Maximum weight matching (MWM) for a predefined number of nodes

1
votes

I am given a weighted graph and want to find a set of edges such that every node is only incident to one edge and such that the sum of the selected edge weights is maximized. As far as I know this problem is generally referred to as maximum weight matching and there exist fast approximations for it: https://web.eecs.umich.edu/~pettie/papers/ApproxMWM-JACM.pdf

However, for my application it would be better if only a certain ratio of nodes is paired. It's more important for my application that the nodes that get paired have a high weight between them. Leaving some nodes unpaired is no big problem.

Currently, I sort the weights between nodes in descending order and always select the edge with the highest weight until I have paired a certain number of nodes. Of course I ensure that pairs of nodes are mutually exclusive. This is only a 1/2 approximation to the original problem and it's probably even worse for the modified problem.

Could you please suggest an algorithm for this issue or tell me how this problem is called?

1
algorithmgraphmatching
The bipartite version is the k-cardinality assignment problem. Is your graph bipartite? – David Eisenstat
Thanks. No it's not bipartite. It is even a complete graph. – creiser
How big is it? Are you open to external libraries (e.g., integer program solver)? – David Eisenstat
It has 20000 nodes. Sure I am open to everything. Speaking of "integer", I forgot to mention that all weights are integers. – creiser

1 Answers

0
votes

Some thoughts.

  1. The greedy algorithm for this problem is a 2-approximation. Imagine that, during the execution of the greedy algorithm, we keep score versus an optimal solution in the following manner. Every time we add an edge to the greedy solution, we delete the incident edges in the optimal solution, which I claim must have weight no greater than the greedy edge. If no edge would be deleted from the optimal solution, we delete the two heaviest edges instead, which also I claim must have weight no greater than the greedy edge. Since the greedy solution must have at least half as many edges as the optimal solution, I claim that there are no optimal edges left at the end, and hence greedy is a 2-approximation because we never deleted more than 2x weight with each greedy edge.

  2. A complete 20,000-vertex graph is right on that line where I don't know whether integer programming would be a good idea. I think it's still worth a try because it's easy enough.

  3. There are polynomial-time algorithms for computing the maximum-weight independent set in the intersection of two matroids (for this problem, the matching matroid and the uniform matroid whose bases have size equal to the size of the desired matching). I don't know if they would be practical.