Handling Session-Based Prediction Task using Gated Graph Attention Mechanism

Abstract

A session-based recommender system (SBRS) focuses on users’ interests depending on their browsing habits. Most existing research predominantly relies on the user's most recent interactions with items to provide appropriate item suggestions during an active session. Session-based recommendation systems are intended to predict user behaviors by evaluating the interactions occurring in anonymous user sessions. Prior models have primarily focused on representing sessions as sequences, concentrating on user representations rather than item representations to generate recommendations. However, these approaches often fail to capture precise user vectors within sessions and overlook the intricate transitions between items. Session-based recommendation models often utilize recurrent neural networks (RNNs) to capture user sessions; still, these methods are mainly focused on short-term session impacts and fail to encapsulate the full scope of session information. To address this limitation, we introduce a novel approach called Session-Based Recommendation with a Gated Graph Attention Network (GGATN). In this approach, session sequences are handled as data with a graph structure, where a Gated Graph Neural Network (GGNN) is first employed to capture user-session information. Following this, a Graph Attention Network (GAT) is used as a message-passing mechanism to enhance the vector representations in the global domain. Each session is then modeled as a combination of global preferences and the specific interests of the current session using an attention mechanism. On the e-commerce dataset, several experiments are carried out to verify the performance of the GGATN model, and found that the proposed model consistently outperforms existing state-of-the-art session-based recommendation methods.