Self-Attention based Multitasking Sequential Recom Mendation

An increasing number of recommender systems use the order of user-item interactions as a key feature to obtain the user’s next behavior. Markov chains generate a state transition matrix by viewing user interaction records, which predicts the next behavior. The recurrent neural network predicts the items that the user is most likely to be interested in next by revealing the semantics of the longer term in the user’s interaction record. The MC-based approach performs well in very sparse datasets,while RNN perform better in more dense datasets with higher model complexity. On the other hand, in real life, there are various relationships between items, these relationships can reveal the detailed knowledge of the item in a closer way. The work in this paper starts from these two aspects and proposes a sequence model based on self-attention that combines the relationships between items to capture both long-term semantics and predictions based on a few interaction records.