Inferring in-home/out-of-home situations unreported in time-use surveys using supervised machine learning

Time-use surveys provide useful data for travel analyses. However, the survey on time use and leisure activities (TULA) Questionnaire A, a representative time-use survey in Japan, does not include questions related to the locations of activities, thus making it difficult to use for travel analyses. This study proposes machine-learning methods to determine the in-home/out-of-home situations of TULA Questionnaire A using TULA Questionnaire B with activity locations as the training data. Random forest performs better than logistic regression and decision trees in the inference. The activity was the most important factor in determining the in-home/out-of-home situations, followed by the accompanying person and time of day. The inferred outputs in the TULA Questionnaire A included the individual-based out-of-home rate profiles and the proportions of mobile persons from 1996 to 2016. Using these outputs, we analyzed trip misreporting in household travel surveys. Comparisons with nationwide and Tokyo person trip (PT) surveys implied soft refusals and trip misreporting in travel surveys. The comparison with the nationwide PT surveys suggested higher soft refusals on weekends than on weekdays. The comparison with the 1998, 2008, and 2018 Tokyo PT surveys implied the increased soft refusal in PT surveys, particularly among the male group aged between 20 and 39 and the female group aged between 35 and 49 during 1998–2018. These results suggest that careful handling of recent household travel survey data may be required. In addition, the proposed machine-learning-based method enables us to utilize the rich sample of Questionnaire A for activity-based travel analysis in future studies.