Improved Prediction of High Taxi Demand: A Deep Spatiotemporal Network for Hyper-imbalanced Data

Abstract

Taxi demand prediction is an intensively studied topic in intelligent transportation research. Recently, deep learning models have been widely applied and have shown good performances. However, these methods overlook the existence of hyper-imbalanced taxi demand, which may result in good indicators in numerical experiments but weak performance in real scenarios. In this paper, we focus on the hyper-imbalance data and improve deep learning abilities for taxi demand prediction. To accomplish this task, slice indicators are introduced to fairly evaluate prediction performance at each taxi demand level. Then, through the lens of the slice indicators, a new form of loss called slice-weighted loss (SWL) is developed to improve the prediction of high taxi demand. Combining the SWL with an improved convolutional long short-term memory (Conv-LSTM) model, a spatiotemporal network called slice-wighted-Conv-LSTM (SW-CLSTM) is proposed. It can overcome the problem of data hyper-imbalance and make considerable improvements in taxi demand prediction. By conducting extensive experiments on large-scale TLC trips, we validate the power of sliceindicators and demonstrate the effectiveness of our approach over state-of-the-art methods.