DiffGCA: A Diffusion Recommendation Model with Global Attention Mechanism

Sequential recommendation aims to predict users' future preferences by analyzing their historical interactions. Compared to traditional recommendation systems, sequential models emphasize capturing the dynamic changes in user interests to generate more personalized recommendations. Diffusion models introduce noise and denoising processes, providing a flexible generative mechanism for sequential recommendation, which better handles the uncertainty and diversity in user preferences. However, existing diffusion models struggle with capturing the global dependencies and multi-dimensional interests in user behaviors, especially when confronted with sparse data and rapidly changing user interests, leading to recommendations that often lack diversity and accuracy. To overcome these challenges, we propose the DiffGCA model, which integrates the Global Grouped Coordinate Attention module to enhance feature representation by leveraging global information. Extensive experiments on multiple real-world recommendation datasets demonstrate that DiffGCA significantly outperforms two representative baseline models in both recommendation accuracy and diversity, validating its effectiveness in addressing issues related to sparse data and dynamic user behavior.