AI and machine learning interview questions

A kernel computes a similarity that matches a dot product in a richer feature space without building that space by hand.

KNN finds the K closest training rows under a distance metric.

Small K is flexible and noisy

Distance is dominated by large-scale features if you skip scaling.

Clustering groups similar unlabeled rows together.

K-means places K centroids, assigns each point to the nearest centroid, then moves each centroid to the mean of its points.

You plot inertia, the within-cluster sum of squares, against K.

Hierarchical clustering returns a dendrogram and does not need K chosen in advance.

PCA finds orthogonal directions of maximum variance so correlated features can be compressed.

A principal component is a linear combination of the original features.

Yes, unless the features already share a meaningful common unit.

Feature engineering turns raw fields into inputs that make the pattern easier to learn.

One-hot encoding turns a categorical column into binary columns, one per category.

Scaling puts numeric features on a comparable range.

Normalization often means min-max scaling into a 0-1 range.

Imbalance means one class is much rarer than another, such as one percent fraud.

If 99 percent of rows are negative, always predicting negative scores 99 percent and finds zero fraud.

Use class weights, oversample the minority, undersample the majority, or create synthetic minority rows.

A Pipeline chains transformers and a final estimator so they fit and predict as one object.

If you scale the whole table before splitting, validation rows leak into the scaler.