A gene is treated as a unit of heredity in a living organism. It resides on a stretch of DNA. Gene Regulatory Network (GRN) is a network of transcription dependency among genes of an organism. A GRN can be inferred from microarray data either by unsupervised or by supervised approach. It has been observed that supervised methods yields more accurate result as compared to unsupervised methods. Supervised methods require both positive and negative data for training. In Biological literature only positive example is available as Biologist are unable to state whether two genes are not interacting. A common adopted solution is to consider a random subset of unlabeled example as negative. Random selection may degrade the performance of the classifier. It is usually expected that, when labeled data are limited, the learning performance can be improved by exploiting unlabeled data. In this paper we propose a novel approach to filter out reliable and strong negative data from unlabeled data, so that a supervised model can be trained properly. We tested this method for predicting regulation in E. Coli and observed better result as compared to other unsupervised and supervised methods. This method is based on the principle of dividing the whole domain into gene clusters and then finds the best informative cluster for further classification. © 2013 Springer-Verlag.