Question

A data scientist is preparing a tabular fraud-detection dataset in which only 1.2% of records are labeled as fraud. Initial models trained on the raw data achieve high accuracy but almost never predict the positive class. The team wants to address the class imbalance during data preparation without discarding legitimate transaction records and without generating unrealistic synthetic samples that could distort the minority distribution. Which approach best meets these requirements?

Accepted Answer

Keep all records and assign class weights (or scale_pos_weight) so the training algorithm penalizes minority-class errors more heavily.

Answer

Apply random undersampling of the majority (non-fraud) class until both classes are roughly balanced before training.

Answer

Use SMOTE to synthesize new minority-class examples until the classes are balanced, then train on the augmented dataset.

Answer

Drop the label column and switch to an unsupervised anomaly-detection algorithm such as Random Cut Forest.