iQIYI Generic Ranking Framework for Video Recommendation

iQIYI's recommendation system is composed of multiple stages (Recall, PreRanking, Ranking, ReRanking) and is applied to various scenarios such as long videos, short videos, topics, and actors. To reduce development cost and ensure stability across scenarios, the company extracted common parts of ranking models, standardized and componentized them into a generic ranking framework.

The framework consists of four modular phases, each with executable standard processes and monitoring mechanisms:

Feature production and loading (Feature_Online)

Feature replay

Training

Ranking

Feature Production & Loading – Features are the foundation of recommendation. They are categorized as attribute or statistical, and as offline (daily) or real‑time (second‑level). Offline features are generated and stored in HDFS, then loaded into high‑performance KV stores (e.g., Couchbase, Redis) by the Feature_Online component, which automatically creates the required table format.

Real‑time features are served via a DUBBO RPC interface, delivering millisecond‑level responses. They are generated by Flink jobs that aggregate user behavior logs (exposures, clicks, watches) and write the results back to KV stores.

Feature Replay – During each request, all features used for scoring are logged together with an event ID. Later, by joining the event ID with user behavior logs, the exact feature snapshot at request time can be reconstructed. This solves two major problems: (1) offline feature latency and version mismatch, and (2) absence of real‑time features in training data.

Feature replay enables real‑time features to participate in model training and guarantees consistency between online and offline feature sets.

Training – The training pipeline includes data preparation and model training using a configurable framework. It supports models such as DNN, DeepFM, DCN, DIN, DSSM, and MMoE. Users only need to modify configuration files; no additional code is required. The pipeline handles label up‑sampling, feature operators, and flexible label‑feature combinations. Training data are converted to TFRecords and fed to TensorFlow‑Serving.

Ranking – Ranking is split into inference (TensorFlow SavedModel via TF‑Serving) and a ranking service that bridges the recommendation engine and inference. The service employs feature pre‑loading, high‑performance local caches, and batch concurrent access to improve latency and throughput. It also includes automated monitoring (Prometheus + Grafana) for feature access latency, success rates, and other KPIs.

The framework has been deployed across iQIYI's multiple platforms (main app, TV, short‑video player, channel waterfall pages) and has yielded significant improvements: over 20% increase in average watch time for short‑video playback and more than double the number of first‑play videos on the homepage feed.

Future work includes supporting more models, refining monitoring, and extending the framework to additional business scenarios.