GATS: Generative Audience Targeting System for Online Advertising

Cong Jiang, Zhongde Chen, Bo Zhang, Yankun Ren, Xin Dong, Lei Cheng, Xinxing Yang, Longfei Li, Jun Zhou, Linjian Mo

Abstract: This paper presents GATS (<u>G</u>enerative <u>A</u>udience <u>T</u>argeting <u>S</u> ystem for Online Advertising), a new framework using large language models (LLMs) to improve audience targeting in online advertising. GATS overcomes the shortcomings of rule-based, look-alike, and graph-based methods by facilitating flexible and interpretable audience criteria expression. The framework integrates intent recognition, knowledge mining, and Data Management Platform (DMP) mapping to translate advertiser demands into actionable user tags and correlate them within a DMP. A small, white-box model called LightGATS (base on QWen-14B), fine-tuned with a high-quality LLM corpus, ensures the framework's safety and efficiency, operating within a scalable hybrid online-offline architecture. GATS's effectiveness is validated through extensive experiments, marking a significant advancement in audience targeting technology.