Test-cost-sensitive attribute reduction on heterogeneous data for adaptive neighborhood model

Abstract

Test-cost-sensitive attribute reduction is an important component in data mining applications, and plays a key role in cost-sensitive learning. Some previous approaches in test-cost-sensitive attribute reduction focus mainly on homogeneous datasets. When heterogeneous datasets must be taken into account, the previous approaches convert nominal attribute to numerical attribute directly. In this paper, we introduce an adaptive neighborhood model for heterogeneous attribute and deal with test-cost-sensitive attribute reduction problem. In the adaptive neighborhood model, the objects with numerical attributes are dealt with classical covering neighborhood, and the objects with nominal attributes are dealt with the overlap metric neighborhood. Compared with the previous approaches, the proposed model can avoid that objects with different values of nominal attribute are classified into one neighborhood. The number of inconsistent objects of a neighborhood reflects the discriminating capability of an attribute subset. With the adaptive neighborhood model, an inconsistent objects-based heuristic reduction algorithm is constructed. The proposed algorithm is compared with the $\lambda$-weighted heuristic reduction algorithm which nominal attribute is normalized. Experimental results demonstrate that the proposed algorithm is more effective and more practical significance than the $\lambda$-weighted heuristic reduction algorithm.