import joblib
import pandas as pd
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(repo_id="yahiaehab10/fraud-ccf-lightgbm", filename="pipeline.pkl")
pipeline = joblib.load(model_path)

# Download threshold
threshold_path = hf_hub_download(repo_id="yahiaehab10/fraud-ccf-lightgbm", filename="threshold.json")
import json
threshold = json.load(open(threshold_path))["threshold"]

# Make predictions
# X should have columns: Amount, V1, V2, ..., V28
probabilities = pipeline.predict_proba(X)[:, 1]
predictions = (probabilities >= threshold).astype(int)

Performance

Source: ULB/Kaggle Credit Card Fraud Dataset
Timeframe: 2 days of transactions
Positive Rate: 0.172% (highly imbalanced)
Features: Amount + V1-V28 (PCA-transformed features)

The model is optimized for fraud detection with a focus on minimizing false positives while maintaining high recall for fraud cases.

Limitations

Educational purposes only - Not intended for production use
Trained on historical data - may not generalize to future fraud patterns
Highly imbalanced dataset - requires careful threshold calibration

License

Educational use only. Please refer to the original dataset license.