Comparison of DR and HD-SDR on HAR data with dimensions pre-selected using PCA

Comparison of DR and HD-SDR using 10 outliers (red crosses) and 5K random points uniformly-distributed in a 20-dimensional hypersphere centered at the origin and with different radii (R=1,2,3,4, and 5). The distances between the outliers and the origin are always 5. Note that t-SNE and UMAP and their sharpened versions fail to preserve the outliers even when the distance between the outliers and the cluster is the largest (row 1). Further note that SRP and SLMDS preserve similar or larger numbers of outliers compared with RP and LMDS, respectively.