Multi-level Feature Extraction in Time-Weighted Graphical Session-Based Recommendation.

Session-based recommendation aims to simulate users’ behavior through a series of anonymous sessions. Recent research work mainly introduces deep learning into the recommender systems, and has achieved relatively good results. Previous research only focused on the clicked item thus ignoring the time information, that is dwell time for each item. It is undeniable that the length of dwell time on an item can reflect the user’s preferences to a certain extent. And they lack the mining latent features of items. In this paper, we propose to explore multi-level feature extraction in time-weighted graphical session-based recommendation, abbreviated as F-TGNN. In F-TGNN, we first construct graphs for session sequences, in which the dwell time between two items is used as the weight of the corresponding edge. Then we use gated Graph Neural Network (GNN) to learn the transitions of items in the session sequence and obtain the embedding of each item. After that, we propose a Feature Extraction Module (FEM) to mine sequential patterns from item-level and contextual information between items from sequence-level. Finally, the predicted score for each item to be the next click is calculated. Extensive experiments conducted on two real datasets show that F-TGNN evidently outperforms the state-of-the-art session-based recommendation methods consistently.