Introduction of high-dimensional interpretable machine learning models and their applications. (Introduction de modèles de machine learning interprétables en grande dimension et leurs applications)

Simon Bussy

Abstract: This dissertation focuses on the introduction of new interpretable machine learning methods in a high-dimensional setting. We developped first the C-mix, a mixture model of censored durations that automatically detects subgroups based on the risk that the event under study occurs early; then the binarsity penalty combining a weighted total variation penalty with a linear constraint per block, that applies on one-hot encoding of continuous features; and finally the binacox model that uses the binarsity penalty within a Cox model to automatically detect cut-points in the continuous features. For each method, theoretical properties are established: algorithm convergence, non-asymptotic oracle inequalities, and comparison studies with state-of-the-art methods are carried out on both simulated and real data. All proposed methods give good results in terms of prediction performances, computing time, as well as interpretability abilities.