Precise Marketing Algorithms and Practices at Hello Mobility

Background and Value

Hello Mobility has expanded from a ride‑hailing service to a broader service‑e‑commerce platform that includes local life, hotels, and e‑bikes. Precise marketing is essential to drive user growth across these new businesses by targeting users throughout their lifecycle.

Marketing Scenarios and Process

The marketing flow follows a classic who‑what‑how sequence: identify target groups (who), decide the content to deliver (what), and choose the delivery method (how). Three primary scenarios are addressed: acquisition (拉新), activation/retention (活跃), and re‑engagement (挽留).

Business Pain Points

Low efficiency in finding precise user groups, requiring extensive manual testing.

Low ROI due to high marketing costs and limited returns.

Limited algorithm coverage and low integration efficiency.

Absence of a systematic closed‑loop for post‑marketing analysis.

Project Value

Efficiency: operators can directly use algorithm‑generated user groups, reducing testing effort and increasing CTR by ~20%.

Revenue: precise marketing is expected to boost order volume by ~20%.

Precise Marketing Framework

The framework consists of three modules:

Feature Processing : offline feature tables stored in Hive/ES and real‑time features computed by Flink and cached in Redis.

Precise Marketing : includes algorithms, a user‑analysis platform, and a delivery platform. Algorithms comprise industry‑wide LookAlike models (Pu‑Learning) and intelligent scaling via graph‑based embeddings.

Algorithm Scenarios : support acquisition, activation, and churn reduction through banner, in‑app messages, or push notifications.

Algorithm Capabilities

1. Pu‑Learning LookAlike Modeling : expands seed users to similar audiences using supervised, semi‑supervised, or unsupervised learning. Challenges such as sparse new‑business features are mitigated with two‑wheel features.

2. TSA Semi‑Supervised Model : first identifies reliable negative samples among unlabeled data, then trains a supervised model on the combined set.

3. Graph Embedding for User Expansion : builds a spatio‑temporal graph where nodes are users and edges represent co‑occurrence in the same location/time. User embeddings are generated via DeepWalk, then refined with EGES to incorporate side‑information and weighted node types.

Data‑driven similarity search is powered by Milvus, offering near‑real‑time queries and multiple indexing options.

Results

Coverage: platform‑wide support for intelligent scaling, covering 60% of scenarios with zero additional cost.

ROI: overall ROI improvement of over 20%.

Future Directions

Improve graph construction by incorporating both public and private click behaviors.

Develop an automated threshold recommendation system to suggest optimal scaling factors based on predicted ROI.

Overall, the talk demonstrates how advanced AI techniques—semi‑supervised learning, graph embeddings, and large‑scale vector search—enable scalable, high‑ROI precise marketing at Hello Mobility.