Interpretable Machine Learning - Online Course

Interpretable Machine Learning refers to techniques and approaches that make machine learning models understandable to humans, ensuring that their behavior and predictions are transparent.

We will start by reviewing intrinsically interpretable models, such as linear and logistic regression. These models offer clear interpretations by providing coefficients that explain the importance of each feature, helping us understand how predictions are made. This will set the foundation for the level of interpretability we aim to achieve in more complex models. We will also explore techniques like Lasso regression, which enhance interpretability by enforcing sparsity in the coefficients.

Next, we’ll examine more complex machine learning models that are not inherently interpretable, such as gradient-boosted trees, BART (Bayesian Additive Regression Trees), or neural networks. We will then introduce a range of post-hoc and model-agnostic interpretation methods that help researchers interpret these models. These include both global methods—which explain how features impact predictions on average—and local methods—which focus on individual predictions.

For global interpretation, we will focus on partial dependence plots, which show how a feature influences the prediction when other features are averaged out. These plots are useful for understanding the overall behavior of a model. Additionally, we’ll cover other techniques like Accumulated Local Effects (ALE) plots, feature interaction measures (H-statistic), functional decomposition, permutation feature importance, and global surrogate models.

For local interpretability, we will delve into individual conditional expectation (ICE) curves, which illustrate how changes in a feature affect the prediction for an individual instance. We’ll also explore local surrogate models (such as LIME), which approximate a complex model with a simpler, interpretable one for specific predictions. Further, we will cover scoped rules (anchors), counterfactual explanations, which reveal how features need to change to alter a prediction, and Shapley values, which fairly attribute a prediction to individual features. We’ll also discuss the widely-used SHAP values, which provide consistent and theoretically sound feature attributions.

The analysis will primarily be conducted in R, but similar methods can be applied using Python or other machine learning software.

For R, you are strongly encouraged to have the latest version of R installed on your computer. It is also recommended to download and install RStudio, a free and user-friendly front-end for R that simplifies the workflow.

If you’d like to take this course but are concerned that you don’t know enough R, there are excellent on-line resources for learning the basics. Here are our recommendations.

This seminar is ideal for anyone interested in gaining practical experience in interpretable machine learning. It is suited for students, applied researchers in academia, industry professionals, and government analysts with a basic understanding of machine learning methods.

The course is also highly valuable for experienced researchers looking to deepen their knowledge of modern techniques for interpreting complex models. While a foundational understanding of machine learning is recommended, those from various fields who are eager to enhance the transparency and interpretability of their models are encouraged to attend.

Foundations of interpretability

• • Introduction to interpretable machine learning
    • Why interpretability matters
      • Understand the importance of interpretability in machine learning, especially for transparency, trust, and fairness in model predictions
    • Core concepts of interpretability
      • Introduction to key principles of interpretability, comparing traditional interpretable models (like linear and logistic regression) with complex machine learning models (like neural networks and decision trees)
  • Interpretable models
    • Understanding intrinsically interpretable models
      • Explore simple models like linear regression, logistic regression, and decision trees, where feature importance is easily extracted and explained
    • Using regularization for interpretability
      • Techniques like lasso regression to enforce sparsity and improve interpretability of coefficients
  • Experimental design of interpretability
    • Benchmarking interpretability
      • Establishing benchmarks for interpretability with simple models before moving on to complex ones

Complex models and post-hoc interpretation

• • Interpreting complex machine learning models
    • Post-hoc and model-agnostic interpretability
      • Introduction to techniques that provide interpretability for black-box models, like random-forests, gradient-boosted trees, and BART
  • Global interpretation methods
    • Partial Dependence Plots (PDPs)
      • Learn how PDPs illustrate the effect of individual features on model predictions by averaging out the influence of other features
    • Accumulated local effects (ALE) and feature interactions
      • Explore ALE plots for more nuanced interpretations of feature impacts and interaction measures like the H-statistic
  • Local interpretation methods
    • Individual conditional expectation (ICE) curves
      • Understanding how feature changes influence predictions for individual data points
    • Local surrogate models (LIME)
      • Learn how to approximate complex models locally with interpretable models for explaining individual predictions
    • Introduction to Shapley Values and SHAP
      • Explaining predictions by attributing outcomes fairly to individual features using Shapley values, and their extension through SHAP

Practical applications and advanced topics

• • Practical applications of interpretability
    • Handling nonlinearity and nonadditivity
      • Understanding how to explain models that capture non-linear relationships or complex interactions between features
    • Addressing data issues in interpretability
      • Dealing with missing data, data leakage, and feature engineering for interpretable results
  • Advanced interpretation techniques
    • Permutation feature importance
      • Learn how to measure feature importance by observing the model’s performance when features are randomly shuffled
    • Global and local surrogate models
      • Using simpler interpretable models to mimic the behavior of complex models, both at a global and local level
  • Ethics and trust in interpretability
    • Ethical considerations in model interpretability
      • Address issues such as fairness, bias detection, and transparency in using interpretable models for sensitive applications

The fee of $995 includes all course materials.

PayPal and all major credit cards are accepted.

Our Tax ID number is 26-4576270.