TL;DR: MLE picks parameters that maximize the likelihood of the observed data; MAP adds a prior and maximizes the posterior, which is equivalent to MLE plus a regularization term (a Gaussian prior gives L2, a Laplace prior gives L1). For the medical test, Bayes shows that even a 99%-accurate test on a rare disease yields a low probability of actually being sick given a positive result, because the base rate dominates. Most people drastically overestimate that posterior.
How to approach it. Give the MLE/MAP contrast and tie MAP to regularization (the connection interviewers want), then work the Bayes numbers explicitly because the base-rate fallacy is the real test and a clean calculation is the strongest possible answer. Do not hand-wave the arithmetic; the interviewer wants to see the false positives swamp the true ones.
A strong answer. MLE vs MAP. Maximum Likelihood Estimation chooses parameters θ that maximize P(data | θ), the probability of the observed data under the model. Maximum A Posteriori maximizes P(θ | data) ∝ P(data | θ)·P(θ), adding a prior belief about θ. Take logs and MAP = MLE + log-prior, so the prior acts as a regularizer: a Gaussian prior on the weights is exactly L2 regularization, a Laplace prior is L1. With abundant data the likelihood dominates and MAP approaches MLE; with little data the prior matters, which is why regularization helps most when data is scarce.
| MLE | MAP | |
|---|---|---|
| Maximizes | P(data|θ) | P(θ|data) ∝ P(data|θ)·P(θ) |
| Prior | none | yes (Gaussian → L2, Laplace → L1) |
| Behavior | overfits with little data | prior shrinks toward sane values |
| Limit | MAP → MLE as data grows | adds bias, cuts variance |
Bayes / base-rate problem. A disease affects 1 in 1,000 people. A test has 99% sensitivity (true-positive rate) and a 5% false-positive rate. You test positive. What is the probability you have the disease?
P(D) = 0.001, P(~D) = 0.999
P(+|D) = 0.99, P(+|~D) = 0.05
P(+) = P(+|D)P(D) + P(+|~D)P(~D)
= 0.99 * 0.001 + 0.05 * 0.999
= 0.00099 + 0.04995 = 0.05094
P(D|+) = P(+|D)P(D) / P(+) = 0.00099 / 0.05094 ≈ 0.0194
So a positive result means roughly a 2% chance of actually having the disease, not 99%. The reason is the base rate: the disease is so rare that the few true positives are swamped by the many false positives drawn from the huge healthy population (0.05 × 999 ≈ 50 false positives vs ~1 true positive). This is the base-rate fallacy, and it is exactly why screening for rare conditions produces mostly false alarms and why precision collapses on rare-positive ML problems.
Key takeaways
- MAP = MLE + log-prior; the prior is regularization (Gaussian → L2, Laplace → L1) and matters most when data is scarce.
- The likelihood
P(data|θ)and the posteriorP(θ|data)are different objects; do not swap them. - For a rare disease, a 99%-sensitive test still gives only ~2% posterior after one positive, because false positives outnumber true ones ~50 to 1.
- This is precision under class imbalance restated, which is why you report PR-AUC and set thresholds by cost.
What interviewers probe next.
- "Connect MAP to a specific regularizer." Gaussian prior → L2 (ridge); Laplace prior → L1 (lasso, sparsity). The prior's strength is the regularization coefficient.
- "When does MLE overfit?" With little data or high model capacity; the prior in MAP pulls toward sane values and reduces variance.
- "How would a second positive test change it?" Apply Bayes again with the first posterior (~2%) as the new prior; independent positives compound and the probability rises sharply.
- "How does this map to ML metrics?" It is precision under class imbalance: even a strong classifier has poor precision when positives are rare, which is why you report PR-AUC and pick the threshold by cost.
Common mistakes.
- Confusing the likelihood
P(data|θ)with the posteriorP(θ|data). - Answering ~99% to the test question, ignoring the base rate.
- Not knowing MAP-with-a-prior is regularization (the connection interviewers are after).
- Treating the false-positive rate and the posterior as the same thing.