How Intelligent UI Boosted Alibaba’s Holiday Sales by 10%+ Through User Preference Modeling

Background

Personalized product recommendations generate significant incremental traffic, but the explosion of item data creates decision fatigue for users, leading to churn. Alibaba’s BUs began exploring intelligent UI in 2019 to shorten decision time by showing users the most relevant content.

Early Success and Challenges

Algorithm‑driven UI personalization in the 2020 March promotion increased PVctr by 12% month‑over‑month. However, the approach suffered from four main issues: cold‑start training, dependence on high traffic, non‑reusable models across diverse scenarios, and a high exploration barrier for teams without algorithm support.

Shift to UI Digitalization

The team re‑thought the problem by focusing on UI digitalization combined with user‑portrait analysis. By dynamically allocating UI variants based on user preferences, they eliminated cold‑start, reduced traffic dependence, and solved model reuse problems, achieving a 10.27% overall PVctr lift, 15.77% in smart‑shelf scenarios, and 14%+ improvements for price‑ and finance‑sensitive groups.

User Preference Modeling

Preferences are split into behavioral (price, service, benefit, origin, finance) and product (new, trending, hot) dimensions. Six binary classification models predict whether a user will interact with items bearing a specific tag within a future window, using features such as 1‑day, 7‑day, 30‑day, and 90‑day behavior aggregates (clicks, adds‑to‑cart, favorites, search terms, chat logs, page flows, and custom business‑specific signals).

Labels are set by weighted thresholds (e.g., click = 1, favorite = 2, purchase = 8). Adjustments address label imbalance caused by uneven tag coverage.

Modeling Steps

Problem definition – predict future interaction with tagged items based on past N‑90 day behavior.

Feature engineering – multi‑temporal behavior, text, tag exposure, page flow, and custom signals.

Label setting – weighted thresholds for each preference.

Model selection – GBDT binary classifiers.

Group Preference Modeling

Users are clustered by identity × region × level . Group TGI scores compare the proportion of a trait in the target group versus the overall population, guiding cold‑start for sparse segments (e.g., mothers, young women).

Fusion of Individual and Group Preferences

Rich‑behavior users rely on individual scores; sparse‑behavior users fall back to group scores, enabling a two‑tier personalized UI allocation.

Experiment Design

Five‑step workflow: split user portrait, develop & assemble materials, label methods, UI & user recommendation, data analysis & attribution. Experiments use three buckets – baseline (10% traffic), random (10%), and targeted (80%) – to validate UI impact.

Results: overall PVctr +10.27%; smart‑shelf PVctr +15.77%; price‑sensitive +14.21%; finance‑sensitive +14.63%.

Productization

Low‑code material development enables one‑time creation with multiple reuses. UI is decomposed into layout and material blocks, identified by IDs and letters. Tagging service links materials to semantic tags (content, type, overall scheme). UI recommendation service matches user‑group tags to material tags, delivering personalized UI at runtime.

Tagging & Attribution Platform

Tag libraries support labeling, black‑listing, and attribution, feeding back performance data to designers for iterative improvement.

Future Roadmap

Plans include expanding intelligent UI to senior users, integrating UI recommendation into broader business domains, building a 2.0 data attribution platform, and offering a design‑guidance portal for UED teams.