Denoising Sequence Embeddings via Contrastive Learning for Micro-video Recommendation

Rui Zhao, Beihong Jin, Yisong Yu, Beibei Li

Abstract: In recent years, micro-video apps such as TikTok, Kwai, etc. have become widely popular, but recommendation models dedicated for micro-videos are relatively few in number. The paper analyses the characteristics of micro-video apps and sequences of interactions between users and micro-videos, and then presents a model CLASRec for micro-video recommendation. CLASRec follows the typical method of sequential recommendation, i.e., learns embeddings of user interests from historical interaction sequences. On this basis, for an interaction sequence of the same user, CLASRec adopts contrastive learning to reduce the noise implicit in interaction embeddings. In particular, two contrastive encoders are presented to align the embeddings of a pair of semantically-equivalent sequences. The extensive experiments are conducted on two real-world datasets. The experimental results show that CLASRec outperforms existing multiple models and achieves significant performance improvements in terms of Recall and NDCG.