A First Look at Predictability and Explainability of Pre-request Passenger Waiting Time in Ridesharing Systems

Jie Wang, Guang Wang

Introduction

In modern ridesharing ecosystems like Uber, Lyft, and DiDi, passenger waiting time is a critical determinant of user satisfaction and platform loyalty. While predicting waiting times after a driver has been assigned (post-request) is relatively straightforward due to known driver locations and traffic conditions, predicting waiting times before a request is made (pre-request) presents a significant challenge. At the pre-request stage, the system must estimate wait times without a specific driver-passenger match, relying instead on aggregate supply-demand dynamics and historical patterns.

This paper introduces FiXGBoost, a novel predictive framework designed to bridge the accuracy gap between pre-request and post-request estimations. By leveraging massive-scale data and sophisticated feature interaction techniques, the researchers demonstrate that pre-request waiting times can be predicted with high precision and interpretability.

Data and Preliminary Analysis

The study is grounded in a massive real-world dataset from Shenzhen, China, encompassing over 30 million trip records from 2019. This dataset includes high-resolution spatiotemporal data, including GPS coordinates, timestamps, and trip characteristics.

To understand the underlying mechanics of waiting times, the authors performed a preliminary analysis using metrics such as:

  • Deficiency: The absolute difference between demand (orders) and supply (available vehicles).
  • Availability: The ratio of supply to demand.
  • Fano Factor: A measure used to quantify the dispersion and predictability of waiting times across different time windows.

One key discovery was the "Friday Surge." Unlike typical weekdays where waiting times peak during morning and evening commutes, Fridays exhibit a unique, sustained increase in waiting times starting in the late afternoon and peaking between 18:00 and 19:00. This is attributed to a cumulative shortage of vehicles as high demand persists throughout the day.

Methodology: The FiXGBoost Model

The core contribution of this work is FiXGBoost, which enhances the Extreme Gradient Boosting (XGBoost) algorithm with a latent feature interaction layer. The methodology consists of three primary stages:

1. Multi-Dimensional Feature Engineering

Features are categorized into four distinct buckets:

  • Spatiotemporal: Region IDs and time-based indicators (e.g., rush hour).
  • Demand-Supply: Real-time order volume and vehicle density.
  • Contextual: Weather conditions (e.g., rain, temperature).
  • Trip Characteristics: Estimated trip distance and duration.

2. Latent Feature Interaction

Inspired by Factorization Machines (FM), the authors implemented a collaborative filtering approach to capture hidden dependencies. By treating geographical regions as "users" and time slots as "items," the model learns latent representations that reflect regional popularity and temporal driving patterns that are not explicitly captured by raw data.

3. Gradient Boosted Ensemble

These engineered and latent features are fed into an XGBoost model. XGBoost's ability to handle non-linear relationships and its robustness against outliers make it ideal for the noisy environment of urban mobility data.

Experimental Results

FiXGBoost was evaluated against several baseline models, including Gradient Boosting Decision Trees (GBDT) and standard XGBoost. The results demonstrate a significant leap in performance:

  • Accuracy: FiXGBoost achieved a Mean Absolute Error (MAE) of 112 seconds, a 28.2% improvement over standard GBDT (165s) and a substantial gain over standard XGBoost (156s).
  • Reliability: Approximately 65% of all pre-request predictions fell within a 2-minute margin of the actual waiting time.
  • Benchmarking: Remarkably, the pre-request accuracy of FiXGBoost (112s MAE) outperformed the post-request accuracy of simpler models like Linear Regression (140s MAE). This suggests that sophisticated feature interaction can effectively compensate for the lack of specific driver-location data.

Explainability and Insights

A major focus of the research was ensuring the model remained a "white box." Feature importance analysis revealed that for pre-request predictions, temporal features (such as isRushHour) and weather conditions are far more influential than they are in post-request scenarios. In post-request models, the physical distance of the assigned driver (pickDistance) typically dominates; however, in the pre-request phase, the system's state (congestion and weather) serves as the primary proxy for driver availability.

Conclusion and Future Work

The FiXGBoost model provides a robust solution for the "blind" prediction of passenger waiting times, offering platforms a way to set more accurate user expectations. While the current model relies on gradient-boosted trees, the authors suggest that future iterations could incorporate Transformers or Large Language Models (LLMs) to better model complex, long-range spatiotemporal dependencies. Additionally, testing the model's generalizability across different urban environments remains a priority for future research.

Summary updated

BibTeX
@inproceedings{wang2025first,
  author = {Jie Wang and Guang Wang},
  title = {{A First Look at Predictability and Explainability of Pre-request Passenger Waiting Time in Ridesharing Systems}},
  booktitle = {SIGSPATIAL/GIS},
  year = {2025},
  doi = {10.1145/3748636.3763224},
}
Figure