Enhancing session-based E-commerce recommendations with self-attention: a transformer-based approach

In e-commerce, suggesting relevant items to users based on their immediate, in-session behavior is a critical task. Session-based recommendation systems (SBRS) aim to infer a user's short-term intent from the sequence of interactions within a single browsing session. Although recurrent and convolutional models (RNN/CNN) have been widely used, their reliance on strictly local or sequential processing restricts their ability to capture global relationships spanning the entire session. To address this, we introduce the Transformer-based Session Recommender (T-SR), which applies multi-head self-attention to evaluate all item interactions simultaneously. This mechanism enables the model to learn long-range and global behavioral patterns. Experiments on the RecSys 2020 e-commerce dataset demonstrate that T-SR significantly outperforms CNN and RNN baselines in ranking performance, achieving a Hit Rate (HR@10) of 0.652 and an NDCG@10 of 0.580, showcasing its effectiveness in generating relevant recommendation lists.