The Story of Mobility: Combining State Space Models and Transformers for Multi-Step Trajectory Prediction

Abstract

Machine learning models for predicting human mobility often require large datasets for training, which are not always available. As a result, methods capable of learning from limited data are essential. The Human Mobility Challenge 2024 was designed to evaluate the effectiveness of various approaches in such constrained scenarios. In this paper, we present a deep learning model that integrates state space models with transformers for multi-city trajectory prediction. Specifically, the model employs the state space model Mamba as an encoder to process long-range trajectories, while a transformer decoder predicts future locations by querying past trajectories with future timestamps. Our results demonstrate the model's effectiveness and suggest strong generalizability across cities. The approach ranked in the top 10 of the challenge, highlighting its competitiveness in limited-data settings.