A Lightweight Method of Knowledge Graph Convolution Network for Collaborative Filtering

Introduction

The advancement of social technology, specifically the widespread integration of the mobile Internet into people's daily lives, has resulted in individuals being exposed to an extensive range of information daily. The overwhelming quantity of data has given rise to information overload, causing individuals to experience feelings of being overwhelmed. Recommendation systems have emerged to alleviate the issue of information overload. The main objective of recommendation systems is to assist individuals in navigating through the extensive data and identifying content that may be relevant or of personal interest.

Recommendation systems are widely used in diverse domains, including e-commerce, short videos, healthcare services, and education (George & Lal, 2021; Salloum & Tekli, 2021; Xiao et al., 2022). For example, a general approach in recommendation systems is ranking, where items are rated according to popularity, and highly popular items are recommended to users. However, this recommendation method may need more attention focused on user preferences and personalized needs. Collaborative filtering is a conventional method that leverages historical user-item interactions to generate personalized recommendations. Extensive literature has demonstrated the significant advantages of collaborative filtering in improving recommendation performance (He et al., 2017; Herlocker et al., 2004). However, collaborative filtering may encounter challenges, such as data sparsity and cold start, in certain recommendation scenarios (Wei & He, 2022).

Knowledge graph is a knowledge database representing the objective world in a graphical form. It is currently widely utilized in various applications (Ji et al., 2021), such as human-computer interaction and intelligent search. Higher-level structures and semantic information extracted from the given entities can effectively alleviate the data sparsity and cold-start issues encountered in traditional recommendation (Li et al., 2022). Several studies have demonstrated the substantial benefits of incorporating knowledge graphs into collaborative filtering (H. Wang et al., 2019; Zhang et al., 2016). Currently, the predominant approach involves constructing knowledge-aware recommendations using graph neural networks, of which knowledge graph convolutional networks (KGCN) (H. Wang et al., 2019) and knowledge graph attention networks (X. Wang et al., 2019a) are two common methods.

Despite the effectiveness of these models based on graph neural networks in enhancing recommendation performance, several challenges still need to be addressed. For instance, many of these models inherit the steps from traditional graph neural networks. Nonetheless, feature transformation and nonlinear activation are ineffective in knowledge-aware recommendation systems and may impede recommendation performance in collaborative filtering. Furthermore, this leads to a substantial increase in the complexity of the recommendation system, thereby complicating the training procedure.

Table 1 presents a summary of the main acronyms used in the paper. It provides a quick reference for readers to understand the abbreviations employed in this paper.