A Rules-based Mode Choice Model using CHAID Decision Trees and Dynamic Transit Accessibility

Abstract

Transportation mode choice models typically represent user decision making using utility-based mode choice models. However, utility models assume that users make compensatory trade-offs between decision variables to maximize their expected utility. The decision process literature raises alternative, non-compensatory theories that suggest people employ simpler, cognitively frugal heuristics in their decision making. Non-compensatory models, including decision tree classifiers, present an opportunity to test the effects of transit accessibility variables on mode choices and improve descriptions of mode choice behaviour. Dynamic forms of transit accessibility, which measure variations in transit service over time, may better capture heuristic perceptions of transit service quality.

This research addresses the need to understand how dynamic transit accessibility (DTA) impacts mode choices, without compensatory decision process assumptions. First, this research develops DTA measures for the Region of Waterloo using General Transit Feed Specification (GTFS) transit schedule information to calculate travel impedance matrices for departures at every 5-minute interval of the day. Zonal mode shares are regressed against alternative DTA measures to analyze the effects of different destination types, time periods of aggregation, and statistical parameters of transit accessibility (i.e., mean and distribution over time). Based on the aggregate mode share predictive performance, a DTA metric is selected for analysis within a binary (transit and not transit) disaggregate mode choice model. Second, this research uses trip diary data to train and score a Chi-squared Automatic Interaction Detection (CHAID) decision tree classifier to represent and predict rules-based mode choice processes. Finally, the selected DTA metric is merged with the trip diary data and applied in another decision tree for comparison. The comparison between the two rules-based mode choice models is based on overall model accuracy, class recall, precision, and interpretability.

Results from the decision tree classifier reveal that users apply heuristics in their transportation mode decision making, including lexicographic and aspiration-level based decision rules. User choices depend primarily on transit pass ownership, and non-transit-pass users consider the trip’s distance thereafter. Including DTA as an independent variable in the decision tree has a small but statistically significant effect: users only seem to consider DTA, a generalized location-based measure, if they do not own a transit pass and only after considering the trip-specific distance. Overall, the rules-based mode choice models report accuracies of roughly 84%; however, low precision in the transit predictions (i.e., many false positives) result in an overestimation of regional transit shares.