Uber Order Cancellation Prediction: A Machine Learning-Based Operational Optimization Solution for Mobility Platforms

This study focuses on the order cancellation issue on the Uber platform. By analyzing 150,000 order data entries, a prediction model was built, key influencing factors were identified, and a data-driven solution was provided for the operational optimization of ride-sharing platforms. Key findings include: order distance, waiting time, and order amount are the three major factors affecting cancellation rates; the random forest model achieves a prediction accuracy of 94.97%, which can support business applications such as real-time risk intervention and regional optimization.

The core competitiveness of ride-sharing platforms lies in supply-demand matching, but the order cancellation rate is as high as 38%, leading to issues such as degraded user experience and increased driver empty mileage. Traditional operations rely on experience-based judgments, which have limitations when facing massive data and complex behavior patterns. Machine learning can learn cancellation patterns from historical data, provide quantitative basis for operational decisions, and solve core problems like supply-demand imbalance.

This study uses 150,000 Uber order data entries (including 21 original features), with order results categorized into completed, driver-canceled, user-canceled, etc. Feature engineering includes missing value handling (median/mode imputation), time feature conversion, binary classification target construction (0 = completed, 1 = canceled/uncompleted), and category encoding. Model comparison experiments cover logistic regression, random forest, XGBoost, etc., with ensemble learning methods (random forest with 94.97% accuracy) performing the best.

Data exploration findings: 1. Geographically, regions like Vinobapuri have cancellation rates of 40-45% (supply-demand imbalance); 2. Go Sedan has a slightly higher cancellation rate (higher user expectations); 3. Time distribution is stable but slightly higher in the early morning/afternoon; 4. Users using digital payments have a slightly higher completion rate. Feature importance shows: order distance (30.59%), average CTAT (24.59%), and order amount (20.49%) are the three key factors. Short-distance orders are more likely to be canceled (low user tolerance, less driver attraction), and long waiting times significantly increase the probability of cancellation.

Based on the model, the following can be implemented: 1. Real-time risk scoring: prioritize dispatching for high-risk orders, dynamic pricing, proactive communication, or recommend alternative solutions; 2. Regional optimization: increase driver incentives in high cancellation rate areas, optimize pricing and matching algorithms; 3. Driver-side optimization: reduce empty mileage, provide demand hot spot information, and incentivize acceptance of short orders. These measures can improve order completion rates and optimize user and driver experiences.

The order cancellation prediction model built in this study has an accuracy rate of 94.97%. Core contributions include identifying key influencing factors, verifying the effectiveness of ensemble learning, and providing directions for operational optimization. Limitations include not considering external factors like weather and potential lag due to historical data. In the future, integrating real-time data, exploring deep learning, conducting causal modeling and A/B testing can further enhance the model's value.