Secant manifold constrained random projections -Improved cluster ensembles

In this paper we present two contributions to the framework of ensemble clustering. Our work expands the robust k nearest neighborhood clustering ensemble to include random projections for further increasing the stochastic exploration of the data set to be clustered. In addition we propose to constrain the random projection ensemble to only contain distance-preserving projection directions. The latter is obtained by constraining the projection directions to be orthogonal to the so-called Secant manifold of the input data set. Promising results are shown on a series of benchmark data sets.