AstroML is a research-driven Python framework for building dynamic graph machine learning models on the Stellar Development Foundation Stellar blockchain.
It treats blockchain data as a multi-asset, time-evolving graph, enabling advanced ML research on transaction networks such as fraud detection, anomaly detection, and behavioral modeling.
AstroML provides end-to-end tooling for:
- Ledger ingestion and normalization
- Dynamic transaction graph construction
- Feature engineering for blockchain accounts
- Graph Neural Networks (GNNs)
- Self-supervised node embeddings
- Anomaly detection
- Temporal modeling
- Reproducible ML experimentation
Blockchain networks are naturally graph-structured systems:
| Blockchain Concept | Graph Representation |
|---|---|
| Accounts | Nodes |
| Transactions | Directed edges |
| Assets | Edge types |
| Time | Dynamic dimension |
Most analytics tools rely on static heuristics or SQL queries.
AstroML instead enables:
- Dynamic graph learning
- Temporal GNNs
- Representation learning
- Research-grade experimentation
AstroML is designed for:
- ML researchers
- Graph ML engineers
- Fraud detection teams
- Blockchain data scientists
Ledger → Ingestion → Normalization → Graph Builder → Features → GNN/ML Models → Experiments
git clone https://github.com/Traqora/astroml.git
cd astromlpython -m venv venv
source venv/bin/activate
pip install -r requirements.txtCreate a PostgreSQL database and update:
config/database.yaml
Backfill ledgers:
python -m astroml.ingestion.backfill \
--start-ledger 1000000 \
--end-ledger 1100000Create a rolling time window graph:
python -m astroml.graph.build_snapshot --window 30dpython -m astroml.training.train_gcn- Fraud / scam detection
- Account clustering
- Transaction risk scoring
- Temporal behavior modeling
- Self-supervised embeddings
- Network anomaly detection
AstroML emphasizes:
- Reproducibility
- Modular experimentation
- Scalable ingestion
- Temporal graph learning
- Production-ready ML pipelines
- Python
- PyTorch / PyTorch Geometric
- PostgreSQL
- NetworkX / graph tooling
- Real-time streaming ingestion
- Temporal GNN models
- Contrastive learning pipelines
- Feature store
- Model benchmarking suite
- Docker deployment
Contributions are welcome!
fork → branch → commit → PRPlease open issues for bugs or feature requests.
MIT License