Self-supervised Recommendation Model Combined with Constrastive Learning Function

Abstract

Aiming at the problem of data deficency and cold start in recommendation algorithm, this paper propose a self-supervised recommendation model combined with constrastive learning function(CLSRec). The model completes self-supervised learning through two positive-and-negative case-based constrastive learning tasks clotfilling prediction and subsequence prediction. Firstly, the algorithm masks a individual item and subsequence of the user interaction respectively according to the corresponding task requirements, and then predicts the missing items through the encoding layer and the self-attention layer. During the process, the model automatically constructs the positive and negative examples of the prediction results, and completes the training according to the constrastive learning loss function corresponding to the task, so as to learn the correct semantic feature representation and space vector coding. Secondly, the model utilizes the trained upstream item coding for the downstream prediction. Finally, the model can obtain user-satisfactory recommended results. By comparison of experimental results on Beauty, Toys and LastFM, the CLSRec works better on multiple metrics.