Interactive online machine learning approach for activity-travel survey

This article proposes a framework for an interactive activity-travel survey method, implementable on mobile devices such as smartphones. The proposed method was developed to reduce the burden (i.e., frequency of questions) on respondents in long-term behavioral surveys, without relying on external data sources. The method employs an online travel context estimation model and an online machine learning method as interactive processes. The estimation model is used for automatically estimating travel contexts during surveys, while the online machine learning method is used for dynamically updating the estimation model, based on answers from respondents. The proposed method was examined by simulations using data obtained from a past probe person survey. The results suggest that the frequency of inputs by respondents in surveys can be significantly reduced, while maintaining high accuracy of the obtained data. For example, the method successfully estimated certain types of trips (e.g., commuting) and the behaviors of certain respondents (e.g., those whose activity-travel pattern is recurrent) because of the learning process and reduced survey burden on them. Meanwhile, although the method could not always precisely estimate some other types of trips, it eventually obtained accurate results because of the interaction process. Therefore, the proposed method could be useful to reduce the burden on respondents in long-term surveys, while maintaining high data quality and capturing traveler heterogeneity.