题目: Using Contextual Embeddings to Predict the Effectiveness of Novel Heterogeneous Treatments
时间: 2025年11月19日10:00-11:30
地点: 管理学院 A523
主讲人: 曾广, Nova School of Business and Economics 助理教授
主持人: 张政，环球网页版“百人计划研究员”
主讲人简介：
Guang Zeng is an Assistant Professor of Management (Quantitative Marketing Group) at Nova School of Business and Economics (NovaSBE), Universidade Nova de Lisboa. Guang joined NovaSBE after graduating from the University of Rochester (Simon Business School) in 2025. His research focuses on Empirical Industrial Organization and Quantitative Marketing, leveraging causal machine learning, structural modeling to solve targeted marketing problems, and information design problems.
讲座摘要：
We propose a framework for causal prediction that enables the design and evaluation of novel treatments based on estimated causal effects. Our approach centers on constructing a response surface that connects unstructured text--captured through contextual embeddings--to economically meaningful outcomes, operationalized as doubly robust scores. This response surface then functions as a critic within an actor-critic algorithm, where a GenAI engine like ChatGPT can serve as the actor. To evaluate our framework, we examine a targeted marketing application involving 3.3 million observations from 34 customer emails over 45 days. Our results highlight the use of contextual embeddings to achieve an accurate prediction of novel treatments while a commonly used encoder, human codification, fails to do so. The accuracy of our framework in prediction of novel treatments is driven by the availability of semantically similar treatments within the contextual manifold. Finally, we illustrate how managers can use our framework to generate novel content, demonstrating that it outperforms a naive application of Generative AI. Moreover, our framework can be extended to account for heterogeneous treatment effects, allowing for segmentation and personalized targeting in developing novel treatments.
欢迎广大师生前来交流！
