The Art of Machine Learning A Hands-On Guide to Machine Learning with R (Norman Matloff) (Z-Library)

# The Art of Machine Learning: A Hands-On Guide to Machine Learning with R ## 【One-Line Pitch】 A practical, math-light introduction to machine learning using R, teaching you not just how to apply methods like k-NN, random forests, and neural networks, but also why they work—perfect for analysts and programmers who want real-world ML skills without a statistics degree. ## 【Book Arc】 - **Opening (~0%–9%)**: Sets the "anti-cookbook" philosophy—understanding the why behind ML methods—and introduces the first method, k-NN, using a bike ridership prediction example. Covers regression functions, dummy variables, and the qeML/regtools package setup. - **Early (~9%–25%)**: Deepens k-NN with real datasets (MLB player weights), explains the bias-variance tradeoff through polling analogies, and introduces hyperparameter tuning, feature scaling (mmscale vs. scale), and the kNN() function's arguments. - **Early (~25%–34%)**: Shifts to classification models with the Telco Churn dataset, clarifying the confusing terminology (numeric-Y vs. classification), and demonstrates multiclass problems with vertebral column data and medical no-show prediction. - **Middle (~34%–44%)**: Covers model evaluation with ROC curves and AUC (using qeROC()), then formalizes the bias-variance tradeoff, overfitting/underfitting, and why too many features (like 42,000 ZIP codes) inflate variance. - **Middle (~44%–47%)**: Introduces dimension reduction with the Million Song Dataset (515,000 rows, 90 features), showing PCA via prcomp() and qePCA() to reduce feature sets, avoid overfitting, and speed up computation. ## 【Key Takeaways】 - **ML is regression in disguise** (Early): All ML methods estimate regression functions—classification is just regression with categorical Y. Understanding this unifies seemingly different techniques and clarifies confusing terminology. - **k-NN is the intuitive foundation** (Opening): Predicting by averaging over similar cases (neighbors) is both simple and powerful. The hyperparameter k is a "Goldilocks" choice—too small overfits, too large underfits. - **Bias-variance tradeoff is central** (Early): Reducing one increases the other. Small k or too many features inflate variance; omitting key features (like height in weight prediction) introduces bias. This tradeoff drives all hyperparameter and feature decisions. - **Feature scaling matters** (Early): Unscaled features give undue influence to variables with large values. The mmscale() function maps to [0,1] for bounded, unitless variables, unlike scale() which produces unbounded values. - **Dirty data and p-hacking are real threats** (Opening): Remove IDs and near-unique identifiers (like patient IDs in medical data) to avoid overfitting. Be aware of how feature engineering can lead to spurious findings. - **ROC curves evaluate classification honestly** (Middle): qeROC() wraps pROC to plot ROC and compute AUC on holdout sets—essential for assessing churn prediction models beyond simple accuracy. - **Dimension reduction fights the curse of dimensionality** (Middle): With 90 features in the Million Song Dataset, PCA creates new features as combinations of originals, reducing candidate sets from 2^90 to just 90—saving computation and preventing overfitting. ## 【Reading Tips】 - **Skim the code-heavy sections** (Early ~25%): The kNN() function arguments and R output dumps are reference material—grasp the concept, then return when you need syntax. - **Deep-read the bias-variance explanations** (Early ~16%–19%): The polling analogy (landline bias, margin of error) is the clearest explanation of variance and bias you'll find—master this and the rest of the book clicks. - **Work through the Telco Churn example** (Early ~28%): This is the book's flagship classification case. Follow the full pipeline: data loading, cleaning, k-NN modeling, then ROC evaluation. - **Don't skip the "Pitfall" sections**: They're marked recurring themes for a reason—they contain hard-won practical advice like removing IDs and handling factor variables with many levels. - **Install packages before starting** (Opening): You need R, qeML, and regtools (version 1.7+) installed and loaded. The book assumes you've done this, so set up early to avoid friction. ## 【Coverage Limits】 This guide covers the book's first half (through dimension reduction with PCA). Excerpts do not cover later chapters on random forests, gradient boosting, logistic models, SVMs, LASSO, neural networks, text/image classification, or time series—though the blurb indicates these follow the same practical, why-it-works approach. ##