Interpretable Machine Learning for Taxi Tipping Behavior Analysis in Urban Mobility Systems

Abstract

Taxi tipping behavior provides a useful case for studying passenger payment decisions in urban mobility services. Using New York City Yellow Taxi trip records from January to March 2025, this study examines two related tasks: predicting tipping occurrence and estimating tip amount for trips with positive tips. The input feature system includes trip and payment variables, temporal variables, spatial and operational variables, and fare-structure variables. Logistic regression and ordinary least squares regression are used as statistical baselines, while GBDT, XGBoost, LightGBM, CatBoost, and FT-Transformer are compared as tabular prediction models. Models trained on January data are evaluated on a January holdout set and on February and March external test sets. The results show that FT-Transformer achieves the highest AUC for tipping occurrence prediction, while CatBoost achieves the strongest January holdout performance for conditional tip amount regression and remains close to the best external-test results. Significance tests support the classification gains of CatBoost over the logistic regression baseline and the GBDT benchmark, whereas regression gains are more limited. Sensitivity analysis shows that credit-card payment is a strong predictive marker for tipping occurrence but has little additional value for tip amount once tipping has occurred. SHAP analysis further indicates that payment context is more important for tipping occurrence, while trip scale and fare-related variables are more relevant to payment intensity. These findings support a dual-layer interpretation of taxi tipping behavior.