tmmurali
commented
10 years ago On Wed, May 14, 2014 at 1:30 PM, Tyler Kahn notifications@github.comwrote:
The good news is that I was able to implement the DINH model from the paper and our library is generate it it in a reasonable time (with some modification of core library classes, see my commits to Node and Hypergraph).
The problem is that generating the transition matrix takes too long. I left my computer running for a half hour and it didn't complete.
You will have to use some sparse matrix approach. I have implemented very similar random walk and power iteration algorithms for undirected graphs in C++. The trick is not to create the full blown matrix but to store the graph in an adjacency list and implement matrix-vector multiplications by iterating over graph edges.
Let me know if this suggestion helps. Murali
I tried both approaches:
a) Matrix multiplication of the incidence, weight, edge degree, and vertex degree matrices. This is _transition_matrix() at the bottom of directedHyperGraph.py. See herehttps://github.com/tmmurali/hypergraph/blob/directed_random_walk/hypergraph/directedHyperGraph.py#L253 .
b) Directly computing each (u, v) cell in P. This is transition_matrix(). See herehttps://github.com/tmmurali/hypergraph/blob/directed_random_walk/hypergraph/directedHyperGraph.py#L303 .
Both never finished. I had to memmap the matrices because Python can't store matrices that large in memory.
The image I tested on was only 154x115 pixels.
Not sure how to remedy this. I spent quite a bit of time on this so it's unfortunate it's not working.
— Reply to this email directly or view it on GitHubhttps://github.com/tmmurali/hypergraph/issues/25 .
tylerkahn
See here.
The good news is that I was able to implement the DINH model from the paper and our library generated it in a reasonable time (with some modification of core library classes, see my commits to Node and Hypergraph).
The problem is that generating the transition matrix takes too long. I left my computer running for a half hour and it didn't complete.
I tried both approaches:
a) Matrix multiplication of the incidence, weight, edge degree, and vertex degree matrices. This is _transition_matrix() at the bottom of directedHyperGraph.py. See here.
b) Directly computing each (u, v) cell in P. This is transition_matrix(). See here.
Both never finished. I had to memmap the matrices because Python can't store matrices that large in memory.
The image I tested on was only 154x115 pixels.
Not sure how to remedy this. I spent quite a bit of time on this so it's unfortunate it's not working.