Friday, December 10, 2010

Spreading the wealth among competing blogs

Thinking some more about the votermedia award algorithm change I made on November 22, I think we should make a more sweeping change to further enhance competition among blogs. Changing the interpolation range from 5% to 10% does encourage the top blogs to compete with each other more, but it still leaves the maximum vote ceiling as an overly firm limit.

In the last few days, I built the capability for the votermedia.org voting platform to have different tally model parameters for different communities. Parameters can be changed at any time, and an archive of past changes is stored. So we can now test the effects of different parameters like the interpolation range, vote decay period etc.

So far only one community has been funding their votermedia blog contest – UBC AMS. My policy is to check with the funder before changing their contest parameters. So the changes I will describe next are not being applied to the UBC AMS ballot. I plan to test them on all the unfunded ballots first, and then propose them for UBC AMS.

The competition problem we are trying to solve is that, in our experience with votermedia so far, there seems to be a tendency for voters to gravitate to just one or two popular blogs. Most observers I have discussed this with agree that it is usually in the interest of the community to spread the awards across four blogs or more. There has been widespread support for limiting award shares to some maximum, so we have implemented that from the beginning of votermedia in 2007.

In its simplest form, the main votermedia algorithm would let each voter vote any share (0% to 100%) to any blog. (See our FAQs for how votes are aggregated to determine awards. It's not a simple average.) So if voters want to spread the awards across four blogs or more, they have the power to do so. A good argument can be made on general principles that we shouldn't limit the power of voters; if they want to fund only one or two blogs, then so be it.

One justification for limiting award shares may be that in these early days of votermedia, many of the voters are friends of the bloggers, and may not represent well the broad interests of the entire community that provides the funding. Thus limiting shares provides a kind of training wheels that help keep things more balanced until votermedia gets rolling well enough not to need this restriction.

However, like most share-the-wealth policies, capping the votable share menus, e.g. at 30% as we are doing for UBC, has the unintended consequence of reducing incentives. Once the top few blogs hit the 30% ceiling, there's not much incentive to keep improving. Bloggers may keep at it for other reasons, but the whole idea of votermedia is to provide financial motivation.

Our current 30% cap with a 10% interpolation range has no effect until a blog's share reaches 25%. Above 25%, the votes for a 30% share get shrunk by one-tenth for each 1% rise in share. For example, a blog with a 27% share would have their votes shrunk by two tenths, so only 8/10 of their votes would be counting. This spreads out across a 10% range what would otherwise be a hard cap at 30%.

It may help intuition to think of the reduction in votes as an income tax. A simple cap at 30% would be like having no taxes until you reach a 30% share, but above 30% everything is taxed away. A voter that wanted to vote you 40% has that vote reduced to 30%. You have no incentive to make anyone want to vote you 40% rather than 30%.

A 10% interpolation range phases in the tax. No tax up to 25%. Tax rate gradually rises from there. By a 27% share, the tax rate is two tenths. By a 35% share, tax rate has become 100% -- all incentive is now taxed away.

Instead of capping the votable share menus, we can get a similar spread-the-wealth effect without so much disincentive effect, by phasing in the "tax" much more gradually across the entire range of shares from 0% through 100%. For our first trial run, I've chosen this exponential function for the tax schedule:

after-tax vote count = (pre-tax vote count) / ( spread ** share)

** means exponent, to the power of

"share" is scaled between 0 and 1, so an award share of 27% means share = 0.27

"spread" is a parameter chosen to give the desired amount of spread-the-wealth effect. spread = 1 means no effect. I plan to try using spread = 4. To reach 100% share, ¾ of votes would be taxed away, so a blogger would need to have 4 times as many votes for 100% as any other competitor has for any positive shares. For example, suppose your strongest competitor has only 60 votes, perhaps only for a 20% share. If you get over 240 votes for 100%, then you would be awarded 100%. So it's still possible to get there, but a lot of voters have to want you to get there.

At a 50% share and spread = 4, we get 1 / (spread ** share) = 0.5, so half your votes are taxed away. At 25% share, about 3/10 of your votes are taxed away.

The idea is that this tax system would take the place of putting a cap on votable share amounts. In addition to improving incentives for the top blogs, this solves other problems caused by caps:

  • Interpolation can cause a blog's share to go above the top votable amount (e.g. AMS Confidential's share is now 31% at votermedia.org/ubc). Voters may be uncertain whether voting 30% actually helps or hurts that blog's award. (It helps.)
  • We've been adjusting the cap according to the number of blogs on the ballot. A 30% cap makes no sense if there are only 2 or 3 blogs. But then the adjustment when new blogs enter can be somewhat arbitrary and discontinuous.
  • Some blogs become dormant at times. Some seasons (e.g. summer at a university) are very quiet, and may only deserve active coverage from 1 or 2 blogs. A cap system would require adjustment, whereas the above "spread" system would respond more smoothly and automatically.

To phase in this test, I have started by eliminating caps in the following sense: I have programmed votermedia.org to always set the maximum votable share at least 5% above the top blog's share:

Share of Top Blog Menu Max
up to 34%40%
35% - 44%50%
45% - 54%60%
55% - 74%80%
75% - 100%100%

So a blog could theoretically make its way up to 100%. The cap keeps rising as the blog's share rises. Voters would have to vote again at each higher cap, to signal that they really want that blog's share to keep rising. I plan to monitor this for a few days, and then implement spread = 4 as discussed above, which will have a moderating effect on the rise of the top blog(s).

Only UBC is still at a 30% cap, and will stay at spread = 1, until I consult with the sponsor.

Feel free to add questions or comments below, or email me (mark[at]votermedia.org).

No comments: