Improving End-to-End Sequential Recommendations with Intent-aware Diversification

Sequential recommenders that capture users' dynamic intents by modeling sequential behavior, are able to accurately recommend items to users. Previous studies on sequential recommendations (SRs) mostly focus on optimizing the recommendation accuracy, thus ignoring the diversity of recommended items. Many existing methods for improving the diversity of recommended items are not applicable to SRs because they assume that user intents are static and rely on post-processing the list of recommended items to promote diversity. We consider both accuracy and diversity by reformulating SRs as a list generation task, and propose an integrated approach with an end-to-end neural model, called intent-aware diversified sequential recommendation (IDSR). Specifically, we introduce an implicit intent mining (IIM) module for SR to capture multiple user intents reflected in sequences of user behavior. We design an intent-aware diversity promoting (IDP) loss function to supervise the learning of the IIM module and guide the model to take diversity into account during training. Extensive experiments on four datasets show that IDSR significantly outperforms state-of-the-art methods in terms of recommendation diversity while yielding comparable or superior recommendation accuracy.