What is conditional probability, and how is it used in classification models?

What you need to know

The intuition: shrink the world

Out of 1,000 emails, 300 are spam. So without knowing anything else, P(spam) = 300 / 1,000 = 0.30. Now you are told the email contains the word "free". Ignore every email without "free" — that is the shrinking. Suppose 200 emails contain "free", and 150 of those are spam. Inside this smaller world, P(spam | "free") = 150 / 200 = 0.75.

The word "free" raised the probability of spam from 0.30 to 0.75. Emails without it drop to 150 / 800 ≈ 0.19.

The same table in pandas. normalize="index" divides each row by its row total, which is exactly "condition on the row":

1import pandas as pd23emails = pd.DataFrame({4    "has_free": [True] * 200 + [False] * 800,5    "is_spam":  [True] * 150 + [False] * 50 + [True] * 150 + [False] * 650,6})7print("P(spam) =", emails["is_spam"].mean())8print(pd.crosstab(emails["has_free"], emails["is_spam"], normalize="index"))

P(spam) = 0.3is_spam    False   Truehas_freeFalse     0.8125  0.1875True      0.2500  0.7500

How classifiers use it

Every probabilistic classifier estimates P(class | features) for a new input:

• Logistic regression directly outputs P(y = 1 | x).
• Naive Bayes builds P(class | features) from P(features | class) using Bayes' theorem.
• Neural networks with softmax output P(class | input) for each class.

The decision is a separate step. You pick a threshold based on costs. A spam filter might require P(spam | email) above 0.9 before hiding an email, because hiding a real email is expensive. A fraud screen might send anything above 0.3 to manual review, because missing fraud is expensive.

Calibration checks whether these conditional probabilities are honest: of all emails the model scored around 0.8, are about 80% really spam?

A real-life example

A telecom company's churn model outputs P(churn | features) for each customer. Customer A: 0.72. Customer B: 0.08. Retention offers cost ₹300 each and a lost customer costs about ₹4,000 in future revenue. Assuming an offer keeps the customer, the expected loss from not acting on customer A is 0.72 × 4,000 = ₹2,880, far more than the ₹300 offer. For customer B it is 0.08 × 4,000 = ₹320, roughly break-even.

So the team sends offers to customers with P(churn | features) above about 0.075, which is 300 / 4,000. The threshold comes straight from the conditional probability and the business costs, not from the default 0.5.

Follow-up questions to expect

• "Is P(A | B) the same as P(B | A)?" — No. P(spam | "free") = 0.75, but P("free" | spam) = 150 / 300 = 0.5. Bayes' theorem converts one into the other.
• "Why not always use a 0.5 threshold?" — 0.5 is only best when both errors cost the same and the classes are balanced; choose the threshold from the cost of false positives versus false negatives.
• "What is a calibrated model?" — One where predicted probabilities match observed frequencies; check with a reliability diagram and fix with Platt scaling, isotonic regression or temperature scaling.