Skip to content

NeuroSymbolic AI for explainable financial fraud detection

Status
Completed
Date
2026
Venue
MS AI project, Bahria University
Tags
Explainable AI · Neuro-symbolic · Imbalanced learning · Ontologies

  • 92.86%

    Precision

  • 84.94%

    ROC-AUC

  • 76.47%

    F1

    vs. 50.67% Random Forest, 44.40% XGBoost alone

Abstract

Fraud models are usually accurate or explainable, rarely both. This project joins an XGBoost classifier, trained with SMOTE to handle extreme class imbalance, to an OWL 2 domain ontology. Model predictions are mapped into the ontology, rule-based reasoning sorts each transaction into a High, Medium or Low risk tier, and every decision carries a natural-language explanation, aimed at the explainability expected by GDPR Article 22 and the EU AI Act.

01Pipeline

  • Statistical layer: XGBoost with inline SMOTE over-sampling, keeping probabilities calibrated.
  • Symbolic layer: owlready2 maps prediction vectors into an OWL 2 ontology built in Protégé; Pellet/HermiT reasoners run on top.
  • SWRL-style rules assign High, Medium and Low risk tiers and can override edge-case neural cut-offs when they find temporal or domain anomalies.
  • Every decision gets a hasExplanation justification in plain language.

02Results

The hybrid reaches 92.86% precision, 84.94% ROC-AUC and 76.47% F1, against 50.67% F1 for Random Forest and 44.40% for XGBoost on its own.