search

IBM / graph_ensemble_learning

Graph Ensemble Learning
Apache License 2.0
37 stars 7 forks source link

Slow Graphene_Smatch #3

Open Zoher15 opened 1 year ago

Zoher15 commented 1 year ago

Hi @lamthuy,

I found that 'vote' and 'graphene' are really quick but 'graphene_smatch' is incredibly slow. It takes hours.... Would you know why that is?

Best, Zoher

lamthuy commented 1 year ago

In graphene_smatch, after modifying each graph in the ensemble similar to the graphene option, it needs to do an extra calculation of the smatch between the modified graph and all other graphs. That step is more expensive if the number of graph in the ensemble is large.

Zoher15 commented 1 year ago

I am only using three parsers for the ensemble....

Zoher15 commented 1 year ago

@lamthuy MBSE also does the same thing right? I am still not sure why graphene_smatch is so much slower in comparison when I am using the same number of graphs in the ensemble for both. Or maybe I am missing something?

Zoher15 commented 1 year ago

@lamthuy Is it because you have to recalculate the pairwise smatch for each pivot o(n^3) whereas MBSE you only need a single o(n^2) calculation?

lamthuy commented 1 year ago

If there are only 3 models it should not be that slow, I have a deadline now, will take a look when I am back.