Temporal Attention Mechanism for Sequential Recommendation

Sequential recommendation, which aims at predicting user preferences based on a sequence of his/her historical behaviors, is one of fundamental tasks of modern recommender systems. Prior work show that user behavior patterns differ in and across sessions (defined by inactive time), indicating the temporal dynamics of user preference. While existing deep learning based approaches (e.g. RNN/CNN/Self-Attention) in sequential recommendation fail to model such temporal dependency, our work directly models the temporal relation among the actions to dynamically reweigh the attended influence from previous events based on their temporal proximity. We show that such temporal attention mechanism based model not only outperforms various state-of-the-art sequential models in both running time and accuracy, but also provides explainability for recommendation from attention correlations.