How User Personas Power Modern Recommendation Systems: From Theory to NetEase Yanxuan

What Is a User Persona

User personas, first introduced by Alan Cooper, are concrete representations of target users built from demographic, browsing, social, and purchasing data, forming virtual models that label users with attributes such as age, gender, location, interests, and consumption ability.

Key Persona Attributes

Common attributes include demographic information (age, gender, region, education, occupation) and behavioral metrics (activity level, loyalty). Different domains emphasize different features: media sites focus on content interests, social platforms on network connections, and e‑commerce sites on shopping preferences and purchasing power.

Persona Construction Process

Define the feature dimensions and data sources.

Collect data using tools like Flume or custom scripts and store it in a Hadoop cluster.

Clean and extract relevant fields from raw logs.

Train models (e.g., logistic regression, Naïve Bayes, LR classifiers) to predict missing attributes.

Predict unknown attributes using the trained models.

Merge features from multiple sources with confidence scores.

Distribute the enriched profiles to downstream systems such as personalized recommendation, CRM, and marketing.

Essence of Recommendation Systems

Recommendations act as a form of advertising that enhances user experience by helping users quickly find desired items, increasing dwell time, retention, and overall GMV. Effective recommendations balance relevance with user satisfaction, avoiding intrusive ad‑like behavior.

Working Principles and Algorithms

Effective algorithms include collaborative filtering, matrix factorization, graph‑based methods, association rules, various embedding techniques, and CTR‑based models. Embedding maps users and items into a shared vector space, enabling similarity calculations for precise recommendations.

Differences Between Movie and E‑Commerce Recommendations

Movies are relatively static cultural products with stable user interests, making embedding straightforward. E‑commerce items are vastly diverse, often short‑term interest‑driven, and include both fast‑moving and durable goods, complicating unified embedding and requiring sophisticated feature engineering.

NetEase Yanxuan Recommendation Practice

NetEase Yanxuan employs a CTR model based on logistic regression as its core. Feature engineering combines user attributes (gender, income, region), behavioral signals (short‑term and long‑term actions), and contextual factors (season, time since last purchase). These are expanded through multiple Cartesian products to create a rich attribute space.

The resulting feature set improves robustness and performance, with key metrics steadily rising since deployment.