Baidu's Practices and Insights on Recommendation Ranking

This article shares Baidu's thoughts and practical experiences on recommendation ranking, covering five main aspects: background, characteristics, algorithms, architecture, and future plans.

Background : Baidu's integrated information flow includes various product types such as immersive feeds, single-column, and double-column layouts, supporting interactions like comments, likes, and collections. The system handles billions of daily impressions and over a hundred million daily active users, requiring high throughput and low latency.

Data Characteristics : Three challenges are highlighted – massive scale (hundreds of billions of impressions per day), high performance requirements (millisecond-level response time), and strong Matthew effect where a small number of active users dominate the traffic. The system must balance memorization of head users/resources with generalization for the long tail.

Algorithmic Strategies : The core algorithmic design includes discrete feature engineering, cross features, bias features (position and system bias), and sequence features. Features are discretized via hashing and one-hot encoding, with emphasis on high discrimination, coverage, and robustness. Explicit cross features are combined with implicit ones, and bias mitigation uses structures like Wide&Deep. Sequence features are modeled with long and short-term components, using gating networks to balance them.

Model Architecture : Baidu employs a multi-stage recommendation funnel (recall → coarse ranking → fine ranking → re-ranking). Coarse ranking focuses on lightweight scoring and sample selection, while fine ranking handles complex cross features and long sequence modeling. The system uses ultra-large sparse embedding matrices (billions of rows) with distributed training across many GPUs. Techniques such as dynamic embedding dimensions, threshold-based embedding creation, and two-stage training (online one-pass for sparse layers and fine-tuning dense layers) are applied to mitigate overfitting.

System Architecture : A hierarchical, divide-and-conquer design is used. Multiple recall channels improve recall rate and diversity. Coarse ranking acts as a lightweight filter, fine ranking performs heavy computation, and re-ranking finalizes the sequence with list-wise modeling and business constraints. Joint training of fine ranking and re-ranking (e.g., CTR3.0) reduces score coupling and improves stability.

Future Plans : Baidu aims to explore large language model (LLM) integration in recommendation systems across three dimensions: (1) upgrading models from prediction to decision-making (e.g., cold-start exploration, immersive sequence feedback); (2) moving from discriminative to generative recommendation (e.g., generating recommendation reasons, data augmentation via prompts); (3) increasing model interpretability (white-box approaches) using causal analysis, fairness, and multi-task learning.

Overall, the presentation provides a comprehensive overview of Baidu's large-scale recommendation ranking system, covering data challenges, feature design, algorithmic choices, system architecture, and forward-looking research directions.