iQIYI Recommendation System: Architecture, Model Evolution, and Ranking Strategies

The article introduces Michael, an assistant researcher in recommendation algorithms at iQIYI, and explains the rapid growth of user‑generated content (UGC) in the mobile Internet era, which drives the need for personalized recommendation.

Personalized recommendation is built on big‑data analysis and artificial‑intelligence techniques to model user interests and alleviate information overload.

The iQIYI recommendation pipeline consists of two stages: a recall stage that selects a candidate set of hundreds to thousands of videos from a massive library based on user interests and history, and a ranking stage that scores these candidates to produce a short list of high‑quality results.

Key components of the system include user profiling, feature engineering, various recall algorithms (collaborative filtering, topic models, content‑based, social networks), and ranking models.

To ensure consistency between training and online prediction, iQIYI stores online features and re‑injects them into training samples, addressing feature drift.

The system is designed for continuous iteration, decoupling data preparation, model training, and online serving, and supports configurable strategies.

Model evolution timeline:

1. Early stage used Logistic Regression (LR) for its interpretability and ease of debugging. Evaluation metrics and optimization objectives were aligned with business goals, emphasizing user engagement over simple CTR.

2. To overcome LR limitations with dense, non‑linear features, the model was upgraded to a GBDT+LR stack, leveraging GBDT’s automatic feature transformation and interaction capabilities.

3. Further improvement replaced LR with Factorization Machines (FM), yielding better support for sparse features and reducing computational complexity to O(N). FM’s architecture combines a linear term, a second‑order interaction term, and a bias term.

4. To capture deep semantic information from embeddings, a DNN was introduced and combined with GBDT and FM in an ensemble (DNN+GBDT+FM). The ensemble uses three hidden layers (1024‑512‑256 units) and pretrained user/video embeddings.

Optimization algorithms explored include SGD, L‑BFGS, and FTRL (which adds L1 regularization, cumulative gradients, and feature‑frequency‑based learning rates). FTRL showed a 4.5% lift on sparse features. The final ensemble model uses the Adam optimizer for faster convergence.

Industrial context: YouTube launched DNN ranking in 2016; PNN, Wide & Deep, and DIN are other notable DNN‑based ranking approaches referenced.

The article concludes that ranking is a critical ML scenario requiring careful feature engineering, architecture design, and pipeline optimization. Future work includes enhancing model architecture, embeddings, diversity, cold‑start handling, and multi‑objective learning.

References are provided for Wide & Deep, PNN, Facebook click prediction, Deep Interest Network, and Factorization Machines.