Classification

Contents

Classification#

Many chemical engineering decisions are inherently discrete: a material is stable or unstable, a sensor reading indicates normal operation or a fault, a candidate molecule is active or inactive. Classification algorithms learn decision boundaries that map continuous inputs to discrete class labels.

This module develops classification from first principles, moving from simple geometric boundaries to flexible non-parametric models and finally to methods that scale to high-dimensional feature spaces. The perovskite stability dataset — a real materials-science problem — is used throughout to ground the methods in a chemical engineering context.

Topics#

  • Topic 3.1 — Classification Basics: Problem setup, decision boundaries, evaluation metrics (accuracy, precision, recall, F1, ROC/AUC), and practical considerations for class imbalance.

  • Topic 3.2 — Generalized Linear Models: Logistic regression, the perceptron, support vector machines, and kernel methods. Covers loss functions, margin maximization, and the role of regularization.

  • Topic 3.3 — Alternate Classification Models: k-nearest neighbors, Naive Bayes, decision trees, and random forests. Introduces ensemble methods and the bias–variance tradeoff in the classification setting.

  • Topic 3.4 — High-dimensional Classification: Applying SVM and tree-based methods to the perovskite dataset. Covers GridSearchCV for hyperparameter tuning, feature importances, and the curse of dimensionality.

Contents