PSR: A Dual-Level Learning Framework for Patching Sequential Recommendation to Capture Local Preference Transitions

Sequential recommendation (SR) systems based on users’ past behavior sequences are widely used on online platforms to provide personalized suggestions effectively. Transformer-based SR models have demonstrated remarkable success by effectively learning the correlations between individual items in users’ item sequences. Although transformer-based SR models perform well overall, they often fail to capture changes in local trends within long interaction sequences, since they primarily deal with sequences on a global scale. Recently, various methods have been proposed to address this issue by emphasizing local preference information. However, these approaches focus solely on enhancing local preference information, often overlooking the transition information between local preferences. Consequently, they struggle to effectively represent dynamic user behavior sequences. To address this limitation, we propose a dual-level learning framework for sequential recommendation called Patching Sequential Recommendation (PSR), which introduces patch sequences derived from local groups of items within original item sequences. The PSR method uses both patch sequences and item sequences as training samples, allowing for simultaneous training at the individual item level and patch level. Moreover, the proposed framework employs a user representation alignment module to improve user representation by aligning users’ item and patch sequence embeddings. Extensive experiments conducted on four real-world datasets demonstrate that our framework outperforms other state-of-the-art baselines. Our code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/kangvic0615/PSR</uri>