01The problem
Suspicious transactions can form patterns across accounts and time. Detection also needs an explanation and a human review workflow.
02The approach
- Extract bounded temporal-graph features and train logistic-regression and boosted-tree models with chronological evaluation.
- Compare graph features with transaction-only and rule baselines, keeping labels out of online scoring.
- Attach evidence graphs to cases with analyst notes, a Java audit trail and separate requester/approver roles.
03The evidence
- The public report evaluates IBM AMLworld HI-Small synthetic transactions.
- The full-window graph/boosted-tree result reports F1 0.576; a separate sensitivity window reports F1 0.417.
- Source code and benchmark reports are publicly available.
04Scope & status
A personal portfolio project using synthetic data. It is not connected to a bank or regulator; benchmark sensitivity and analyst-review limits are retained in the report.
Public source is available for independent inspection.