Applying Reinforcement Learning and Graph Embedding for Intelligent User Operations in Didi Ride‑Sharing

Didi Ride‑Sharing has built a comprehensive user‑operation platform consisting of a traffic distribution system, tagging, profiling, target audience targeting, marketing strategy, and reach‑optimization modules to support user growth.

Traditional manual and supervised‑learning approaches suffer from coarse segmentation, experience‑driven decisions, and message fatigue, prompting the adoption of reinforcement learning (RL) to model the interaction between the platform (agent) and users (environment).

The RL formulation defines State (user features and predictions), Action (coupon issuance, message push, or no‑action), and Reward (user responses such as fuel purchase or coupon usage). A Double Deep Q‑Network (Double DQN) is employed to mitigate Q‑value overestimation, with periodic target‑network updates and negative‑sampling for class imbalance.

To personalize message delivery, graph‑embedding methods—LINE for homogeneous graphs, TransE for heterogeneous graphs, and GraphSAGE for combined structural and feature information—are applied, enabling fine‑grained matching of user demand profiles with platform supply.

Experimental results show that the RL‑driven approach consistently outperforms the control group in ROI, achieving roughly 8% higher acquisition and recall rates while halving costs. Graph‑embedding‑based personalized messaging further improves open rates by 7‑11% and conversion rates by 10‑16% compared to baseline methods.

Overall, the integration of reinforcement learning and graph embedding has enabled end‑to‑end intelligent marketing at scale, delivering higher user engagement, lower acquisition costs, and a robust, data‑driven operation framework.