Conversation
|
Isn't page rank almost the same algorithm but just with a bit of regularization? |
Related to erikbern#2 Simple removal of all edges from Go to * makes: | 22.76% | java | | 15.73% | go | | 14.31% | c | | 13.46% | c++ | | 6.83% | c# | | 6.12% | python | | 5.70% | scala | | 3.98% | rust | | 2.93% | php | | 2.02% | objective c | | 1.54% | ruby | | 1.07% | swift | | 1.00% | node | | 0.65% | kotlin | | 0.40% | r | | 0.30% | matlab | | 0.28% | haskell | | 0.24% | pascal | | 0.21% | cobol | | 0.13% | erlang | | 0.11% | fortran | | 0.10% | visual basic | | 0.07% | lua | | 0.05% | clojure | | 0.02% | perl | | 0.01% | lisp |
|
This is related to http://snap.stanford.edu/class/cs246-2013/slides/09-pagerank.pdf page 35. This is not the regularization but rather a mere necessity. And for god sake don't even try vanilla PI on large stochastic graphs - it does not converge at all. I tried many times :-) |
|
I'm pretty sure this is equivalent to Tikhonov regularization: https://en.wikipedia.org/wiki/Tikhonov_regularization And from a Bayesian point of view it's equivalent to putting normal priors on the stationary distribution. I could be wrong though! |
|
I am unable to check this - I am not familiar with Bayesian inference (shame on me). All right, let it be the regularization. As we know, it is better to have a regularized model than not :) . I suggest to change the unregularized PI code to page_rank code. Do you approve? |
|
i'm curious to see how much it changes the results... probably a little bit. let me run a side by side comparison and check |
|
I think that regarding the referencing PR, page_rank does not change it's result while direct PI changes dramatically. Since I did tiny changes to the matrix, I call it instability. |
2 participants
No description provided.