In recent years, classifier combination has been of great interest for the pattern recognition community as a method to improve classification performance. Several combination rules have been proposed based on maximizing the accuracy and the Area under the ROC curve (AUC). Taking into account that there are several applications which focus only on a part of the ROC curve, i.e. the one most relevant for the problem, we recently proposed a new algorithm aimed at finding the linear combination of dichotomizers which maximizes only the interesting part of the AUC. Since the algorithm uses a greedy approach, in this paper we define and evaluate some possible strategies which select the dichotomizers to combine at each step of the greedy approach. An experimental comparison is drawn on a multibiometric database.
Selection Strategies for pAUC-based Combination of Dichotomizers
RICAMATO, Maria Teresa;MOLINARA, Mario;TORTORELLA, Francesco
2011-01-01
Abstract
In recent years, classifier combination has been of great interest for the pattern recognition community as a method to improve classification performance. Several combination rules have been proposed based on maximizing the accuracy and the Area under the ROC curve (AUC). Taking into account that there are several applications which focus only on a part of the ROC curve, i.e. the one most relevant for the problem, we recently proposed a new algorithm aimed at finding the linear combination of dichotomizers which maximizes only the interesting part of the AUC. Since the algorithm uses a greedy approach, in this paper we define and evaluate some possible strategies which select the dichotomizers to combine at each step of the greedy approach. An experimental comparison is drawn on a multibiometric database.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
