Neural Defect Predictors: Training Dynamics Under Data Quality Issues

Software defect prediction is crucial for maintaining software quality, guiding decisions in testing, release risk, and quality monitoring. Traditional methods often face challenges from data quality issues, specifically class imbalance and class overlap, which frequently occur together in real-world datasets. While previous studies focused on how these issues impact the final performance of defect prediction models, this research delves deeper into the internal workings of neural networks during training. The study aims to understand how these coupled data quality problems manifest within the neural network's training dynamics, such as gradients, weights, biases, and error trajectories. By conducting a controlled intervention study on specific datasets, researchers will train a fixed Multi-Layer Perceptron (MLP) under various conditions: imbalance only, overlap only, and a combination of both. This approach will allow for a detailed characterization of training patterns using effect sizes, trajectories, and sensitivity analyses.