ROC-Based Evolutionary Learning: Application to Medical Data Mining
Résumé
A novel way of comparing supervised learning algorithms has been introduced since the late 90’s, based on Receiver Operating Characteristics (ROC) curves. From this approach is derived a NP complete optimization criterion for supervised learning, the area under the ROC curve. This optimization criterion, tackled with evolution strategies, is experimentally compared to the structural risk criterion tackled by quadratic optimization in Support Vector Machines. Comparable results are obtained on a set of benchmark problems in the Irvine repository, within a fraction of the SVM computational cost. Additionally, the variety of solutions provided by evolutionary computation can be exploited for visually inspecting the contributing factors of the phenomenon under study. The impact study and sensitivity analysis facilities offered by ROGER (ROC-based Genetic LearneR) are demonstrated on a medical application, the identification of Atherosclerosis Risk Factors.
Domaines
Informatique [cs]Origine | Fichiers produits par l'(les) auteur(s) |
---|