| |
| Predicting citywide human mobility is of great importance to several related fields, such as urban planning and traffic engineering. However, mobility data collected in a whole metropolis always faces challenges of label noise and few ground truth annotation. Blindly training with these noisy trajectories could certainly introduce inappropriate bias to model parameters, and reduce the performance of mobility prediction. This paperproposes a prediction with calibration framework, to quantify the quality and importance of each trajectory. The main module of proposed approach is a pre-trained calibration network, which is designed to be model-independent and can be trained in an unsupervised manner. It takes trajectories of a mobile user as input, evaluates the quality of them by learning intrinsic regularity and periodicity, and finally returns a numerical score. In this way, trajectories with strong regularity and periodicity for prediction could get higher scores, while the irregular movements with weak predictability score lower. Finally, a neural prediction model is trained with instance weighting strategy, which integrates the results of calibration network into the parameter updating process of mobility prediction model. Experiments conducted on citywide mobility dataset demonstrate the effectiveness of proposed approach, when dealing with massive noisy trajectories in real world. |
|
| Keywords:mobility prediction, label noise, calibration network, instance weighting |
|
