[Paper Review] A Unified Neural Network Approach for Estimating Travel Time and Distance for a Taxi Trip

Ishan Jindal, Tony Tony+4 more|arXiv (Cornell University)|Oct 12, 2017

Traffic Prediction and Management Techniques15 references62 citations

TL;DR

The paper proposes ST-NN, a unified deep neural network that jointly estimates travel time and travel distance between origin-destination pairs using only GPS bin coordinates and time-of-day, achieving better performance than state-of-the-art methods on NYC taxi data.

ABSTRACT

In building intelligent transportation systems such as taxi or rideshare services, accurate prediction of travel time and distance is crucial for customer experience and resource management. Using the NYC taxi dataset, which contains taxi trips data collected from GPS-enabled taxis [23], this paper investigates the use of deep neural networks to jointly predict taxi trip time and distance. We propose a model, called ST-NN (Spatio-Temporal Neural Network), which first predicts the travel distance between an origin and a destination GPS coordinate, then combines this prediction with the time of day to predict the travel time. The beauty of ST-NN is that it uses only the raw trips data without requiring further feature engineering and provides a joint estimate of travel time and distance. We compare the performance of ST-NN to that of state-of-the-art travel time estimation methods, and we observe that the proposed approach generalizes better than state-of-the-art methods. We show that ST-NN approach significantly reduces the mean absolute error for both predicted travel time and distance, about 17% for travel time prediction. We also observe that the proposed approach is more robust to outliers present in the dataset by testing the performance of ST-NN on the datasets with and without outliers.

Motivation & Objective

• Motivate accurate travel time and distance estimation for urban taxi/ride-hailing systems.
• Leverage large-scale NYC taxi data to develop a regression-based, feature-engineered-free model.
• Demonstrate that a joint ST-NN architecture improves travel time prediction over baselines.
• Show robustness of ST-NN to outliers and compare with existing methods.

Proposed method

• Represent origin and destination as 2-D GPS bins and time-of-day as 1-D time bins.
• Develop Dist DNN and Time DNN modules (three-layer MLPs) to predict distance and time respectively.
• Fuse the modules in a unified ST-NN that jointly minimizes the sum of travel-time and travel-distance losses.
• Train via stochastic gradient descent on the NYC taxi dataset with 80/20 train/test split.
• Compare against linear regression baselines and standalone TimeNN/DistNN modules, as well as a reference method (BTE).
• Perform pre-processing to map GPS coordinates to discrete cells and split time-of-day into weekday/weekend bins.

Experimental results

Research questions

• RQ1 Can a neural network jointly predict travel time and distance from origin-destination GPS coordinates and time-of-day without route information?
• RQ2 Does joint learning (ST-NN) improve travel time predictions compared to time-only models and linear baselines?
• RQ3 How does the model perform in the presence of outliers and across different times of day (weekday vs weekend)?
• RQ4 How does predicted travel distance influence time estimation in urban taxi data?

Key findings

• ST-NN outperforms simple linear regression baselines for travel time estimation.
• Joint learning in ST-NN improves travel time MAE by about 13 seconds over TimeNN on NYC data.
• DistNN outperforms simple linear regression for travel distance; ST-NN achieves near top distance performance with MAE around 0.27 (in the reported units) and high R2 (~0.943).
• ST-NN shows robustness to outliers, maintaining superior performance relative to a baseline method when evaluated on data with anomalous trips.
• Compared to a Manhattan-only benchmark, ST-NN reduces travel time MAE by about 17%.