iQIYI Effect Advertising: Architecture, Click & Conversion Rate Estimation, and Intelligent Bidding

The presentation, authored by senior iQIYI engineer Wang Hui and edited by Wang Hao, introduces iQIYI's effect advertising platform, which leverages a dual‑engine architecture consisting of feed (information flow) and in‑frame slots, as well as additional placements such as "You May Like" and video‑associated positions.

Advertising is billed using the oCPX model, an evolution of CPX that optimizes for post‑click conversions (install, payment, etc.) rather than clicks alone. Advertisers set conversion goals and bids, while the system’s algorithms handle optimization.

Key algorithmic difficulties include extremely high‑dimensional and sparse conversion samples, the need to support multiple conversion types, and the requirement for low‑latency, high‑throughput computation on massive traffic volumes.

The personalized recommendation workflow comprises three stages: recall (candidate ad retrieval based on targeting), coarse ranking (lightweight models with cold‑start and exploration to mitigate Matthew‑effect), and fine ranking (high‑precision models estimating click‑through rate, conversion rate, and applying intelligent bidding factors, with budget smoothing).

Ranking logic is driven by eCPM = CTR × CVR × bid × intelligent‑bidding factor, ensuring ads are ordered by expected revenue per thousand impressions.

Online training and inference follow a pipeline where real‑time features (scene, time, feedback) are combined with short‑term signals (viewing interests, search, social behavior) and long‑term attributes (demographics, interests, ad metadata). Models include FM (per‑day, online learning), deep learning (Wide&Deep), and reinforcement learning, updated on a minute‑level basis using streaming data.

Click‑and‑conversion rate estimation faces challenges such as high‑dimensional sparsity, label lag, and conversion spikes. Solutions involve dynamic bucket sizing, treating clicks as negative samples until conversion occurs (with weight adjustment), and incorporating non‑target conversions as weighted positives.

Intelligent bidding balances cost control and traffic quality. A bidding factor, derived from the ratio of actual to target cost, scales eCPM to either lower cost by using cheaper traffic or increase competitiveness when cost is below target. Further refinement introduces traffic‑quality‑aware bidding functions to avoid over‑spending on low‑quality impressions.