Análise comparativa de modelos de aprendizado de máquina para classificação de corridas rentáveis em plataformas de transporte.

Abstract

The growth of digital urban transportation platforms has increased the amount of data generated about trips, fares, distances, schedules and operational characteristics. Although these data are widely available, drivers and other actors involved in platform-based transportation still face difficulties in converting information into effective decision support, especially when the purpose is to evaluate whether a ride tends to be profitable or not. In this context, this study aims to analyze and compare machine learning (ML) models for classifying profitable and non-profitable rides using two distinct datasets: an Uber/NCR-type ride-booking dataset and a public New York City Yellow Taxi dataset. The methodology was structured according to the CRISP-DM process, covering problem understanding, data understanding, data preparation, modeling and evaluation. Derived variables for estimated cost, estimated profit and profitability were created, together with a methodological discussion of data leakage risks, especially regarding financial variables. Sensitivity analyses were also included to assess the impact of variables directly related to the profitability calculation. The evaluated models included Logistic Regression, Decision Tree, Random Forest, Extra Trees, Gradient Boosting, AdaBoost, baseline models and a simple TensorFlow/Keras neural network as a complementary experiment. For the Uber/NCR dataset, Gradient Boosting achieved an approximate F1-score of 0.99, a result strongly associated with the presence of booking_value and ride_distance. For the Yellow Taxi NYC dataset, the complementary neural network and Random Forest achieved F1-scores close to 0.93, decreasing to 0.7735 when duration_min and avg_speed_mph were removed. The results indicate that ML models can support the identification of patterns associated with ride profitability, provided that dataset limitations, the estimated nature of the target variable and variable availability at decision time are explicitly considered.