RiskBlend Prioritizes ML Regression Test Inputs with Multi-Signal Framework

Retraining machine learning classifiers can introduce regression faults, where previously correct predictions become incorrect. Detecting these faults is costly, often requiring human annotation or expensive simulations. This paper introduces RiskBlend, a novel framework designed to prioritize test inputs for regression testing, aiming to uncover as many faults as possible within a limited verification budget. RiskBlend distinguishes itself by moving beyond single-model confidence scores, which are the primary focus of existing methods. Instead, it integrates four complementary risk signals: historical failure patterns, changes in prediction, shifts in decision boundaries, and alterations in local neighborhoods between model versions. These signals are combined using validation-learned weighting. Across extensive experiments involving various datasets, classifiers, and scenarios, RiskBlend consistently achieved the highest average APFD (Average Percentage of Faults Detected), demonstrating substantial improvements over baseline methods, particularly for non-linear classifiers.