Topological similarity as a proxy to content similarity

Vertices in networks often have external classifications. Vertices with similar classifications may have similar connection patterns in the network, even if they reside in remote regions of the network. We thus introduce a novel method for the detection of groups of non-adjacent vertices with similar classifications in networks through the similarity of measures on the network surrounding them. In the algorithm, vertices are characterized by a large set of structural properties of local and global scale, composing a network attributes vector for each vertex. This characterization is used to construct an affinity dual graph, where clustering is applied. When tested in several real-world networks with ground truth classifications, the groups detected by our algorithm had significantly more homogenous groups than those found by common community detection algorithms. The algorithm allows the clustering of non-adjacent vertices in remote network locations, as shown in two networks. When used in a supervised context, precise predictions of vertices content are accomplished.