metadata
library_name: xgboost
tags:
- fraud-detection
- phishing
- xgboost
- nexus
license: mit
NEXUS Fraud Detection — model artifacts
Trained heads for NEXUS Fraud Detection
(forked from Rishit1769/NEXUS-Fraud-Detection).
Each folder is a drop-in ml/artifacts/<task>/<version>/ directory: copy it into the
repo's ml/artifacts/ (or point MODEL_ARTIFACT_ROOT at this repo) and the
FastAPI model service (ml/service.py) picks it up with no code changes.
| Head | Folder | Model | Unseen-data result |
|---|---|---|---|
| Transaction fraud (XGBoost + isotonic, Optuna-tuned) | transaction/ |
model.json + calibration_model.joblib |
Fresh-seed PR-AUC 0.689; future-years slice 0.847 |
| Message phishing (TF-IDF char+word + lexical, XGB) | message-tfidf/ |
model.json + vectorizer.joblib |
18/18 hand-crafted probes |
| Message count baseline (XGB, fallback) | message/ |
model.json |
advisory-only fallback |
| URL phishing (lexical XGB) | url/ |
model.json |
19/20 probes |
eval_unseen.json is the full unseen-data ledger. Per-head metrics.json,
thresholds.json (recall>=0.80 operating points), feature_manifest.json,
split_manifest.json, and training_manifest.json document provenance.
Note: deterministic rules stay authoritative in NEXUS — these heads are advisory and never override a critical rule. Transaction data is synthetic (IBM); retrain on real adjudicated cases before production use.