[Paper Review] Neural Network Multitask Learning for Traffic Flow Forecasting

Feng Jin, Shiliang Sun|arXiv (Cornell University)|Dec 24, 2017

Traffic Prediction and Management Techniques7 references4 citations

TL;DR

This paper proposes a multitask learning (MTL) backpropagation neural network for traffic flow forecasting, where multiple time-step predictions are trained jointly to improve generalization. By leveraging correlated traffic flow data across adjacent time points as auxiliary tasks, the MTL model reduces prediction error by 9.31% on average compared to single-task learning, demonstrating superior accuracy and robustness in urban traffic forecasting.

ABSTRACT

Traditional neural network approaches for traffic flow forecasting are usually single task learning (STL) models, which do not take advantage of the information provided by related tasks. In contrast to STL, multitask learning (MTL) has the potential to improve generalization by transferring information in training signals of extra tasks. In this paper, MTL based neural networks are used for traffic flow forecasting. For neural network MTL, a backpropagation (BP) network is constructed by incorporating traffic flows at several contiguous time instants into an output layer. Nodes in the output layer can be seen as outputs of different but closely related STL tasks. Comprehensive experiments on urban vehicular traffic flow data and comparisons with STL show that MTL in BP neural networks is a promising and effective approach for traffic flow forecasting.

Motivation & Objective

• To address the limitation of single-task learning (STL) in traffic flow forecasting, which ignores useful information from related tasks.
• To investigate whether multitask learning (MTL) can improve generalization and prediction accuracy in neural network-based traffic flow forecasting.
• To develop and evaluate a neural network MTL framework that jointly predicts traffic flows at multiple future time steps using shared hidden representations.
• To compare the performance of MTL-based neural networks against conventional STL models using real urban traffic flow data.

Proposed method

• A backpropagation (BP) neural network is designed with multiple outputs, where each output corresponds to a different future time-step prediction, effectively modeling multiple related tasks simultaneously.
• The network architecture shares hidden layers across all time-step predictions, enabling knowledge transfer between tasks through a common representation.
• Training signals from adjacent time-step flows are used as extra tasks, providing inductive bias that improves generalization for the main forecasting task.
• The model is trained using a fixed goal of 0.006 for training error, with a maximum of 300 epochs to ensure convergence.
• Root mean square error (RMSE) is used as the primary evaluation metric to compare forecasting performance between MTL and STL models.
• Experiments are conducted on real-world urban traffic flow data from six road links, with training on 22 days and testing on the final 3 days.

Experimental results

Research questions

• RQ1 Can multitask learning improve the generalization performance of neural networks in traffic flow forecasting compared to single-task learning?
• RQ2 How does joint training on multiple time-step predictions affect forecasting accuracy and model robustness?
• RQ3 What is the extent of performance improvement when using MTL with correlated traffic flow data from adjacent time points as auxiliary tasks?
• RQ4 Does the MTL approach reduce prediction error compared to STL across diverse urban road links?

Key findings

• The MTL neural network achieved a 9.31% improvement in forecasting accuracy over STL on the B_b road link, with RMSE decreasing from 77.46 to 70.25.
• On all six tested road links, MTL consistently reduced RMSE compared to STL, with improvements ranging from 2.34% to 9.31%.
• The training error of the MTL network stabilized after approximately 125 epochs, indicating effective convergence with the chosen hyperparameters.
• The MTL model demonstrated better generalization by leveraging inductive bias from related time-step predictions, leading to more accurate forecasts.
• The performance gain was statistically significant across all test links, confirming the effectiveness of MTL in time-series traffic flow forecasting.
• The results suggest that MTL is a promising approach for enhancing Intelligent Transportation Systems (ITS) by enabling earlier and more accurate congestion prediction.