Portable Parallel Design of Weighted Multi-Dimensional Scaling for Real-Time Data Analysis

Projecting a high-dimensional dataset onto a lower dimensional space can improve the efficiency of knowledge discovery and facilitate real-time data analysis. One technique for dimension reduction, weighted multi-dimensional scaling (WMDS), approximately preserves pairwise weighted distances during the transformation; but its O(f(n)d) algorithm impedes real-time performance on large datasets. Thus, we present CLARET, our fast and portable parallel WMDS tool that combines algorithmic concepts adapted and extended from the stochastic force-based MDS (SF-MDS) and Glimmer. To further improve Claret's performance for real-time data analysis, we propose a preprocessing step that computes approximate weighted Euclidean distances by combining a novel data mapping called stretching and Johnson Lindestrauss' lemma in O(log d) time in place of the original O(d) time. This preprocessing step reduces the complexity of WMDS from O(f(n)d) to O(f(n) log d), which for large d is a significant computational gain. Finally, we present a case study of Claret by integrating it into an interactive visualization tool called V2PI to facilitate real-time analytics. To ensure the quality of the projections, we propose a geometric shape matching-based alignment process and a quality metric.