Contrastive Learning‑based Simplified Graph Convolutional Network Recommendation

[Purposes] Considering the problems of the existing Graph Convolutional Network (GCN) recommendation models, such as low model convergence efficiency, over-smoothing, and deteriorative recommendations for long-tail items caused by the effect of high-degree nodes on presentation learning, a Contrastive Learning-based Simplified Graph Convolutional Network recommendation algorithm (SGCN-CL) is presented. [Methods] The self-supervised learning method was used to generate multiple views for the user and item nodes for contrastive learning, in order to improve the accuracy of model recommendation and effectively improve the recommendation of long-tail items. For each view, the same feature extraction task for different inputs was carried out, and an improved message propagation model SGCN was proposed to carry out feature extraction and enhance model efficiency. The algorithm was evaluated on Amazon-Book, Yelp2018, and Gowalla datasets. [Results] The results show that the recall rates of the above three datasets are increased by 15.4%, 4.3%, 1.4%, and NDCG increased by 17.8%, 4.1%, 1.6%, respectively. Additionally, the efficiency of model has increased more than 55%. After the introduction of Contrastive Learning method, the recommendation effect of non-popular long-tail items is also improved.