Cross-city-aware Spatiotemporal BERT

Meisaku Suzuki, Yusuke Fukushima, Ryo Koyama, Hayato Kumagai, Tomohiro Mimura, Keiichi Ochiai

Abstract: Predicting human mobility has been actively studied for the past decade because of its various possible applications, such as traffic optimization and urban planning. Despite the increasing interest in human mobility prediction, the training and evaluation of prediction methods are often constrained by the use of different datasets (i.e., each study uses their own dataset for an evaluation). In considering these, the Human Mobility Prediction Challenge (HuMob Challenge) 2024 was held aiming at evaluating state-of-the-art models for the prediction of human mobility patterns using large-scale open dataset.In this paper, we present our solution that ranked in the top 10 among over 100 participating teams. Our method uses LP-BERT as the base model, incorporating a component to account for cross-city and incorporating LSTM, which is suitable for series forecasting. The final prediciton is selected by ensembling these models and a rule-based method that considers the periodicity of human movement based on the probability of each model.The proposed method is trained using data from all four cities (City A, B, C, and D) and evaluated on the target city data using metrics of GEOBLUE and DTW. As a result, we achieved the accuracy scores of GEOBLEU: 0.3008 and DTW: 23.75 for City B, GEOBLEU: 0.3027 and DTW: 18.71 for City C and GEOBLEU: 0.3305 and DTW: 49.75 for City D.