基于知识图谱注意力与残差增强图卷积的电子商务推荐算法

E-Commerce Recommendation Algorithm Based on Knowledge Graph Attention and Residual Enhanced Graph Convolution

摘要: 随着电子商务的快速发展，推荐系统已成为连接用户需求与平台资源的核心桥梁。传统推荐模型普遍存在语义关联挖掘不充分与特征传递易衰减等问题，难以满足用户对精准化、个性化推荐的需求。对此，本文提出一种基于知识图谱注意力与残差增强图卷积的电子商务推荐模型(E-KGARConv)，该模型通过注意力机制和残差连接实现缓解深层图卷积中的梯度消失问题与充分挖掘知识图谱中对用户偏好更具指示性的核心关联。在ML-100K数据集上的实验表明，E-KGARConv的准确率(0.7043)、精准率(0.7135)、召回率(0.7828)均优于CKE、KGCN、KGAT等主流模型，相较于推荐性能较好的KGAT分别提升了0.4%、0.43%、0.89%，消融实验进一步验证残差连接与注意力机制的有效性。研究结果表明本文提出的E-KGARConv模型能够充分挖掘用户与商品的潜在关联，有效提升推荐性能，可为电子商务场景提供更贴合用户需求的优质推荐服务，同时也为后续知识图谱推荐算法的模块设计提供参考。

Abstract: With the rapid development of e-commerce, the recommendation system has become the core bridge connecting user needs and platform resources. The traditional recommendation model generally has problems such as insufficient semantic association mining and easy attenuation of feature transmission, which is difficult to meet the needs of users for accurate and personalized recommendation. In this regard, this paper proposes an e-commerce recommendation model (E-KGARConv) based on knowledge graph attention and residual enhanced graph convolution. The model alleviates the gradient disappearance problem in deep graph convolution through attention mechanism and residual connection and fully exploits the core association in the knowledge graph that is more indicative of user preferences. Experiments on the ML-100 K dataset show that the accuracy (0.7043), precision (0.7135), and recall (0.7828) of E-KGARConv are better than mainstream models such as CKE, KGCN, and KGAT. Compared with KGAT, which has better recommendation performance, it has increased by 0.4%, 0.43%, and 0.89%, respectively. The ablation experiment further verifies the effectiveness of the residual connection and attention mechanism. The research results show that the E-KGARConv model proposed in this paper can fully tap the potential association between users and commodities, effectively improve the recommendation performance, and provide a high-quality recommendation service that is more in line with user needs for e-commerce scenarios. It also provides a reference for the module design of subsequent knowledge graph recommendation algorithms.