Forecasting Urban Mobility using Sparse Data: A Gradient Boosted Fusion Tree Approach

Haoyu He, Xinhua Wu, Qi Wang

Abstract: Predicting human mobility in urban landscape poses complex challenges, especially when navigating sporadic mobility datasets. This paper introduces a robust predictive model founded on gradient boosted decision trees, designed to forecast human mobility patterns up to 15 days. We propose a unique feature augmentation approach that stresses two pivotal elements: the extraction of user behavior features from past individual mobility trajectories and the integration of temporal variation features based on the mobility patterns of other users during the same periods. Harnessing the capabilities of cutting-edge tree-based algorithms, our model handles missing data and ensures both computational efficiency and model transparency---qualities essential for instantaneous mobility predictions and strategic urban initiatives. A key facet of our methodology is the fusion model tactic, merging the merits of XGBoost and Light-GBM, thereby fortifying prediction reliability, curtailing overfitting, and boosting prediction precision. This paper presents a holistic framework for forecasting extended urban human mobility, providing invaluable perspectives for urban development, transportation planning, and disaster preparedness strategies. The source code can be accessed at https://github.com/he-h/HuMob2023.