Precise Marketing Algorithms and Practices at Hello Mobility

Guest: Yu Liping, Algorithm Engineer at Hello Mobility. Editor: Gao Chao, Beijing Xinyi Hua Technology. Platform: DataFunTalk.

Overview: The talk shares the algorithm and practice of precise marketing scenarios at Hello Mobility, covering background and value, framework, algorithm capabilities, business pain points, project value, future directions, and technical details.

1. Background and Value of Precise Marketing

Hello Mobility is evolving from a ride‑hailing service to a service e‑commerce platform, including local life, hotels, and e‑bikes. Precise marketing is needed to drive user growth across these new businesses by targeting users throughout their lifecycle, aiming to improve the north‑star growth metric.

1.1 Business Background

The company needs to acquire new users (拉新), activate existing users (活跃), and retain churned users (挽留) through targeted marketing campaigns.

1.2 Marketing Scenarios and Process

Who : Define the target user group.

What : Choose the content to deliver.

How : Determine the delivery method.

1.3 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; solutions are highly customized.

Lack of a systematic closed‑loop for post‑marketing analysis and optimization.

1.4 Project Value

Efficiency : Operators can directly use algorithm‑generated user groups, reducing manual testing and increasing conversion rates (estimated CTR uplift of ~20%).

Revenue : Precise marketing is expected to increase order volume by about 20%.

2. Precise Marketing Framework

Before building the framework, the team analyzed the characteristics of Hello’s precise marketing scenarios and devised corresponding solutions.

2.1 Scenario Characteristics and Solutions

Many customized scenarios → move from modular to component‑based and eventually platform‑level algorithms.

Need to expand high‑quality user groups → adopt the semi‑supervised Pu‑Learning framework.

Insufficient seed users for modeling → use unsupervised methods for intelligent scaling.

2.2 Business Framework

Feature Processing : Offline feature tables stored in Hive and online real‑time features computed by Flink and cached in Redis.

Precise Marketing Module : Consists of algorithm, user analysis platform, and delivery platform. Algorithm: Industry‑package (Pu‑Learning LookAlike) and intelligent scaling via graph embedding. User analysis: Operators create seed user groups, optionally trigger intelligent scaling to obtain expanded target groups. Delivery: Create tasks, select target groups, launch campaigns, and collect A/B results.

Algorithm Scenarios : Acquisition, activation, and retention, using resources such as banners, in‑app messages, or push notifications.

2.3 Technical Framework

Marketing tasks are built on two types of target groups:

Industry‑package groups generated offline via model training, stored in Hive → ES tags.

Intelligent scaling groups generated online: behavior data collected in Kafka, processed by Flink, real‑time features stored in Redis, then consumed by the scaling service.

3. Precise Marketing Algorithm Capabilities

3.1 LookAlike Modeling under Pu‑Learning

LookAlike is a concept rather than a specific algorithm: it expands a seed user set by finding similar users. Methods include supervised, semi‑supervised, and unsupervised learning. Semi‑supervised learning (Pu‑Learning) leverages a small labeled set and a large unlabeled set.

3.2 Challenges and Solutions for LookAlike

Sparse features for new‑business users → use two‑round feature engineering.

Limited usable features → extract cross‑business common features.

Need to expand high‑quality groups → adopt advanced Pu‑Learning framework.

3.3 TSA Semi‑Supervised Model

Two‑step process: first identify reliable negative samples from unlabeled data, then train a supervised model on the combined positive and reliable negative sets. This improves ROI while expanding the user pool 3‑10×.

3.4 Graph Embedding for Industrial‑Scale User Expansion

Graph embedding builds a user relationship graph based on spatio‑temporal co‑occurrence (e.g., same location and time). DeepWalk generates random walks, which are fed into Skip‑Gram to learn user vectors. EGES further incorporates side information with weighted contributions.

Embedding similarity is computed using Milvus, a vector engine offering near‑real‑time queries and multiple indexing options.

3.5 Results

Coverage: Platform‑level solution supports intelligent scaling for 60% of scenarios at zero cost.

ROI: Improved by more than 20%.

4. Future Directions

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

Develop an automated threshold recommendation system for intelligent scaling, optimizing ROI while meeting operator expansion goals.

Thank you for listening.