How aditya-ailsinghani/multimodal-fraud-detection works
This repository is a notebook-based multimodal fraud detection project. Its workflow loads and explores IEEE-CIS transaction data, builds tabular, text-derived, and transaction-network signals in parallel, then combines them in a LightGBM fusion model for a final fraud score and feature-importance analysis. The README documents these relationships; no source excerpts were available to verify finer-grained wiring.
Main entry points: 05_fusion_explainability.ipynb Fusion model · 01_data_loading_eda.ipynb Load and explore · 02_tabular_model.ipynb Tabular model · 03_nlp_module.ipynb NLP features
Components 8
Data preparation
- Load and explore
notebooks/01_data_loading_eda.ipynb
Detection signals
- Tabular model
notebooks/02_tabular_model.ipynb - NLP features
notebooks/03_nlp_module.ipynb - Graph features
notebooks/04_graph_module.ipynb
Fusion and analysis
- Fusion model
notebooks/05_fusion_explainability.ipynb - Feature analysis
notebooks/05_fusion_explainability.ipynb
Other components
- IEEE-CIS data
- Fraud assessment
Connections 9
- IEEE-CIS data → Load and explore: loads
- Load and explore → Tabular model: provides data
- Load and explore → NLP features: provides data
- Load and explore → Graph features: provides data
- Tabular model → Fusion model: provides score
- NLP features → Fusion model: provides features
- Graph features → Fusion model: provides features
- Fusion model → Fraud assessment: produces score
- Fusion model → Feature analysis: analyzes importance