k sets(): this function picks k of the constraint sets in the list of constraints to use in order to pick initial cluste

[answerhappygod]

k sets(): this function picks k of the constraint sets in thelist of constraints to use in order to pickinitial cluster centers. By default, it will pick the first kconstraint sets appearing in the list. In theunconstrained case, this is basically picking k arbitrary points inthe training set to use as the clustercenters. However, this is likely to perform poorly in theconstrained case. In particular, a bad selectionof initial clusters may result in some cluster being unused
Consider the implementation of the k sets() function, where wepick the initial cluster centers, fromthe list of constraints. When we have no “must-link” constraints,then it is typical to just select kimages from the dataset at random, and use the locations as initialcluster centers. However, whenwe use “must-link” constraints, the “must-link” constraintsprovides some partial information aboutthe clusters, which we could take advantage of when selectinginitial cluster centers. Consider whathappens as we increase the number of “must-link” constraints thatwe provide. With few constraints,most images are independent, and the ones that are linked togetherare linked to a small numberof other images. With many constraints, few images are independent,and the ones that are linkedtogether are linked to a large number of other images. Based onthis observation, provide an argumentthat we should pick the k-largest sets of linked images to serve asinitial cluster centers