side information of items is aggregated from higher-order neighbor nodes in the CKG.

6. High-Order Connection Learning: By stacking multiple layers, users and items can aggregate information from higher-order neighbors. Through multiple layers of stacking, higher-order connections are captured, and the outputs of the previous layers are concatenated as the final embeddings for users and items.

7. Inference Layer: Representations are predicted through a linear layer and activation functions to obtain the recommendation score function used for ranking recommendations.

Experiments and Results

The paper conducts extensive experiments to verify the model’s effectiveness and demonstrates through experimental data that the recommendation system outperforms other state-of-the-art methods in various scenarios.

1. Experimental Results: In multiple Top-K recommendation tasks, the model proposed in this paper achieves the best performance on HR, NDCG, Recall, and Precision metrics. Particularly in user groupings, whether for high interaction frequency users or sparse interaction users, our model shows stable and excellent performance on both last-fm and yelp2018 datasets, especially prominent in the high interaction frequency user group, proving that the model can effectively capture users’ true preferences.

2. Component Impact Analysis: By designing different model variants, the importance of the multi-level attention algorithm and knowledge graph embeddings in the model is verified. Results show that the multi-level attention mechanism and aggregation of higher-order neighbor relationships significantly improve model performance.

3. Specific Case Analysis: By choosing paths in the social graphs CKG and USG, the model can accurately capture users’ true preferences. For example, through the path u25→i580→e45819→i45, the model successfully captures user u25’s preference for a particular restaurant and accurately recommends new items through social relationships.

Conclusion

This paper proposes a novel socially enhanced knowledge graph attention network (SKGAN) for explainable recommendation. By comprehensively utilizing the Common Knowledge Graph (CKG) and User Social Graph (USG), this paper enriches the embedding representations of users and items and significantly enhances the model’s ability to capture higher-order semantic information through GCNs. Experimental results validate the model’s effectiveness and superiority in practical tasks, demonstrating that mining higher-order connection information from social graphs can significantly improve the performance of recommendation systems. Future research will explore more user and item-side information to further optimize recommendation system performance.

Significance and Value of the Research

The method proposed in this paper not only advances the academic research of knowledge graph-based recommendation systems but also provides effective solutions for practical applications. It can be widely applied in precise recommendations in e-commerce, social media, and other fields, truly realizing the need for personalized user recommendations. Through a series of innovative algorithm designs and experimental validations, the method significantly enhances recommendation performance, possessing broad application value and research significance.