[Paper Review] Predicting origin-destination ride-sourcing demand with a spatio-temporal encoder-decoder residual multi-graph convolutional network

Jintao Ke, Xiaoran Qin+4 more|arXiv (Cornell University)|Oct 17, 2019

Transportation and Mobility Innovations43 references208 citations

TL;DR

This paper proposes ST-ED-RMGC, a deep learning model that predicts origin-destination (OD) ride-sourcing demand using a spatio-temporal encoder-decoder framework with residual multi-graph convolutional networks. By constructing multiple OD graphs capturing geographical, functional, and demand correlation relationships, and combining them with LSTM for temporal modeling, the model achieves state-of-the-art performance, reducing MAPE by 22% compared to prior methods on Manhattan data.

ABSTRACT

With the rapid development of mobile-internet technologies, on-demand ride-sourcing services have become increasingly popular and largely reshaped the way people travel. Demand prediction is one of the most fundamental components in supply-demand management systems of ride-sourcing platforms. With accurate short-term prediction for origin-destination (OD) demand, the platforms make precise and timely decisions on real-time matching, idle vehicle reallocations and ride-sharing vehicle routing, etc. Compared to zone-based demand prediction that has been examined by many previous studies, OD-based demand prediction is more challenging. This is mainly due to the complicated spatial and temporal dependencies among demand of different OD pairs. To overcome this challenge, we propose the Spatio-Temporal Encoder-Decoder Residual Multi-Graph Convolutional network (ST-ED-RMGC), a novel deep learning model for predicting ride-sourcing demand of various OD pairs. Firstly, the model constructs OD graphs, which utilize adjacent matrices to characterize the non-Euclidean pair-wise geographical and semantic correlations among different OD pairs. Secondly, based on the constructed graphs, a residual multi-graph convolutional (RMGC) network is designed to encode the contextual-aware spatial dependencies, and a long-short term memory (LSTM) network is used to encode the temporal dependencies, into a dense vector space. Finally, we reuse the RMGC networks to decode the compressed vector back to OD graphs and predict the future OD demand. Through extensive experiments on the for-hire-vehicles datasets in Manhattan, New York City, we show that our proposed deep learning framework outperforms the state-of-arts by a significant margin.

Motivation & Objective

• Address the challenge of predicting origin-destination (OD) ride-sourcing demand, which is more complex than zone-based demand prediction due to intricate spatial-temporal dependencies.
• Model non-Euclidean spatial correlations among OD pairs using multiple graph structures based on proximity, functional similarity, distance, and historical demand patterns.
• Develop an end-to-end encoder-decoder framework that jointly captures spatial dependencies across OD pairs and temporal dynamics within each OD pair.
• Improve real-time decision-making in ride-sourcing platforms by enabling accurate short-term OD demand forecasts for vehicle reallocation, dispatching, and shared ride optimization.

Proposed method

• Construct multiple OD graphs where each OD pair is a node, and adjacent matrices encode relationships such as geographical proximity, functional similarity, distance, and historical demand correlation.
• Design a residual multi-graph convolutional (RMGC) network to extract contextual-aware spatial dependencies from the constructed OD graphs.
• Integrate a long-short-term memory (LSTM) network to model the temporal dynamics of individual OD pairs.
• Implement an encoder-decoder architecture: the RMGC and LSTM networks encode spatial and temporal features into a dense latent vector space, which is then decoded back into OD graphs using the same RMGC layers for future demand prediction.
• Train the end-to-end model using real-world for-hire-vehicle data from Manhattan, optimizing for RMSE, MAE, and MAPE with a focus on non-zero demand OD pairs.
• Fine-tune hyperparameters of all baseline models (e.g., XGBoost, LSTM, GCNs) for fair comparison, using the same data split and evaluation metrics.

Experimental results

Research questions

• RQ1 Can modeling multi-aspect spatial correlations—geographical, functional, and demand-based—improve OD demand prediction accuracy compared to single-structure or non-graph-based methods?
• RQ2 To what extent does the proposed encoder-decoder structure with shared RMGC layers enhance feature learning and prediction performance compared to standalone encoders or decoders?
• RQ3 How does the integration of residual connections and multi-graph convolutions improve the modeling of complex, non-Euclidean spatial dependencies in OD demand?
• RQ4 Does the ST-ED-RMGC model significantly outperform state-of-the-art methods in predicting short-term OD demand across diverse temporal and spatial patterns in urban mobility?

Key findings

• The ST-ED-RMGC model achieves a MAPE of 0.38 on non-zero OD pairs, representing a 22% relative improvement over the second-best baseline (Spatial LSTM, MAPE = 0.42).
• Graph-based models (MGC, RMGC, ST-ED-RMGC) significantly outperform traditional machine learning and standard deep learning models like XGBoost, MLP, and LSTM, demonstrating the importance of spatial correlation modeling.
• The residual block in the RMGC network provides marginal but consistent gains over the standard MGC, indicating improved gradient flow and feature learning stability.
• The encoder-decoder structure further improves performance over the RMGC baseline (MAPE 0.40 vs. 0.38), showing that joint encoding and decoding of spatial and temporal features enhances predictive generalization.
• Visual analysis confirms that the model accurately captures complex human mobility patterns, such as morning commutes to commercial zones and evening returns, with low relative error across spatial OD matrices.
• The model effectively handles temporal variability, including multi-peak and irregular demand patterns across different days, outperforming baselines that over- or under-react to fluctuations.