How Alibaba Uses AI Models to Optimize Double 11 Consumer Benefits

Business Analysis

During the pre‑sale and the Double 11 day, consumers can receive various user benefits such as shopping vouchers, red packets, store coupons, and category coupons on Taobao and Tmall. Optimizing the issuance of these benefits balances the interests of consumers, merchants and the platform, increasing conversion, sales and overall GMV.

Key Benefit Types

Shopping vouchers are universal discount coupons for the Double 11 event, set with tiered thresholds by category and can be used across stores. They promote transactions and raise average order value.

Red packets are another common benefit, providing promotional and social effects. Various interactive red‑packet games (torch red packet, carnival city red packet, cut red packet) were launched in 2017 to boost activity and visibility.

Problems to Solve

Estimate consumer spending on Double 11 to determine appropriate voucher amounts.

Analyze consumer discount sensitivity to allocate different red‑packet amounts.

Assess consumer spread ability to give high‑spread users higher chances of receiving unlit red packets.

Analyze consumer category preference and churn to target specific groups with tailored red packets.

Spending Forecast Model

Based on historical Double 11 purchase records, a model predicts each consumer’s spending on the day. Features include short‑term, long‑term and promotion‑period behavior. Two core models are used: a regression model for general consumers and a classification model for high‑spending users.

Accurate forecasts increased the proportion of consumers using vouchers by 51 % and the voucher‑related GMV by 72 %.

Discount Sensitivity Model

An XGBoost model uses user features derived from historical benefit usage to predict sensitivity to different discount amounts. A/B tests showed a 17.7 % increase in red‑packet usage rate and a 16.4 % rise in per‑user spending for targeted groups.

Spread Ability Model

By analyzing interaction behavior, spread and activity factors are computed and users are segmented into four groups (high‑spread/high‑activity, high‑spread/low‑activity, low‑spread/high‑activity, random). High‑spread/high‑activity users drive the torch red‑packet’s viral effect.

Category Preference Model

The model predicts whether a consumer will click a specific category within the next seven days. It incorporates an Ebbinghaus forgetting curve to model decay of interest, using behavior counts over multiple time windows. Logistic regression with the decay factor achieves the reported performance.

Churn Prediction Model

A random‑forest model predicts user churn using platform activity features (browsing, adding to cart, favoriting, purchasing, brand follows) across various time windows on both mobile and PC.

Future Outlook

Benefit issuance will become more intelligent, automatically selecting benefit combinations, settings, target audiences and timing based on merchant, brand and industry characteristics, while reducing user friction.