An encoder enhanced sequence recommendation algorithm

Sequence recommendation algorithms using self-supervised learning models usually use the same encoder to generate different embedding vectors of the original sequences, which can easily lead to the lack of the information they carry and the formation of embedding collapse. Meanwhile, when the model pays too much attention to the data details and noise, it will lead to a decrease in model performance and insufficiently accurate recommendation results. Aiming at the above problems, this paper proposes a sequence recommendation algorithm based on an enhanced encoder. Firstly, different data enhancement methods are used to construct views of the original sequences, and then the contrast loss function is used to make the views of the same training samples more similar. In contrast learning, this paper uses an optimized version of the two-tower model for training to obtain the embedding vectors, so that the difference between different augmented views of the same user sequences is minimized and the difference between augmented views of different user sequences is maximized. In addition, after the encoder, the network layer is further modeled and selective discarding is achieved by means of fully-connected layer reconstruction, which improves the accuracy of the system recommendation. Experiments on the Beauty and Sports datasets demonstrate the effectiveness of the augmented encoder model, outperforming the compared models in terms of Hit and NDCG metrics.