Predicting Trip Cancellations and No-Shows in Paratransit Operations

Abstract

The productivity of paratransit systems could be improved if transit agencies had the tools to accurately predict which trip reservations are likely to result in trips. A potentially useful approach to this prediction task is the use of machine learning algorithms, which are routinely applied in, for example, the airline and hotel industries to make predictions on reservation outcomes. In this study, the application of machine learning (ML) algorithms is examined for two prediction problems that are of interest to paratransit operations. In the first problem the operator is only concerned with predicting which reservations will result in trips and which ones will not, while in the second prediction problem the operator is interested in more than two reservation outcomes. Logistic regression, random forest, gradient boosting, and extreme gradient boosting were the main machine learning algorithms applied in this study. In addition, a clustering-based approach was developed to assign outcome probabilities to trip reservations. Using trip reservation data provided by the Metropolitan Bus Authority of Puerto Rico, tests were conducted to examine the predictive accuracy of the selected algorithms. The gradient boosting and extreme gradient boosting algorithms were the best performing methods in the classification tests. In addition, to illustrate an application of the algorithms, demand forecasting models were generated and shown to be a promising approach for predicting daily trips in paratransit systems. The best performing method in this exercise was a regression model that optimally combined the demand predictions generated by the machine learning algorithms considered in this study.