Distill n' Explain

dc.contributorAalto Universityen
dc.contributor.authorPereira, Tamaraen_US
dc.contributor.authorNascimento, Eriken_US
dc.contributor.authorResck, Lucas E.en_US
dc.contributor.authorMesquita, Diegoen_US
dc.contributor.authorSouza, Amaurien_US
dc.contributor.departmentFederal Institute of Education, Science and Technology of Cearáen_US
dc.contributor.departmentFundação Getúlio Vargasen_US
dc.contributor.departmentDepartment of Computer Scienceen_US
dc.contributor.editorRuiz, Franciscoen_US
dc.contributor.editorDy, Jenniferen_US
dc.contributor.editorvan de Meent, Jan-Willemen_US
dc.descriptionFunding Information: This work was supported by the Silicon Valley Community Foundation (SVCF) through the Ripple impact fund, the Fundac¸ão de Amparo à Pesquisa do Estado do Rio de Janeiro (FAPERJ), the Fundac¸ão Cearense de Apoio ao Desenvolvimento Científico e Tecnológico (FUNCAP), the Coordenac¸ão de Aperfeic¸oamento de Pessoal de Nível Superior (CAPES), and the Getulio Vargas Foundation’s school of applied mathematics (FGV EMAp). Publisher Copyright: Copyright © 2023 by the author(s)
dc.description.abstractExplaining node predictions in graph neural networks (GNNs) often boils down to finding graph substructures that preserve predictions. Finding these structures usually implies back-propagating through the GNN, bonding the complexity (e.g., number of layers) of the GNN to the cost of explaining it. This naturally begs the question: Can we break this bond by explaining a simpler surrogate GNN? To answer the question, we propose Distill n' Explain (DnX). First, DnX learns a surrogate GNN via knowledge distillation. Then, DnX extracts node or edge-level explanations by solving a simple convex program. We also propose FastDnX, a faster version of DnX that leverages the linear decomposition of our surrogate model. Experiments show that DnX and FastDnX often outperform state-of-the-art GNN explainers while being orders of magnitude faster. Additionally, we support our empirical findings with theoretical results linking the quality of the surrogate model (i.e., distillation error) to the faithfulness of explanations.en
dc.description.versionPeer revieweden
dc.identifier.citationPereira , T , Nascimento , E , Resck , L E , Mesquita , D & Souza , A 2023 , Distill n' Explain : explaining graph neural networks using simple surrogates . in F Ruiz , J Dy & J-W van de Meent (eds) , Proceedings of The 26th International Conference on Artificial Intelligence and Statistics (AISTATS) 2023 . Proceedings of Machine Learning Research , vol. 206 , JMLR , pp. 6199-6214 , International Conference on Artificial Intelligence and Statistics , Valencia , Spain , 25/04/2023 . < https://proceedings.mlr.press/v206/pereira23a.html >en
dc.identifier.otherPURE UUID: 67a21ecf-41f9-45e3-b2ca-4244c4841c5aen_US
dc.identifier.otherPURE ITEMURL: https://research.aalto.fi/en/publications/67a21ecf-41f9-45e3-b2ca-4244c4841c5aen_US
dc.identifier.otherPURE LINK: http://www.scopus.com/inward/record.url?scp=85164203717&partnerID=8YFLogxKen_US
dc.identifier.otherPURE LINK: https://proceedings.mlr.press/v206/pereira23a.htmlen_US
dc.identifier.otherPURE FILEURL: https://research.aalto.fi/files/125301716/SCI_Pereira_etal_AISTATS_2023.pdfen_US
dc.relation.ispartofInternational Conference on Artificial Intelligence and Statisticsen
dc.relation.ispartofseriesProceedings of The 26th International Conference on Artificial Intelligence and Statistics (AISTATS) 2023en
dc.relation.ispartofseriesProceedings of Machine Learning Researchen
dc.relation.ispartofseriesVolume 206en
dc.titleDistill n' Explainen
dc.typeConference article in proceedingsfi