Unified conservation-based anomaly detection across all domains — one pattern (Laplacian → CR → threshold) applied to music, finance, climate, social, protein, neural, and PX4 data.
The core insight: anomalies violate the smooth structure captured by the graph Laplacian. Build a Laplacian, compute the conservation ratio, threshold on deviation. This atlas demonstrates the same detector architecture working across 7 domains with domain-specific graph construction.
- Universal anomaly pattern — Laplacian → conservation ratio → threshold
- 7 domains — Music, Finance, Climate, Social, Protein, Neural, PX4
- Domain-specific graphs — transition matrices, correlation networks, contact maps
- Conservation ratio as anomaly score — low CR = anomalous
- Visualization suite — multi-panel domain comparison plots
- Noise robustness — tested across noise levels
pip install numpy matplotlib
python anomaly_atlas.pyOutputs go to figures/ with multi-panel PNG plots.
Part of the SuperInstance ecosystem:
- conservation-spectral-python — Core SDK
- px4-conservation-poc — PX4-specific version
- anomaly-atlas — Unified cross-domain atlas (this repo)
MIT