Fraud detection in consumer-credit baskets

Problem

BNP Paribas Personal Finance wants to flag fraudulent baskets before a credit is approved (ENS Challenge Data no. 104). Each basket holds up to 24 items described across 147 columns, and only 1.42% of the 92,790 training baskets are fraudulent, so the metric is precision–recall AUC.

Method

We flattened the item slots into basket-level aggregates (item count, total price, category indicators), one-hot encoded categories and kept the fraud rate identical across an 80/20 split. A decision tree, logistic regression, a random forest and gradient boosting were compared; the boosted model was tuned with a randomised hyperparameter search (RandomizedSearchCV) under cross-validation.

Results

Gradient boosting placed first on both leaderboards. The most informative features were basket size and total price, a few high-risk categories, and brand × model interactions such as repeated high-end phones.