Correlation-based similarity metrics in MBR for ordered data

Abstract

To improve the precision of memory-based reasoning (MBR) for ordinal data, this article presents correlation-based similarity metrics. The basic idea of this research is an intuitive assumption: if the correlation between the answer and one feature in a sample data set is large, the weight of this feature for predication should be large. To validate this proposal, we promote “leave-one-out cross-validation” for 53 examples which were collected from Japanese client companies who outsource software development to vendor companies. Three measures, including mean absolute error, variance of error, and precision, are compared among the proposed methods: the per-category feature importance (PCF), the per-feature category importance (PFC), the averaged category feature importance (ACF), and the cross-category feature importance (CCF). The Wilcoxon matched-pairs signed ranks test is also discussed.