On High Generalization Ability of Test Feature Classifiers

抄録

Test feature classifiers are generated directly from training samples and have a 100% recognition rate on training data. Although this perfect learnability is an important feature of the classifiers, it does not guarantee a good generalization. In this paper, we concentrate on the performance of classifiers on test data, and describe cases when a 100% recognition rate can be achieved. We show that training data can contain information about possible discriminant boundaries between entire classes. In general, it is impossible to extract this information, although we propose a heuristic algorithm which could lead to a 100% recognition rate. To test the performance of the classifiers, we apply them to both artificial and real data. For the real data, we use the well-known breast cancer and satellite image databases. Our experimental results show that the proposed classifiers have not only a high recognition ability, but also confirm the ability of a 100% recognition rate in real classification problems.