Contrastive Learning-based Meta-Learning Sequential Recommendation

Contrastive Learning (CL) has significant practical value in the current field of recommendation systems. Contrastive learning involves learning representations by comparing and contrasting the similarities and differences between different data points or samples. Currently, contrastive learning (CL) is emerging as a promising alternative method to tackle challenges of data with low density and high levels of noise in recommendation systems. However, existing CL methods often focus on finding different augmentation operations for different datasets, which leads to a lack of generalization in the models. Additionally, these CL methods may exhibit abnormal behavior when dealing with limited resources due to the uncertainty introduced by data augmentation. Moreover, in recommendation systems, there is a well-known phenomenon where a small number of popular items account for a large portion of user interactions, while numerous cold-start items have extremely few interaction data points. Therefore, this paper proposes a more versatile contrastive learning model that mitigates the issues of data sparsity and noise, termed the Contrastive Learning with Meta-learning Framework (CLMLRec). Building upon the existing model MCLRec, this paper introduces a pre-training warm-up framework (PRWF) that includes Meta Scaling and Shift Networks, which create adjustment and displacement individually for each item. Through this framework, the paper aims to learn to pre-activate cold item embeddings and demonstrates the effectiveness of CLMLRec on commonly used datasets.