AttributeError: 'Heat' object has no attribute 'analysis'

[sgrady3]

I've been trying to use the graphwave function and there seems to be a bug. When I run the function graphwave on a networkx graph I get the error mentioned in the title: AttributeError: 'Heat' object has no attribute 'analysis' I've tracked this to the file heat_diffusion.py line 53: Sf_vec = Hk.analysis(np.eye(n_nodes))

I looked that the pygsp filter class documentation for a function called analysis but only found analyze. When trying to replace analysis with analyze in heat_diffusion.py line 53 I get an index range error so I don't think this is the fix.

I hope this helps, I'm very interested in using this technique and would like to be able to use this code.

-Thank you

[donnate]

Hi,

Thank you for picking up this mistake. This is due to an update in the pygsp Library. I have fixed this and made the code independent from pygsp (all you now need is numpy, scipy, pandas and networkx). Please let me know if you encounter further issues!

[sgrady3]

Thank you! I do have one question. The current implementation of graphwave_alg as an argument for a list of time points. Do you have an intuition for what would be a good selection of time points?

[donnate]

Hi,

For the list of time points, I typically used 25 evenly spaced points in 0:100 (np.linspace(0,100,25)). In order to "guide" the choice of the range and number of samples, I found it useful to plot a few parametric curve induced by the characteristic functions, and visually assess if they are all clamped together, etc. This gives you a way to assess visually the amount of signal in the embeddings. Hope this helps! Do let me know if you have any additional issues!

[sgrady3]

Great, thank you the advice. I do have one other question. I saw in the paper that the embeddings are independent of graph size. Are the embeddings always of size 700?

[donnate]

Hi, No, the embedding size varies: each node structural embedding is of size equal to: 2 x nb_timepoints x nb_scales where:

• the 2 comes from the fact that we have a real and imaginary part,
• nb_timepoints is the number of time points that you use to sample along the parametric curve induced by each characteristic function
• and nb_scales is the number of different scales that you stack up in your multi-resolution signature.

[sgrady3]

Okay, that clears up a lot. Thank you!