DRM: An Interpretable Dual-view Reasoning Model for POI Retrieval

With the widespread application of Location-Based Services (LBS), Point-of-Interest (POI) retrieval has demonstrated significant value in business intelligence and social network analysis. This study focuses on ad-hoc POI retrieval, a fundamental task in LBS that aims to retrieve matching POIs based solely on the current user query and location, without relying on user preferences. This task is challenging due to ambiguous user intentions and sparse spatial representations. Existing research relies on large black-box models, which achieve high performance at the cost of decision interpretability. To address this trade-off, this study proposes a POI retrieval reasoning model named DRM that integrates text semantics and spatial proximity. First, we decompose the multimodal POI retrieval problem into two independent views: geographic and semantic. Second, we propose a dual-view joint training paradigm to enhance the model's generalization ability by replaying memories from two views. Finally, we designed a step-by-step reasoning decision module to improve retrieval performance and the interpretability of decisions. Extensive experiments on two real-world datasets demonstrate that our proposed method significantly outperforms state-of-the-art (SOTA) baselines. Furthermore, through a novel decision transition matrix and case studies, we demonstrate the model's interpretability. The implementation of the DRM model is available at https://github.com/simplehx/DRM .