Triple apply on a multigraph ?

User 942 | 3/26/2015, 12:01:33 PM

I am trying to implement a recommender system using GraphLab Create open source. It is mostly similar to NMF.

Q1) In terms of efficiency, is the python toolkit even competitive or should I really use the C++ SDK ? I don't mind loosing 3x in terms of speed but not 1000x. One key thing is to not copy data back and forth between iterations which leads to my second question.

I need to store 2 types of edges in the graph and triple-apply a function on each type respectively. So I have a type attribute for my edges, I can create a subgraph by selecting the relevant type but I need the vertex data to be shared and not copied between the 2 different triple_apply. Is it possible ?

Thank you


User 1190 | 3/27/2015, 2:05:01 AM

If you are using triple apply, the SDK version is much better at least 10x faster and memory efficient.

For your second question, vertex data cannot be shared between two graphs. But you can fuse two triple apply into one, and have a if statement to switch code branch for different edge type.

User 942 | 3/27/2015, 9:09:10 AM

That's what I thought. Thank you.