An Integrated Prototype-Based Learning Algorithm
Gennady Agre
Institute of Information Technologies - Bulgarian Academy of Sciences, 1113 Sofia,
E-mail: agre@iinf.bas.bg
Abstract: The paper discusses a problem for reduction of stored examples needed for instance-based classifiers to run without significant degradation of their classification accuracy. Based on the detailed analysis of the existing approaches, a new prototype based learning classification algorithm (IPBL) is proposed. The algorithm integrates a noise removal component with an example averaging component. The algorithm has been tested on fifteen real world benchmark databases. It has been shown that IPBL predictive accuracy is higher than that of the nearest neighbour algorithm run on the whole training set. The IPBL needs to store as class descriptions some artificially constructed examples (prototypes), which number is only about 3% (in average) of the size the training set. Comparing these parameters with similar results reported for such algorithms allows concluding that the proposed IPBL algorithm may be seen as one of the best existing prototype-based learning algorithms.