spider
Detects change points in regularly sampled measurements and links them into chains of change.
Spider is a system that analyzes regularly sampled measurements to detect change points and connect them into chains: ordered sequences of changes in which one shift in a signal can plausibly trigger or precede the next. Its goal is not only to identify when a series’ behavior changes, but also to reconstruct how those changes propagate over time and across related series.
Components
- Change-point detection. Online and offline methods identify shifts. Because the measurements are regularly spaced, the system can use windowed statistics and FFT-based features without requiring interpolation.
- Event characterization. Each detected change is described by a compact signature that includes the type of change (level shift, spike, drift, or regime change), its magnitude, its duration, and its confidence level.
- Chain construction. Events are connected in a directed graph using temporal ordering, lag analysis, and a maximum-lag constraint. Paths through the graph represent the “chains of change” and are evaluated according to their statistical support.
- Dynamic updating. The pipeline operates in streaming mode. New samples incrementally update detections, while chains are extended, merged, split, or pruned as evidence accumulates, without needing to recompute the entire process.
Outcomes
- Benchmarks using synthetic data that include known ground-truth chains, as well as real-world datasets (industrial sensors, energy consumption, or market data).
- Metrics: detection delay, false-alarm rate, and precision and recall of the detected chains.