User Modeling for Search Ranking: Practices, Model Design, and Experimental Analysis at Alibaba

Background and Significance User modeling is a core technology for search and recommendation systems. In Taobao search, the ranking target is a <user, query, item> triple, where item features are dense while user features are sparse, requiring extensive generalized features.

Information Abstraction The authors enrich user modeling through static user profiles, preference tags mined from behavior, and fine‑grained real‑time behavior modeling to capture current interests.

Information Processing User behavior data are organized by behavior cycle (short‑term vs. long‑term) and content (explicit vs. implicit signals). Short‑term sequences are filtered by predicted query categories, while long‑term data span two years and are divided into quarterly sub‑sequences.

Model Architecture The overall model concatenates user profile, multiple behavior sequence features, item features, and contextual signals (e.g., weather, network) before feeding them into a DNN classifier. Sequence modeling uses an optimized self‑attention mechanism with cosine similarity (replacing dot‑product) and a query‑attention pooling layer to emphasize behavior relevant to the current query.

Short‑Term Sequence Improvements Cosine‑based similarity with scaling improves softmax logits, and query‑attention pooling further aligns historical actions with the current intent, avoiding the heavier target‑attention computation.

Long‑Term Sequence Modeling Quarterly sequences are embedded, masked, and processed through multi‑layer self‑attention and attention‑pooling, then concatenated to form a comprehensive long‑term preference representation, aiding seasonal personalization.

Client‑Side Click and Exposure Modeling Click sequences captured on the device provide millisecond‑level latency and richer signals (e.g., dwell time, button clicks). Exposure sequences (items shown but not clicked) are modeled via mean pooling and an auxiliary loss that enforces a margin between positive and negative item distances, improving the representation of user disinterest.

Experiments and Analysis Using online exposure and click data, the authors evaluate models on AUC. Incorporating the proposed sequence improvements yields up to 0.3% absolute AUC gain, while adding new sequence features contributes an additional ~0.7% gain. Attention weight visualizations show the model learns both positional importance and inter‑item relationships.

Conclusion and Outlook The presented user modeling techniques significantly boost Taobao search performance during major events, and future work will focus on finer user data perception, more scientific data organization, and further model refinements.