Clustering and classification of cases using learned global feature weights

Case-base Reasoning (CBR) systems have attracted great attention in the last few years. It is a system that allows user to store, share and reuse what has been stored inside the system (i.e., the previous problem solving experiences/cases). It is similar to the way we solve unknown problems by using our experience and knowledge. When the case-base size increases, it is very difficulty to maintain the case-base, for example when similar cases have accumulated, anomalies such as redundant cases, conflicting cases, ambiguous cases, subsumed cases and unreachable cases may exist in the case-base. This is the so called case-base maintenance (CBM) problem. In this paper we propose a method to improve the performance of clustering and classification of cases in a large-scale case-base by using a learned global feature weight methodology. This methodology is based on the idea that we could use similarity measure to find several concepts (clusters) in the problem-domain such that those cases in a cluster are closely related among themselves while among different clusters those cases are farther apart. It has been demonstrated in the experiment that the performance of clustering with learned global feature weights is much better than the performance without global feature weights in terms of the retrieval efficiency and accuracy of solution provided by the system.