A Comprehensive Review of Industrial-Scale Deep Learning for Click-Through Rate Prediction in Online Advertising

The author reflects on the development of click‑through‑rate (CTR) estimation techniques from the early shallow‑learning era (2005‑2015) to the current industrial‑scale deep learning era, highlighting key milestones, challenges, and lessons learned.

In the shallow‑learning period, linear models, GBDT, and SVM dominated, with heavy reliance on massive feature engineering and distributed training. Limitations of these models prompted a shift toward non‑linear approaches such as MLR, FM, and hybrid pipelines.

With the advent of deep learning around 2015, the team adopted end‑to‑end neural models, leading to a rapid “red‑line” period (Deep Learning 1.0) where architectures like attention‑based DIN/DIEN, memory networks (MIMN), and SIM dramatically improved CTR performance.

Subsequent innovations focused on full‑library recall technologies: tree‑based TDM, low‑precision vectorized recall, and the hybrid "二向箔" approach that balances model complexity and compute constraints.

Co‑design of algorithms, compute, and system architecture became central, exemplified by the COLD coarse‑ranking framework and the flexible "二向箔" system, which trade off precision and latency to maximize business impact.

The paper also discusses multi‑objective modeling (e.g., ESMM, STAR) and the evolution of bidding and auction mechanisms, including learning‑based auction designs (Deep GSP, DNA) that ensure incentive compatibility while optimizing platform revenue.

Beyond algorithms, the author describes three generations of compute‑efficiency techniques (OFI), a streaming training engine (Bernoulli) that dramatically reduces resource usage, and a unified online AI service architecture for recall and ranking.

Finally, the author reflects on the current “dark period” of diminishing returns, proposes future directions such as fully end‑to‑end DNN pipelines spanning recall to auction, and emphasizes the importance of continued co‑design across algorithms, hardware, and system infrastructure.