Algorithmic Optimization of Information‑Flow Advertising for Hallo Mobility

Guest: Zhou Bingqian, Senior Algorithm Engineer at Hallo Mobility (DataFunTalk).

1. Information‑flow ad landscape – Information‑flow ads embed native content (images, text, video) into media streams, offering value‑added experiences. Since 2006 they have grown from Facebook to TikTok, Weibo, Toutiao, etc., reaching ~40% market share by 2022 and becoming a primary channel for user acquisition.

2. Hallo business background – The ad‑tech stack evolved through four stages: exploration, initial scaling, cost‑reduction, and mature automation. The current system integrates a unified API gateway, storage (Redis, MySQL, HBase, Elasticsearch), and an application layer handling decision‑making, automated operation, and attribution.

3. Advertiser‑side algorithmic optimization

3.1 Plan dimension – The goal is to predict whether a user who submits certification will complete an order ("completion"). Due to limited exposure data, the problem is reframed as low‑quality traffic detection and completion‑rate estimation using LightGBM. Samples are built from users who complete within 7 days; data augmentation creates order‑level samples.

Features include user, environment, ad, and time attributes; the model uses LightGBM for binary classification.

3.2 Creative dimension – To avoid waste from similar low‑performing creatives, a model predicts whether a new creative can generate volume. Feature engineering combines ID embeddings (word2vec on creative/plan IDs), numeric feature cross‑features, and configuration parameters, feeding a multi‑class classifier.

3.3 Pre‑bid uplift mechanism – An uplift model (causal inference) scores users from 0‑5 to decide exposure or bidding adjustments, separating naturally converting users from those needing ad incentives, thereby reducing cost per acquisition.

4. Future directions – Plans include more precise pre‑exposure screening using exposure data, full‑automation of online delivery with algorithm‑driven creative/plan generation, budget allocation, and integration of advanced AI techniques (CTR, reinforcement learning, CV‑based creative generation).

5. Q&A highlights – Discussed handling multi‑business competition, why similar creatives behave differently across accounts, and the relationship between RTB and RTA in future pipelines.

Thank you for attending.