About perplexity about tsnet HOT 4 CLOSED

Hello, and thanks for your interest!

I don't really know what's happening there. All I know is that it's happening in the pre-processing stage where the standard deviations of the Gaussians are sought, to attain the requested perplexity.

You can change the perplexity with the -p flag. I assume you've tried increasing it?

Is there anything worth mentioning about the dataset you're using? Is it large? Disconnected? Sparse?

from tsnet.

Thanks for your answer! It's much faster than I thought.

Yes, I have tried changing the perplexity with the -p flag.

The dataset is disconnected. And the range of perplexity is very very narrow, just like I can only choose exactly one value. But when I deleted the disconnected nodes, the problem has gone.
Maybe the problem is about the disconnected network?

from tsnet.

Alright, then it's probably related to how distances are set to non-connected vertices.

Right now, those distances are set to k times the largest shortest-path distance between any two connected vertices. k is set to 10 (relevant code in modules/distance_matrix.py), which may be a bit excessive. You can try playing around with that value. I hope that helps!

I'd also gladly accept a pull request where the k parameter is exposed to the CLI.

from tsnet.

Yes, I have tried adjusting the k parameter. I also tested other disconnected datasets. I found the problem only happens when there exist nodes whose degree are zero.( In this situation, adjusting k has little effect)
I'll try to add the k to the CLI.
Thank you!

from tsnet.