UCSL : A Machine Learning Expectation-Maximization framework for Unsupervised Clustering driven by Supervised Learning

Subtype Discovery consists in finding interpretable and consistent sub-parts\nof a dataset, which are also relevant to a certain supervised task. From a\nmathematical point of view, this can be defined as a clustering task driven by\nsupervised learning in order to uncover subgroups in line with the supervised\nprediction. In this paper, we propose a general Expectation-Maximization\nensemble framework entitled UCSL (Unsupervised Clustering driven by Supervised\nLearning). Our method is generic, it can integrate any clustering method and\ncan be driven by both binary classification and regression. We propose to\nconstruct a non-linear model by merging multiple linear estimators, one per\ncluster. Each hyperplane is estimated so that it correctly discriminates - or\npredict - only one cluster. We use SVC or Logistic Regression for\nclassification and SVR for regression. Furthermore, to perform cluster analysis\nwithin a more suitable space, we also propose a dimension-reduction algorithm\nthat projects the data onto an orthonormal space relevant to the supervised\ntask. We analyze the robustness and generalization capability of our algorithm\nusing synthetic and experimental datasets. In particular, we validate its\nability to identify suitable consistent sub-types by conducting a\npsychiatric-diseases cluster analysis with known ground-truth labels. The gain\nof the proposed method over previous state-of-the-art techniques is about +1.9\npoints in terms of balanced accuracy. Finally, we make codes and examples\navailable in a scikit-learn-compatible Python package at\nhttps://github.com/neurospin-projects/2021_rlouiset_ucsl\n