On Effective Decomposition of Training Data Sets for Min-Max Modular Classifier

Our previous work shows that traditional randomization partition method for min-max modular (M) classifier can not ensure stable generalization accuracy when the number of two-class problems increases. To overcome this drawback, we consider how to effectively decompose the training data set for a two-class problem in this paper. We propose four basic clustering and anti-clustering strategies and their combinations for partitioning training data sets. These four basic strategies are hyperplane decomposition, K-means algorithm, anti-K-means algorithm, and scatter procedure. Our experimental results show that all the proposed clustering and anti-clustering strategies are superior to the traditional random partition method.