abdu-l7hman/aml-isolation-forest

AML Isolation Forest (unsupervised)

Isolation Forest trained only on normal transactions over 5 behavioural features (amount_sar, txn_count_24h, txn_count_7d, amount_sum_24h, distinct_counterparties_7d). Alert threshold set at the 99% quantile of validation scores (target false-positive rate 1%).

Overall (unseen typologies): ROC-AUC 0.855, recall 4.0% at FPR 1.16%. Per-typology results are in config.json. drift_reference.json holds the PSI reference bins.

Trained on a synthetic AML dataset. Not for production use on real customer data without validation. Load with joblib.load("model.joblib"); score = -model.score_samples(X); flag if score > threshold in config.json.