How iQIYI Boosted Click‑Through Rates with AI‑Powered Personalized Poster Generation

Background

Personalized poster recommendation can significantly increase click‑through rate (CTR) on video platforms. Inspired by Netflix’s 2014 artwork personalization experiments, a system was built to generate, rank, and serve personalized posters at scale.

AI Poster Production

The production engine creates a large candidate pool of posters through three main modules:

Smart Cropping : AI models detect faces, bodies, objects, and text regions to generate semantically meaningful crops.

Smart Frame Extraction : Video streams are segmented into scenes; a dynamic sampling rate selects high‑information frames (higher for trailers) while limiting redundancy.

ZoomAI Enhancement : Denoising and color‑enhancement algorithms improve visual quality.

All candidates undergo a manual review step before entering the personalized poster pool.

Personalized Poster Distribution

Offline Stage

Poster images are uploaded to a CDN and indexed in a database. A feature‑extraction service computes image‑level attributes such as quality scores, celebrity presence, sentiment, and embedding vectors, as well as behavior statistics. These features are combined with user portrait data to train ranking models, including Factorization Machines (FM), DeepFM, Logistic Regression (LR), and Gradient Boosted Decision Trees (GBDT).

Online Stage

During recommendation, candidate posters are recalled based on scene configuration, then re‑ranked by the trained model. Business rules (e.g., exposure caps) are applied before the final poster is served.

Exploration via Multi‑Armed Bandits

Because early feedback for new posters is unavailable, Multi‑Armed Bandit (MAB) algorithms are used to explore effectiveness:

Context‑free MABs: Epsilon‑Greedy, Upper Confidence Bound (UCB), Thompson Sampling with a sliding‑window adaptation.

Context‑aware MABs: Linear models that predict reward from combined user portrait and poster features, selecting the arm with the highest upper confidence bound.

Application Strategies

Richness Control : Limit the number of posters with similar composition or tag distribution; use Max‑Marginal Relevance (MMR) to diversify.

Display Deweighting : Reduce exposure frequency for the same video to increase stylistic variety.

Poster Retirement : Gradually increase exposure for new posters, monitor CTR, and demote or retire under‑performing items.

Scene Matching : Align poster tags with textual cues in the surrounding UI (e.g., starring actor names) to improve relevance.

Online Experiment Results

Small‑traffic A/B tests on TV and mobile devices showed consistent CTR and UCTR improvements across placement, video, and poster dimensions. The uplift was observed in both overall exposure and per‑poster performance metrics.

Future Directions

Research aims to develop end‑to‑end models that jointly generate and rank posters, and to explore generative techniques that automatically compose semantic elements (characters, objects) while respecting composition constraints.

References

https://medium.com/netflix-techblog/artwork-personalization-c589f074ad76

https://www.jianshu.com/p/558d38c62579