AI and machine learning interview questions

It is linear in the log-odds, not in the probability.

A small change in data can pick a different first split and rearrange the whole tree.

Some implementations learn surrogate splits that mimic the primary split when a value is missing.

max_depth caps the longest path and is a coarse global budget.

Forest error falls when individual trees are strong and when their errors are weakly correlated.

Vote fractions from deep trees tend to pile up near 0 and 1.

A second-order, Newton-style step uses the Hessian of the loss, not only the gradient.

Subsampling rows each round, as in stochastic gradient boosting, makes trees less correlated and faster.

A monotonic constraint forces the prediction to rise or fall as a feature rises, matching a business rule such as risk increasing with debt.

Hold out a validation set, or use a CV fold, and stop when the evaluation metric stops improving for a patience window.

The dual writes the decision function as a weighted sum of kernel evaluations against training points.

Training complexity grows poorly with n, often between quadratic and cubic depending on the solver.

One-vs-rest trains K binary models and is cheaper when K is large.

Use approximate nearest neighbor indexes such as HNSW, IVF, or Annoy to trade a little recall for latency.

Cosine ignores vector length and fits sparse text or l2-normalized embeddings.

A GMM treats clusters as Gaussians with their own covariances and uses soft assignments.

The gap statistic compares observed within-cluster dispersion with that of a null reference distribution.

In high dimensions, Euclidean K-means suffers from distance concentration, so a PCA or autoencoder projection can help.

If X is centered, the right singular vectors of X are the principal axes and the squared singular values relate to explained variance.