ARUNAGIRINATHAN/SentinelX_v1.0_LGBM

SentinelX v1.0 — LightGBM Fraud Detector

A LightGBM-based binary classifier for tabular transaction fraud detection. Trained via run_pipeline.py, with features engineered in features/build_features.py. Outputs per-transaction fraud probability.

Usage

from huggingface_hub import hf_hub_download
import joblib
import pandas as pd

repo_id = "ARUNAGIRINATHAN/SentinelX_v1.0_LGBM"

model_path = hf_hub_download(repo_id=repo_id, filename="model.joblib")
model = joblib.load(model_path)

feat_path = hf_hub_download(repo_id=repo_id, filename="features.txt")
with open(feat_path, "r") as f:
    features = [line.strip() for line in f]

Intended Use

• Use for ranking or thresholded classification of potentially fraudulent transactions.
• Not production-ready without calibration, fairness checks, drift monitoring, and human-in-the-loop review.

Model Details

• Algorithm: LightGBM (binary objective)
• Input: Tabular features saved in features.txt
• Output: Probability of fraud (predict_proba[:, 1])
• Artifacts: model.joblib, lightgbm_model.txt, features.txt, shap_summary.png, submission.csv

Data & Training

• Training data: Dataset/train_transaction.csv (local file, not included on the Hub)
• Process: ingestion/load_data.py → features/build_features.py → models/train_lgbm.py → models/evaluate.py
• Run: python run_pipeline.py (artifacts saved under artifacts/)

Evaluation

• Metrics (validation): ROC AUC, Precision, Recall, F1.
• Update with your results:
  • ROC AUC: TBD
  • Precision/Recall/F1: TBD

Limitations

• Performance depends on data quality and drift; retrain periodically.
• Threshold selection requires business calibration and cost analysis.