Learning Beyond Proximity: Causal Reasoning with LLMs for Robust POI Prediction

ICLR 2026 Conference Submission16091 Authors

Keywords: Large Language Models, Point of Interests, Spatial, Causal Inference

TL;DR: To the best of our knowledge, this is the first work to embed causal inference into LLM-based POI prediction, explicitly mitigating exposure bias from a geographic perspective for more robust performance.

Abstract: Point-of-Interest (POI) prediction forecasts a user’s next destination from mobility history. A key challenge is geographic exposure bias, where users often visit nearby or popular places out of convenience rather than genuine interest. Such convenience-driven behaviors create spurious correlations that obscure true preferences, leading models to misinterpret frequent check-ins as strong signals of interest. Traditional sequential/graph models rely on surface-level statistical correlations, and recent Large Language Model (LLM)-based methods improve semantic coverage but still inherits exposure bias from observational logs. We address this with causal inference, explicitly modeling the data-generating process and distinguishes preference-driven behaviors from convenience-driven ones. In particular, we estimate geographic propensity scores that quantify the likelihood of a visit due to spatial exposure, and use them to reweight check-ins and align trajectory retrieval in exposure-consistent space. Towards this end, we propose Causal Geographic Prediction (CGP), a unified framework that integrates causal inference with LLM-based trajectory modeling. It employs exposure-aware trajectory prompting, causal-geographic similarity alignment, and supervised fine-tuning to separate genuine preferences from convenience-driven behaviors. Experiments on real-world datasets show that CGP outperforms state-of-the-art baselines.

Primary Area: causal reasoning

Submission Number: 16091