Improving Positive Unlabeled Learning Algorithms for Protein Interaction Prediction

Abstract

In binary classification, it is sometimes difficult to label two training samples as negative. The aforementioned difficulty in obtaining true negative samples created a need for learning algorithms which does not use negative samples. This study aims to improve upon two PU learning algorithms, AGPS[2] and Roc-SVM[3] for protein interaction prediction. Two extensions to these algorithms is proposed; the first one is to use Random Forests as the classifier instead of support vector machines and the second is to combine the results of AGPS and Roc-SVM using a voting system. After these two approaches are implemented, their results was compared to the original algorithms as well as two well-known learning algorithms, ARACNE [9] and CLR [10]. In the comparisons, both the Random Forest ( called AGPS-RF and Roc-RF) and the Hybrid algorithm performed well against the original SVM-classified ones. The improved algorithms also performed well against ARACNE and CLR.