管理科学系学术讲座（4月14日）

时   间：2026年4月14日（周二） 15:00-16:00

地   点：管理学院史带楼502室

主   题： Generating Input Distributions for Explaining Data-Driven Decision Pipelines

主讲人：Ilker Birbil  阿姆斯特丹大学教授

主持人：戴悦 复旦大学管理学院教授

Abstract：The increasing deployment of AI in high-stakes domains has given rise to data-driven decision pipelines — systems in which a predictive model, trained on historical data, feeds its outputs directly into a large-scale constrained optimization process. While explainable AI has made considerable progress in interpreting individual model predictions, explaining the decisions that emerge from the full pipeline remains an open challenge. Understanding why an optimizer produces a particular allocation, plan, or assignment requires going beyond local feature attribution and reasoning about the pipeline as an integrated whole.

We propose a predict-optimize-explain framework that addresses this gap through probing: generating input distributions that answer targeted what-if questions at the decision level. A decision-level event of interest is encoded as a decision-based loss within a Bayesian updating scheme, yielding a Gibbs distribution that is the unique minimizer of an entropy-regularized expected-loss problem. We illustrate the framework on a strategic asset allocation pipeline, showing how probing reveals the macroeconomic conditions consistent with specific portfolio behaviors. We close with directions toward certification and auditing of data-driven decision systems.

Bio：Ilker Birbil is a professor of AI & Optimization Techniques for Business & Society in University of Amsterdam. In the past, he had served for three years as a professor of Data Science and Optimization at the Department of Econometrics of Erasmus University, and before that he had been a professor of optimization at the Industrial Engineering Department of Sabancı University for more than a decade. His research interests center around mathematical programming methods in data science and decision making. Lately, he is working on explainable artificial intelligence and optimization.