Interpretable machine-learning models for estimating trip purpose in smart card data

Abstract

Investigating trip purposes of transit passengers is crucial in assessing current urban transportation systems and prioritising investments in the public transportation infrastructure. Smart card data provide day-to-day information on passengers’ boardings and alightings, but the lack of information on trip purposes leads to restrictions on the use of these data. This paper focuses on estimating trip purposes of transit passengers in smart card data, using a machine-learning model that is trained by household travel survey data. To accomplish this objective, a random forest model coupled with interpretable machine-learning methods – that is, feature importance, feature interactions and accumulated local effects plot is proposed. This approach can be used to estimate trip purposes and to explain the decision-making process of the models. The models include the spatiotemporal features that can be extracted from both the smart card data and the geographic information data, which can be collected sustainably and cost-effectively. The proposed model achieves an 83% overall accuracy in its estimation of the validation data. The interpretation methods show that temporal features are the dominant factors in estimating the purposes of trips, and the spatial features influence the estimates mainly through cross-effects with the temporal features.