Date : June 1, 2023, 1:30 p.m. - Room :Amphi 3 - Pôle commun
Computing Shapley Values as Explanation Scores in Data Management and Machine Learning
Leopoldo BERTOSSI, Professeur - SKEMA Business School, Montreal
The Shapley value, originally introduced in coalition game theory, has found applications in data management and machine learning, as an explanation score. With it one can assign numerical values to database tuples of feature values, depending on whether we observe a result of a query or a classification, to quantify the relevance of the tuple or the feature value in relation to this result. The Shapley value has a high computational complexity in general. In this presentation we will show relevant cases where its computation turns out to be tractable, and we will show applications of the efficient computation algorithms, in particular to explain outcomes from neural networks used as classifiers.
Bio: Leopoldo Bertossi is a professor at the SKEMA Business School (Montreal, Canada). He is an Emeritus Professor of the School of Computer Science, Carleton University (Ottawa, Canada), a Senior Researcher at the "Millennium Institute on Foundations of Data » (IMFD, Chile), and a a Principal Investigator at the « National Basal Center for AI Research" (Cenia, Chile). He has a previous affiliation with the "Universidad Adolfo Ibanez » (UAI, Santiago, Chile) as a Full Professor at the Faculty of Engineering and Sciences, and a Senior UAI Fellow. He has also been a Senior Computer Scientist at RelationalAI Inc., and a professor at the Pontifical Catholic University of Chile (PUC, Chile). He has PhD in Mathematics by the PUC, Chile. His broad research interests are related to Data Science and Artificial Intelligence; more specifically, he has done research on explainable AI, logical and uncertain knowledge representation and reasoning, data management, computational logic, ontological representations and reasoning, causality, and machine learning.
Personal web page: www.scs.carleton.ca/~bertossi