Who You Are Adds Nothing Detectable to Where You Go Next: Sociodemographic Conditioning in LLM Next-Location Prediction

Research investigating the incremental predictive value of sociodemographic attributes in Large Language Model (LLM) next-location prediction has found that including age, gender, occupation, and income offers minimal to no detectable improvement in prediction accuracy. This conclusion is drawn from a study linking sociodemographic records with mobility data of 5,000 Shenzhen residents, where models ranking 100 candidate destinations showed negligible changes in top-1 accuracy whether these attributes were included or excluded.

Why it matters

This finding challenges the assumption that personal demographic data significantly enhances the accuracy of LLM-driven behavioral predictions, particularly in location forecasting. It suggests that resources invested in collecting and processing such sensitive data for this specific purpose may not yield proportional strategic benefits, prompting a re-evaluation of data collection priorities and model design.

Key insights

• Sociodemographic attributes (age, gender, occupation, income) provide negligible incremental predictive value for LLM-based next-location prediction.
• The study used passively sensed mobility data and sociodemographic records from 5,000 Shenzhen residents.
• Prediction instances were evaluated with and without sociodemographic data, while holding mobility history and other prompt content constant.
• Across various history lengths, the change in top-1 accuracy was found to range from -0.8 to essentially zero, indicating no significant benefit from these attributes.
• The research directly tested the contribution of sociodemographic conditioning, which is common in LLM-based travel simulation.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.09609