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

1. Business Background

Since 2013, O2O platforms such as food delivery, ride‑hailing and real‑estate have dramatically changed social operation, introducing new offline interaction challenges for user experience and governance. Didi has built a powerful governance algorithm system focused on improving driver‑passenger experience.

The governance problem aims to reduce and resolve various trip‑related disputes caused by platform issues, expectation gaps, or personal problems, including cancellation disputes, fee anomalies and service issues.

Solutions are divided into two dimensions: order‑level governance (covering the whole order lifecycle) and person‑level governance (covering driver and passenger lifecycle).

2. Governance Algorithm Challenges

High business complexity: dozens of categories and dozens of scenarios require maintaining up to a hundred models.

Technical challenges include:

Scarcity of high‑quality labeled samples across many categories and strategy nodes.

Need for high interpretability because judgment results directly affect driver‑passenger experience.

Incorporating multimodal features such as communication texts, complaint texts and vehicle‑camera data.

3. Governance Algorithm Framework

Business Layer : three sub‑solutions for fee‑dispute handling – pre‑dispute prevention, real‑time intervention during dispute, and post‑dispute control and compensation.

System Layer :

Online service engine with visual strategy workflow configuration.

Model engine supporting LR, XGB, DNN online deployment and prediction.

Rule engine parsing DSL‑based rules to lower iteration cost.

Strategy basic‑ability library with text‑algorithm toolkits and an offline data warehouse.

Annotation platform combining offline sampling and online labeling, with T+1 model update for automatic model refresh.

4. Model Exploration

4.1 Small‑Sample Learning : Self‑learning algorithm expands limited high‑quality samples, improving AUC by 0.4 pp and recall by 2 pp.

4.2 Multi‑Task Learning : Uses ESMM structure, feature discretization with embedding, and multi‑model fusion of ASR‑derived navigation semantics; achieves +0.6 pp accuracy and +4.2 pp recall over baseline XGB, and further gains over hard‑shared MTL.

4.3 Modeling Target Evolution : Shifts from outcome modeling to process (meta‑ability) modeling, enabling decomposition of raw features → meta‑abilities → root causes → final responsibility, thus improving interpretability.

5. Feature Exploration

5.1 Initial Stage : Business basic features, spatio‑temporal features, and driver‑passenger statistical features (≈300 + 1000 dimensions).

5.2 Large‑Scale Multimodal Features : Vehicle‑camera data (over 50 % of orders) and full‑trip audio recordings provide rich multimodal signals; applied via end‑to‑end joint training or two‑stage models.

5.3 Streaming Feature Exploration : Real‑time trajectory, audio and video streams are transformed into semantic features; supervised sub‑network embedding with LSTM variants (Bi‑LSTM > Vanilla‑LSTM > Stacked‑LSTM) shows best performance.

6. Summary

Governance algorithms constitute a new AI‑driven field emerging with O2O platforms. Didi’s work has delivered significant business improvements in NPS and CPO metrics and built a deep technical stack. Future work will continue to tackle challenges in sample efficiency, model robustness, and feature richness.