You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently edge probabilities are "normalized" (made to sum to 1) as a pre-processing stage to running Osrank. However if a trusted seed set is used, then the final graph that Osrank runs on will be a subgraph of the original graph.
In that case we need to add a new step to our Osrank implementation where we re-normalize the edge probabilities in the graph after applying the Trustrank-style filter and before simulating the graph traversals.
The text was updated successfully, but these errors were encountered:
I will crosscheck with @MeBrei once she's back, but I suspect you are spot on. Even if we consider the example graph in the paper:
Supposing we run the TrustRank phase and we realise that P1 and A1 need to be pruned, we will now end up in a situation like this:
This is clearly incorrect, because now the outgoing edges from P3 do not sum all to 1.
I suspect that doing this properly with our current graph implementation might not be trivial, but I will start thinking about this in preparation for Merle to be back 😉
Currently edge probabilities are "normalized" (made to sum to 1) as a pre-processing stage to running Osrank. However if a trusted seed set is used, then the final graph that Osrank runs on will be a subgraph of the original graph.
In that case we need to add a new step to our Osrank implementation where we re-normalize the edge probabilities in the graph after applying the Trustrank-style filter and before simulating the graph traversals.
The text was updated successfully, but these errors were encountered: