Deep Multi-task Learning Models for Short-term Supply-demand Forecasting in Ride-hailing System

Abstract

Transportation network companies (TNCs) are providing on-demand door-to-door ridehailing services in many cities around the world. In order to reduce passenger waiting time and driver search friction, TNCs need to conduct spatio-temporal forecasting of demand and supply-demand gap. However, due to spatio-temporal dependencies pertaining to demand and supply-demand gap in a ride-hailing system, making accurate forecasts for demand and supply-demand gap is a difficult task. Although spatio-temporal deep learning methods have recently proven to be successful in detecting the spatio-temporal dependencies, there exist few more challenges in implementing these methods that require further attention. One of the assumptions in spatio-temporal online taxi-hailing demand forecasting with deep learning is that spatio-temporal dependencies rely on spatial structure and therefore, the zonewise historical data is processed as image pixels. However, spatio-temporal dependencies among the zones are complex and do not capture patterns like image pixels. Furthermore, commonly applied two-dimensional convolution and long-short term memory (LSTM) to detect spatio-temporal patterns from zone-wise historical data increases model complexity, which is not justified with respect to a spatio-temporal deep learning baseline. Therefore, this study applies one-dimesional convolution to historical data with flattened zones to investigate the effectiveness of preserving spatial structure in spatio-temporal forecasting of on-demand taxis and develops a spatio-temporal deep learning baseline architecture containing onedimensional convolutional recurrent layers for justifying increased model complexity in architectures containing two-dimensional convolutional recurrent layers. Experiments with real-world online taxi-hailing data of Didi Chuxing from Chengdu show that convolutional recurrent model implementing one-dimensional convolution in vanilla recurrent neural network with rectified linear activation outperforms convolutional recurrent models combining two-dimensional convolution and long short-term memory network. The findings indicate that preserving spatial structure in online taxi-hailing demand forecasting can be sometimes redundant and complex spatio-temporal deep learning models should be compared to a spatio-temporal deep learning baseline in order to build more computationally efficient architectures for spatio-temporal forecasting. xv Due to confidentiality and privacy issues, ride-hailing data are sometimes released to the researchers by removing spatial adjacency information of the zones, which hinders the detection of spatio-temporal dependencies in deep learning. To that end, a novel spatiotemporal deep learning architecture is proposed in this study for forecasting demand and supply-demand gap in a ride-hailing system with anonymized spatial adjacency information, which integrates feature importance layer with a spatio-temporal deep learning architecture containing one-dimensional convolutional neural network (CNN) and zone-distributed independently recurrent neural network (IndRNN). The developed architecture is tested with real-world datasets of Didi Chuxing, which shows that models developed based on the proposed architecture can outperform machine learning models (e.g., gradient boosting machine, distributed random forest, generalized linear model, artificial neural network). Additionally, the feature importance layer provides an interpretation of the model by revealing the contribution of the input features utilized in prediction. Designing and maintaining the spatio-temporal forecasting models separately in a task-wise and city-wise manner is a challenging task for the continuously expanding TNCs. therefore, a deep multi-task learning architecture is proposed in this study by developing a gated ensemble of spatio-temporal mixture of experts network (GESME-Net) containing convolutional recurrent neural network (CRNN), convolutional neural network (CNN), and recurrent neural network (RNN), which is capable of simultaneously forecasting different spatio-temporal tasks in a city as well as same task across different cities. The proposed architecture is tested with real-world data of Didi Chuxing for: (i) simultaneous forecasting of demand and supply-demand gap in Beijing, and (ii) simultaneous forecasting of demand for Chengdu and Xian. In both scenarios, models from the proposed architecture outperformed the multi-task learning benchmark (e.g., shared bottom model, single-gated spatio-temporal mixture of experts), task-wise and city-wise spatio-temporal deep learning models, and machine learning algorithms (e.g., gradient boosting, random forest, and generalized linear models). The developed architecture provides a basis for spatio-temporal multi-task learning in smart cities. The developments made in this study bridges the gap between applying advanced deep learning methods and maintaining the privacy of the TNCs’ data at the same time, provides a way for reducing the computational cost of spatio-temporal deep learning models for the TNCs, and importantly, shows the viability of utilizing spatio-temporal multi-task learning architecture for jointly forecasting demand and supply-demand gap in a ride-hailing system.