Add PageRank #788
Add PageRank #788
Conversation
Pull Request Test Coverage Report for Build 5086986067
💛 - Coveralls |
|
Everything should be working now |
There was a problem hiding this comment.
This is great, thanks for writing this! I really like using sprs for this, it also avoids us having to figure out how to link against and leverage blas/lapack to compute the eigenvectors of the google matrix from an ndarray. I need to refresh my memory on the algorithm before I do a detailed review on the algorithm code. I just has some quick high level comments from a quick scan of the code.
The algorithm is approximating the eigencector of the transition matrix. You might want to check NetworkX’s Python code directly because they have some quirks on how they handle dangling nodes etc. |
Co-authored-by: Matthew Treinish <mtreinish@kortar.org>
| /// | ||
| /// :returns: a read-only dict-like object whose keys are the node indices and values are the | ||
| /// PageRank score for that node. | ||
| /// :rtype: CentralityMapping |
There was a problem hiding this comment.
Do you want to document that it will raise FailedToConverge if max_iter is reached? It's something people might want to catch.
There was a problem hiding this comment.
I think I will add the notes for this, HITS, eigenvector centrality and all the centralities that cannot converge in a separate PR
|
gonna put a pin in it and double check everything and maybe make a new issue about it, but somehow i'm getting totally uniform results from this where networkx is (a lot) slower but produces some results that make more sense |
Please open a new issue. If you could give the graph that triggers the difference that would be great as well. We use the Power Method and NetworkX uses SVD decomposition to approximate the eigenvector so there might be some discrepancies. |
Related to #315
Adds an implementation of the PageRank algorithm using sparse matrices. It uses the
sprscrate combined withndarrayto implement a Power Method approach of finding the PageRank.Also, we test this implementation against NetworkX's implementation of the PageRank. We accept all the arguments that NetworkX accepts: tolerance, max_iter, personalization, dangling, etc.
n.b: it's ready for review