Abstract

Credit fraud and billing anomalies impose significant economic losses on financial institutions worldwide. This paper introduces Credit Sentinel, a synergistic multi-layered ML architecture that combines unsupervised anomaly detectors, supervised classifiers, and rule-based filters in a unified inference pipeline.

Architecture

The system operates in three stages:

  1. Tier 1 — Unsupervised Pre-screening: Isolation Forest and Local Outlier Factor flag statistical outliers in transaction feature space.
  2. Tier 2 — Supervised Classification: XGBoost and LightGBM models trained on labeled fraud instances refine the candidate set.
  3. Tier 3 — Rule Engine Post-filter: Business logic constraints remove false positives based on account history and merchant context.

Performance

Achieves F1 score of 0.943 on the benchmark dataset, outperforming all single-model baselines while maintaining sub-20ms inference latency.

Citation

Journal: SCIENTIFIC CULTURE
DOI: 10.5281/zenodo.18681376