User Growth, Full‑Stack Growth System, and Deep Learning Applications in Search at Alibaba 1688

Alibaba 1688 measures product success by user volume, with search serving as the primary traffic source; the talk outlines the platform’s user‑lifecycle analysis, full‑link growth system, and deep‑learning‑driven search improvements.

User Growth – The growth pipeline follows the AARRR model, emphasizing user recall over acquisition. Users are segmented into potential users (pre‑download), new users (first three orders), mature users (content engagement), and churned users (re‑engagement via churn models and targeted offers).

Installation Control – A budget‑constrained optimization problem maximizes installation volume, login rate, and next‑day retention. Operators input budget data, and the system predicts installation outcomes across platforms, using constrained inequality solving and Pareto‑front analysis to guide spend allocation.

Intelligent Coupons – The platform distinguishes between no‑threshold red packets, full‑reduction coupons, and merchant coupons. Algorithms decide who receives coupons, the optimal amount, and timing, balancing financial controls with conversion goals through offline (pre‑computed) and online (real‑time) optimization, modeled as knapsack and multi‑objective problems.

Smart Push – Push notifications are optimized for relevance, timing, and fatigue constraints. The pipeline includes product matching, copy generation, global allocation under fatigue and industry fairness constraints, and timing modeling based on user activity patterns to maximize open rates and long‑term engagement.

Information‑Flow Placement – Low‑activity users are targeted with personalized ads; image processing removes intrusive overlays, and budget‑driven bidding models determine exposure and click‑through rates across channels.

Search System – The search workflow comprises query input, classification, rewriting, term and semantic recall (using DSSM), coarse ranking, deep‑learning‑based fine ranking, re‑ranking to avoid homogeneity, and content‑centric extensions (articles, guides). Recent advances include long‑short‑term user behavior modeling, multi‑task learning (CTR, CVR, category prediction), and graph‑based representations to address long‑tail items.

Graph‑Based Sequence Representation – User behavior sequences are linked to queries via attention mechanisms, allowing the model to weight relevant historical interactions and improve query embeddings, especially for ambiguous or multi‑intent searches.

Order‑Aggregation Control – A central control module aggregates demand across scenarios, allocates coefficients based on scene performance, and uses reinforcement learning to optimize daily distribution before midnight, ensuring efficient order consolidation for merchants.

The presentation concludes with an invitation for algorithmic talent to join Alibaba 1688’s research and product teams.