Interpretable machine learning assessment

Abstract

With the surge of machine learning in AI and data science, there remains an urgent need to not only compare the performance of different methods across diverse datasets but also to analyze machine learning behaviors with sensitivity using an explainable approach. In this study, we introduce a uniquely designed diagnostic index: d-index to tackle this challenge. This tool integrates classification effectiveness from multiple dimensions, delivering a transparent and comprehensive assessment that transcends the limitations of traditional evaluation methods in classification. We propose two innovative concepts: breakeven states and imbalanced points in this study. Integrated with the d-index, these concepts afford a more profound understanding of the learning behaviors across different machine learning models compared to the existing classification metrics. Significantly, the d-index excels as a powerful tool, identifying learning singularity problems (LSPs) that remain elusive to most current machine learning models and imbalanced learning techniques. Furthermore, leveraging the d-index, we unravel the mechanisms behind imbalanced point generation in binary and multiclass classification. We also put forth a novel technique: identifying a priori informative kernels to optimize support vector machine learning, ensuring outstanding d-index values with the fewest necessary support vectors. Moreover, we address a seldom-discussed state of overfitting in deep learning, where overfitting occurs despite the training and testing loss curves exhibiting favorable trends throughout the epochs. To the best of our knowledge, this work represents a pioneering stride in the realm of explainable machine learning assessments and will inspire further studies in this area.