Meta-Paths Aware Dynamic Multi-Interest Learning for Sequential Recommendation

The multi-interest sequential recommendation system analyzes users' historical behaviors to construct multiple representations of interests, thus better capturing their diverse interests. Although existing methods perform well in recommendation tasks, most of them construct multiple user interests based on static graphs. They fail to consider the evolving nature of user interests and the semantic relations between consecutive items. They overlook the evolving nature of user interests and the dependencies between consecutive items. Therefore, we propose Meta-Path aware Dynamic Multi-Interest Learning for sequential recommendation (MPDMI). MPDMI constructs multi-interest representations of users that evolve with behavior sequences by capturing meta-paths between consecutive interactions in a dynamic graph. This method not only accurately captures the dynamic changes in user preferences but also extracts dependencies between consecutive items. We conduct extensive experiments on three benchmark datasets, and the results show that MPDMI significantly outperforms several state-of-the-art methods, demonstrating its great potential in handling the dynamic evolution of user multi-interests. The implementations are available at https://github.com/MPDMI/MPDMI.