Data Science from Scratch First Principles with Python (Joel Grus)（Z-Library）

【One-Line Pitch】 A hands-on guide that rebuilds the core toolkit of data science—statistics, probability, linear algebra, machine learning, and data wrangling—from first principles in pure Python, so you understand what the libraries actually do. Best for programmers and analysts who want conceptual depth rather than a tour of `pandas` and `scikit-learn` APIs. 【Book Arc】 - **Opening (~0%–11%)**: Sets the philosophy—build small, readable, "illuminating but impractical" implementations yourself, then learn which library scales them. Uses a fictional social network (DataSciencester) to motivate problems, and covers Python essentials plus type annotations. - **Early (~11%–33%)**: The mathematical foundations. Visualization, linear algebra (vectors, matrices), statistics (central tendency, dispersion, correlation, Simpson's paradox), probability, conditional probability and Bayes's theorem, the central limit theorem, and hypothesis testing. - **Early–Middle (~26%–44%)**: Working with real data—reading files, CSV, JSON, web APIs, and scraping—then the core modeling loop: gradient descent, minibatch variants, and simple linear regression with least squares. - **Middle (~44%–60%)**: Multiple regression, regularization (ridge), and the move into classification: nearest neighbors, the curse of dimensionality, and Naive Bayes (with a spam classifier built and unit-tested). - **Late (~52%–80%)**: Tree-based methods (entropy, decision trees), then the neural-network and deep-learning chapters, plus clustering and other unsupervised techniques. - **Ending (~80%–100%)**: Natural language processing, network analysis, recommender systems, databases/SQL, and MapReduce-style data processing—closing the loop from raw data to deployed techniques. 【Key Takeaways】 - **The book deliberately builds "toy" implementations** (Opening): the goal is understanding, not production scale; each chapter points to the library you'd actually use on web-scale data. - **Python is chosen for clarity, not preference** (Opening): the author admits other languages are more pleasant, but Python's simplicity and ecosystem make it the teaching vehicle. - **Statistics is about distilling data, not displaying it** (Early): correlation, dispersion, and the outlier/Simpson's-paradox examples show how summary numbers can mislead without visual inspection. - **Bayes's theorem is framed as a data scientist's best friend** (Early): reversing conditional probabilities is presented as a recurring, practical reasoning tool, not just a formula. - **Gradient descent is the workhorse of modeling** (Early–Middle): from simple linear regression to minibatch variants, the same optimization loop underlies the book's models. - **Classification starts with intuition, then formalizes it** (Middle): nearest neighbors (with the curse of dimensionality) and Naive Bayes (built and unit-tested on spam) show the progression from idea to working classifier. - **Entropy drives decision-tree splits** (Late): the book explains information gain conceptually before wiring it into a tree structure. - **Data wrangling is treated as first-class** (Early): CSV, JSON, APIs, and scraping get dedicated attention because real data rarely arrives clean. 【Reading Tips】 - **Deep-read the math chapters** (statistics, probability, linear algebra): these are the foundation the rest of the book leans on; skimming them makes later modeling chapters feel like magic. - **Skim the Python crash course** if you already know the language—focus on the type-annotation and f-string sections, which the book uses throughout. - **Type along with the code**: the value is in building the implementations yourself; reading alone loses the "first principles" payoff. - **Use the "For Further Exploration" notes** as your bridge to real libraries—treat the book's code as a learning scaffold, not a toolkit to ship. - **Don't skip the testing examples** (e.g., the Naive Bayes unit tests): they model how to validate a model you built yourself. 【Coverage Limits】 The excerpts cover the book's opening through roughly the middle and late chapters, with the ending chapters (NLP, networks, recommenders, databases, MapReduce) represented only by their position markers. Specific chapter titles and detailed content for the final sections are not fully covered in the excerpts.