Hasler foundation (http://www.haslerstiftung.ch/)









Model averaging approaches for credible classification
A general problem in the design of any classifier is feature selection, i.e., the identification of the supposedly optimal set of attributes; it has been shown that using Bayesian Model Averaging (BMA) to statistically aggregate the output of several classifiers, characterized by different feature sets, leads to better accuracy than relying on a single classifier, characterized by the supposedly bestfeature set. Yet, currently there exist no BMA techique to aggregate credal classifiers. The first goal of the project is hence the development of BMA for the Naive Credal Classifier (NCC); this will allow for statistically aggregating the output of different NCCs, characterized by different feature sets. BMA for NCC aims hence at avoid the overfitting arising from the choice of a single feature set (thanks to BMA) and at dealing robustly with the specification of the prior of each single classifier (thanks to credal approach). BMA for NCC is going to be developed in an analytical way, i.e. its formulas will be exact and its computation will be fast. The second goal of the project is to address a general issue of BMA, i.e. its need for the specification of a single, precise prior distribution over the different classifiers (which actually represents the a priori belief of the investigator about the credibility of the different classifiers). In fact, on small data sets, the classification issued via BMA might be sensitive on the specification of such prior. To overcome this drawback, we plan to extend BMA itself to imprecise probabilities, thus developing Credal Model Averaging (CMA). CMA will hence regard as feasible a set of priors over the models, rather than a single prior. Eventually, the developed methodologies will be extensively validated on real world data sets.
People Giorgio Corani

Marco Zaffalon



Copyright © 2003 - 2012 IDSIA. All Rights Reserved.