An Advanced Sequence Recommendation Model Based on Temporal Mixed Attention and Dynamic Fusion

In the realm of natural language processing (NLP), the advent of cutting-edge models like BERT has revolutionized the field, significantly enhancing the efficiency of sequence recommendation systems. However, current models such as BERT4Rec often overlook crucial emotional and contextual information embedded within user sequences, primarily focusing on historical interaction data for analysis. This gap in understanding user preferences and behavior is problematic as it limits the depth and personalization of recommendations. To address this, we introduce a novel model called MAY, specifically tailored for sequence recommendations. MAY aims to integrate user sequences with relevant data, enabling a deeper comprehension of user preferences and behaviors, including emotional and contextual factors. At the heart of the MAY model lies a unique time-mixed attention mechanism that enhances the self-attention function in sequence processing. This mechanism allows MAY to capture a broader range of user preferences and behavior patterns, leading to more accurate and personalized recommendation results. To validate the effectiveness of MAY, we conducted extensive experiments on multiple public datasets. The experimental results demonstrate that MAY significantly outperforms traditional benchmark models, such as BERT4Rec, across various metrics. This underscores the superiority and practical potential of MAY in the domain of sequence recommendations. By addressing the limitations of existing models and introducing a novel approach, MAY offers a promising step forward in enhancing the quality and personalization of sequence recommendations.