A Review of Learning Paradigms and Artifact Representations for Software Defect Risk Prediction

Abstract

The software quality risk prediction helps in resource allocation for software testing and maintenance by predicting the risk of software components. Recently, the research has moved from statistical and machine-learning classifiers to feature selection, class-imbalance handling, ensemble learning, source-code representations, temporal models and pretrained code language models. There are still numerous studies, however, that only present software metrics as 1D tabular vectors or only consider source code as linear semantic sequences thus not capturing complementary relationships between size, com-plexity, coupling, cohesion and other product characteristics. The results of recent soft-ware defect prediction studies are classified into five groups: metric-based machine learn-ing, ensemble and optimization techniques, source-code deep learning, cross-project and temporal prediction, and language-model pre-training. It also delves into multi-view rep-resentations through metric heatmaps, correlation matrices and distribution-oriented imag-es. Class imbalance, dataset representativeness, cross-project generalization, inconsistent validation, computational cost, lack of interpretability, data leakage, and lack of reproduc-ibility are among the key challenges. The multi-branch deep-learning architecture is a conceptual architecture that combines complementary views for better risk classification.