Multi-information Self-attention Autoencoder Sequential recommendation

Abstract

The field of sequential recommendation plays a crucial role in personalized recommendation systems, aiming to model users' past interactions and predict their future interactions with items or behaviors. Traditional methods in sequential recommendation typically rely on user behavior history and item attributes for making recommendations, but they overlook the internal relationships and contextual information among items within a sequence. Moreover, existing autoencoder models face limitations in capturing long-term dependencies and effectively modeling contextual information for sequential recommendation tasks. To address these issues, we propose a novel framework called miSAASRec that leverages a multi-information autoencoder with a self-attention module to capture internal relevance and contextual features of the data. This enables miSAASRec to achieve more accurate and comprehensive data encoding, thereby enhancing the performance of the autoencoder. We conducted several experiments to demonstrate the superior performance of our miSAASRec model compared to existing methods, achieving improvements ranging from 7.92% and 32.20% in MRR (Mean Reciprocal Rank) and 9.00% and 28.90% in Recall@10.