B1076
Title: Decision surface Markov chain
Authors: Nicolas Wicker - University of Lille (France) [presenting]
Amael Broustet - University of Lille (France)
Abstract: The purpose is to provide a diagnostic tool for supervised learning algorithms once a predictor is obtained. More specifically, we want to assess how complicated a decision surface is. A lot of theory has been developed to assess the complexity of a class of functions (Rademacher complexity, VC dimension, covering numbers), but to our knowledge, tools are lacking to assess the complexity of a particular decision function. The main idea is to use a Markov chain, sampling asymptotically a uniform distribution and hindering its progress whenever it encounters the decision surface, so that the slower it converges, the more complex the surface is supposed to be. This approach is shown in a theoretical example as well as in real data sets.