In our formulation, we still require a complete graph as
input, but every edge is associated with a weight, which is
a number in [0, 1]. Weights represent similarity between data
points and the extent to which data points should be assigned
to the same cluster. For defining distances between sets of
labels, we consider two measures: a set-intersection indicator
function and the Jaccard coefficient. We also constrain the
maximum number of cluster labels allowed, either globally or
per data point. These alternatives, together with the possibility
of having fractional or binary edge weights, produces a whole
family of problems.