Deep Learning Technologies Applied to Sogou Search Advertising

Shupeng, a researcher from Sogou Search Advertising, introduces the background of search advertising and the overall workflow, from user query and commercial intent to ad recall, CTR estimation, ranking, and rendering.

The presentation focuses on two typical applications of deep learning: automated creative generation and click‑through‑rate (CTR) prediction.

For automated creative, the process includes creative source mining (user behavior analysis, landing‑page analysis, purchase behavior analysis) and creative generation (using models such as GAN, seq2seq, CVAE). A CVAE‑based framework is described for keyword expansion, with domain classification and reinforcement‑learning‑based reward estimation to control diversity.

Creative source mining also involves session segmentation using an LSTM‑Attention model, which was published at SIGIR 2018.

Text relevance computation is discussed, covering traditional string matching (TF‑IDF, BM25), semantic matching (LDA, entity recognition), intent matching (click‑behavior graphs, SVD++), and deep‑learning approaches (seq2seq, neural machine translation).

CTR prediction models have evolved from early liblinear‑based linear models (2008) to LR, FTRL, FM, GBDT, DNN, and later to wide‑&‑deep and NFM architectures. Model training uses massive query and click logs, with feature extraction for queries, ads, and matching signals.

The speaker describes the online deployment pipeline, including feature extraction, model inference, and ranking, as well as model fusion techniques such as LR + DNN cascades and model‑feature integration.

Evaluation metrics include AUC improvement and online revenue gain, with emphasis on debugging, survival bias, and feature coverage when moving from offline to online.

Future plans involve scaling wide‑&‑deep models with low‑power GPUs, exploring more complex architectures like NFM, and improving data utilization and system infrastructure.