MintSR: Multi-Intent Perception with Time-Based Segmentation for Session-Based Recommendation

In an anonymous session recommendation scenario, the key to predicting the user’s next interaction item is accurately recognizing the user’s intent. Traditional methods focusing on the item Identity Document (ID) modeling have limitations. They often only consider the current preference based on the target session and the last clicked item, neglecting the user’s long-term intention tendencies and dynamic changes in intent over time. Moreover, solely relying on ID modeling can result in biased recommendations towards popular items, overlooking subtle shifts in user preferences. To address these issues, this paper introduces the MintSR session recommendation algorithm. MintSR utilizes a hierarchical decoupling model based on time dimensions, categorizing user intent into long-term, short-term, and instantaneous levels. The global context information is mapped to long-term intent, and in the long-term intent, we use the Transformer encoder to eliminate the problem of long-distance dependency and design a unique intent capture module that captures the intent through the category similarity, which can be used to capture the user’s intent through category similarity. The same capture module is also used to capture the user’s short-term and transient intent through category similarity. Through this hierarchical division, the algorithm can fully explore the user’s potential intent and deeply integrate the item category information for modeling. Based on the integration of the three levels of intent, the category comparison learning mechanism is introduced to improve the recommendation performance. Finally, the MintSR algorithm is tested on three real datasets, and the results show that its accuracy is better than other mainstream algorithms.