Random projection tree similarity metric for SpectralNet

SpectralNet is a graph clustering method that uses neural network to find an embedding separates the data. So far it was only used with k-nn graphs, which are usually constructed using distance metric (e.g., Euclidean distance). graphs restrict points have fixed number of neighbors regardless local statistics around them. We proposed new similarity based on random projection trees (rpTrees). Our experiments revealed produces better accuracy rpTree compared metric. Also, we found out parameters do not affect accuracy. These include leaf size and selection direction. It computationally efficient keep in order log(n), project onto direction instead trying maximum dispersion.