Geography-Aware Large Language Models for Next POI Recommendation

TL;DR

The paper tackles the challenge of applying large language models to next POI recommendation by addressing two core gaps: geographic awareness of precise coordinates and POI transition knowledge. It introduces GA-LLM, which combines the Geographic Coordinate Injection Module (GCIM) and the Point-of-Interest Alignment Module (PAM) to embed geographic context and align POI transitions into the LLM semantic space, using quadkey-based geographic encoding and learnable Fourier embeddings, along with a cross-modal POI alignment mechanism. Empirical results on three real-world datasets show GA-LLM achieves state-of-the-art accuracy (Acc@1) and demonstrates robustness in cross-city cold-start scenarios, with ablations confirming the contribution of GCIM, PAM, and Fourier encoding. The framework also delivers efficiency benefits during training and inference, highlighting its practicality for scalable location-based services and real-time recommendations.

Abstract

The next Point-of-Interest (POI) recommendation task aims to predict users' next destinations based on their historical movement data and plays a key role in location-based services and personalized applications. Accurate next POI recommendation depends on effectively modeling geographic information and POI transition relations, which are crucial for capturing spatial dependencies and user movement patterns. While Large Language Models (LLMs) exhibit strong capabilities in semantic understanding and contextual reasoning, applying them to spatial tasks like next POI recommendation remains challenging. First, the infrequent nature of specific GPS coordinates makes it difficult for LLMs to model precise spatial contexts. Second, the lack of knowledge about POI transitions limits their ability to capture potential POI-POI relationships. To address these issues, we propose GA-LLM (Geography-Aware Large Language Model), a novel framework that enhances LLMs with two specialized components. The Geographic Coordinate Injection Module (GCIM) transforms GPS coordinates into spatial representations using hierarchical and Fourier-based positional encoding, enabling the model to understand geographic features from multiple perspectives. The POI Alignment Module (PAM) incorporates POI transition relations into the LLM's semantic space, allowing it to infer global POI relationships and generalize to unseen POIs. Experiments on three real-world datasets demonstrate the state-of-the-art performance of GA-LLM.

Figures (7)

• Figure 1: Illustration of text-only LLM limitations in POI recommendation. (a) Average error distances of incorrectly predicted POIs to their correct POIs for text-only LLMs versus sequential models on NYC, CA, and TKY datasets. (b) Case study where text-only LLM fails to predict POI 404 without input context, while our model leverages transition relations for accurate prediction.
• Figure 2: Overview of the GA-LLM framework. The left illustrates the workflow, where user trajectories are transformed into queries for LLM fine-tuning. The center highlights the GCIM module, encoding GPS data via a quadkey system and aligning it with the LLM’s semantic space. The right shows the PAM module, aligning POI embeddings from sequential models (e.g., MTNet) to the LLM’s semantic space.
• Figure 3: Comparison of methods handling geographic coordinates.
• Figure 4: Cumulative distribution function (CDF) of distances between incorrectly predicted and ground truth POIs on two datasets.
• Figure 5: Average error distances between incorrectly predicted and ground truth POIs on CA and TKY.