Class majority in designing euclidean fuzzy local approximation NN for overlapping data in pattern classification

To deal with a problem of overlapping data in pattern classification, a class majority method in designing hidden units of Fuzzy Local Approximation NN is proposed. Moreover, to improve the output confidence of the networks, Euclidean fuzzy similarity is proposed as hidden unit operator. For each cluster formed by the adaptive clustering of F-SONIA, the number of vectors that belong to the same class is calculated. Therefore the fractions of each class are known. One class should have a single class majority in the cluster. Then, the cluster with no single majority is broken down into two clusters. In experiments, the real-world benchmark datasets, e.g., 2-D vowel, Iris, and thyroid data that have different challenges to the networks in terms of overlapping and size are used to test the networks. Experiments show that the proposed methods improve the classification performance as well as the output confidence of the networks.