Modelling of Activity Concentrations Using Machine Learning Technique

Abstract

This study develops a modelling framework for activity concentration prediction following a Machine Learning (ML) approach. The study aims to understand the spatial and temporal evolution of activity considering the effects of the built environment, time of day, and past activities. This study utilizes the Nova Scotia Travel Activity (NovaTRAC) survey and Info Canada Business Establishment dataset to extract activity and built environment information. An Artificial Neural Network (ANN) modelling approach is adopted for prediction modelling as it provides advantages over other modelling approaches in dealing with assumptions, non-linearity, and model accuracy. The ANN model is trained and validated through a trial-and-error process using various model architectures and evaluation metrices, such as mean square error (MSE) and mean absolute error (MAE). Results show that model fitness is satisfactory with significantly smaller values of MSE (0.37) and MAE (0.23). Furthermore, results indicate high prediction accuracy. The outcomes of this study provide insights into factors affecting activity density which will further aid in designing public spaces, transportation infrastructures and safety protocols.