A fast matrix autoregression algorithm based on Tucker decomposition for online prediction of nonlinear real-time taxi-hailing demand without pre-training

Abstract

Online prediction of real-time taxi-hailing demand generally provides better real-time decision support for passengers and taxi drivers compared with offline prediction. Current studies focused on using deep spatial-temporal models to predict complex nonlinear taxi-hailing demand. However, whether these models can be used for online prediction of real-time taxi-hailing demand through online training or offline pre-training is hardly discussed. Generally, deep models are not lightweight enough for online training, and pre-training these models requires some time and computational resources. Therefore, a lightweight Fast Matrix Autoregression algorithm based on Tucker Decomposition (FMAR-TD) is proposed for online real-time training and prediction of nonlinear taxi-hailing demand without pre-training. The experimental results show that FMAR-TD achieves millisecond-level online prediction of real-time taxi-hailing demand. Compared with baselines, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of FMAR-TD marginally increase by 2.51 % and 2.56 %, while the computation time (sum of training time and prediction time) significantly reduces by 86.16 %. Open-source link: https://github.com/qdu318/FMAR-TD.