Analysing Rough Sets weighting methods for Case-Based Reasoning Systems

Case-Based Reasoning systems retrieve cases using a similarity function based on the K-NN or some derivatives. These functions are sensitive to irrelevant, interacting or noisy features. Many similarity functions weigh the relevance of features to avoid this problem. This article proposes two weighting methods based on Rough Sets theory: Proportional Rough Sets and Dependence Rough Sets. Both weighting methods use the representative knowledge extracted from the original data to compute the feature relevance using two different policies. The first one computes the proportional participation of the features in the representative knowledge. The second one computes the dependence of each feature in the representative knowledge. This dependence denotes if a feature is superfluous within the knowledge. Experiments using different domains show that weighting methods based on Rough Sets maintain or even improve the classification accuracy of Case-Based Reasoning Systems, compared to non-weighting approaches or well-known weighting methods.