A constructive approach for classification of Semi-Labeled data by extending the BLTA algorithm

Abstract

In this paper BLTA is extended to tackle the classification of Semi-Labeled data. BLTA works for Labeled data and perceptron based 4-layered neural network structure is formed. In our proposed extension, this 4-layered neural network structure works for classification of Semi-Labeled data, some samples are labeled and some are unlabeled. Learning algorithm is modified to tackle with such samples. The proposed method works in two phases. In first phase labeled samples are used for learning and another phase makes use of unlabeled samples to properly learn them in decided neuron. The proposed algorithm is tested with various benchmark datasets. Results are presented in the form of number of neurons and generalization accuracies. The accuracies are varying from 45 to 98% for different values of M-circle. © 2010 IEEE.