Maximal Compatibility Matching for Preference-Aware Ride-Hailing Systems

TL;DR

Maximal Compatibility Matching (MCM) addresses the gap in ride-hailing matchings by incorporating passenger comfort and driver behavior into the assignment process. It learns personalized passenger comfort zones with gradient-boosted decision trees and defines driver operating envelopes as quantile-based axis-aligned hyperrectangles, then computes compatibility as the volume of overlap $A_{ij}$ between these regions. A tunable utility-based optimization maximizes $\sum_i\sum_j (\alpha A_{ij} - (1-\alpha) D_{ij}) X_{ij}$, enabling a continuum from purely distance-based to fully preference-driven matching. Validation in a Unity-based simulator with real-time feedback demonstrates improved personalization and social acceptability without sacrificing operational performance, highlighting the practical potential of MCM for scalable, preference-aware ride-hailing systems.

Abstract

This paper presents the Maximal Compatibility Matching (MCM) framework, a novel assignment strategy for ride-hailing systems that explicitly incorporates passenger comfort into the matching process. Traditional assignment methods prioritize spatial efficiency, but often overlook behavioral alignment between passengers and drivers, which can significantly impact user satisfaction. MCM addresses this gap by learning personalized passenger comfort zones using gradient-boosted decision tree classifiers trained on labeled ride data, and by modeling driver behavior through empirical operating profiles constructed from time-series driving features. Compatibility between a passenger and a driver is computed as the closed-form volume of intersection between their respective feature-space regions. These compatibility scores are integrated into a utility-based matching algorithm that balances comfort and proximity through a tunable trade-off parameter. We validate the framework using a Unity-based driving simulator with real-time passenger feedback, demonstrating that MCM enables more personalized and socially acceptable matchings while maintaining high levels of operational performance.

Figures (5)

• Figure 1: Geometric illustration of compatibility in a two-dimensional feature space. The passenger's comfort zone $G_i(\mathbf{x})$ is depicted as a downward staircase-shaped region, while the driver's operating zone $f_j$ is shown as a red rectangle. The shaded intersection area $A_{ij}$, highlighted with a hatched pattern, represents the region of mutual compatibility between passenger $i$ and driver $j$.
• Figure 2: Top 5 features based on gain values computed by the gradient-boosted decision tree model. The gain metric reflects each feature's relative contribution to improving model performance, with higher values indicating greater predictive importance.
• Figure 3: Comparison of passenger-driver matching under three assignment strategies: distance-based matching (left), which minimizes total travel distance; comfort-based matching (center), which maximizes affinity based on passenger preferences; and maximal compatibility matching (right), which optimizes a utility function combining comfort and distance with trade-off parameter $\alpha = 0.5$. The utility-based approach captures a balance between proximity and personalized satisfaction.
• Figure 4: Visualization of passenger and driver compatibility regions in a 2D feature space. The red region with decision boundaries represents the passenger's comfort zone learned using a gradient-boosted decision tree classifier. The blue rectangle denotes the operating zone of a candidate driver. Blue scatter points correspond to the driver’s historical ride samples, used to estimate their typical operating behavior. The area of the driver operating zone that falls within the passenger comfort zone represents the passenger-driver compatibility.
• Figure 5: Jaccard similarity between the utility-based matching and the baseline matchings (distance-based and comfort-based) as a function of the trade-off parameter $\alpha$. As $\alpha$ increases, the utility-based matching becomes more similar to the comfort-based matching and less similar to the distance-based one, reflecting the shift in optimization priority from spatial proximity to user preference.