Explainable next POI recommendation based on spatial–temporal disentanglement representation and pseudo profile generation

Abstract

The current research in Point-of-Interest (POI) recommendation primarily aims to decipher users’ transitional patterns to predict their future location visits. Traditional approaches often intertwine various features to model these check-in transitions, which inadvertently compromises the quality of the resulting representations. This issue is compounded in both industrial and academic settings, where user-generated textual data is frequently inaccessible or restricted due to privacy concerns. Such limitations in user profiles pose significant challenges to the effectiveness of subsequent applications. In response to these challenges, the recent rise of Large Language Models (LLMs) offers a novel perspective. Diverging from the conventional approach of leveraging LLMs for semantic-based next check-in predictions, our research investigates the potential of integrating LLMs with sequential recommendation systems. This integration aims to augment feature dimensions and facilitate the generation of explicit explanations. To this end, we introduce CrossDR-Gen, a Cross-sequence Location Disentanglement Representation methodology. CrossDR-Gen is specifically designed for next POI recommendation and explanation generation. It uniquely considers spatial and temporal factors in shaping check-in behaviors, offering a comprehensive global view of location transitions. Crucially, CrossDR-Gen utilizes LLMs for pseudo profile generation in scenarios with limited semantic context, thereby enriching user features without relying on additional textual profiles or conversational data. Our experiments on real-world datasets demonstrate that CrossDR-Gen not only excels in addressing cold-start scenarios but also showcases robust recommendation capabilities. These findings validate the effectiveness of our proposed cooperative paradigm between LLMs and sequential recommendation models, highlighting a promising avenue for future research in POI recommendation systems.