Architecture-aware multiclass design-level fault classification using KISS, DRY, and SOLID principle-based metrics

Traditional software fault classification approaches assign faults or faulty software instances to predefined categories using source-code characteristics and software metrics. However, conventional classification approaches provide limited characterization of design-level violations and do not explicitly incorporate software design principles when distinguishing fault categories. This paper proposes an architecture-aware software analytics framework that integrates KISS, DRY, and SOLID principle-based metrics for fine-grained design-level fault classification. A unified 36-dimensional feature space is constructed through static source-code analysis. A two-phase hierarchical machine-learning architecture is introduced in which the first phase classifies high-level fault families and the second phase identifies fault subtypes within the corresponding family. To provide ground-truth labels for the classification task, controlled design-level faults were injected into four widely used open-source Java projects: Apache Ant, Apache Lucene, Apache Camel, and JUnit. Labels were assigned during fault injection independently of metric extraction, thereby avoiding circular label construction and reducing the potential for label leakage arising from metric-derived labeling. The resulting dataset contains controlled design-level fault instances organized into predefined fault families and subtypes. Random Forest and XGBoost are used to evaluate the proposed representation against conventional object-oriented, complexity, and process metric representations under the same experimental setting. The results demonstrate the classification capability of the proposed architecture-aware representation for fault families and subtypes and show the performance of hierarchical classification relative to equivalent flat multiclass classification. The study focuses on controlled injected design-level faults and does not claim equivalence between these faults and naturally occurring software defects.