A composite classification scheme is proposed by combining several classifiers with distinctly different design methodologies. The classifiers are selected from a number of state of the art pattern classification schemes with a view to obtain superior performance. In this scheme, no a priori information except a set of pre-classified data is assumed to be available. By using distinctly different classifiers, the common mode data misclassification is reduced. Traditionally, after the design and evaluation phase, the pre-classified data is discarded. In this scheme, however, the misclassified data from each classifier in the training set is tagged and stored with a view to weight the decisions of the classifiers. If a given test sample is close to a misclassified data cluster of a particular classifier, then the decision made by this classifier is given a lower weighting. The final decision is made by analysing the weighted combination of individual classifier decisions. The proposed algorithm is evaluated on both simulated data and on a biological cell classification problem and it is shown that improved accuracy is obtained when compared to that of the most accurate classifier.
[1]
Peter E. Hart,et al.
Pattern classification and scene analysis
,
1974,
A Wiley-Interscience publication.
[2]
Sandro Ridella,et al.
Circular backpropagation networks for classification
,
1997,
IEEE Trans. Neural Networks.
[3]
Roberto Battiti,et al.
Democracy in neural nets: Voting schemes for classification
,
1994,
Neural Networks.
[4]
David B. Skalak,et al.
Prototype Selection for Composite Nearest Neighbor Classifiers
,
1995
.