Large Language Models are Zero-Shot Next Location Predictors

TL;DR

This paper tackles zero-shot next-location prediction by leveraging large language models to overcome the geographic transferability limitations of traditional DL approaches. By evaluating 15 LLMs across three mobility datasets, the authors show that zero-shot predictions can reach up to $ACC@5$ of about 0.362, substantially outperforming DL baselines and suggesting strong language-based generalization for mobility tasks. They also analyze in-context learning, the role of contextual and historical visits, data contamination, and the ability of LLMs to provide textual explanations for their predictions, finding that larger models generally perform best, prompt strategy effects vary, and data leakage is not driving results. The work highlights potential policy and urban-planning implications of deployable LLM-based mobility predictors while calling attention to ethical considerations such as bias and privacy. Overall, the study demonstrates that LLMs can serve as effective zero-shot next-location predictors in data-scarce scenarios, with explainability capabilities and practical impact for transportation and public policy.

Abstract

Predicting the locations an individual will visit in the future is crucial for solving many societal issues like disease diffusion and reduction of pollution. However, next-location predictors require a significant amount of individual-level information that may be scarce or unavailable in some scenarios (e.g., cold-start). Large Language Models (LLMs) have shown good generalization and reasoning capabilities and are rich in geographical knowledge, allowing us to believe that these models can act as zero-shot next-location predictors. We tested more than 15 LLMs on three real-world mobility datasets and we found that LLMs can obtain accuracies up to 36.2%, a significant relative improvement of almost 640% when compared to other models specifically designed for human mobility. We also test for data contamination and explored the possibility of using LLMs as text-based explainers for next-location prediction, showing that, regardless of the model size, LLMs can explain their decision.

Figures (7)

• Figure 1: Framework for predicting the next location using LLMs. In panel A), we graphically describe the problem of next-location prediction. Given a set of historical (green) and contextual (yellow) user visits, a next location predictor is a model that predicts the next place (red) they will visit. In panel B), an example of how we query an LLM and an example of response.
• Figure 2: ACC@5 values of LLMs with zero-shot (purple), one-shot (dark blue) and few-shot (light blue) prompts in New York. The other cities have similar results and are reported in Supplementary Information 7 and 9 files. Overall, the best performances are reached by larger models regardless of the prompt adopted. In particular, Llama 3 70B Instruct is the best-performing model, with GPT-4o, GPT-4, Llama 2 Chat 70B and GPT-3.5 reaching similar performances. Smaller models, especially non-instruct or non-chat ones, have lower performances. Performances may vary significantly depending on the prompt adopted.
• Figure 3: Results in terms of relative improvements and drops with respect to the scenario with $C=6, H=15$ for the city of New York when we modify the availability of contextual $C$ (left) and historical $H$ (right) information. Adding locations to $H$ or $C$ ($C=12$ or $H=30$) led to an improvement of ACC@5 while limiting such information led to a drop of performances that can be more ($C=0$ or $H=0$) or less ($C=3$ or $H=7$) severe.
• Figure SI 1: ACC@5 of LLMs with zero shot (purple), one shot (dark blue) and few shot (light blue) prompts in Tokyo.
• Figure SI 2: ACC@5 of LLMs with zero shot (purple), one shot (dark blue) and few shot (light blue) prompts in Ferrara.

Theorems & Definitions (3)

• Definition 1: Trajectory
• Definition 2: Historical visits
• Definition 3: Contextual visits