Governance Algorithms for O2O Ride-Hailing Platforms: Challenges, Framework, and Model Exploration

Since 2013, O2O platforms such as food delivery, ride‑hailing, and real‑estate have dramatically changed social operation patterns. Compared with the previous generation of Internet companies, the new generation faces offline interactions between drivers and passengers, bringing new challenges in experience and governance. Didi has built a powerful governance‑algorithm system focused on improving driver‑passenger experience.

Business Background : The governance problem aims to reduce and resolve various trip disputes caused by platform issues, driver‑passenger expectations, or personal problems, including cancellation disputes, fee anomalies, and service problems. Governance is divided into order‑level and person‑level dimensions. Order‑level governance covers the entire order lifecycle (pre‑incident, early incident, incident, post‑incident). Person‑level governance focuses on comprehensive driver and passenger management (service scores, education, control).

Algorithm Challenges :

High business complexity: dozens of categories and hundreds of scenarios require maintaining dozens of models.

Scarce high‑quality samples: Limited labeled data per scenario; need to fuse offline annotation with online responsibility data.

Strong interpretability requirements: Responsibility decisions directly affect experience, but machine‑learning models are inherently correlation‑based.

Multimodal features: Besides order, spatio‑temporal, and driver‑passenger statistics, communication texts, complaint texts, and vehicle‑camera features must be leveraged efficiently.

Governance Algorithm Framework (illustrated in the original diagrams) consists of three layers:

Business Layer : Designs product solutions for dispute prevention and handling, including pre‑incident mitigation, real‑time intelligent acceptance, and post‑incident compensation (service‑score deduction, refunds, fines).

System Layer : Provides a mature online service engine with a visual strategy‑flow engine, model engine supporting LR, XGB, DNN, and a rule engine with DSL parsing. It also includes a strategy capability library (text‑algorithm toolbox, offline data warehouse) and an annotation workbench that combines online and offline labeling.

Model Layer : Explores large‑scale business models for anomaly detection, small‑sample learning, multi‑task learning, and meta‑capability modeling to improve both performance and interpretability.

Model Exploration :

Small‑sample learning: Self‑learning algorithms expand limited high‑quality samples, achieving a 0.4 pp AUC increase and 2 pp recall gain.

Transfer learning across multiple ride‑hailing categories (fast‑car, pooled‑car, premium, enterprise) to reduce data collection and labeling costs.

Multi‑task learning (e.g., ESMM, ESM2, MMOE) leverages related supervision signals, improving accuracy (+0.2‑0.6 pp) and recall (+1.1‑4.2 pp) over single‑task baselines.

Feature Exploration progresses through three stages:

Initial stage : Basic business, spatio‑temporal, and driver‑passenger statistical features (≈300 + 1,000 dimensions).

Large‑scale multimodal features : Utilizes Didi’s in‑vehicle camera (JueShi) covering >50 % of orders and continuous audio recordings, enabling end‑to‑end multimodal training or two‑stage models.

Streaming feature exploration : Extracts semantic features from real‑time streams (trajectory, audio, video). Supervised sub‑network embedding with LSTM variants (Bi‑LSTM > Vanilla‑LSTM > Stacked‑LSTM) shows the best AUC.

The paper concludes that governance algorithms constitute a new direction driven by O2O platforms, delivering significant business gains in NPS and CPO metrics. Ongoing challenges remain in sample acquisition, model robustness, and feature richness, and the team will continue to explore self‑learning, multi‑task learning, and model interpretability.