Learning Daily Activity Sequences of Population Groups using Random Forest Theory

The choice of daily activity sequences differs between individuals based on their socio-demographic characteristics and their health and/or mobility status. The aim of this paper is to provide an improved methodology for learning and modeling the daily activity engagement patterns of individuals using a state-of-the-art machine learning algorithm. The dependencies between activity type, activity frequency, activity sequence, and socio-demographic characteristics of individuals are taken into account by employing a random forest model. In order to capture the heterogeneity and diversity among the predictor variables, we employed two different methods for split selection in the random forest algorithm: Classification and Regression Tree (CART) and curvature search. These two methods were examined under two different layer settings. In the first setting, the algorithm grows trees using all alternative predictor variables, whereas in the second setting the importance of the predictor variables is estimated and then the algorithm grows trees using only high-score predictor variables. The models were applied to time use data from the large Halifax Space-Time Activity Research (STAR) household travel diary survey. We evaluated the estimation accuracy of the proposed models using confusion matrix, transition matrix, and sequential alignment techniques. Results show that the random forest model with CART split selection using the first layer setting has the best accuracy in replicating activity agendas and activity sequences of individuals. The results of this paper are expected to be implemented within the activity-based travel demand model, Scheduler for Activities, Locations, and Travel (SALT).