A novelty detection machine and its application to bank failure prediction

Abstract

Novelty detection has been well-studied for many years and has found a wide range of applications, but correctly identifying the outliers is still a hard problem because of the diverse variation and the small quantity of such outliers. We address the problem using several distinct characteristics of the outliers and the normal patterns. First, normal patterns are usually grouped together, forming clusters in the high density regions of the data space. Second, outliers are characteristically very different from the normal patterns, and hence tend to be located far away from the normal patterns in the data space. Third, the number of outliers is generally very small in a given dataset. Based on these observations, we can envisage that the appropriate decision boundary segregating the outliers and the normal patterns usually lies in some low density regions of the data space. This is referred to as cluster assumption. The resultant optimization problem to learn the decision function can be solved using the mixed integer programming approach. Following that, we present a cutting plane algorithm together with a multiple kernel learning technique to solve the convex relaxation of the optimization problem. Specifically, we make use of the scarcity of the outliers to find a violating solution to the cutting plane algorithm. Experimental results with several benchmark datasets show that our proposed novelty detection method outperforms existing hyperplane and density estimation-based novelty detection techniques. We subsequently apply our method to the prediction of banking failures to identify potential bank failures or high risk banks through the traits of financial distress.