The Hundred-page Machine Learning Book

The Hundred-Page Machine Learning Book condenses the vast body of machine learning knowledge into a lean, readable guide covering supervised, unsupervised, and other learning paradigms. Andriy Burkov walks readers from the mathematical notation and fundamental algorithms (linear and logistic regression, decision trees, SVM, kNN) through neural networks and deep learning, then into the practical craft of feature engineering, regularization, model assessment, and hyperparameter tuning. Rather than exhaustively cataloging every technique, it curates the parts with proven practical value and teaches readers how to think about whether a problem is 'machine-learnable' and which methods to try. It is ideal both for beginners seeking a comfortable grasp of the field and for experienced practitioners wanting a concise reference and a springboard for further self-improvement.

What it argues

A causal framework linking data and design levers (dataset quality, feature engineering, model complexity, regularization, hyperparameter tuning) through intermediate model behaviors (fit to training data, generalization) to outcomes (predictive performance, practical utility).

Key ideas it contributes

• Dataset Quality and Size — The extent to which the training data is sufficiently large, correctly labeled, low in noise, and representative of the distribution from which future inputs will be drawn.
• Feature Engineering Quality — The quality of the transformation of raw data into informative feature vectors that carry high predictive power for the target.
• Model Complexity — The representational capacity of the chosen model and algorithm to fit intricate relationships in the data.
• Regularization Strength — The degree to which techniques are applied to constrain model complexity and reduce variance.
• Hyperparameter Tuning Effort — The extent and sophistication of the search over non-learned algorithm settings to optimize validation performance.
• Fit to Training Data (Bias Level) — How accurately the trained model reproduces the labels of the training examples, inversely related to bias/underfitting.
• Generalization (Variance Control) — The model's capacity to predict accurately on unseen examples from the same distribution, reflecting controlled variance.
• Predictive Performance — The measured quality of the model's predictions on holdout data using metrics appropriate to the task and its cost structure.