Development of An Enhanced Extreme Learning Machine for Software Defect Prediction.

Abstract

Software Defect Prediction (SDP) enables development teams to focus limited testing and code review effort on the modules most likely to contain faults, improving software quality while controlling costs. Extreme Learning Machine (ELM) is well suited to this task because its single-hidden-layer feedforward structure is trained by analytically solving for output weights rather than by iterative backpropagation, giving very fast training. This paper proposes a hybrid model, SMOTE-OOA-ELM, that addresses both problems jointly: the Synthetic Minority Over-sampling Technique (SMOTE) is applied to the training data to correct class imbalance before model construction, and the Osprey Optimisation Algorithm (OOA), a two-phase nature-inspired metaheuristic based on osprey hunting behaviour, searches for near-optimal ELM input weights and hidden biases in place of random initialization. The paper details the architecture of the combined model, a preprocessing and optimization pipeline, and a full experimental protocol built around NASA/PROMISE benchmark datasets, stratified cross-validation, and imbalance-aware metrics (F1-score, AUC, and Matthews Correlation Coefficient (MCC)) rather than raw accuracy. The model is positioned against plain ELM, SMOTE-ELM without metaheuristic tuning, and OOA-ELM without oversampling, isolating the individual and combined contribution of each component. Illustrative performance patterns consistent with prior swarm-optimized ELM and oversampling literature are presented to demonstrate the intended evaluation format, indicating that the combined use of SMOTE and OOA yielded larger gains in minority-class detection than either technique applied alone