<title>Expanding context against weighted voting of classifiers</title>

The paper describes the use of a context to improve the classification accuracy. The two - dimensional context is considered based on contextual features and contextual examples. The quality function is defined which evaluates the context in the interval (0,1) from the point of view of its effect to classificatio n accuracy. The key idea presented in the paper is the idea of expanding context: if one puts the predictions of the classifiers in order according to the quality of their contexts then it is assumed the existence of a trend among these predictions. This trend can be successfully used to extrapolate the opinions and find the integrated one outside all the opinions in the point of optimal quality. The approach fits well for a continuous outcome classification (regression). One can create the ensemble of classifiers using the same learning algorithm several times in the con texts of different quality. For some families of classifiers, for example Nearest Neighbor, or linear regression, classifier is constructed from its context and is equal to the context. When classifying a new instance one should build the sample set of exa mples (quality - prediction) based on predictions of the classifiers. By extrapolating this set of points one can obtain the value of prediction for maximal (equal to 1) quality. It is experimentally shown that for some datasets the extrapolation of expand ing contexts performs better than voting.