Liftago On-Demand Transport Dataset and Market Formation Algorithm Based on Machine Learning

Jan Mrkos, Jan Drchal, Malcolm Egan, Michal JakobarXiv 2016 · arXiv

Introduction

In the competitive landscape of on-demand transport services like Uber, Lyft, and the Czech-based Liftago, "market formation"—the process of matching passengers with drivers—is a critical operational challenge. Traditionally, these systems rely on manual heuristics, such as notifying all drivers within a fixed radius. However, this often leads to "notification spam," where drivers are overwhelmed by requests they are unlikely to accept, or a lack of matches for the passenger.

Researchers have developed a data-driven algorithm called SIDMAF (SImple Data-driven MArket Formation) to address these inefficiencies. By leveraging machine learning to predict the probability of a driver accepting a specific trip, SIDMAF reduces driver pings while increasing the overall success rate of matches.

Methodology

The proposed system operates in two distinct stages:

  1. The Acceptance Model (ACM): Utilizing a Random Decision Forest Ensemble (RDFE), the researchers built a binary classifier to estimate the probability that a specific driver will accept a given ride request. Unlike simple binary outcomes, this model provides a nuanced probability score for every potential match.
  2. SIDMAF Algorithm: This core algorithm uses the ACM scores to rank available drivers. Instead of a fixed geographic radius, SIDMAF selects the smallest possible subset of drivers required to meet a target cumulative probability threshold (e.g., a 99.9% chance that at least one driver will accept).

Dataset Overview

The study utilized a comprehensive real-world dataset provided by Liftago, capturing operations in Prague, Czech Republic, throughout 2015. The dataset includes:

  • 31,787 ride orders, resulting in 253,687 individual requests.
  • 390 active drivers and 5,647 active customers.
  • Features: Spatio-temporal data including pickup/drop-off coordinates, timestamps, driver IDs, and high-frequency GPS traces (sampled every 20 seconds).

Key Insights: What Drives Acceptance?

The research highlights that driver behavior is not solely dictated by proximity. While distance to the passenger is a primary factor, the model revealed that historical behavior is nearly as significant.

Feature importance for the driver acceptance model, highlighting the significance of pickup distance and historical acceptance rates.

As shown in the feature importance analysis, pickup_distance and mean_accept_rate (the driver's historical tendency to accept) are the dominant predictors. This suggests that some drivers are inherently more selective, and accounting for this "pickiness" allows the system to target notifications more effectively.

Performance Results

The researchers compared SIDMAF against Liftago’s existing expert-designed heuristics using two Key Performance Indicators: KPI1 (matchmaking success rate) and KPI2 (average number of drivers notified per request).

Comparison of SIDMAF and Liftago heuristics on success rate (KPI1) and notification count (KPI2).

As illustrated in the results table, SIDMAF significantly outperformed the baseline. It achieved a success ratio of 0.867 (nearly double the heuristic's 0.476) while simultaneously reducing the average number of notifications from 7.67 to 3.56. This demonstrates an "efficiency paradox": by sending fewer, better-targeted notifications, the system achieves higher overall success.

Limitations and Future Directions

While the simulation results are promising, the current model has limitations. It assumes that predicted acceptance leads to a match without fully modeling the price bidding process or the final customer choice between multiple offers. Future work will involve real-world A/B testing to observe how human drivers respond to these optimized notification patterns in real-time and how psychological factors like "driver fatigue" might influence the system's long-term performance.

Summary updated

BibTeX
@misc{mrkos2016liftago,
  author = {Jan Mrkos and Jan Drchal and Malcolm Egan and Michal Jakob},
  title = {{Liftago On-Demand Transport Dataset and Market Formation Algorithm Based on Machine Learning}},
  year = {2016},
  eprint = {1608.02858},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1608.02858},
}
Figure