Abstract
This paper focuses on Wrapper-based feature selection in a credit scoring context. Wrapper Feature selection methods are guided by two important aspects: the evaluation classifier and the search strategy. Here we focus on improving Wrapper approach from the perspective of search strategy. Different search strategies can be applied in Wrapper process such as exhaustive and heuristic. Exhaustive search give an optimal subset, but it is time consuming. Heuristic in the other hand are easier to put into practice but don't always achieve an optimal result. Hence, we propose a more sophisticated version that balance complete and heuristic search shortcomings by combining them. Four real-world datasets were used to demonstrate the benefits and properties of our approach.