Influence of clustering coefficient on network embedding in link prediction

Abstract Multiple network embedding algorithms have been proposed to perform the prediction of missing or future links in complex networks. However, we lack understanding how topology affects their performance, which are more likely better given topological properties network. In this paper, investigate clustering coefficient a network, i.e., probability that neighbours node also connected, algorithms’ performance link prediction, terms AUC (area under ROC curve). We evaluate classic algorithms, Matrix Factorisation, Laplacian Eigenmaps and node2vec, both synthetic networks (rewired) real-world with variable coefficient. Specifically, rewiring algorithm is applied each change while keeping key properties. find higher tends lead except for not sensitive To understand such influence coefficient, (1) explore relation between rating (probability pair link) derived from aforementioned number common pair, (2) these ability reconstruct original training (sub)network. All tested tend assign likelihood connection pairs share an intermediate high neighbours, independently Then, predicted will triangles, thus As increases, all but Factorisation could These two observations may partially explain why increasing improves performance.