Predicting Ride Hailing Service Demand Using Autoencoder and Convolutional Neural Network

Abstract

Ride hailing services, such as Uber, Lyft, and Grab, have become a major transportation mode in the last decade. The number of current passenger requests is one of the important factors for such services routing and pricing algorithms. Therefore, predicting future passenger request for ride hailing services can boost the efficiency of the service for both drivers and riders by pre-planning the allocation of vehicles and avoiding traffic congestions. Demand forecasting for ride hailing services relies upon the spatial and temporal correlations of its features. The existing literatures mostly divide the target area into rectangular grids (based on the longitude and latitude), consider only adjacent grids for spatial correlation, and calculate demand for each grid independently. An individual grid can contain different regions with high and low demand or have a major part of it outside the land area, which obscures the granularity and precision of estimations and predictions. This paper attempts to mitigate the limitations of grid-based methods by estimating and predicting ride hailing service demand between geographic regions as pickup and destination zones. For predicting demand, a convolutional neural network is integrated with a recurrent autoencoder network to best capture the spatial–temporal correlations of features, including time of the day, month, year, weekend, holiday, pickup zone, destination, and demand. In our experiments, we forecast the demand for each pickup–destination pair for the next day at a certain hour by observing the demands over the past 2 weeks during the same hour in the New York City hire vehicle data set. Using the same model (CNN-biLSTM-AE) to predict demand for geographical regions, it achieved an [Formula: see text] of 0.984, while predicting demand for cells in the grid achieved an [Formula: see text] 0.545. While using the geographical regions instead of grids for partitioning the space, we compared our deep learning model with LSTM, CNN, CNN-LSTM, and LSTM-AE models and observed an improvement in [Formula: see text] from 0.632 to 0.767 and an improvement in RMSE from 20.53 to 16.33 against CNN.