Skip to main content
. 2020 May 19;8:2120. Originally published 2019 Dec 18. [Version 2] doi: 10.12688/f1000research.21642.2

Figure 2. Performance of EmbedSOM variants compared with other dimensionality reduction methods.

Figure 2.

The speed is represented in cells per second. EmbedSOM-based algorithms show almost perfect linear scaling with growing dataset size, and even minor speed improvements when sufficient data is available for saturating the parallel computation. As expected from their asymptotic complexities, performance of UMAP, TriMap and t-SNE decreased with additional data. t-SNE was not executed on datasets larger than 50 thousand cells because of time constrains.