Hello's Automated Growth Algorithm Loop: C‑Side Scenarios, Challenges, and Active Growth Strategies

This article introduces Hello's practice of building an automated growth algorithm loop, focusing on the C‑side algorithm scenarios and the challenges they face.

Hello's C‑side algorithm aims to support both ride‑hailing growth and lifestyle services. Traffic sources include in‑app banners, pop‑ups, various components, external advertising on platforms such as public ad networks and carrier systems, as well as free traffic from public (WeChat, Douyin) and private domains, and even offline stores and two‑wheel vehicles connected via the AloT platform.

Because of the large number of scenarios and limited development resources, Hello encounters a contradiction between extensive business needs and scarce engineering capacity. The team therefore adopts a methodology of avoiding duplicated effort, combining industry‑leading techniques with proprietary innovations to improve algorithm development efficiency.

Initially, simple custom models such as tree‑based models and classic CTR estimation were used, primarily for the largest ride‑hailing business. In 2021, Hello launched a unified C‑side algorithm system built on its self‑developed Javis AI platform. The system supports both passive growth (traffic efficiency, eCPM) and active growth (user and revenue growth, ROI) using a common technology stack of multi‑stage recall and ranking algorithms, with special emphasis on causal inference and L2R techniques.

Active growth algorithms follow a 4W2H framework: precise targeting, creative material, intelligent timing (push), and task‑based marketing activities, with subsidies playing a crucial role.

Precise Targeting includes two modes: industry packages (selecting a small subset from a known pool of ~500 million registered users) and Look‑Alike. Because negative samples are ambiguous, Hello employs semi‑supervised PU‑Learning. The process starts with a TSA (spy sample) technique, iteratively refining a classifier to identify true negative samples. Improvements include EM‑based threshold estimation and replacing the baseline GBM with DeepFM to mitigate information‑cocoon effects.

For Look‑Alike, Hello uses graph embedding (EGES) to construct user graphs based on spatiotemporal interactions and app‑wide event sequences. Milvus is used as the vector engine, enabling near‑real‑time similarity search for tens of millions of active users.

Smart Subsidy leverages causal inference via uplift modeling. Two uplift models are described: TreeLift (a tree‑based uplift model with a revenue‑maximizing split criterion) and DeepLift (an enhanced DragonNet that simulates randomized experiments and adds a propensity‑score sub‑network). These models achieve up to 4.7 % revenue lift compared to traditional response models.

Smart Copy improves click‑through rates by 30‑100 % using a template‑plus‑fluency approach, later enhanced with an epsilon‑greedy exploration (EE) module to diversify copy and reduce user fatigue.

Smart External Advertising applies similar uplift and causal techniques to optimize ad plan quality, using ID‑heavy features embedded with word2vec, FM models for wide‑deep prediction, and multi‑class classification for plan quality estimation. The system can automatically shut down low‑quality plans and allocate budget to high‑quality ones, reducing acquisition cost on external platforms.

The article concludes by announcing that the next part will cover passive growth algorithms and the overall growth engine.