User Profile Tagging: Construction, Feature Processing, and Evaluation

The presentation begins with an overview of user profile tags, emphasizing the growing need for deep user understanding in digital transformation. It introduces three main sections: (1) basic attribute tags—attributes like gender, age, OS, and city—constructed via user input, event tracking, model prediction, or third‑party data, and applied in daily analysis and modeling.

(2) Business‑oriented tags are closely linked to KPI goals, categorized into strong and weak KPI associations. Construction methods include KPI‑based distance definitions and composite behavior calculations, while usage focuses on monitoring KPI progress, drilling down user groups, and implementing differentiated operational strategies.

(3) Strategy‑oriented tags target specific interventions (e.g., coupons) and are evaluated using uplift models, repurchase cycle predictions, or binary classifiers to maximize ROI.

The article then details feature processing techniques essential for reliable tags: anomaly detection and filling (using box plots, AVF, cap/floor percentiles), handling missing values based on metric nature, time‑decay weighting aligned with RFM dimensions (Recency, Frequency, Monetary), and logarithmic smoothing to mitigate long‑tail effects.

Evaluation of tag quality is discussed through two primary criteria: cohesion (measured by Silhouette Coefficient) and stability (measured by Coefficient of Variation). High cohesion indicates similar users within a segment, while low variation across time reflects stable segment behavior.

A Q&A segment addresses practical concerns such as calculating cohesion for different activity levels, defining activity thresholds, computational complexity of time decay, perspectives for basic business strategies, and scenarios where tag evaluation is applied.

The session concludes with acknowledgments to the speaker, Wu Zihua from ByteDance Data Science, and the organizing team.