DATAVERSE: P2P Payment Scam Detection (Graph-Based)
Team CodeBlitz
KIET A+ GROUP OF INSTITUTIONS
1. Presentation Outline and Script
This script is designed for a high-impact presentation, covering all six slides and points in a
logical flow, suitable for FinTech judges.
Slide 1: Title Page (DATAVERSE)
Content Problem Statement Title: ML for fraud detection & credit scoring
Content Theme: FinTech (Smart Finance with AI)
Content Team Name: CodeBlitz
Speaker Script:
(Open with energy and clear objective) Good morning, everyone. We are Team CodeBlitz,
and we are here to talk about a growing crisis in the FinTech space: P2P payment fraud. Our
project, **DATAVERSE**, introduces a next-generation solution using Machine Learning for
real-time fraud detection. We are focused squarely on the Smart Finance with AI theme.
Slide 2: Proposed Solution (The Core Problem and Our USP)
Content Idea Title: DATAVERSE
Content Proposed Solution: Develop an AI-driven fraud detection framework that models Peer-to-
Peer (P2P) transactions and detects scam rings, money-mule clusters, and account takeovers
in real time.
Content System Components: Admin Dashboard, Transaction Engine, Data Layer (Graph Con-
struction), AI/ML Component (GNNs + Anomaly Detection), Real-Time Detection & Alert
System.
Content USP (Unique Selling Proposition):
• Understands multi-user relationships unlike flat transaction models.
• Detects network-level fraud behavior that rule-based or tree models cannot visualize.
• Novel: Very few production systems currently implement graph ML in P2P fraud de-
tection.
Speaker Script:
(Hook: Define the Problem) The problem is that traditional fraud detection fails against
syndicated P2P fraud. Why? Because existing systems look at transactions in isolation: *User
A paid $50 to User B*. They miss the full story.
(Introduce the Solution) Our solution, DATAVERSE, revolutionizes this by shifting the per-
spective. We don’t see transactions; we see a Graph .
(Focus on the USP) This is our core innovation, our USP: We model every user as a Node and
every payment as an Edge. A single suspicious payment might look fine, but a *network*
of suspicious payments—like a new node rapidly receiving money from 50 unique users and
cashing out—reveals a hidden scam ring. We are using **Graph Neural Networks (GNNs)**,
a modern, powerful technique, specifically designed to detect these network-level conspira-
cies that standard ML models simply cannot visualize.
2
Slide 3: Technical Approach and Process Flow
Content Process Flow: USER INTERACTION → DATA PROCESSING PIPELINE → AI AND DETECTION
→ ANALYST INTERACTION (with a Feedback Loop).
Content Data Processing Pipeline: Data Ingestion → Graph Construction → Feature Extraction.
Content AI and Detection: GNN Model Training → Anomaly Detection.
Content Technologies:
• Languages: Python, SQL.
• Frameworks: PyTorch Geometric, NetworkX, Neo4j, Scikit-learn.
• Deployment: FastAPI for real-time inference, ngrok.
Speaker Script:
(Explain the Tech Flow) Let’s walk through the architecture . When a user logs in and makes a
transaction (User Interaction), that raw data is ingested. Crucially, it’s immediately converted
into our Graph structure using **NetworkX and Neo4j**.
(Explain the AI) This transaction graph feeds our **AI and Detection** module, powered by
**PyTorch Geometric**, which is optimized for GNNs. The GNN learns the ’normal’ behav-
ior of user relationships, allowing our **Anomaly Detection** layer (like Isolation Forest) to
instantly flag deviations—the sudden appearance of a money mule or an account takeover.
(Show the Outcome) The resulting risk score is then pushed to the **Fraud Dashboard**,
where an analyst can review it. The key here is the Feedback Loop: analyst decisions are
fed back into the GNN for continuous, adaptive learning, ensuring our model stays ahead of
evolving fraud tactics.
Slide 4: Feasibility and Viability
Content Feasibility: Works on existing transaction and user metadata (no additional data collec-
tion required). Highly scalable using cloud-based graph databases (e.g., Neo4j, Tiger-
Graph). Compatible with P2P apps (UPI, PayPal, Venmo, GPay).
Content Potential Challenges & Risks:
• Data imbalance (genuine users far outnumber fraud cases).
• Graph drift (user behaviors evolve over time).
• Computational cost of training large dynamic graphs.
Content Strategies to Overcome:
• Oversampling for balance.
• Online/continual learning to adapt to graph drift.
• Optimized model inference using mini-batch training and edge sampling.
Speaker Script:
(Address Practicality) DATAVERSE is not just a concept; it’s highly viable. We leverage existing
transaction data—no complex new infrastructure is needed for data capture. We’ve chosen
scalable technologies like **Neo4j** and TigerGraph to ensure we can handle millions of
real-time transactions across platforms like UPI, PayPal, and Venmo.
(Show Domain Knowledge - Challenges) Of course, advanced ML presents challenges. We
acknowledge data imbalance is a major factor, as are graph drift and computational cost.
(Provide Confidence - Mitigation) But we have clear strategies to mitigate these risks: **Over-
sampling** to balance the data, Online/Continual Learning so our GNN can adapt to new
behaviors instantly, and optimized techniques like **mini-batch training** to control cost.
This demonstrates not just an idea, but a production-ready system blueprint.
3
Slide 5: Impacts and Benefits
Content Impacts:
• Early Detection: Identifies fraud rings within minutes instead of days.
• Scalable Protection: Can handle millions of real-time transactions.
• Trust Enhancement: Increases user confidence in P2P platforms.
• Industry Advancement: Introduces graph learning to mainstream FinTech fraud pre-
vention.
Content Categorized Benefits: ECONOMIC, TECHNOLOGICAL, SOCIAL, OPERATIONAL.
Speaker Script:
(Maximize Impact Statement) The benefits are transformative, touching four key pillars.
Technologically, we are pushing the industry forward by moving from archaic rule-sets to
cutting-edge graph learning.
Operationally, we provide Early Detection—identifying fraudulent networks in minutes, not
days. This minimizes losses and speeds up intervention.
Economically, this scalable protection means platforms can handle exponential growth in
P2P volume without a corresponding rise in fraud loss.
And finally, Socially, in an era of digital mistrust, DATAVERSE fundamentally enhances user
confidence and trust in their payment apps, securing the financial ecosystem for everyone.
Slide 6: Research and References
Content Research Papers & Studies:
• ”Graph Neural Networks for Fraud Detection” - ACM Transactions, 2023
• ”Dynamic Graph Learning for Transactional Anomalies” - IEEE BigData 2022
• ”GNNExplainer: Interpretable Graph Neural Networks” - NeurIPS 2021
Content Datasets Used / Referenced:
• IEEE-CIS Fraud Detection Dataset
• Kaggle Datasets for Fraud Detection
• PaySim (Simulated Mobile Money Transactions)
• Open Graph Benchmark (OGB) Datasets
Speaker Script:
(Establish Credibility) Our solution is deeply grounded in current academic and industry
research. We have leveraged insights from leading papers on GNNs and dynamic graph
learning, as shown here. This is not just a hypothesis; it’s an informed, research-backed
engineering project.
(Show Practicality) We have a clear path to execution, having identified and referenced sev-
eral key datasets, including the IEEE-CIS and PaySim, which will be instrumental in training
and validating our model.
(Conclusion) DATAVERSE is the necessary evolution of fraud detection—from looking at sin-
gle users to securing the entire payment network. Thank you. We are now open for ques-
tions.