>
Please read section IX of this book chapter:
Will do. Or, at least, try
>
That is exactly what this does, but in a numerically stable manner.
Not seeing how the function you linked is incremental. It seems to require collecting every conditional probability in memory for every vertex pair. I'm collecting for ~6m pairs on my laptop, that would blow my RAM. (If I'm understanding correctly)
>
If you don't supply anything, a flat prior will be used, which means the
inference will be guided by the network structure alone.
When I use the conditional probability method it too only uses the network structure, correct?
This smaller graph I'm testing on has
225,150 possible edges. There are 20,911 actual edges. So, there are
204,239 possible additional edges. Finding 13 potential edges out of 204,239 seems extremely off (even using network structure alone; indeed all the other link prediction methods I'm using only use network structure). I could test this by removing 2000 and then seeing how many more it predicts, but I'm fairly certain already it will be similarly low.
In the context of exploratory link prediction, recommender systems for social networks, etc I don't think there is a principled way to give a "deleted fraction". I don't know how many links are "missing" but it's certainly not 13. I assume I could just assert something more aggressive arbitrarily and get a better result. Would it not be more principled to have a single 'non-edge aggressiveness parameter' and then make a uniform prior (hyperprior?) for it over some reasonable range, perhaps that f ratio you mention? Say, uniformly between [0.01, 0.2]. Is that doable with graph-tool?
Thanks for your help, as always