Forecasting Of Automatic Relevance Determination for Feature Selection (FARD-FS) In Financial Fraud Detection

摘要

Financial Fraud Detection (FFD) is a momentous topic that is high significance in the areas like academic, research and industries. The Ripper method is used to create rule sets for huge datasets with more features. But the classification performance is less due to the occurrence of missing data. Automatic relevance determination is a feature selection technique that improves the classification performance. But the disadvantage in this method is over fitting. In this article an innovative method is introduced called Forecasting of Automatic Relevance Determination based Feature Selection (FARD-FS) for enhancing the classification performance. FARD-FS is an effectual algorithm for feature selection and sparse learning. This FARD method improves the classification performance based on computation of the predictive performance of the classifier. The convex expectation propagation (CEP) algorithm is used to forecast the performance as a side-effect of its iterations. This method divides the features into a number of components. The features are divided as core features and augmented features, select all the core features and take potentials from the augmented features and combine them. This can be accomplished by using the full entropy function H. This method is used in the ripper classification to do feature and to choose data points in a spare Bayesian kernel classifier. An experimental result shows that the proposed FARD method achieves high classification accuracy when compared to the existing ARD feature selection method.