Classification Intuition

Sorting things into buckets

Spam or ham? Churn or loyal? Malignant or benign? Whenever the answer is a category, you're classifying. The workhorse starter is logistic regression — a confusing name for a clean idea: model the probability of each class, then pick the highest. Despite "regression" in the name, it outputs categories.

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report
model = LogisticRegression(max_iter=1000).fit(Xtr, ytr)
print(classification_report(yte, model.predict(Xte)))

Accuracy is a liar (when classes imbalance)

A fraud detector facing 99% legitimate transactions scores 99% accuracy by predicting "legit" always — perfect score, worthless model. The honest report card:

• Precision: of the flagged ones, how many were real? (False alarms cost trust.)
• Recall: of the real ones, how many did we catch? (Misses cost money/lives.)
• F1: their harmonic mean — one number that refuses to be gamed by imbalance.

Always report these for the rare class you actually care about — the fraud, the churners, the diagnoses.

Thresholds are business decisions

Models output probabilities; you choose the cutoff. Flag fraud above 0.3 if missing fraud is expensive; require 0.9 if false alarms wake people at night. Moving the threshold trades precision against recall — a product decision wearing math clothes, and interviewers love asking about it.

Remember this

• Categories out → classification. Imbalanced data → precision/recall/F1, never raw accuracy.
• Thresholds encode business costs; choose them deliberately, not by default 0.5.