A Novel Effective Distributed Dimensionality Reduction Algorithm

Dimensionality reduction algorithms are extremely useful in various disciplines, especially related to data processing in high dimensional spaces. However, most algorithms proposed in the literature assume total knowledge of data usually residing in a centralized location. While this still suffices for several applications, there is an increasing need for management of vast data collections in a distributed way, since the assembly of data centrally is often infeasible. Towards this end, in this paper, a novel distributed dimensionality reduction (DDR) algorithm is proposed. The algorithm is compared with other effective centralized dimensionality reduction techniques and approximates the quality of FastMap, considered as one of the most effective algorithms, while its central execution outperforms FastMap. We prove our claims through experiments on a high dimensional synthetic dataset.