Day 15: Supervised vs unsupervised learning; the ML mental model
What machine learning actually is
You've spent your career writing explicit rules: if order.total > 5000 then flag. Machine learning inverts that. Instead of writing the rule, you show the machine thousands of examples of inputs paired with the correct outputs, and it fits a function that maps input → output. The rule is discovered from data, not coded by you. For FitXpert, that's the difference between hand-writing 'a shirt has two sleeves and a collar' (brittle, endless edge cases) and showing a model 10,000 labelled garment photos until it learns 'shirt-ness' on its own.
Interview vocabulary, then stop
This first week of classical ML exists to give you the *vocabulary* — supervised, overfitting, precision/recall — that every later stage and every interview assumes you already have. It is deliberately shallow. Learn the words, do one project, then move to deep learning where the real FitXpert work happens.
Supervised vs unsupervised
Supervised learning trains on labelled examples — each input carries its correct answer. Classification predicts a category ('shirt' / 'dress'); regression predicts a number ('return probability = 0.73'). Unsupervised learning finds structure in *unlabelled* data — clustering similar garments together, or reducing high-dimensional data to a few meaningful axes. Almost everything FitXpert does is supervised; the Stage 4 recommender's embedding similarity is the closest you'll get to unsupervised.
Key terms
- Supervised learning
- Training on input/output pairs where the correct answer (label) is known for each example.
- Classification
- Supervised task predicting a discrete category (e.g. garment type).
- Regression
- Supervised task predicting a continuous number (e.g. probability, price).
- Unsupervised learning
- Finding structure (clusters, low-dimensional axes) in data that has no labels.
You want to predict the exact probability (0.0–1.0) that a given order will be returned. What kind of task is this?
