A New View of Conformity and Its Application to Classification

摘要

The paper discusses a new way of estimating the conformity of an item, described in terms of Boolean-valued features, with respect to a class of items. A usual view of conformity is to compare, feature by feature, the item value with the corresponding distribution of values observed over the class. Then combining the comparison results for the different features yields a global conformity measure. In this paper, the item is rather compared to triples of elements taken in the class: it is checked if the item conforms, over a maximal number of features, to the majority of the elements in each triple. Based on the idea that a new item should be allocated the class to which it conforms the most, a simple classification algorithm is proposed. Experiments on a set of benchmarks show that it is competitive with classical methods.