How Alibaba’s AI Powers Real‑Time Customer Segmentation and Personalized Shopping

Background

Ma teacher described three technological revolutions and emphasized that data has become the core resource of a new energy era. The speaker also noted that big data has empowered Double 11 and aims to empower all merchants to achieve the mission of making business easier.

Customer Operation Platform

The platform provides merchants with data‑driven, AI‑enabled fine‑grained customer operation capabilities, shifting focus from short‑term transactions to long‑term customer loyalty. In 2016’s Double 11, over 230,000 merchants used the platform for personalized store operations and precise fan‑member marketing, significantly boosting conversion.

Visitor operation uses AI clustering to segment visitors and deliver personalized homepages, such as tailored layouts for users preferring moisturizers versus masks, improving experience and conversion rates.

AI Real‑Time Clustering

The AI clustering engine has three main characteristics:

Industry‑level models: the same feature may have different importance across industries.

Combination of long‑term, recent, and real‑time feature systems to capture stable preferences and immediate context.

Shop‑level adaptive clustering: models adjust to each shop’s product mix using shop sales distribution priors and multi‑armed bandit reinforcement learning.

Compared with static demographic clustering, AI clustering offers higher real‑time accuracy and shop adaptability, leading to over 10% conversion improvement and a 40% increase in personalization penetration.

Double 11 Promotion Page Personalization

For Double 11, Alibaba achieved full‑site personalization, including the promotion landing page that bridges the event venue and individual stores. The AI personalization engine consists of a matching framework and a ranking framework.

Matching methods include graph‑based mining (adsorption, adar, Jaccard, SimRank++), hashing (MinHash, SimHash), graph embedding, semantic matching, streaming computation, and scenario‑specific re‑ranking.

Alibaba introduced a large‑scale distributed Graph Embedding algorithm that represents each product as low‑dimensional vectors, captures asymmetric relationships with dual embeddings, and processes billions of nodes and edges on the ODPS Graph platform. This algorithm was accepted at AAAI 2017.

Ranking employs massive sparse models (LR, FTRL, DNN) trained on billions of samples, integrating long‑term, short‑term, and real‑time features, as well as context features. Distributed GPU training and optimized inference achieve prediction speeds ten times faster than baseline. The personalized landing page yields more than a 20% lift in conversion, translating to billions of additional transactions.

QianNiu Headlines (千牛头条) Technology

QianNiu Headlines is a commercial media platform that delivers personalized news to merchants. Its recommendation system comprises offline, near‑real‑time, and real‑time components.

Offline builds comprehensive user profiles with expected preference distributions smoothed by a Gamma‑Poisson model. Near‑real‑time analyzes newly published articles, extracting keywords via TextRank, Mutual Information, and Log Odds Ratio, and generates industry tags using a multi‑task semantic vector model.

The online component trains an Online Bayesian Logistic Regression (BLR) model per article, improving CTR by about 20%.

High‑order statistical features are created by cross‑combining user and article attributes with business‑driven statistics to avoid feature explosion while maintaining generalization.

Cold‑start for new articles leverages historical keyword and topic performance to predict outcomes.

Service Market Personalization

The Service Market connects over 90% of active Taobao merchants but suffers from low visit frequency and sparse behavior. A personalized framework was built to improve search and recommendation.

Key technologies include real‑time preference identification (using time‑decay aggregation), semantic matching with query expansion, and model ranking that combines user, item, and cross features. Offline updates merchant and service features, while online components handle real‑time behavior analysis and personalized ranking.

Results show click‑through rate improvements of 10%–90%, empty‑result reduction by 400%, and conversion gains of 70%–200% across various metrics.