Zhihu Recommendation Page Ranking: Architecture, Feature Design, Model Evolution, and Practical Insights

The article introduces Zhihu's recommendation page and the history of its ranking system, outlining the three‑stage request flow: recall (topic‑based and content‑based), ranking (rule‑based and model‑based using GBDT and DNN), and re‑ranking for product considerations.

It describes the evolution of ranking algorithms through four stages: simple time‑ordering, EdgeRank inspired by Facebook, Feed Ranking using GBDT, and Global Ranking employing deep‑learning models (DNN, embedding).

Feature engineering is divided into three categories—user profile, content profile, and cross features—with various representations such as numeric, one‑hot, multi‑hot, and value‑weighted encodings. Design principles emphasize feature completeness, raw value preservation, high coverage, and consistency between online and offline pipelines.

New feature directions include explicit cross features, business‑driven insights, and content embeddings derived from skip‑gram models trained on massive search‑behavior sessions (85 billion samples) using nce‑loss.

The CTR model uses cross‑entropy loss with a sigmoid output. Initial DNN architecture separates user and content blocks, concatenates hidden layers, and feeds them into fully connected layers, achieving an AUC of 0.7618. Subsequent optimizations—block‑wise DNN, DeepFM (adding FM and first‑order terms), DIN with topic‑based attention, GRU‑based Last Display, and multi‑task learning—incrementally improve AUC up to 0.7678.

Practical experience sharing covers handling time‑varying statistical features, ensuring offline‑online feature consistency, feature normalization (log‑scaling large values), illegal value checks, caching user‑side computations, large‑scale data handling with FlatBuffer on HDFS, and automated model updates.

Challenges identified include differences between recommendation and search pages, pointwise CTR models ignoring item interactions, and user fatigue from repetitive content. Future directions propose reinforcement learning with actor‑critic architecture to capture user feedback and provide whole‑screen recommendations.