This is the third in a series of posts about the development of socelect.org. Find the first here, introducing socelect.

Davenport’s algorithm, developed in the early oughts by Andrew Davenport in the math department at IBM TJ Watson Research center in Yorktown Heights, New York, provides a practical way to compute a Kemeny-Young preference ranking from some number of individual preference rankings.

Kemeny-Young Preference

Suppose you have half a dozen alternatives ranked in terms of preference, most to least, by any number of interested parties. The Kemeny-Young method selects an overall ranking that contains the least number of pair-wise disagreements with the individual rankings.

Here is an illustration of “least number of pair-wise disagreements.” Suppose, given three alternatives, Budd, Ace, and Rickie, you receive preferences as follows:

  • 29 have Ace before Budd and then Rickie
  • 31 have Budd before Ace and then Rickie
  • 40 have Rickie before Ace and then Budd

preference graph

  • Placing Ace before Budd and then Rickie scores 111 disagreements:
    • 31 disagreements with those that placed Budd before Ace plus
    • 40 disagreements with those who placed Rickie before Budd plus
    • 40 disagreements with those who placed Rickie before Ace
  • Placing Budd before Ace and then Rickie scores 169 disagreements:
    • 69 disagreements with those that placed Ace before Budd plus
    • 60 disagreements with those who placed Budd before Rickie plus
    • 40 disagreements with those who placed Rickie before Ace
  • Placing Rickie before Ace and then Budd scores 131 disagreements:
    • 31 disagreements with those that placed Budd before Ace plus
    • 60 disagreements with those who placed Budd before Rickie plus
    • 40 disagreements with those who placed Rickie before Ace

There are three more possible orderings. You can count them up yourself and see that they have 131, 169, and 189 disagreements.

The Kemeny-Young preference is thus Ace before Budd and then Rickie, because that ordering gives the least number of disagreements. This makes intuitive sense. The least number of people are offended by anything in that ranking. To put it the other way around, it is the ranking that satisfies the most people.

John Kemeny at Dartmouth College in Hanover, New Hampshire developed the concept and published it in 1959. Peyton Young when at the Systems and Decision Sciences Division at the Institute for Applied Systems Analysis, Austria (now at Oxford University in England), proved in 1978 that the Kemeny order is a maximum likelihood estimator of the true preference. In plain terms, that means it is the ranking most likely to represent the preference of the group.

As a side note, if the above was an election and you were selecting one winner, plurality vote would select Rickie (the least preferred). A runoff election between the top two would eliminate Ace. The Ace followers specified Budd next, and would likely transfer their votes to him in the runoff, electing Budd. An instant runoff election that transferred Ace’s votes to Budd would choose Budd. The candidate most preferred, that most would be happy with, by examination and by Young’s proof, is Ace.

Do you see why I think Kemeny-Young preference ranking is the best thing since popcorn? The bee’s knees? Cake and ice cream? Almost as good as apple pie?

The contribution of Davenport was to develop an algorithm that doesn’t require trying and measuring every possible ranking. With three alternatives we have six possible rankings. With only a dozen alternatives there are nearly half a billion. The numbers fairly quickly become astronomical, Carl Sagan numbers.

Davenport’s Algorithm

The algorithm makes finding Kemeny-Young preferences tractable by quickly ruling-out prefixes of the rankings. If you choose the first, second, and third ranked alternatives and see that it will already lead to a number of disagreements higher than the lowest numbered solution you’ve yet found, there’s no need to look at any more rankings that begin with those three.

The technical term for this ruling-out process is “bounding”, or “pruning search using a lower bound.”

The second thing that makes Davenport’s algorithm work well is use of a very good heuristic (rule of thumb) for selecting which alternatives to try first. It tries alternatives first that have the greatest majority of preference over all of the others.

majority graph

In the example, Ace has a majority of (60 - 40 = 20) over Rickie and (69 - 31 = 38) over Budd for a total of 58. Budd’s majority is 20, Rickie’s is zero. You see how strong this heuristic is.

The lower bound computation goes like this. Once you reach the possible rankings with Rickie as the first choice, how far do you have to go before seeing that rankings starting with Rickie will never have a lower number of disagreements than the one found with 111 disagreements (Ace before Budd and then Rickie)?

Rickie disagreements

When you choose Rickie first, you immediately have 120 disagreements. There’s no need to explore any further. All rankings that start with Rickie may be pruned, eliminated from consideration.

Budd disagreements

When you choose Budd first, you immediately have 109 disagreements. To that, you can add the minimum number of disagreements that will result from ordering Ace and Rickie after Budd. That number is 40. Thus, starting with Budd you have a lower bound of 149. All rankings that start with Budd may be pruned.

The pruning is not such a big deal with three alternatives, but with dozens of alternatives, it is essential.

The next contribution that Davenport made with his algorithm was computation of a strong lower bound. By “strong,” we mean it is as low as you can possibly make it without being lower than the true lowest valued result, as in, not wrong.

majority graph

Davenport used a method from capacity planning for computing maximum flow. He modeled the problem using a majority graph, which has nodes (ovals) corresponding to alternatives and directed edges (arrows) weighted with the majority preference for one alternative over the other. Computing maximum flow through all three-cycles in this graph gives the minimum majority that must be broken in order to break the cycles and rank the alternatives.

Two years later, in 2006, Davenport teamed up with IBM Post-doctoral visiting scientist Vincent Conitzer (Ph.D. at Carnegie Mellon University in Pittsburgh, Pennsylvania; now a professor at Duke University in Durham, North Carolina) to improve the lower bound (make it stronger) by accounting for three-cycles that share an edge in the graph.

Closing

Here I went using all of the space in this post to explain Kemeny-Young preference aggregation and Davenport’s algorithm. Next time I’ll write about the implementation.