Alibaba 1688 User Growth, Full‑Chain Growth System, and Deep‑Learning Applications in Search and Promotion

The talk, delivered by Alibaba algorithm expert Qian Wang, begins with an overview of 1688's user‑growth metrics, highlighting a significant rise in DAU and buyer numbers since April and emphasizing the importance of the B‑class buyer mindset and content‑driven traffic.

It then outlines the user‑lifecycle framework used by the growth team: potential users (pre‑download acquisition via SEO and app‑store ads), new users (first‑order conversion with onboarding incentives), mature users (retention through content‑rich experiences and periodic push of offers), and churned users (reactivation via churn‑rate models, targeted red‑packets, and information‑flow partnerships).

Next, the installation central control module is described. Given limited advertising budgets across platforms (e.g., Xiaomi), the team formulates a constrained optimization problem that maximizes installation volume, login rate, and next‑day retention while respecting budget limits. An interactive tool allows operators to input a budget and receive predicted installation outcomes for each placement.

The presentation then moves to intelligent rights (red‑packet) strategies . Three types of coupons are discussed—no‑threshold red‑packets, threshold‑based discounts, and merchant‑issued coupons—each governed by strict financial controls. The allocation problem is modeled as a knapsack problem where the total bonus pool is the capacity and individual red‑packet amounts are item weights, aiming to maximize total conversion probability.

Following that, the smart push system is introduced. Push messages must balance immediate app‑open rates with long‑term user engagement, avoid fatigue, and respect industry‑level fairness constraints. The pipeline includes product matching, intelligent copy generation, global allocation with fatigue and industry constraints, timing modeling based on user activity patterns, and venue‑specific landing‑page optimization.

The information‑flow advertising workflow is explained: low‑activity users are targeted, external data sources enrich feature engineering, CTR‑based bidding determines placement, and image processing removes intrusive watermarks before delivery.

Search is a core focus. The end‑to‑end search stack includes query input , query classification , query rewriting (handling typos and synonyms with seq2seq models), recall (term‑based followed by DSSM semantic recall and emerging query‑to‑image matching), coarse ranking (high‑throughput filters), fine ranking (state‑of‑the‑art deep models), re‑ranking (de‑duplicating homogeneous items), and integrated search (adding articles and guides to keep users browsing longer).

A recent graph‑based sequence representation model is highlighted. User behavior sequences are linked to queries, and attention mechanisms weight historically relevant items (e.g., “multi‑succulent plants”). Query embeddings are refined by distinguishing important tokens (e.g., “拆机” in “笔记本电脑拆机工具”). Long‑tail items are represented via a product co‑occurrence graph, allowing supervised and unsupervised signals to improve training.

Finally, the talk covers the order‑aggregation central control system, which balances fixed production costs across merchants by aggregating orders, allocating central‑control coefficients based on venue performance, and using reinforcement learning to optimize the allocation before midnight each day.

The session concludes with thanks and references to further reading on real‑time data engineering and intelligent operations at Alibaba 1688.