Activity Schedule Modeling Using Machine Learning

A novel data-driven approach for activity schedule modeling is presented in this paper. The paper's contribution is twofold. First, the activity schedule is modeled as a time series to facilitate simultaneous prediction of activity participation, start times, and duration. Simultaneous prediction helps avoid assuming a predefined decision structure and allows all possible interdependencies among these choice facets to be modeled. The time series representation also ensures time budget constraints are automatically satisfied. Second, a machine learning tool called long short-term memory (LSTM) network is used to model the time series. The LSTM's ability to model long-term dependencies ensures that activity patterns are generated considering the influence of distant and recent past. A bidirectional LSTM is used to capture the effect of (planned) future activities on the present activity participation. The model derives all the relations from the data without requiring assumptions by the modeler on the decision-making behavior. Further, the problems arising from class imbalance in the schedule caused due to less frequently performed activities are also explored and addressed. The models are calibrated and validated using the activity-travel diary data from the OViN 2016 dataset. To evaluate the robustness of the model, it is also tested on a time budget dataset with 23 different activity types. The results indicate that the proposed method can predict the distributions of activity start times and duration with reasonable accuracy. The results demonstrate that the proposed method can efficiently model activity schedules and can be a useful tool for travel demand modeling.