Module 01: Machine Learning Fundamentals

Goal

Understand how classical machine learning works, why models fail, how to evaluate them, and how to turn a model into a usable system.

Core Topics

• Linear algebra: vectors, matrices, dot products, norms, eigenvectors, dimensionality reduction.
• Probability and statistics: distributions, conditional probability, Bayes theorem, sampling, confidence intervals, hypothesis tests.
• Optimization: derivatives, gradients, gradient descent, learning rate, loss functions.
• Algorithms: linear regression, logistic regression, decision trees, random forest, gradient boosting, XGBoost, SVM, K-Means, PCA.
• Evaluation: train/validation/test split, cross-validation, accuracy, precision, recall, F1, ROC-AUC, PR-AUC, MAE, MSE, RMSE, calibration, leakage, overfitting, underfitting.

Study Tasks

• Implement linear regression with NumPy.
• Implement logistic regression with NumPy.
• Train tree-based models with scikit-learn.
• Compare at least five algorithms on the same dataset.
• Create an evaluation report with tables and charts.
• Explain overfitting using a real experiment.
• Write notes for every metric and when to use it.

Mini Labs

• Build a notebook that visualizes gradient descent.
• Create a synthetic classification dataset and intentionally add noise.
• Create a data leakage example and explain why it is dangerous.
• Compare random split vs time-based split for a temporal dataset.

Project

Build projects/01-business-sales-prediction/.

Tips

• Do not skip evaluation metrics. Model selection without metrics is guessing.
• Always create a dumb baseline before training a complex model.
• Track assumptions about data quality, feature availability, and deployment constraints.
• Learn the math enough to debug, not to memorize formulas.

Recommended Resources

• scikit-learn user guide.
• StatQuest videos for intuition.
• Hands-On Machine Learning for practical workflows.
• Kaggle datasets for experimentation.

Completion Checklist

• Can explain bias, variance, overfitting, and underfitting.
• Can choose metrics for classification and regression.
• Can build a clean training pipeline.
• Can produce a model comparison table.
• Can identify leakage risks.