How Would You Tackle a Severely Imbalanced Classification Dataset?

Data level: Oversampling / undersampling, SMOTE-style synthetic samples. Always split first, then sample, to avoid leakage.

Algorithm level: class_weight / scale_pos_weight, focal loss to downweight easy samples, threshold moving (tune the decision threshold by business cost, not always 0.5).

Evaluation level: Do not look at accuracy. Use PR-AUC, F1, recall at fixed precision, and report the confusion matrix.

First separate sample-count imbalance from difficulty imbalance. Often the problem is not the ratio - the minority class has too few samples to learn patterns. More data fixes the root cause better than resampling tricks.