Interpretability Methods in Machine Learning: A Brief Survey

This post surveys existing interpretability methods in machine learning, including Partial Dependence Plots (PDP), Individual Conditional Expectation (ICE), Permuted Feature Importance, Global Surrogates, and Local Surrogates (LIME). It explains the concepts behind each method, illustrates them with examples (including a time-series dataset and image classification), and discusses their respective advantages and disadvantages, such as handling heterogeneous effects, computational cost, and potential biases.