iQIYI Effect Advertising: Product Overview, Ranking Algorithm Practices, and Business Strategy

The presentation is divided into four parts: an introduction to iQIYI effect advertising products and their resource slots, a practical overview of the ranking algorithms (including coarse‑ranking and fine‑ranking models), business‑level strategies such as cold‑start handling and smart bidding, and finally a summary and outlook.

Effect advertising resources are classified into in‑frame slots (pre‑mid‑post‑roll and pause ads) and out‑of‑frame slots (near‑feed, pure feed, focus images, startup screens, video‑related ads, etc.). The main selling types are CPX (cost per click) and oCPX (optimized CPX), both using real‑time bidding based on eCPM.

In CPX mode, eCPM = bid × PCTR. Advertisers face three pain points: unstable conversion cost, heavy manual optimization, and difficulty scaling. oCPX introduces a richer billing formula: bid × PCTR × PCVR × smart‑bidding factor, allowing advertisers to set a target cost while the platform optimizes delivery.

The ranking funnel consists of three stages. First, the recall stage (targeted and intelligent recall) selects ads based on advertiser‑defined criteria and similarity/estimated conversion. Next, the coarse‑ranking stage performs lightweight CTR/CVR estimation, creative selection, and cold‑start handling. Finally, the fine‑ranking stage applies budget smoothing, detailed CTR and CVR models, and smart bidding.

Data flow: real‑time tracking logs generate on‑site features, which are stored in Kafka and Hive. Some features feed HTTP services for smart‑bidding calculations, while others are used for offline model training. Feature monitoring ensures coverage and stability.

Offline training includes coarse‑ranking and fine‑ranking models. Coarse‑ranking initially used FM models, later upgraded to a dual‑tower architecture that outputs user and creative vectors for efficient dot‑product inference. Fine‑ranking employs FM, wide‑deep, DNN, and DCN models; the final choice was a wide‑deep model optimized for learning rate, optimizer, and layer size. Model calibration uses monotonic regression for PCTR and additive bias for PCVR to align predictions with actual rates.

Conversion‑rate estimation faces challenges: multiple conversion goals, sparse conversion data, sample bias, and varying rates across slots. A multi‑task learning framework jointly trains a primary PCVR network and an auxiliary PCTR network, sharing features to improve sparsity handling.

Business strategy modules include cold‑start handling (adding new ads into the coarse‑to‑fine pipeline, exposure and conversion support), and a sophisticated smart‑bidding system. Smart bidding adjusts the bidding factor based on remaining budget, expected conversions, and cost dynamics, with additional tactics such as exploratory bidding, holiday/peak‑hour scaling, and smooth scaling to avoid sudden spikes.

Traffic selection further refines bidding by comparing predicted PCTCVR with historical averages, boosting bids for high‑quality traffic and reducing them for low‑quality traffic. Resource‑level bidding computes separate factors per ad slot to respect differing cost and conversion characteristics.

The final strategy combines dual‑objective bidding (front‑end goals like downloads/activations and back‑end goals like payments) by calculating separate smart‑bidding factors and using the larger one for scaling, weighted by back‑end conversion volume.

Overall, the system aims for increased automation, personalization, and stable back‑end costs, with future work focusing on deeper business scenario understanding, richer data collection, continuous model iteration, multi‑objective transfer learning, and conversion‑delay mitigation.