We projected the multiscale KNN graph to two dimensions using a modified t-SNE algorithm we call graph t-SNE (gt-SNE). In contrast to standard t-SNE, which is based on distance measures, we directly projected the multiscale KNN, which is based on multiscale weighted ranks. We achieved this by replacing the distance matrix P in regular t-SNE with the distance matrix of the weighted multiscale KNN graph. The result was a more accurate projection of the graph itself, with more compact and well-defined neighborhoods. We stored the gt-SNE embedding as column attributes _X and _Y in the Loom files.
