OCMTL: Transfer learning by orthogonal core-extraction of a matrix and its application in cross-project defect prediction

Due to the scarcity of labeled defect data, cross-project defect prediction (CPDP) has become popular nowadays. Most of the successful CPDP methods are based on different transfer learning approaches. However, these methods are mostly sensitive to parameters, dependent on pseudo-labels, and user-defined feature dimensions. To address the issues, we first propose orthogonal core-extraction of a matrix (OCM). OCM is a low-cost and parameter-free iterative method that captures the maximum achievable variance and alleviates the multicollinearity in data by extracting orthogonal cores. We then extend it and propose OCM for transfer learning (OCMTL). OCMTL achieves transfer learning by capturing shared components between source and target without the use of pseudo-labels. Moreover, we propose a method to select the important features. OCMTL is compared with eleven state-of-the-art methods on four benchmark datasets. It achieves 71% average AUC across all datasets, around 2% improvement to its closest competitor (JCSL). Specifically, it achieves average AUC improvements of 4.2% in RELINK against TCA, 1.8% in NASA against JCSL, 0.05% in SOFTLAB against TCA+ and equal average AUC in AEEEM against DMDA_JFR, representing a generalized prediction ability across all datasets. We also conduct significance tests, where OCMTL ranks first in most of the cases.