Comparison of Two Dimension-Reduction Methods for Network Simulation Models

Experimenters characterize the behavior of simulation models for data communications networks by measuring multiple responses under selected parameter combinations. The resulting multivariate data may include redundant responses reflecting aspects of a smaller number of underlying behaviors. Reducing the dimension of multivariate responses can reveal the most significant model behaviors, allowing subsequent analyses to focus on one response per behavior. This paper investigates two methods for reducing dimension in multivariate data generated from simulation models. One method combines correlation analysis and clustering. The second method uses principal components analysis. We apply both methods to reduce a 22-dimensional dataset generated by a network simulator. We identify issues that an analyst must decide, and we compare the reductions suggested by the methods. We have used these methods to identify significant behaviors in simulated networks, and we suspect they may be applied to reduce the dimension of empirical data measured from real networks.