Developing a Significant Nearest Neighbor Search Method for Effective Case Retrieval in a CBR System

Abstract

Case-based reasoning is a problem-solving technique commonly seen in artificial intelligence. A successful CBR system highly depends on how to design an effective case retrieval mechanism. The K-nearest neighbor (KNN) search method which selects the K most similar prior cases for a new case has been extensively used in the case retrieval phase of CBR. Although KNN can be simply implemented, the choice of the K value is quite subjective and will influence the performance of a CBR system. To eliminate the disadvantage, this research proposes a significant nearest neighbor (SNN) search method. In SNN, the probability density function of the dissimilarity distribution is estimated by the expectation maximization algorithm. Accordingly, the case selection can be conducted by determining whether the dissimilarity between a prior case and the new case is significant low based on the estimated dissimilarity distribution. The SNN search avoids human involvement in deciding the number of retrieved prior cases and makes the retrieval result objective and meaningful in statistics. The performance of the proposed SNN search method is demonstrated through a set of experiments.