Credit scoring is one of the most challenging research topics that have been a source of many innovative works in banking field. Choosing the appropriate set of features is one of the most interesting and difficult tasks that have a key effect on the performance of credit scoring models. With the huge amount of feature selection techniques and specially ranking techniques for feature selection, rank aggregation techniques become indispensable tools for fusing individual ranked lists into a single consensus list with better performance. However, in some cases the obtained ranking may be noisy or incomplete witch lead to an unsatisfactory final rank. We investigate on this issue by proposing a similarity based algorithm that extends two standard methods of rank aggregation namely majority vote and mean aggregation based on the similarity between the features in the dataset. Evaluations on four credit datasets show that feature subsets selected by the aggregation based similarity technique give superior results to those selected by individual filters and the standard aggregation techniques.
展开▼
