A Bayesian Linear Regression Approach to Predict Traffic Congestion

Abstract

Regression-based traffic modelling can estimate traffic congestion as a response variable by incorporating explanatory spatiotemporal components. Bayesian inference is widely used in traffic modelling as it has advantages over a frequentist approach. Previous approaches mainly focused on offsetting Bayesian inference by incorporating supervised feature extraction, data redistribution and competitive expectation-maximization techniques to achieve better accuracy in traffic forecasting. Unlike the frequentist approach, these combined Bayesian inference approaches lack interpretability. This paper proposes a simple Bayesian Linear Regression approach for spatiotemporal traffic congestion prediction that leverages Bayesian inference to facilitate model interpretability and quantify model uncertainty. The model is evaluated in terms of mean absolute error (MAE) and root mean squared error (RMSE). The experiment shows that Bayesian linear regression modelling can be trained on small data observations to quantify model uncertainty and predict traffic congestion without sacrificing interpretability and accuracy in comparison with the frequentist approach.