Supervised learning strategies in multi-classifier systems

This paper presents three strategies in order to re-train classifiers in a multi-expert scenario when new labeled data become available. The simplest possibility is the use of the entire new dataset. The second possibility is related to the consideration that each single classifier is able to select new patterns starting from those on which it performs a miss-classification. Finally, the multi expert system behavior can be inspected to select profitable samples. More specifically a misclassified sample, for a particular classifier, is used to update that classifier only if it produces a miss-classification by the ensemble of classifiers. The three approaches are compared under different conditions on two different state of the art performing classifiers by considering the CEDAR (handwritten digit) database. It is shown how results depend by the amount of the new training samples, as well as by the specific combination decision schema.