<scp>FIPRec</scp> : Fusing User Main Intention and Long and Short‐Term Preferences for Sequential Recommendation

ABSTRACT User behaviours often consist of multi‐intention, and fully leveraging this intention information enables accurate preference modelling in sequential recommendation. However, most existing models struggle to identify the main intention driving the current behaviour in multi‐intention scenarios. They fail to effectively leverage multi‐intention information to enhance preference modelling and even lack the capability to perceive multi‐intention in user behaviours. To this end, we propose FIPRec, a novel sequential recommendation framework that dynamically identifies and utilises the main intention to guide long and short‐term preference modelling. First, we introduce a dual‐perspective intention encoder to generate moment‐level intention representations and aggregate them into a dynamic main intention representation. Second, we propose a behaviour identification module to evaluate the alignment between behaviours and the main intention, and select main behaviours for refined preference extraction. Third, we construct a two‐stage fusion module that integrates long and short‐term preferences under the guidance of the main intention. Finally, we design a self‐supervised learning module to optimise the results of the behaviour identification module. We conducted extensive experiments on four real‐world datasets. The results demonstrate that the FIPRec model significantly improves recommendation performance compared to mainstream models across multiple metrics.