Keynote Talk
"Understanding Feature Importance in Explainable AI"
Abstract: Explaining a prediction often means identifying which features matter. Popular methods such as SHAP address this question by assigning numerical scores, while formal approaches identify conditions that guarantee the prediction. These perspectives can lead to surprisingly different conclusions.
In this talk, I will present a mathematical framework that connects these approaches and helps explain their disagreements. Drawing on ideas from Fourier analysis, I will discuss what feature importance scores tell us about a prediction and how they relate to logical explanations. I will also present recent results on when these approaches agree, how extreme their disagreements can become, and how frequently they arise.
Biography
Marcelo Arenas is a Professor in the Department of Computer Science and the Institute for Mathematical and Computational Engineering at Pontificia Universidad Católica de Chile. He is an ACM Fellow and a former director of the Millennium Institute for Foundational Research on Data and the Center for Semantic Web Research. He received his Ph.D. in Computer Science from the University of Toronto in 2005. His research interests include data management, applications of logic in computer science, and explainable artificial intelligence. His research has received several distinctions, including an IBM Ph.D. Fellowship, a SIGMOD Jim Gray Doctoral Dissertation Award Honorable Mention, the Semantic Web Science Association Ten-Year Award, an ACM SIGMOD Research Highlight Award, and nine best paper awards at major conferences. He has also served on numerous program committees and editorial boards and has participated as an invited expert in the W3C and the OECD.