Gene selection is an important issue in microarray data processing. Microarray gene expression data usually consists of a large amount of genes. Spectral Bi-clustering is used for the selection of publically available datasets. The existing work semi unsupervised gene selection method finds much smaller and informative gene subsets without class information a priori. It uses gene ranking and gene combination selection methods of semi unsupervised gene selection where the gene combinations are selected based on the similarity between genes. The time efficiency in this method is very high. Compared to the previous work, our method can make accurate predictions with smaller gene subsets and able to identify the cancer in a single or two gene combinations. In this paper, we study a new and efficient semi-unsupervised gene selection method which results in much smaller gene subsets without prior subtype knowledge. Also it reduces the number of genes combinations.