>>17784327>>17784375I propose that we use a spectral clustering algorithm to do dimensionality reduction of the /vsj+/ chuuba set by computing the eigendecomposition of the similarity matrix of Twitch's chuuba-connectivity graph. We could experiment by running the Shi-Malik algorithm recursively to generate an appropriate hierarchy or maybe use the LOBPCG method (Locally Optimal Block Preconditioned Conjugate Gradient), just in case our iterative eigenvalue solver runs into matrix conditioning problems.
