Binary Classification with Supervised-like Biclustering and Adaboost

Biclustering based classification methods have achieved great success. In existing biclustering based classification methods the residue score threshold used for searching biclusters is fixed and the number of mined useful biclusters is small. To solve the two problems, in this study we proposed a novel binary classification method based on supervised-like biclustering and adaboost. Supervised-like means the biclusters is searched twice with two different bicluster quality indicators MSD and weights instead of unique indicator MSR in other biclustering based classification methods. In initial search, the residue score threshold is identical for all biclusters. In the second search, some biclusters of low quality are searched again with smaller residue threshold. Besides, due to some limitations, many biclusters cannot be found. To obtain more biclusters, two additional operations are adopted. In the proposed method, we initially mine column constant biclusters from datasets, then transform the biclusters to weak classifiers. Through adaboost, the initial strong binary classifier can be constructed. Finally, with supervised-like strategy, the better final strong binary classifier can be obtained. To verify the performance of the proposed method, it is compared with seven binary classification methods on two datasets. Experimental results demonstrated that the proposed method outperformed other binary classification methods.