A Multi-class Ride-hailing Service Subsidy System Utilizing Deep Causal Networks

Zhe Yu, Chi Xia, Shaosheng Cao, Lin Zhou

Abstract

In the ride-hailing industry, subsidies are predominantly employed to incentivize consumers to place more orders, thereby fostering market growth. Causal inference techniques are employed to estimate the consumer elasticity with different subsidy levels. However, the presence of confounding effects poses challenges in achieving an unbiased estimate of the uplift effect. We introduce a consumer subsidizing system to capture relationships between subsidy propensity and the treatment effect, which proves effective while maintaining a lightweight online environment.

Introduction

In the competitive ride-hailing landscape, platforms must strategically distribute subsidies across various service tiers (e.g., "Economy" vs. "Luxury") to maximize growth within a fixed budget. The core challenge is estimating "uplift"—the incremental change in user behavior specifically caused by a subsidy.

Traditional models often suffer from confounding bias because historical subsidies are typically targeted at low-activity users. Without proper causal modeling, a system might incorrectly conclude that subsidies decrease booking rates simply because they are correlated with lower-activity segments. Furthermore, most existing uplift models are limited to binary scenarios, failing to address the multi-dimensional nature of real-world service classes and varying subsidy amounts.

Methodology: The MulTeNet Architecture

To address these challenges, we introduce the Multi-class Treatment Generalized Propensity Score Network (MulTeNet). As illustrated in Figure 1, the architecture utilizes a shared feature network that branches into three specialized components:

  1. Generalized Propensity Score Net: This estimates the probability of a user receiving a specific subsidy level, helping the model account for selection bias in the training data.
  2. $Y_0$ Net: This predicts the baseline conversion rate—the probability of a user booking if they receive zero subsidy.
  3. Monotone Net: This component estimates the uplift effect. By employing a softplus activation function, it enforces monotonicity, ensuring the model understands that increasing a subsidy should never decrease the probability of a booking.
MulTeNet Architecture

To bridge the gap between prediction and action, the system employs a Model Predictive Control (MPC) framework. This allows for daily adjustments to subsidy allocation, ensuring the platform remains within budget while accounting for real-world randomness. To maintain efficiency in production, complex optimizations are performed offline to create a lookup dictionary for the online server.

Experimental Results

The MulTeNet system was evaluated using a massive real-world dataset of approximately 2 million queries from a major Chinese ride-hailing platform. In offline evaluations, MulTeNet achieved superior performance in AUUC (Area Under the Uplift Curve) and the QINI coefficient compared to baselines like DragonNet, indicating a higher precision in identifying "persuadable" users.

In a live A/B experiment, MulTeNet demonstrated significant real-world impact. As detailed in the performance comparison below, the system achieved a Return on Investment (ROI) of 1.05, representing a 34% improvement over the best-performing baseline. Total orders increased by 5.47%, and revenue grew by 2.98%.

Real-world Performance Comparison

Discussion and Future Work

While MulTeNet showed a slight trade-off in standard AUC (raw prediction accuracy), this is expected when regularizing for causal effects rather than simple correlations. The model prioritizes decision-making utility over pure forecasting.

Future research will focus on expanding this framework to a two-sided marketplace, incorporating driver-side incentives alongside consumer subsidies. Additionally, investigating the long-term effects of subsidies on user habits remains a key area for exploration.

Summary updated

Abstract

In the ride-hailing industry, subsidies are predominantly employed to incentivize consumers to place more orders, thereby fostering market growth. Causal inference techniques are employed to estimate the consumer elasticity with different subsidy levels. However, the presence of confounding effects poses challenges in achieving an unbiased estimate of the uplift effect. We introduce a consumer subsidizing system to capture relationships between subsidy propensity and the treatment effect, which proves effective while maintaining a lightweight online environment.
BibTeX
@misc{yu2024multi,
  author = {Zhe Yu and Chi Xia and Shaosheng Cao and Lin Zhou},
  title = {{A Multi-class Ride-hailing Service Subsidy System Utilizing Deep Causal Networks}},
  year = {2024},
  eprint = {2408.02065},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.02065},
}
Figure