Due to the vast amount of application generated high-dimensional data
and their distribution among network nodes the fields of Distributed
Knowledge Discovery (DKD) and Distributed Dimensionality Reduction (DDR)
have emerged as a necessity in many application areas. While there
exists a variety of centralized dimensionality reduction algorithms,
only few have been proposed for distributed environments and they are
mainly adaptations of centralized approaches. In this paper, we
introduce K-Landmarks, a new DDR algorithm and we evaluate its
comparative performance against a set of well known distributed and
centralized dimensionality reduction algorithms. We primarily
concentrate on each algorithm's ability in maintaining clustering
quality throughout the projection, while retaining low stress values.
Our algorithm exhibits better performance in most cases, showing both
its superiority as well as its suitability for highly distributed
environments.