Predicting internet network distance using ISOMAP

Since coordinate-based methods for network distance prediction can estimate distances more accurately and effectively than previously proposed methods, they have been widely studied and used in Internet applications. However, there still exist at least three problems unsolved: to find a embedding low-dimensional Euclidean space best preserving distance information, to determine the dimension of embedded Euclidean space, and to reduce time and parametric complexity derived from iterative optimizing process. This paper proposes a new coordinate-based method using ISOMAP to address these problems. ISOMAP estimates distances between nodes by their shortest path distance and employs Multidimensional Scaling (MDS) which uses matrix decomposition to find nodes' coordinates in embedding Euclidean space best preserving distances. MDS avoids the complexity of optimization and helps exploit the dimension size of embedding space according to information preservation. Discussion and experiments have proved that the proposed method performs faster and more accurately than the Global Network Positioning (GNP) does.