Healthcare Fraud Detection via Knowledge Graphs

A workshop project for exploring open-source Python knowledge graph tools to detect fraud in healthcare claims data.

What's in here

data/
  generate_dataset.py      — synthetic data generator (run this first)
  patients.csv             — 300 patients (4 deceased)
  providers.csv            — 60 providers (5 fraudulent)
  claims.csv               — all claims WITH fraud labels  ← instructor only
  claims_public.csv        — all claims WITHOUT labels     ← distribute this
  referrals.csv            — provider-to-provider referrals

notebooks/
  01_solution_fraud_detection.ipynb  — full solution with all 5 patterns  ← instructor only
  02_workshop_challenge.ipynb        — competition notebook for participants

Five fraud patterns embedded in the data

#	Pattern	Fraudster	Signal
1	Ghost Billing	PRV-047	Claims filed after patient date_of_death
2	Referral Ring	PRV-031/032/033	Closed Louvain community, inflated amounts (2–3×)
3	Impossible Travel	PAT-0089	Same-day claims in cities 1,300 miles apart
4	Upcoding	PRV-022	99215 billing rate 8× above peers (3σ outlier)
5	Duplicate Billing	PRV-055	Parallel edges in multigraph (same patient/procedure/date)

Quick start

# Install dependencies
uv sync

# Generate the dataset
uv run python data/generate_dataset.py

# Launch Jupyter
uv run jupyter notebook

Then open notebooks/01_solution_fraud_detection.ipynb (instructor) or notebooks/02_workshop_challenge.ipynb (participants).

Key packages

Package	Role
networkx	Graph construction, traversal, subgraph analysis
python-louvain	Louvain community detection (import community)
pyvis	Interactive HTML graph visualisation
pandas / numpy	Data wrangling
matplotlib / seaborn	Static plots

Workshop format

1. Distribute data/claims_public.csv, data/patients.csv, data/providers.csv, data/referrals.csv
2. Each team works through notebooks/02_workshop_challenge.ipynb
3. Teams submit findings by calling score_team('Team Name') — max 100 pts
4. Run notebooks/01_solution_fraud_detection.ipynb for the debrief