LPDA: Cross-Project Software Defect Prediction Approach via Locality Preserving and Distribution Alignment

Abstract

Cross-Project Defect Prediction (CPDP) based on domain adaptation aims to achieve defect prediction tasks in an unlabeled target software project by borrowing the defect knowledge extracted from well-annotated source software projects. Most existing CPDP approaches enhance transferability between projects but struggle with misalignments due to limited exploration of class-specific features and inability to preserve original local relationships in transformed features. In order to tackle these challenges, The article introduces a novel Cross-Project Defect Prediction (CPDP) approach called Local Preserving and Distribution Alignment (LPDA). This approach addresses the challenge of misalignments in CPDP due to limited exploration of discriminative feature representations and the failure to preserve original local relationship consistency. LPDA combines transferability and discriminability for CPDP tasks. It uses locality-preserving projection to maintain module consistency and distribution alignment, which includes transferable and discriminant distribution alignment. The former narrows the distributions of both source and target projects, while the latter increases the discrepancy between different classes across projects. The effectiveness of LPDA was tested through 118 crossproject prediction tasks involving 22 software projects from four distinct repositories. The results showed that LPDA outperforms baseline CPDP methods by efficiently learning representations that integrate transferability and discriminability while preserving local geometry to optimize distances within and between categories.