Parallel Implementation of Multidimensional Scaling Algorithm Based on Particle Dynamics

We propose here a parallel implementation of multidimensional scaling (MDS) method which can be used for visualization of large datasets of multidimensional data. Unlike in traditional approaches, which employ classical minimization methods for finding the global optimum of the “stress function”, we use a heuristic based on particle dynamics. This method allows avoiding local minima and is convergent to the global one. However, due to its O(N) complexity, the application of this method in data mining problems involving large datasets requires efficient parallel codes. We show that employing both optimized Taylor’s algorithm and hybridized model of parallel computations, our solver is efficient enough to visualize multidimensional data sets consisting of 10 feature vectors in time of minutes.