Supervised Fraud Detection Data Flow

Labeled data and features through rebalancing, training, cost-based thresholding, to a decision

Supervised Fraud Detection Data Flow Labeled data and features through rebalancing, training, cost-based thresholding, to a decision 01 / Sources 02 / Rebalance 03 / Train 04 / Calibrate 05 / Decide Labeled Txns · matured labels · 01 / Sources · rare positives Labeled Txns matured labels rare positives Feature Store · leak-audited · 01 / Sources · point-in-time Feature Store leak-audited point-in-time Rebalance · class weights · 02 / Rebalance · weighted Rebalance class weights weighted Train Classifier · gradient boosting · 03 / Train · fit Train Classifier gradient boosting fit Cost Matrix · FP vs FN · 03 / Train · $ asymmetry Cost Matrix FP vs FN $ asymmetry Calibrate + Threshold · min expected cost · 04 / Calibrate · isotonic Calibrate + Threshold min expected cost isotonic Fraud Score · calibrated prob · 05 / Decide · PR-AUC Fraud Score calibrated prob PR-AUC Approve / Review / Block · two cutoffs · 05 / Decide · policy Approve / Review / Block two cutoffs policy labeled events supervised feature vectors point-in-time weighted batches in-fold only raw scores ranked FP / FN costs business calibrated prob actionable score + cutoffs policy Legend primary data policy / PII async batch data store

Training Path

  • • Labeled transactions and features merge into a single rebalancing step
  • • Class weights reshape the loss without discarding data
  • • Resampling, if used, stays strictly inside the training fold

Cost-Driven Threshold

  • • The cost matrix encodes asymmetric FP versus FN dollars
  • • Calibration makes the score a usable probability
  • • The threshold minimizes expected monetary cost, not error count

Decision Policy

  • • Two cutoffs split traffic into approve, review, and block
  • • Scores are evaluated with PR-AUC and recall-at-precision
  • • Automated precision stays high while margins go to review