On the Scarcity of Labeled Data

Scarcity of labeled data can be encountered in various engineering applications due to several factors. This raises the question of how to generate sufficient amounts of labeled data when it is sparse in order to build effective learning tools. One approach to overcome this problem is to use unlabeled data. In this paper, we propose two approaches, each is a two-step process for learning from data that is dominantly unlabeled. In the first approach, the k-NN algorithm is applied to pre-label the unlabeled data. A multi-layer perceptron is then used to classify the pre-labeled data. In the second approach, a prototypicality rule based on FCM is used to pre-label unlabeled data before training the MLP classifier. The evaluation, conducted on three data sets, shows how unlabeled data enhances the accuracy of the neural classifier