| ---
|
| tags:
|
| - fraud-detection
|
| - tabular-classification
|
| - lightgbm
|
| - shap
|
| library_name: sklearn
|
| ---
|
|
|
| # FraudLens fraud detector (lightgbm, version 2)
|
|
|
| Credit card fraud scoring model from **FraudLens** (https://github.com/YASHR2002/fraudlens): leakage-free behavioural
|
| features, a cost-based decision threshold, SHAP explanations and LLM analyst notes.
|
| This repo is what the public demo API downloads at startup.
|
|
|
| | | Test set (fraudTest, used once) | Validation |
|
| |---|---|---|
|
| | PR-AUC | 0.9745 | 0.9853 |
|
| | Recall at threshold | 95.2% | 97.6% |
|
| | Precision at threshold | 86.1% | 90.1% |
|
| | Total cost (missed fraud + $5 per false alarm) | $25,150 | $2,975 |
|
|
|
| Decision threshold: **0.4326** (minimises total cost on validation).
|
|
|
| ## Contents
|
|
|
| - `champion/model/`: scikit-learn pipeline in MLflow format, serialised with **skops**
|
| (load with the trusted types listed in `fraudlens.models.estimators.SKOPS_TRUSTED_TYPES`).
|
| - `champion/metadata.json`: version, threshold, metrics, feature importance, fairness.
|
| - `features.json`: the 18 input features with plain-English descriptions.
|
| - `state/`: per-card state snapshot (end of the training data) and demo transactions.
|
|
|
| ## Data and limitations
|
|
|
| Trained on the synthetic Sparkov dataset (CC0), so scores are higher than real fraud detection
|
| would achieve; card numbers and personal fields in the demo files are synthetic. Gender is not a
|
| model input; age is, and fairness gaps by gender and age are reported in the project's model card.
|
| Not for production use.
|
|
|
