In this repository we present two demo notebooks how spectral clustering can be applied to extract relevant dynamics out of a variety of monitored metrics.
The toy example notebook sketches how spectral clustering works on a small perturbed block matrix with three (quasi-)blocks.
In the real data example the algorithm is applied to real masked data. For this purpose is specified a similarity function, then computed a cluster assignment and reordered similarity matrix according to the identified quasi-blocks.
Once we have identified strongly connected clusters of metrics, we can compute a representative for each cluster by averaging over all assigned metrics.
Eventually we give an visualization that provides a better insight, which metrics enter into the particular clusters together with a confidence corridor for each represantative.
For real data example we used the scikit-learn implementation of Spectral Clustering