ML Methods

Classical machine learning worked through one algorithm at a time: what it assumes, where it breaks, and a runnable example for each.
Author

Benedict Thekkel

This is the pre-deep-learning half of machine learning, the part that still solves most problems that arrive as a table of numbers. Each page takes one algorithm, states the intuition in a line, lists what it is good and bad at, then works an example in scikit-learn.

Several algorithms get two pages: one for the method and one marked TUT applying it to a real dataset. Read them as a pair.

For neural networks see DL Methods, and for forecasting see Time Series.


Foundations

The parts that apply no matter which model you pick.

Page Covers
Training and Testing Why you hold data back at all, and how train_test_split does it
Loss and Cost Functions The common losses, and which ones suit classification against regression
L1 and L2 Regularization Overfitting shown on a plain regression, then Lasso and Ridge as the fix
Hyper-parameter Optimization Three escalating approaches: manual trial and error over a split, then K-fold cross validation, then GridSearchCV

Regression

Predicting a continuous value.

Page Covers
Linear Regression The single-variable case, ending in generated house price predictions
Multi-Linear Regression The same with several input variables
Polynomial Regression Fitting curves, second order and higher
Kernel Ridge Regression Ridge regression combined with the kernel trick, for non-linear relationships
Support Vector Regression SVM applied to a continuous target

Classification

Predicting a category.

Page Covers
K-NN Take any point, it is what it is close to. Distance metrics, choosing k, then a first model and a GridSearchCV search for the optimum
Decision Tree If-statements placed where entropy drops fastest, plus the trade-offs of CART
Logistic Regression Binary and multiclass, and the classification types each suits
Naive Bayes The Bayesian approach and Gaussian processes
Naive Bayes Tutorial The same method built into an email spam filter
Random Forest Many trees voting

Ensembles

Page Covers
Ensemble Methods Bagging against boosting, and when each helps
Ensemble Tutorial Bagging applied to heart disease prediction, then boosting
Bagging: Diabetes Prediction A second worked bagging example

Unsupervised Learning

Finding structure without labels.

Page Covers
Principal Component Analysis Dimensionality reduction
PCA Tutorial PCA applied to heart disease prediction
Gaussian Mixture Models Modelling data as a mixture of Gaussians, the expectation-maximization algorithm behind it, and its use for clustering, density estimation, and anomaly detection
K-Means Clustering The algorithm step by step, choosing K, and the alternatives when it does not fit

GPU Programming

Page Covers
CUDA The odd one out, and the longest page here. Getting a CUDA toolchain working, then writing kernels: RGB to greyscale, then matrix multiplication. Useful background for why the deep learning libraries are shaped the way they are

Not Covered Yet

  • Three pages are placeholders marked “coming soon” with no content: CNNs, Generative Models, and Back Propagation. All three are really deep learning topics and are covered properly in DL Methods.
  • Nothing on gradient boosting libraries. No XGBoost, LightGBM, or CatBoost page, despite them being the usual winners on tabular data. XGBoost is covered from the forecasting angle in Time Series.
  • No feature engineering page, and no treatment of missing data, categorical encoding, or scaling as topics in their own right.
  • No model interpretability: nothing on SHAP, permutation importance, or partial dependence.
  • Evaluation is spread through the tutorials rather than collected anywhere, so there is no single page on precision, recall, ROC, or calibration.

Back to top