Zero-shot learning for predicting transportation traffic volume

Abstract

Traffic prediction is essential for the success of Intelligent Transport Systems (ITS), as accurate and reliable traffic information directly impacts stakeholders' ability to make informed decisions regarding route selection. A range of statistical and deep learning methods have been employed for these predictions; however, these approaches are often time-consuming due to the extensive training, validation, and testing required on historical datasets. Zero-shot learning—a category of machine learning algorithms—has emerged to overcome these challenges. Unlike conventional models, zero-shot techniques are pre-trained on extensive historical time series data from diverse domains, so they do not require manual training. This study evaluates the prediction accuracy of various zero-shot learning models with statistical and deep learning models. Through comprehensive experiments involving seven types of machine learning models and multiple frequencies of time series data, our findings reveal that zero-shot learning models excel in predicting traffic volume.