MSAM: Cross-Domain Recommendation Based on Multi-Layer Self-Attentive Mechanism

XiaoBing Song, Jiayu Bao, Yicheng Di, Yuan Liu

Abstract: In recent years, recommendation systems have been extensively implemented across multiple platforms. It can extract useful information from vast amounts of data and recommend appropriate products to users based on their preferences. Typically, recommendation systems are plagued by data scarcity and chilly start issues, which is a serious challenge. To solve these problems, cross-domain recommendations (CDR) have received a lot of attention. Typically, CDR aims to leverage data from other domains or map user preferences from other domains to the target domain to improve recommendation quality and alleviate data sparsity and chilly start problems for new services. However, the majority of existing cross-domain recommendation methods are based on matrix decomposition, which can only learn shallow linear features and cannot adequately address the challenges associated with cross-domain recommendation. Therefore, this paper proposes a multi-layer self-attentive mechanism (MSAM) cross-domain recommendation method to make more accurate predictions by fusing and passing information between different domains. The framework is composed primarily of a feature extraction layer, a multilayer perceptron, and a feature fusion layer that discovers the potential factors of users and items and fuses the potential factors of users from various domains. In addition, we employ the Wasserstein self-attentive mechanism and the multi-headed self-attentive mechanism in the feature extraction layer and the feature fusion layer, respectively, to better extract key user features and learn the affinity of user potential factors across domains. we conducted multiple experimental validations on two actual datasets to demonstrate the model's efficacy and superiority.