Fuzzy C-means clustering based uncertainty measure for sample weighting boosts pattern classification efficiency

The paper presents a fuzzy c-means clustering based fuzzy measure for weighting samples in a dataset for pattern classification. The method improves classification efficiency. The fuzzy c-means generated membership grades of a sample for belonging to different clusters are interpreted as measures of uncertainty for assigning specific crisp class label to this sample. The fuzzy measure of total uncertainty for a sample is defined as U = -Σk=1c Mk log2 Mk where Mk denotes the membership grade in k-th cluster, and the summation extends is over all the c clusters. The data samples in feature space are then transformed according to X → (1 + U)X. By using a radial basis function neural network classifier the classification efficiency is compared based on the original and the transformed feature vectors. Several data sets collected from open sources were used for validation.