Machine learning finds patterns in observations and data to produce useful answers, especially predictions. It overlaps heavily with statistics but emphasizes tuning algorithms, predictive performance, and results over complete model explainability. section 1.5 section 1.8
A practical data-driven workflow connects questions to predictions:
- Form a hypothesis about a pattern or relationship. section 1.8
- Gather representative data rather than trusting a convenient slice. section 1.8
- Fit a model that captures relationships in the data. section 1.8
- Evaluate performance on unseen data to check generalization. section 1.5 section 1.8
The unseen-data check
Train a predictive model with labeled data, then test it on a labeled dataset it has never seen. Strong performance on new data matters more than memorizing the training examples, but a highly accurate black box may still be difficult to explain or audit for bias. section 1.5 section 1.8
