Data Science from Scratch (Joel Grus)（Z-Library）

# Data Science from Scratch — Reading Guide ## 【One-Line Pitch】 A hands-on, code-first tour through the fundamental math, statistics, and algorithms of data science, teaching you to build core tools from scratch rather than blindly relying on libraries. Ideal for programmers with basic Python skills who want to truly understand what's under the hood before using scikit-learn, pandas, or TensorFlow. ## 【Book Arc】 - **Opening (~0%–10%)**: Sets up the "data science Venn diagram" (hacking skills + math/statistics + substantive expertise) and explains why the book focuses on the first two. Introduces a running example—a fictional social network called DataSciencester—and starts manipulating friendship data with basic Python structures. - **Early (~10%–23%)**: A crash course in Python essentials (functions, dictionaries, defaultdict, classes, and "dunder" methods), followed by visualization with matplotlib and the foundational building blocks of linear algebra: vectors, matrices, and their operations. - **Early-to-Middle (~23%–39%)**: Covers descriptive statistics (mean, median, quantiles, variance, correlation), then probability theory (PDFs, CDFs, normal distribution), hypothesis testing, p-hacking, and gradient descent as the workhorse optimization technique for machine learning. - **Middle (~39%–48%)**: Moves into practical data wrangling—reading/writing CSV files, scraping HTML/XML with Beautiful Soup, working with APIs (GitHub example), and cleaning/transforming data. Introduces dataclasses and more advanced data structures. - **Late (~48%–end)**: Dives into machine learning proper: PCA for dimensionality reduction, the critical distinction between training/validation/test sets, and why "accuracy" is a misleading metric for classification (the "Luke leukemia test" example). Excerpts suggest the book continues into regression, classification, and beyond. ## 【Key Takeaways】 - **Data science = hacking skills + math/statistics** (Opening): The author deliberately skips "substantive expertise" to keep the book focused; you learn how to code and think statistically, not domain knowledge. (Early) - **Python fundamentals are the foundation** (Early): Master dictionaries, defaultdict, Counter, classes, and first-class functions before touching any data science library—these are the tools you'll use to build everything else. (Early) - **Vectors and matrices are the language of data** (Early): Understanding shape, row/column operations, and matrix construction via functions like `make_matrix` is prerequisite for every algorithm later in the book. (Early) - **Mean vs. median is a real trade-off** (Early): The mean is smooth and calculus-friendly but outlier-sensitive (the Michael Jordan salary example); the median is robust but requires sorting. Choose based on whether outliers are signal or noise. (Early) - **Probability distributions are building blocks** (Early): The uniform and normal distributions, their PDFs and CDFs, underpin everything from random sampling to hypothesis testing—and you'll implement them yourself. (Early) - **p-hacking is a real danger** (Early): If you test enough hypotheses, 5% will appear "significant" by chance alone; the book demonstrates this with a coin-flip experiment that rejects fairness 46 out of 1000 times. (Early) - **Gradient descent is the workhorse** (Early): Maximizing/minimizing functions by stepping in the direction of the gradient is the core of most ML algorithms; minibatch variants make it practical for large datasets. (Middle) - **Data wrangling is unglamorous but essential** (Middle): CSV handling, HTML/XML scraping, and API calls (with date parsing via `dateutil`) are the messy realities of real data science work. (Middle) - **Accuracy is a trap for classification** (Late): A test that predicts "leukemia if named Luke" is >98% accurate but useless—you need precision, recall, and proper train/validation/test splits. (Late) ## 【Reading Tips】 1. **Skim the Python crash course if you're experienced** (Early): Chapters 2–3 cover basics you may know; focus instead on the defaultdict/Counter patterns and the vector/matrix implementations, which appear throughout. 2. **Deep-read the statistics and probability chapters** (Early): These are the conceptual core—don't skip the discussions of mean vs. median, quantiles, and the normal distribution even if the code seems simple. 3. **Code along with the DataSciencester example** (Opening–Early): The running social-network dataset makes abstract concepts concrete; implement the friendship-graph manipulations yourself. 4. **Watch for the "gotcha" examples** (Early–Late): The p-hacking experiment and the "Luke test" are memorable warnings about statistical pitfalls—internalize these lessons even if you skim the surrounding code. 5. **Expect to install third-party libraries** (Middle): The book uses matplotlib, requests, Beautiful Soup, and python-dateutil; set up a clean Python environment before starting the data-wrangling chapters. ## 【Coverage Limits】 This guide is based on excerpts covering roughly the first half of the book (through PCA and model evaluation). The later chapters on regression, classification algorithms, neural networks, and natural language processing are not covered in the sampled material. ##