Multi-interest sequence recommendation algorithm based on BERT
Fei Wang, Weisen Feng · 2021
Many studies use a fixed vector to represent the user ' s various interest preferences, and this embedding vector does not have enough ability to effectively capture the user ' s different interests, resulting in a lot of user interest representation missing. Therefore, A model based on BERT multi-interest sequence recommendation (BMISREC) is proposed. The model uses BERT for pre-training, and then divides the users historical interaction sequence into t time windows according to the timestamp. By fusing the cavity convolution and forward attention, the local preference within each time window and the global preference between time windows are obtained. Fusion local and global preferences for each window. Multiple different interest embeddings are generated for each user. In the target item preference module, multiple interests learned can adaptively perceive the correlation between different target items through the attention mechanism, and the target item preference changes with the different target items. Then, the deep neural network is used to recommend the fusion of user and target item preferences in a nonlinear manner. Finally, the effectiveness of the model is verified on two real datasets.