When Recommender Systems Meet Fleet Management: A Practical Study on Online Driver Repositioning

The article provides a detailed reading of the WWW 2020 research paper "When Recommender Systems Meet Fleet Management: Practical Study in Online Driver Repositioning System" published by the Didi ride‑hailing team. It addresses the classic supply‑demand imbalance problem in ride‑hailing platforms, where passengers often cannot find a car while idle drivers waste time cruising.

Background : Empty‑car cruising consumes more than 50% of a taxi driver’s working time. In ride‑hailing, drivers receive real‑time location updates from both passengers and the platform, allowing the platform to intervene and guide idle drivers to areas with higher order‑receiving probability.

What is a dispatch task? A dispatch task is a real‑time instruction sent to an idle driver, recommending a specific destination (a POI within a grid cell). The driver can accept the task, navigate to the destination, and continue receiving orders en route. The task can end in four states: (1) driver ignores the task and drives away, (2) driver accepts and receives an order before reaching the destination, (3) driver accepts, reaches the destination, then receives an order within a time window, (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 and triggers compensation.

Framework Overview (three stages) :

Candidate Generation : Identify idle drivers whose idle time exceeds a threshold. For each driver, generate candidate destination grids from three sources: nearby cells, frequently visited cells in historical trajectories, and city‑wide hotspot cells. Select a POI within each grid and set an expiration time based on the estimated time of arrival (ETA). A failure‑probability prediction model filters out tasks with high failure risk, and compensation is proportional to travel distance.

Task Scoring : Use a piecewise linear function to model the relationship between response rate and supply‑demand ratio. The marginal gain of adding one idle driver to a spatio‑temporal cell is computed, yielding a score for each candidate task. The model also estimates the driver shortage (capacity gap) needed to achieve a target response rate.

Planning (Optimization) : Convert the selection of final dispatch tasks into a maximum‑benefit problem with driver‑experience constraints. The problem is transformed into a minimum‑cost flow formulation, enabling efficient computation of the optimal set of tasks that maximizes platform-wide utility while respecting driver constraints.

Experimental Results : The framework was evaluated in an online A/B test on Didi’s platform. Multiple rounds of experiments showed that the proposed system improves driver efficiency, experience, and total earnings compared with self‑driven cruising. The marginal‑gain function, minimum‑cost‑flow module, and failure‑compensation design all contributed positively. A post‑experiment survey reported a 64.6% acceptance rate for dispatch tasks and an NPS of 27.0%, indicating strong driver satisfaction.

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