This special topics course is of interest to students who wish to explore Machine Learning concepts beyond the introductory course, COMP 3202.

Prerequisites:  COMP3202

Availability: This course will be offered occasionally.

Course Objectives

This is an advanced course that explores the production of human-understandable explanations for the behaviour of Machine Learning (ML) models. The course introduces and explores the concepts of explainability and interpretability of ML models, and the notion of locality and globality of explanation. It explores the behaviour of ML models, ranging from fully interpretable models to absolutely non-interpretable models.

Representative Workload

• Midterm exam / Quizzes 20%
• Assignments / Project 50%
• Final Exam 30%

Representative Course Outline

• Introduction
• Inherent interpretability and model-agnostic explanation
• Locality and Globality of explanation
• Building explanations for the behaviour of interpretable or partially interpretable models, including:
  • Linear/Logistic Regression, Generalized Linear and Additive Models (GLMs and GAMs), Decision Trees, Decision Rules, RuleFit, Naive Bayes Classifier, K-Nearest Neighbours, Random Forest
• Global model-agnostic explanatory tools
  • Global Surrogate Models
  • Partial Dependence Plots (PDP)
  • Accumulated Local Effects (ALE)
  • H Statistic
  • Permutation Feature Importance
• Local model-agnostic explanatory tools
  • Local Interpretable Model-agnostic Explanations (LIME) / Anchors
  • Individual Conditional Expectation (ICE)
  • Ceteris Paribus Plots
  • Breakdown and Interaction Breakdown Plots
  • SHAP values
  • Example-Based Explanations
• Tools for explaining clusters
  • In Cluster Analysis
  • In Topological Data Analysis
• Feature engineering in Explainable Machine Learning
• An introduction to building explanations for Deep Learning models