Abstract

Case-based reasoning (CBR) is a methodology for problem solving and decision-making in complex and changing business environments. Many CBR algorithms are derivatives of the k-nearest neighbor (k-NN) method, which has a similarity function to generate classification from stored cases. Several studies have shown that k-NN performance is highly sensitive to the definition of its similarity function. Many k-NN methods have been proposed to reduce this sensitivity by using various distance functions with feature weights.

This paper proposes an analogical reasoning structure for feature weighting using a new framework called the analytic hierarchy process (AHP)-weighted k-NN algorithm. The paper also introduces AHP methodology for assigning relative importance in case indexing and retrieving. The AHP model is a methodology effective in obtaining domain knowledge from numerous experts and representing knowledge-guided indexing. The proposed AHP weighted k-NN algorithm has been shown to achieve classification accuracy higher than the pure k-NN algorithm. This approach is applied to bankruptcy prediction involves the examination of several criteria, both quantitative (financial ratios) and qualitative (non-financial variables).

Author Keywords: Case-based reasoning; Analytic hierarchy process; Feature weights; Bankruptcy prediction

Article Outline

1. Introduction

2. Research background

2.1. Classification techniques for case-based reasoning

2.1.1. Case retrieval and indexing using k-nearest neighbor

2.1.2. Review of previous feature weighting methods

2.2. The analytic hierarchy process approach

3. k-Nearest neighbor approach with AHP feature weights

3.1. The AHP modeling for bankruptcy prediction

3.2. k-Nearest neighbor retrieval and indexing with AHP feature weights

4. The bankruptcy prediction

4.1. Experimental design

4.2. Data and variables

4.3. Empirical results

4.3.1. Result of feature weights using AHP

4.3.2. The result of bankruptcy prediction

5. Conclusion and remarks

Acknowledgements

References

Corresponding author. Tel.: +82-2-958-3673; fax: +82-2-958-3604; email: cspark@kgsm.kaist.ac.kr