Card Fraud Detection Under Review Capacity

Preview mode. Browse milestones and edit starter code here. Sign in to run and submit.

0/8EDA

Python

advanced~8h8 milestones

The problem

A payments company loses money to card fraud, but its manual review team can only investigate a small fraction of transactions per day. The model you build decides which transactions enter that queue. Flag too few and fraud goes through. Flag too many and the queue backs up while good customers get declined. Your job is to find the operating point that saves the most money inside the capacity the business actually has.

Your task · Measure the Imbalance

Establish the fraud rate overall and, separately, for accounts with no spend history.

Implement explore_fraud(train_df) returning a dict with:

row_count: number of training rows

fraud_count: number of fraudulent transactions

fraud_rate: fraction of transactions that are fraud, rounded to 6 places

missing_avg_amount_30d: count of NaN values in avg_amount_30d

fraud_rate_new_accounts: fraud rate among rows where account_age_days < 30, rounded to 6 places

fraud_rate_established: fraud rate among the rest, rounded to 6 places

The last two numbers are the point of this milestone. Compare them before you move on.

Constraints

▪About 1.5% of transactions are fraudulent, so accuracy is useless
▪avg_amount_30d is missing for accounts under 30 days old, and that cohort is roughly twice as risky
▪Manual review capacity is 2% of transactions per day
▪Recovering a fraudulent transaction is worth $180, a false alarm costs $6
▪Hidden test set is scored server-side, so you cannot tune against it

Graded on

PR-AUC >= 0.22 and ROC-AUC >= 0.88 on the hidden test set
Recall >= 0.32 while alerting on no more than 2% of transactions
Net benefit >= $22,000 with alerts capped at 3%
Batch scoring completes within 8 seconds

Evaluated server-side against a hidden test set.

Skills you'll demonstrate

Class ImbalanceInformative MissingnessPrecision-Recall AnalysisThreshold SelectionCost-Sensitive LearningGradient BoostingBatch Inference