WeChat "Look" Recommendation System: Architecture, Modeling, and Engineering Challenges

Background – WeChat’s public platform publishes millions of articles daily, making personalized recommendation essential. The "Look" feature leverages user reading, following, and sharing signals combined with deep learning to recommend articles, videos, and short clips.

Overall Architecture – The system follows a classic three‑stage pipeline: recall, coarse ranking, and fine ranking, with an additional heterogeneous mixing layer after fine ranking that incorporates a reinforcement‑learning model for long‑term reward optimization.

Data and Recall – Data is split into user and content attributes. Recall reduces a million‑scale candidate pool to tens of thousands using interest‑based, collaborative, public‑account, and social recall strategies, followed by diverse exploratory recall for cold‑start scenarios.

Ranking – Coarse ranking processes ~10,000 candidates with lightweight linear models, while fine ranking scores ~1,000 candidates using CTR prediction models. The fine‑ranking stage predicts click‑through‑rate (CTR) for each user‑item pair.

Modeling Progression – Initially, large‑scale Logistic Regression (LR) was trained on billions of features using the PanguX distributed platform. To address LR’s memorization limits, Factorization Machines (FM) were introduced to learn feature embeddings. Subsequently, the wide&deep model combined linear memorization with deep neural networks, and DeepFM further integrated FM’s low‑order interactions with deep layers for higher‑order feature learning.

Multi‑Objective Optimization – Beyond CTR, the system optimizes for watch time, shares, likes, and comments. Multi‑task joint modeling shares embeddings across tasks while allowing task‑specific networks, and online weight‑fusion balances these objectives based on offline AUC trade‑offs and online user engagement metrics.

Diversity – Diversity is treated as a key metric alongside CTR. Heuristic limits on category/Topic/Tag counts were first used, later replaced by submodular optimization and finally reinforced‑learning where diversity is added as a reward, achieving a win‑win for CTR and diversity.

Engineering Challenges – Major challenges included scaling the algorithm platform (moving from Spark MLlib to a custom PS‑lite based PanguX), overcoming online serving memory and latency bottlenecks (feature KV store, model sharding, TF‑Serving replacement with DeepX), and redesigning feature extraction for performance and configurability.

Conclusion – The "Look" ranking pipeline evolved from linear to deep models, incorporated reinforcement learning and multi‑objective learning, and delivered significant improvements in CTR, dwell time, and diversity, highlighting the importance of robust algorithmic engineering at scale.