Features matter

Garbage in, garbage out: if inputs are uninformative, no algorithm can rescue you.

Features are the numbers (or encodings) you feed the learner. Domain knowledge suggests what might matter; experiments compare feature sets. Some models ignore noisy features; others collapse when you include junk dimensions.

Predicting house price

Using only “day of week the ad was posted” will fail. Square metres, location, and condition carry signal. Feature choice is modelling — often more important than swapping classifier A for classifier B.

Exam tip

Connect to the intro: unsupervised categorization also lives or dies by which properties you look at.

Key takeaways

Feature design is part of ML.
Background knowledge helps.
Match preprocessing to the algorithm.