Interactive machine learning framework for predicting non-performing bank loans

The aim of this study is to develop an Interactive Machine Learning framework, based on outlier hunting techniques to predict non-performing bank loans. Multivariate Gaussian Anomaly Detection technique was used for the outlier hunting to ensure that every incoming data observation to be learnt (as an update to the interactive classifier) during usage, is not anomalous, malicious, fictitious, or having erroneous measurements. The framework was applied to predict credit facilities such as loans in financial institutions that are not non-performing, and to enable the system keep learning new trends while in use. The loan classification model was trained using Support Vector Machine (SVM) algorithm because of its ability to perform well as observed in literature. The model built from the experiments was evaluated using standard machine learning evaluation metrics which include precision, recall, accuracy, area under the curve (AUC) and was also benchmarked against other models built with different algorithms. Results of the experiments show that SVM model has 96% accuracy which is the highest percentage of AUC obtained among the models trained. This study would enable users to verify the validity of data or feedback in an interactive machine learning environment before deciding whether or not the observations will be used in updating the model. This method would also help reduce the training time in the learning process. This study is significant in guaranteeing financial security of banks for the future and helping to maintain stability in the credit lending rates.