Autoregressive models for session-based recommendations using set expansion.

Abstract

With the rapid growth of internet technologies, session-based recommendation systems have emerged as a key paradigm in delivering personalized recommendations by capturing users' dynamic and short-term preferences. Traditional methods predominantly rely on modeling the sequential order of user interactions, deep learning approaches like recurrent neural networks and Transformer architectures. However, these sequence-based models often struggle in scenarios where the order of interactions is ambiguous or unreliable, limiting their real-world applicability. To address this challenge, we propose a novel session-based recommendation model, Deep Set Session-based Recommendation (DSETRec), which approaches the problem from a set-based perspective, eliminating dependence on the interaction sequence. By conceptualizing session data as unordered sets, our model captures the coupling relationships and co-occurrence patterns between items, enhancing prediction accuracy in settings where sequential information is either unavailable or noisy. The model is implemented using a deep autoregressive framework that iteratively masks known elements within a session, predicting and reconstructing additional items based on set data characteristics. Extensive experiments on benchmark datasets show that DSETRec achieves outperforms state-of-the-art baselines. DSETRec achieves a 13.2% and 11.85% improvement in P@20 and MRR@20, respectively, over its sequence-based variant on Yoochoose. Additionally, DSETRec generalizes effectively across both further short and long sessions. These results highlight the robustness of the set-based approach in capturing unordered interaction patterns and adapting to diverse session lengths. This finding provides a foundation for developing more flexible and generalized session-based recommendation systems.