Semi-Supervised Algorithms for Approximately Optimal and Accurate Clustering

We study k-means clustering in a semi-supervised setting. Given an oracle that returns whether two given points belong to the same cluster in a fixed optimal clustering, we investigate the following question: how many oracle queries are sufficient to efficiently recover a clustering that, with probability at least (1 − δ), simultaneously has a cost of at most (1 + ǫ) times the optimal cost and an accuracy of at least (1− ǫ)? We show how to achieve such a clustering on n points with O((k log n) · m(Q, ǫ, δ/(k log n))) oracle queries, when the k clusters can be learned with an ǫ′ error and a failure probability δ′ using m(Q, ǫ′, δ′) labeled samples, where Q is the set of candidate cluster centers. We show that m(Q, ǫ′, δ′) is small both for k-means instances in Euclidean space and for those in finite metric spaces. We further show that, for the Euclidean k-means instances, we can avoid the dependency on n in the query complexity at the expense of an increased dependency on k: specifically, we give a slightly more involved algorithm that uses O(k/(ǫδ) + (k/ǫ) log(1/δ) + k ·m(Q, ǫ/k, δ)) oracle queries. Finally, we show that the number of queries required for (1− ǫ)-accuracy in Euclidean k-means must linearly depend on the dimension of the underlying Euclidean space, whereas, for finite metric space k-means, this number must at least be logarithmic in the number of candidate centers. This shows that our query complexities capture the right dependencies on the respective parameters.