A Co-evolutionary Framework for Nearest Neighbor Enhancement: Combining Instance and Feature Weighting with Instance Selection

The nearest neighbor rule is one of the most representative methods in data mining. In recent years, a great amount of proposals have arisen for improving its performance. Among them, instance selection is highlighted due to its capabilities for improving the accuracy of the classifier and its efficiency simultaneously, by editing noise and reducing considerably the size of the training set. It is also possible to remark the role of feature and instance weighting as outstanding methodologies for improving further the performance of the nearest neighbor rule. In this work we present a new co-evolutionary algorithm for combining the former techniques. Its performance is compared with evolutionary approaches performing instance selection, instance weighting and feature weighting in isolation, as well as with the nearest neighbor classifier. The results obtained, contrasted through nonparametric statistical tests, supports the capabilities of co-evolution as a outstanding strategy for joining several proposals for enhancing the nearest neighbor rule.