You have a fraud dataset where 0.3 percent of rows are positive. How do you handle that imbalance?

First change the metric, since accuracy is useless at 0.3 percent: use precision recall AUC, or recall at a fixed alert budget. Then try class weights before resampling, because weighting is simpler and does not fabricate rows. If you do resample, do it inside the training fold only. Finally recalibrate probabilities, since weighting and oversampling both distort the predicted scores.

The first thing I change is the scoreboard, because a model that predicts no fraud every time is 99.7 percent accurate and completely worthless. I use precision recall AUC and, more usefully, recall at whatever alert volume the team can actually review. Then I start simple with class weights in the loss, which costs nothing and often gets most of the benefit. I only reach for SMOTE or undersampling if weighting stalls, and honestly I have found synthetic minority samples underwhelming on tabular fraud data, because the interpolated points sit in regions of feature space that do not exist. Two things I am strict about. Resampling happens inside the training fold, never before the split, or the validation set ends up holding near duplicates of training rows and the score is fiction. And after any of this the predicted probabilities are no longer calibrated, so if anyone downstream multiplies the score by a dollar amount, I fit a calibration step on a clean holdout. On my last fraud model that shifted the expected loss estimate more than the model change did.