DR vs. HD-SDR for synthetic data

Comparison of DR and HD-SDR using five different types of synthetic data. DR methods are ordered from left to right based on how well clusters are separated. The samples are colored by the ground-truth labels of five different clusters for visual examination purposes. The results for SRP and SLMDS have been obtained with α=0.04 and for St-SNE (perplexity=50) and SUMAP with α=0.01. Note that our sharpening method enhances the separation of clusters for DR methods with less ability to separate clusters (i.e., RP and LMDS). The HD-SDR is however less effective for separating subclusters, as shown in the fourth row.