Semi-Supervised Pattern Classification Utilizing Fuzzy Clustering and Nonlinear Mapping of Data

We present a semi-supervised algorithm for classification of arbitrarily distributed patterns. We project data into a classification space through two stages, first is a nonlinear mapping with radial basis functions and second is a linear projection with a semi-supervised locality preserving projection. Radial basis functions are arranged by fuzzy clustering of training data. This fuzzy clustering is also exploited for selection of data to be labeled for semi-supervised learning. We devise a simple semi-supervised algorithm in which data similarity is multiplicatively modulated on the basis of label information. We examine performance of the proposed classifier with experiments for synthetic and some real data and show that our method outperforms similar graph spectral algorithms and kernel semi-supervised methods.