Phoenix News Feed Recommendation System: Architecture, Modeling, and Feature Engineering

As a user‑oriented news product, Phoenix News continuously explores applying AI across the entire content production and consumption chain, aiming to maximize human value by guiding users rather than merely catering to them.

The presentation covers business characteristics, overall recommendation architecture, feature engineering, and extended applications, highlighting key challenges such as item scale, user scale, product form, content timeliness, user satisfaction metrics, content quality assessment, cold‑start problems, and the balance between manual rules and machine algorithms.

The system follows a classic recall‑and‑ranking framework: a massive item pool is filtered through recall, coarse ranking, and fine‑grained multi‑objective ranking, finally refined by re‑ranking and manual rules to produce a shortlist for users.

Model evolution moved from traditional LR/GBDT to FM, FFM, and ultimately deep learning models, with embedding layers for users and items, FM for low‑order feature interactions, DNN for high‑order patterns, and attention mechanisms for session modeling. Multi‑objective learning optimizes CTR, duration, likes, comments, and other business metrics, with plans to explore reinforcement learning for hyper‑parameter tuning.

Feature pipelines integrate offline logs, real‑time logs, and online request data, ensuring consistency between training and inference. Features are categorized into user profiles, content profiles, and request context, and are often crossed to capture nuanced signals such as short‑term interests, demographic preferences, and contextual behaviors.

Advanced techniques include using FFM for optimized recall, addressing item cold‑start via vector similarity, and mitigating the “Matthew effect” by improving diversity and surprise. Additional extensions cover title‑clickbait detection, thumbnail aesthetic cropping, and intelligent thumbnail selection using fine‑tuned aesthetic ranking models.

Metric balancing is crucial: exposure, click‑through rate, reading duration, completion ratio, interaction frequency, diversity, and retention are jointly considered. Sample re‑weighting based on bounce rate further refines model training, leading to measurable online improvements.

Q&A sections discuss practical aspects such as real‑time model updates, feature real‑time vs. model real‑time trade‑offs, handling video and image features, and advice for practitioners transitioning into recommendation engineering.