Feature Engineering for Machine Learning: Encoding, Scaling, Binning & Leakage

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Find every missing income (blanks and a text placeholder), add a missing-indicator feature, and fill the gaps with the median and with the mean of the applicant's employment group.
• Turn employment type into one-hot dummy columns, add a count (frequency) encoding, map the credit grade A–E to an ordinal rank, and avoid the dummy-variable trap.
• Fit three scalers on the training rows, apply them to every applicant including the test rows, and measure how one 210,000 income squeezes min-max scaling.
• Cut years-in-job into fixed bands with an approximate LOOKUP, cut income into quartile buckets, compute the default rate per band and chart it.
• Take the natural log of a right-skewed income, compare the skewness before and after, and cap extreme incomes at the 5th and 95th percentiles.
• Build two ratio features and one interaction feature from the raw columns and show with CORREL that the ratios track default far better than the raw loan amount.
• Encode loan purpose by its default rate three ways: naive on all rows (leaky), on the training rows only, and smoothed towards the prior, and you see why only the last one is safe for unseen and rare categories.
• Screen five engineered features against default with CORREL and a pairwise AUC, turn the AUC into a direction-free strength and chart the ranking.