Research on Software Defect Prediction Based on Transfer Learning by Deep Neural Network

Software defect prediction (SDP) in real-world settings is challenged by distribution shift across versions and projects, scarcity and imbalance of labeled data, cross-language heterogeneity, and the need for actionable targets beyond binary labels. We present a unified framework that addresses these challenges through four complementary components: (i) process aware cross-version transfer, which augments static code metrics with version-history signals to encode software evolution; (ii) imbalance-robust learning, which couples ensemble under 65sampling with a minority-weighted transfer strategy to emphasize defect-prone modules without overfitting; (iii) heterogeneous cross-language alignment, which combines Bagging-enhanced Canonical Correlation Analysis (CCA) with Stacking to exploit disparate source domains; and (iv) actionable multi-class defect-count prediction, which moves beyond binary labels to support finer-grained test prioritization. To further mitigate domain discrepancy and stabilize optimization, the framework integrates Transfer Component Analysis and a Beetle Antennae Search–Artificial Bee Colony hybrid swarm intelligence algorithm for hyperparameter optimization. Extensive studies on widely used benchmarks demonstrate consistent and practically meaningful gains over strong baselines in cross-version, cross-dataset, and cross-language scenarios, with particularly notable improvements on minority (defective) classes that matter most for screening. Ablation analyses confirm that each component contributes additively, underscoring the value of a holistic design rather than isolated techniques. By coupling transfer learning with history-aware features, imbalance-robust optimization, heterogeneous knowledge fusion, and multi-granularity targets, the proposed approach offers a reliable and extensible recipe for SDP in data-sparse, language diverse, and safety-critical environments, enabling more precise test prioritization and more efficient allocation of quality assurance resources.