RTiSR: a review-driven time interval-aware sequential recommendation method

The emerging topic of sequential recommender (SR) has attracted increasing attention in recent years, which focuses on understanding and learning the sequential dependencies of user behaviors hidden in the user-item interactions. Previous methods focus on capturing the point-wise sequential dependencies with considering the time evenly spaced. However, in the real world, the time and semantic irregularities are hidden in the user’s successive actions. Meanwhile, with the tremendous increase of users and items, the hardness of modeling user interests from spare explicit feedback. To this end, we seek to explore the influence of item-aspect reviews sequence with varied time intervals on sequential modeling. We present RTiSR, a review-driven time interval-aware sequential recommendation framework, to predict the user’s next purchase item by jointly modeling the sequence dependencies from aspect-aware reviews. The main idea is twofold: (1) explicitly learning user and item representation from reviews by assigning different weights, and (2) leveraging a hybrid neural network to capture the collective sequence patterns with a flexible order from aspect-aware review sequences. We conduct extensive experiments on industrial datasets to evaluate the effectiveness of RTiSR. Experimental results demonstrate the superior performance of RTiSR in different evaluation metrics, compared to the state-of-the-art competitors.