Feature-Level Attentive Neural Model for Session-Based Recommendation

Abstract

The main goal of session-based recommendation is to predict a user's next click based on historical anonymous session data. One important aspect is capturing the interest drift that occurs in a user's click sequences. Recent studies have mainly exploited the attention mechanism to extract the main intentions of users and reduce the influence of unintentional actions on the performance of recommendation systems. However, the previous works attempt to extract a user's interests at the item level; they ignore the inherent correlations between a user's interest drift. Therefore, we propose a novel Feature-level Attentive Neural Model (FANM) as a solution capable of capturing a user's current interests by considering interest drift at the feature level. Our model exploits a gated recurrent unit (GRU) and a multihead attention mechanism to extract a user's current interests at the feature level from click sequences; then, it integrates the user's long-term and short-term interests with their last actions (e.g., clicks) to predict the next action. Our proposed model effectively captures a user's interest drift by performing sufficient modeling of a user's sequence data, resulting in increased recommendation accuracy. The experiments on the two real-world datasets show that FANM performs significantly better than baseline methods.