Une approche visuelle pour l'exploration des résultats d'un grand nombre de classifieurs
Abstract
We propose a visual method to analyze a large number of supervised learning results, and we illustrate it on a real world application (detection of autistic disorders from eye tracking data) in which we were able to obtain more than 30,000 classification results. Each "representation-classifier-parameters" triplet is positioned in 2D with a dimension reduction algorithm (MDS or t-SNE) on the basis of a distance using the probability of the classes predicted for each data. With different settings of this visualization, we show that the user can visually observe and evaluate the effectiveness of the representations, classifiers and their settings.