Can AI-Driven Driver Repositioning Solve Ride-Hailing Supply-Demand Gaps?

1. Background

With the rise of mobile internet, ride‑hailing has become a common travel option, but platforms often face supply‑demand imbalance: passengers cannot get rides while drivers idle. Empty cruising can occupy more than half of a driver’s working time, wasting capacity.

Scheduling idle online drivers to high‑demand areas is a natural solution. The following is a detailed interpretation of the WWW 2020 paper “When Recommender Systems Meet Fleet Management: Practical Study in Online Driver Repositioning System”.

2. What Is a “Dispatch Task”

In practice, idle drivers rely on personal experience to choose cruising destinations, which can be sub‑optimal. The paper proposes sending real‑time dispatch tasks from the platform to guide drivers to the best cruising locations.

A dispatch task appears as a card in the driver’s app; clicking the navigation button opens a route to the dispatch destination, and the driver can still receive orders en route.

Each dispatch task consists of three steps: (1) inform the driver of a clear destination; (2) track the driver’s behavior to determine success; (3) provide compensation if the task fails.

3. Task End States

Four possible end states are defined (see Figure 3):

State 1: driver ignores the task and drives away.

State 2: driver accepts and heads to the destination, receiving an order en route.

State 3: driver accepts, reaches the destination, and gets an order within a time window.

State 4: driver accepts, reaches the destination, but receives no order within the window.

States 2 and 3 are considered successful; state 4 is a failure, triggering compensation to maintain driver trust.

4. Method

The proposed framework has three stages.

Stage 1: Generate Candidate Dispatch Tasks

A task includes driver, destination, expiration time, and compensation amount. Candidates are selected among drivers idle longer than a threshold. For each driver, three types of candidate destinations are considered: nearby grid cells, frequently visited cells from historical trajectories, and city‑wide hotspots. A POI within each cell is chosen, and the ETA determines the task’s expiration.

A failure‑probability prediction model filters out tasks with high failure risk; compensation is proportional to travel distance.

Stage 2: Score Tasks

Each candidate is scored by estimating the marginal benefit of adding one idle driver to a spatio‑temporal cell, using a piecewise linear function that relates response rate to supply‑demand ratio. The score reflects the expected increase in platform efficiency.

The model also derives the required number of drivers to meet a target response rate (driver shortage).

Stage 3: Planning

After scoring, a planning algorithm selects a subset of tasks that maximizes global platform benefit while respecting driver‑experience constraints. The optimization is transformed into a minimum‑cost flow problem (see Figure 5).

5. Experimental Results

Online A/B experiments on the Didi platform show that the framework improves driver efficiency, experience, and total income compared with self‑driven cruising. Key components such as marginal‑gain function, min‑cost flow, and failure compensation all contribute positive gains.

Post‑experiment surveys indicate 64.6 % of drivers would accept a future dispatch task, with an NPS of 27.0 %.

6. Conclusion and Future Work

The framework effectively balances supply and demand, enhances driver efficiency, and has been deployed to serve millions of drivers daily. Future directions include reinforcement‑learning‑based end‑to‑end solutions and direct route optimization using road‑network data.