Theory/ACO Seminar - Matching in Lopsided Bipartite Graphs and a New Matching Polytope

Other Talks
Friday, August 20, 2010 - 2:00pm
1 hour (actually 50 minutes)
Klaus 1447
Microsoft Research, Redmond, WA

This talk should be non-technical except the last few slides. The talk is
based on a work done in collaboration with Denis Charles, Max Chickering,
Nikhil Devanur, and Manan Sanghi, all from Microsoft.

Lopsided bipartite graphs naturally appear in advertising setting. One side is all the eyeballs and the other side is all the advertisers. An edge is when an advertiser wants to reach an eyeball, aka, ad targeting. Such a bipartite graph is lopsided because there are only a small number of advertisers but a large number of eyeballs. We give algorithms which have running time proportional to the size of the smaller side, i.e., the number of advertisers. One of the main ideas behind our algorithm and as well as the analysis is a property, which we call, monotonic quality bounds. Our algorithm is flexible as it could easily be adapted for different kinds of objective functions. Towards the end of the talk we will describe a new matching polytope. We show that our matching polytope is not only a new linear program describing the classical matching polytope, but is a new polytope together with a new linear program. This part of the talk is still theoretical as we only know how to solve the new linear program via an ellipsoid algorithm. One feature of the polytope, besides being intriguing, is that it has some notion of fairness built in. This is important for advertising since if an advertiser wants to reach 10 million users of type A or type B, advertiser won't necessarily be happy if we show the ad to 10 million users of type A only (though it fulfills the advertising contract in a technical sense).