Prediction of Telecommuting Engagement Through Machine Learning to Enhance Travel Survey Data

This paper presents a novel approach to enhance household travel survey (HTS) data by predicting telecommuting engagement using machine learning (ML) classification techniques. The study aims to address the debate surrounding the impact of telecommuting on overall travel behavior, considering rebound effects and latent demand. While previous research has primarily relied on questionnaires or HTS data for analysis, few studies have successfully integrated telecommuting behavior into travel demand models. The intricate relationship between telecommuting and travel behavior has been a challenge, limiting the incorporation of telecommuting data into existing models. This study fills this research gap by utilizing ML algorithms to predict telecommuting engagement based on one-day HTS data, employing features such as daily distances traveled and time spent at home. Three feature selection algorithms, Boruta, VSURF, and Recursive Feature Elimination (RFE), were applied to identify the most relevant features for the ML models. Among the five classification methods tested, the Random Forest (RF) model utilizing features selected by the Boruta algorithm demonstrated superior performance, achieving high accuracy, specificity, F1-Score, and Matthew’s Correlation Coefficient (MCC). The Bayesian Network (BN) model, while performing well in sensitivity, underperformed in other metrics due to the unsuitability of continuous data. To evaluate the proposed approach’s generalization, the RF model was applied to a separate HTS dataset from the German Mobility Panel (MOP). The out-of-sample prediction achieved promising results, with a 76% accuracy in predicting telecommuting days. The approach presented in this study has potential applications in enhancing HTS data and can be extended to other data sources to improve activity-based models.