Analysing similarity assessment in feature-vector case representations

Abstract

Case-Based Reasoning (CBR) is a good technique to solve new problems based in previous experience. Main assumption in CBR relies in the hypothesis that similar problems should have similar solutions. CBR systems retrieve the most similar cases or experiences among those stored in the Case Base. Then, previous solutions given to these most similar past-solved cases can be adapted to fit new solutions for new cases or problems in a particular domain, instead of derive them from scratch. Thus, similarity measures are key elements in obtaining reliable similar cases, which will be used to derive solutions for new cases. This paper describes a comparative analysis of several commonly used similarity measures, including a measure previously developed by the authors, and a study on its performance in the CBR retrieval step for feature-vector case representations. The testing has been done using six-teen data sets from the UCI Machine Learning Database Repository, plus two complex environmental databases.