Jalapeno Bootyhole Posted:

Or, the non-sarcastic answer: We still haven’t determined a way to address dumpvoting. As far as I can tell, there’s no simple and effective method to force people to vote a certain way. But we are thinking about it.

Scale votes by the history of prior winning votes. I.e., people who routinely vote for winners have bigger votes than people who routinely vote for losers. This is vaguely similar to the Google PageRank algorithm — pages that largely link to relevant pages are strong indicators of relevance. A result in voting is then the sum of the weights of the names who voted for a submission.

A user who dumpvotes is then reducing their ability to reinforce spamming behaviour. It’s less draconian than banning a dumpvoter, but asymptotically reduces their ability to distort the outcome to whatever “good decision rate” they establish. It establishes a new equilibrium point in the game/meta-game combination where voting for winners is an improved long-term strategy than the current game/meta-game combo.

Recommendations in an implementation:

• Winners are not penalized for having to vote for someone else.
• No one’s vote weight can ever go to zero. So, a less draconian way to do this would be to weight by log(winner votes – loser votes) similar to the current tracking in Forumbuildr or Itembuildr, but scaled by a log() to avoid absurd dominated voting schemes.
• Newbies start with the median vote weight.
• Vote weight can be increased for BP.

Open questions:

• Do all alts of a user get a lumped weight, or are alts imagined to be distinct individuals with distinct voting habits?
• Is log() the right scaling function? Log base 2? 10? Is sqrt() better? (This may require modeling based on real data. A 50% winner voter over a set of n contests is expected to have a zero win-loss with a standard deviation of sqrt(n). So half of all 50% win voters have a negative or zero score and half have a positive or zero score. Depending on the desired dynamics wins and losses perhaps will “count differently”, e.g. the score is based on 3*winer votes – loser votes or similar. Perhaps even winner/loser. Some modeling is prescribed here.)

Or does someone have a better idea?