3

I am using t-SNE to visualize cytometry data. Most of guides I found (https://distill.pub/2016/misread-tsne/) warn how the choice of perplexity hyperparameter can influence the result.

However, my data set size is really small, always expecting 10-30 points since visualizing one point per cluster only. In this case, does a constant, reliable value of perplexity exist OR a mehod how to estimate it.

ludgo
  • 465
  • 1
  • 6
  • 13
  • It's more about the data-statistics itself than the size. Go by your own eye! – sascha Nov 12 '17 at 17:37
  • I am stuck on the exact same problem. Small dataset size, getting different plots for different perplexities and learning rates..I know the data is separable because a linear classifier can separate the data, but if i pick the tsne with the lowest kl div, this does not give me any clusters. Some random values do. I am struggling to find a reliable method. – Tarun Mar 13 '18 at 18:45

0 Answers0