Table 2 Main features of the model-agnostic interpretability techniques used in this study
From: On the interpretability of machine learning-based model for predicting hypertension
Technique | Global | Local | Advantages | Disadvantages |
|---|---|---|---|---|
Feature Importance | ✓ | • Highly compressed global interpretation • Consider interactions between features | Unclear whether it can be used on training dataset or testing dataset | |
Partial Dependence Plot | ✓ | Intuitive and clear interpretation | Assumption of independence between features | |
Individual Conditional Expectation | ✓ | Intuitive and easy to understand | Plot can become overcrowded to understand | |
Feature Interaction | ✓ | Detects all interactions been features | Computationally expensive | |
Global Surrogate Models | ✓ | Easy to measure the goodness of your surrogate model using R-squared measure | Not clear what is the best cut-off for R-squared to trust the resulted surrogate model | |
Local Surrogate Model (LIME) | ✓ | • Short and comprehensible explanation. • Explains different types of data (tabular, text and image) | • Instability of the explanation • Very close points may have totally different explanations | |
Shapley Value Explanations | ✓ | Explanation is based on strong game theory theorem | Computationally very expensive |