Applied Machine Learning (Jason Hodson)（Z-Library）

【One-Line Pitch】 A hands-on, business-first introduction to applied machine learning, this book walks analysts through three real-world use cases—from data preparation and model selection to evaluation, implementation, and monitoring—making it ideal for data analysts and career switchers who want practical skills without a master's degree. 【Book Arc】 - **Opening (~0%–11%)**: Sets the stage by framing the book as a "mini-master's class" for working analysts, emphasizing that understanding the business problem and data matters more than mastering algorithms. It also introduces practical setup tools like GitHub and Anaconda, plus the value of finding a technical mentor. - **Early (~11%–33%)**: Dives deep into data preparation, covering data sources (CSV vs. Excel), exploration via visualization and descriptive statistics, and cleaning techniques like handling missing data, dummy coding, and consolidating high-cardinality columns. Three use cases—orders, customer behavior, and LA crime data—are used to illustrate these steps. - **Middle (~33%–52%)**: Transitions to model selection, introducing a decision framework and core algorithms like regression, decision trees, and random forest. Key techniques include time-series-aware train/test splits, lagging variables, feature engineering (e.g., extracting month), and using metrics like MAE, accuracy, and ROC AUC to evaluate initial models. - **Late (~52%–70%)**: Focuses on model tuning and interpretation, showing how to use grid search for hyperparameter optimization (e.g., max_depth, min_samples_split) and how to visualize decision trees to understand what the model prioritizes. The emphasis is on balancing performance with interpretability. - **Ending (~70%–100%)**: Covers implementation and monitoring, guiding readers on deploying models to make predictions, tracking performance over time, and measuring business impact. The book closes with advice on iterating through feature engineering and data adjustments to sustain model value. 【Key Takeaways】 - **Business context beats algorithmic mastery** (Opening): A model is useless if it doesn't answer the business problem; prioritize understanding the data and context over technical sophistication. This reframes ML as a tool for decision-making, not a purely technical exercise. - **Data preparation is the bulk of the work** (Early): Cleaning involves identifying errors, handling missing values (remove, impute, or build a second model), and dummy coding categorical variables—but always watch cardinality to avoid exploding column counts. Practical judgment, not rigid rules, guides these choices. - **Visualization and descriptive stats are your first model** (Early): Exploring data through plots (e.g., distributions of victim age, sex, race) and summary statistics reveals issues like incorrectly coded zeros or unclear categories, informing consolidation decisions before modeling begins. - **Time-series data demands date-aware splitting** (Middle): For predicting future values, never randomly split train/test; instead, use a date cutoff (e.g., train on years 1–4, test on year 5) to honestly evaluate predictive performance. Lagging variables and extracting month features give the model temporal context. - **Start simple, then tune with grid search** (Middle–Late): Begin with a baseline model (e.g., linear regression or a shallow decision tree) to gauge error rates, then use grid search over parameters like criterion, max_depth, and min_samples_split to systematically improve. This avoids over-engineering early iterations. - **Interpretability is a feature, not a luxury** (Late): Visualizing decision trees (e.g., with max_depth=4) shows which variables drive predictions, helping you catch biases and explain results to stakeholders before moving to "black box" models like random forests. - **Evaluation metrics need context** (Middle): A MAE of 6.5 is hard to judge alone; comparing it to the average target value (e.g., ~25% error) gives a rough accuracy percentage, making results actionable. Use multiple metrics like accuracy and ROC AUC for a fuller picture. 【Reading Tips】 - **Skim the setup chapters** (~0%–11%) if you're already comfortable with Python and Git; focus instead on the business-context framing and the three use-case introductions, which anchor the rest of the book. - **Deep-read the data preparation sections** (~11%–33%)—they're the most detailed and transferable. Pay special attention to the dummy coding and cardinality reduction examples, as these are common pain points in real projects. - **Treat the code listings as templates, not scripture** (~33%–52%): The book uses consistent patterns (e.g., pd.merge, get_dummies, model.fit), so focus on understanding the *why* behind each step (e.g., why lag by 3 months) rather than memorizing syntax. - **Revisit the model tuning sections** (~52%–70%) with your own dataset in mind: The grid search and decision tree visualization examples are best learned by adapting them to a problem you care about, not just reading along. - **Don't skip the "Real-World Story" asides**—they offer rare, candid insights (e.g., using business context to reduce columns) that bridge the gap between textbook ML and messy organizational realities. 【Coverage Limits】 This guide synthesizes the first ~70% of the book (through model tuning and interpretation); the final sections on implementation, monitoring, and measuring business impact are only partially covered in the excerpts, so readers should expect additional depth there.