Modeling daily in-home activities using machine learning techniques

Abstract

Activity-based modeling for travel demand forecasting have mainly focused on out-of-home activities. However, out-of-home (OH) and in-home (IH) activities are closely related decisions. For example, individuals working-from-home for longer duration are less likely to make commute trips to the workplace. Furthermore, COVID-19 has forced individuals to replace their OH with IH activities, which further indicates the need for in-home activity modelling. To address the inter-dependencies and better predict travel demand, efforts are required to model in-home activities. This paper develops machine learning (ML) models to investigate in-home activities. We have considered in-home activities in the four thematic categories: sleeping, leisure and discretionary, household and personal maintenance, and mandatory activities. Models were developed for: 1) activity participation and 2) activity duration. The former is focused on modeling what type of in-home activities individuals would perform during a typical weekday, where the latter models the duration for each in-home activity. Several machine learning models were applied, including artificial neural network (ANN), regression trees (RT), Ensembles, support vector machine (SVM), k-nearest neighbor (KNN), and Gaussian process regression (GPR). For each technique, we seek the best model by fine-tuning the modeling architecture. We also compared the prediction speed of the models to understand how they would perform in practice. The results of the participation models had an overall accuracy above 95%, where the activity duration models had an R2 between 0.74 and 0.94. This research demonstrates how machine learning models are robust and can be adopted to accurately predict activity participation and duration.