Advertising Inventory Estimation and Allocation Techniques at iQIYI: From ARIMA to Deep Learning and AI Tagging

iQIYI’s core commercial model combines high‑quality video content with paid membership services, with online advertising serving as the second largest revenue source (approximately CNY 2.1 billion in Q2 2019).

In China, advertising is generally divided into brand ads, which focus on exposure and long‑term brand value, and performance ads, which aim at short‑term conversion and direct user actions. Brand ads tend to appear in premium content (e.g., popular dramas, variety shows) and occupy high‑visibility slots such as pre‑roll and splash screens, while performance ads are more common in information‑flow slots and on long‑tail content.

The presentation describes the end‑to‑end workflow of brand‑advertising orders, which consists of a pre‑sale stage (inquiry, reservation, approval) and an execution stage (delivery, settlement). Algorithms support both stages by providing inventory‑locking services, price‑reasonable inquiry, and real‑time inventory allocation.

Inventory estimation is a critical time‑series problem. Early solutions used simple statistical methods (moving average, exponential smoothing) and later evolved to classic ARIMA and Seasonal‑ARIMA models. To improve robustness, iQIYI introduced Adaptive‑ARIMA by adding holiday factors as external regressors and applying extensive data preprocessing (outlier handling, truncation).

Since 2018, deep‑learning models have been explored. Initial attempts with LSTM faced challenges due to daily granularity, limited sample size, and high dimensional targeting. A hierarchical‑clustering LSTM model was proposed, grouping similar dimensions into clusters and training a shared parameter set per cluster, which mitigates sample‑size and resource constraints.

For pre‑roll inventory, a factorization‑machine (FM) regression model is used. Enhancements include removing the bias term, enforcing non‑negative parameters, and applying cost‑sensitive learning to penalize under‑prediction more heavily. These improvements increase self‑consistency of inquiry results and align the model with business priorities (avoiding under‑selling).

Online inventory allocation is formulated as a large‑scale bipartite‑graph optimization problem, aiming to minimize overall shortage while respecting supply and demand constraints. The problem is NP‑hard; solutions such as the Shale algorithm and UID‑hash bucket A/B testing are employed to evaluate allocation strategies.

Precise targeting introduces a massive set of long‑tail dimensions (gender, age, video tags, audience packs). To keep computation tractable, iQIYI builds a hierarchical inventory model: core dimensions (platform, region, channel) are allocated first, followed by independent handling of long‑tail dimensions using orthogonal approximations.

Frequency capping further complicates allocation. The system models the effective available inventory under different frequency limits and incorporates these constraints into the deduction order, ensuring that high‑frequency‑capped orders do not exhaust inventory needed by other orders.

The “AI Band‑Aid” dynamic point solution uses AI to tag video frames with entities such as celebrities, scenes, and spoken words. A two‑stage pipeline first extracts key frames and generates tags via deep models, storing them in HBase. A pre‑allocation service then filters tags based on business rules (e.g., max one ad per minute, no duplicate exposure), commercial value, historical sell‑through, and inventory. The filtered tags are fed to a real‑time allocation service that optimizes for minimal shortage (inventory side) and maximal inquiry volume (advertiser side). Pre‑allocation runs in batch (2‑3 hours for full data, ~30 minutes for incremental updates) and uses greedy or dynamic‑programming algorithms to maximize total tag value.

Overall, the talk covers the full stack of iQIYI’s brand‑advertising technology: from business processes and statistical/ML models for inventory forecasting, through scalable allocation algorithms, to AI‑driven dynamic ad placement.