4 ms·
On a hunch: Run Pagerank (the original thing, i.e. just the stable distribution of the random walk) on the graph of who-upvotes-whom. Then tally upvotes weighte
by danvet 15y ago
On a hunch: Run Pagerank (the original thing, i.e. just the stable distribution of the random walk) on the graph of who-upvotes-whom. Then tally upvotes weighted with the upvoters pagerank. It's probably safe to assume that the pagerank of a given user changes rather slowly, so this shouldn't be a too big problem to actually compute (and needs to be update at most every few days). Also, to favour recent history a bit one could decay the wheight of older upvotes in the who-upvotes-whom graph.
Now that upvotes are non-public information, people can't know their own rank (at least not easily), so this looks at least halfway resistant to gaming.
- chalst 15y agoAdvogato has used an attack-resistant trust metric with similar features; Raph Levien called such systems attack-resistent trust metrics. I'm sure they would do the task you describe successfully. Check out: 1. http://www.levien.com/free/tmetric-HOWTO.html http://www.levien.com/free/tmetric-HOWTO.html 2. http://www.advogato.org/person/chalst/diary/189.html http://www.advogato.org/person/chalst/diary/189.html
- bermanoid 15y agoWhoops, missed a lot of responses here, sorry about that! Yes, vanilla Pagerank (where the "sites" are accounts and "links" are upvotes) is the first thing that came to my mind, mainly because it's been pretty well battle tested; running it on the comment graph is actually a much simpler problem than running it behind a search engine because you don't need the additional refinements based on search keyword. A comment's Pagerank is, by itself, enough to set an ordering. If that's too much work, though, even a simple vote weighting based on some function of the upvoter's karma would be a decent approximation. Say every user's votes had an impact based on some sigmoid function of their karma (maybe a Gompertz function?), tuned so that the plateau is hit for the top 10% of users or something like that.