Case-based reasoning ensemble and business application: A computational approach from multiple case representations driven by randomness

Abstract

Case-based reasoning (CBR) holds the unique capability of making predictions as well as suggestions to corporate executives and organizational decision-makers. How to improve its predictive performance is critical. This research aims to explore an ensemble of CBR from multiple case representations as an alternative to traditional approaches, which aims to produce lower errors than its member CBR predictors and independent CBR predictors and produce better performance in business failure prediction (BFP). This method is to base the member CBR predictors on randomly generated feature subsets in order to produce diversity in them. As a result, the CBR ensemble needs not to consider the two difficult/challenging tasks in BFP, i.e., the optimization of a single CBR and the search of optimal single CBR for a specific problem. We statistically validated the results of the CBR ensemble by comparing them with those of multivariate discriminant analysis, logistic regression, and the classical CBR algorithm. The results from Chinese short-term BFP indicate that the CBR ensemble significantly improves predictive ability of CBR.