A Comprehensive Overview of Computational Advertising: Architecture, Deep‑Learning Evolution, and Future Directions

The author, Wang Zhe, reflects on his four‑year experience in advertising systems and four years in recommendation systems, using this perspective to give a systematic review of computational advertising and its recent deep‑learning driven changes.

He proposes that understanding the oCPM (Optimized Cost Per Mille) system—where the cost per impression is calculated as CPA × pCTR × pCVR × pacingFactor —provides a sufficient “superset” view of advertising technology, encompassing CTR, CVR, and pacing modules that are essential for effective ad delivery.

A classic advertising architecture is presented, highlighting three major challenges: integration with external exchanges and data sources, the need for highly accurate CTR/CVR predictions, and complex budget‑flow matching. The architecture is divided into four parts: data engineering, algorithm flow (including targeting, budgeting, pacing, and bidding), core models (CTR, CVR, pacing), and model engineering.

The evolution of each module in the deep‑learning era is discussed. Ad ranking has shifted from large‑scale model replacement to fine‑grained optimization, with trends toward Transformer‑based encoders (e.g., BERT4Rec) and graph‑based models (e.g., RippleNet, KGAT). Model compression (knowledge distillation) and multi‑task learning are identified as current opportunities.

Federated learning is introduced as a privacy‑preserving approach for training models across data owners, with examples from Alibaba (FederatedScope) and WeBank (FATE). The main drawback is the high technical barrier for smaller participants.

Pacing, described as the “hidden core” of ad systems, controls the spend rate using a PID‑style controller. The classic pacing formula is shown and explained, emphasizing the need to balance budget consumption against real‑time spend.

Bidding is treated as a game‑theoretic problem. The article outlines the construction of a bid‑landscape, the use of deep models (DLF) to predict win probabilities, and a profit‑maximization objective under budget constraints, noting that practical solutions often rely on greedy or heuristic strategies.

Calibration corrects systematic bias in CTR/CVR predictions by applying monotonic regression or post‑training adjustments, ensuring that model outputs align with observed post‑click conversion rates.

Conversion‑delay handling is explored, including delayed‑feedback models (DFM) that jointly train CVR and delay estimators via EM, as well as simpler pipelines that treat early clicks as negatives and later conversions as positives, followed by calibration to mitigate bias.

The author concludes with thoughts on the future growth of advertising algorithm engineers, emphasizing the importance of mastering both high‑level system design and low‑level data‑driven details, and encourages continuous, problem‑oriented learning.

Brief author bio: Wang Zhe is a well‑known Zhihu author and the writer of “Deep Learning Recommender Systems,” with extensive experience in both recommendation and advertising algorithms.

DataFun, the platform hosting this series, celebrates its 5‑year anniversary by publishing technical articles on big data and AI, and will launch a data‑intelligence knowledge map video series.