Cost-Sensitive Local Collaborative Representation for Software Defect Prediction

Abstract

Recently, representative sparse representation based classifiers, namely dictionary learning and collaborative representation based classifier (CRC), has been introduced into software defect prediction (SDP) and demonstrated to be effective for SDP. The dictionary learning based SDP method needs relatively large computational cost, while collaborative representation based method can significantly reduce the computational cost and achieve comparable prediction effects as the former. In this paper, we aim to preserve the desirable efficiency of collaborative representation based SDP method and further improve its prediction effect. We propose a cost-sensitive local collaborative representation (CLCR) approach for SDP. CLCR firstly efficiently finds the neighboring modules of a given test (query) module using CRC. Then CLCR represents the test module as a linear combination of its neighbors and uses the representation error for prediction. To solve the class-imbalance problem, CLCR further incorporates the cost-sensitive factor into the representation coefficients in the prediction phase. Experiments on five projects of the NASA dataset demonstrate the effectiveness of the proposed approach as compared with several related SDP methods.