How Optimizing Information Disclosure Boosts Efficiency in Mobility‑On‑Demand Platforms

Research Background

Hulala currently uses a global broadcast method for order assignment: within a specified time/distance range, all orders are pushed to all drivers, who then browse the list and select orders that meet their expectations. When multiple drivers respond to the same order, the algorithm selects the most suitable driver based on platform efficiency, user experience, and driver experience.

In this context, order push to drivers is considered a form of information disclosure, and the paper proposes an algorithmic improvement at the broadcast stage.

Drivers have varying states (spatial‑temporal position, supply‑demand balance, accumulated work hours, earnings, etc.) and consequently different income expectations. Example scenarios include:

Supply‑constrained: far more orders than available drivers; drivers tend to accept higher‑priced orders.

Demand‑constrained: far fewer orders than drivers; drivers prefer non‑idle trips and are less price‑sensitive.

Unrestricted broadcast: drivers see city‑wide or nationwide order lists, leading to decision difficulty when supply exceeds demand.

Strongly restricted broadcast: drivers see at most one order, causing anxiety when no suitable order appears.

Accurately estimating driver decision behavior and optimizing information disclosure (which orders are shown to which drivers) aim to improve driver experience, user experience, and overall platform efficiency.

Research Value

Unlike traditional recommendation systems (e.g., Douyin, Taobao) where items are virtually unlimited resources, the Hulala scenario imposes resource constraints: orders are limited by time and space, and drivers can view only a finite set of orders on their screens. This class of resource‑constrained problems lacks mature solutions.

The crowdsourced dispatch model introduces additional complexity and driver competition. For m drivers and n orders, traditional dispatch solves a 1‑vs‑1 matching problem, whereas broadcast addresses an n‑vs‑m information disclosure problem, leading to exponential growth in search space as problem size increases.

Algorithm Framework

1. Prediction

Based on conditional dependencies in the business scenario, the authors extend a Multinormal Logit Model to estimate driver decisions in two steps:

Driver decides whether to accept an order or not.

If accepting, driver selects a specific order from the candidate list.

Maximum likelihood estimation on historical behavior data is used to fit model parameters.

The model’s rationality is analyzed in two levels: (1) drivers may reject the current list and wait for a more satisfactory one; (2) once satisfied, they choose the most preferred order, and reducing disclosure of a particular order can increase the acceptance probability of other orders while also raising the probability of rejecting the list.

2. Planning

Objective Function

Problem Solving

i. Global information disclosure

ii. Local information disclosure

iii. Original edge‑cut algorithm

iv. Minimal Loss Edge Cutting algorithm

Overall Algorithm

Experiments

Offline: Historical data from three cities were used to train the driver decision estimation model.

Online: AB tests were conducted across several time slots in three cities, comparing:

A – Hulala’s existing global information disclosure method.

B – The proposed MLEC (Maximum Likelihood Edge Cutting) algorithm framework based on driver decision predictions.

Quantitative results show a significant increase in overall response rate and driver utilization when using the new framework.

Qualitative results indicate that darker colors (representing more severe order‑response issues) are mitigated by the proposed framework, alleviating local supply‑demand imbalances.

The paper “Improving the Information Disclosure in Mobility‑on‑Demand Systems” presents a framework that optimizes information disclosure through modeling and solving, achieving business metric improvements. While demonstrated on Hulala’s logistics platform, the approach is also applicable to other resource‑constrained recommendation systems.