Abstract
Many classification techniques have been successfully applied to credit scoring tasks. However, using them blindly may lead to unsatisfactory results. Generally, credit datasets are large and are characterized by redundant features and nonrelevant data. Hence, classification techniques and model accuracy could be hampered. To overcome this problem, this study explores a variety of filter and wrapper feature selection methods for reducing nonrelevant features. We argue that these two types of selection techniques are complementary to each other. A fusion strategy is then proposed to sequentially combine the ranking criteria of multiple filters and a wrapper method. Evaluations on three credit datasets show that feature subsets selected by fusion methods are either superior to or at least as adequate as those selected by individual methods.