DbISCDP: An Empirical Study on Distance-Based Nearest Neighbor Approaches for Credit Default Prediction

Financial institutions usually use expert systems based on artificial intelligence to make credit lending a successful venture. However, the problem is made worse because these systems are typically not tailored to the lender's requirements and the instances are frequently quite diverse. Because of this, understanding the existing data is crucial for better insight. Since distance measures play a pivotal role in machine learning analytics, the focus has been given to the distance metrics utilized in the k-Nearest Neighbour (k-NN) algorithm. As the underlying structure of the data used to differ for datasets, focus has been given primarily to credit risk datasets. As no prior studies have been done on finding optimal distance metrics for credit risk datasets, the objective of the study is to find it. An appropriate distance metric for credit risk datasets can accurately determine the closest neighboring points which will help in decision making. For this, seven distance metric families with thirteen different distance measures have been considered in this scope. To get better insight, an expert system was proposed named Distance-based Intelligent System for Credit Default Prediction or DbISCDP utilizing lO-fold cross-validation for training and testing. In terms of average accuracy, Cosine and Hassanat distances provide better results for German and Australian benchmark datasets, respectively.