Team : LFI
Arrival date : 02/01/2017 Localisation : Campus Pierre et Marie CurieSorbonne Université - LIP6 Boîte courrier 169 Couloir 26-00, Étage 5, Bureau 516 4 place Jussieu 75252 PARIS CEDEX 05 FRANCE Tel: +33 1 44 27 87 26, Thibault.Laugel (at) nulllip6.fr http://webia.lip6.fr/~laugel/
Supervision : Christophe MARSALA Co-supervision : LESOT Marie-Jeanne, DETYNIECKI Marcin
Interpretability for human-friendly machine learning models
In this thesis, we consider the problem of generating predictive explanations for a supervised classification model. The paradigm considered, called "post-hoc", studies the particular case where the model to be explained is treated as a "black box", i.e. no information is available about the classifier nor the data used to train it. Although practical since it allows a more flexible use (the same explanation system can be applied to any classification algorithm), the question arises as to the link between the generated explanations and the training data, to which the system did not have access. In this context, we are interested in different problems that these systems may face: the locality of the explanations, their link with the truth on the ground, and their stability.
This analysis is performed through the study of a particular type of approach: counterfactuals, which aim at identifying the minimum perturbation to apply to an observation in order to change the class predicted by the model.
Th. Laugel, M.‑J. Lesot, Ch. Marsala, X. Renard, M. Detyniecki : “Comparison-based Inverse Classification for Interpretability in Machine Learning”, 17th International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems (IPMU 2018), Information Processing and Management of Uncertainty in Knowledge-Based Systems. Theory and Foundations, Cadix, Spain, pp. 100-111, (Springer Verlag) (2018)