Business Context

PayLink, a digital payments platform processing 12M transactions per month, wants to improve account-level fraud detection. In addition to standard tabular features, the risk team believes relationships between users, devices, cards, merchants, and IPs contain useful signal that is not captured by flat features alone.

Dataset

You are given a heterogeneous graph built from 6 months of payment activity. Nodes represent accounts, devices, cards, merchants, emails, phone numbers, and IP addresses. Edges represent relationships such as used_device, used_card, transacted_with, logged_in_from_ip, and shares_phone.

• Target: Predict whether an account will be confirmed fraudulent within 30 days
• Class balance: 2.4% positive, 97.6% negative
• Missing data: 8% missing in account profile fields, 20% missing in some merchant metadata, sparse graph neighborhoods for new accounts

Success Criteria

A good solution should show that graph-derived embeddings or encodings improve fraud ranking over a tabular-only baseline. Target at least AUC-ROC > 0.88, PR-AUC > 0.30, and recall > 0.70 at precision >= 0.25.

Constraints

• Nightly batch scoring for 420K accounts; end-to-end inference under 30 minutes
• Risk analysts need interpretable signals alongside embeddings
• New accounts and unseen entities must be handled without failing inference

Deliverables

1. Build graph embeddings or graph encodings for account nodes
2. Combine graph features with tabular account features in a fraud classifier
3. Compare at least one embedding-based approach against a tabular baseline
4. Explain how you handle temporal leakage, cold-start nodes, and class imbalance
5. Report evaluation metrics and feature importance or diagnostic analysis