Dongyun Yi

Identifying missing and spurious interactions in directed networks

Recent years, the studies of link prediction have been overwhelmingly emphasizing on undirected networks. Compared with it, how to identify missing and spurious interactions in directed networks has received less attention and still is not well understood. In this paper, we make use of classical link prediction indices for undirected networks, adapt them to directed version which could predict both the existence and direction of an arc between two nodes, and investigate their prediction ability on six real-world directed networks. Experimental results demonstrate that those modified indices perform quite well in directed networks. Compared with bifan predictor, some of them can provide more accurate predictions.

Identifying Missing and Spurious Interactions in Directed Networks

Recent years, the studies of link prediction have been overwhelmingly emphasizing on undirected networks. Compared with it, how to identify missing and spurious interactions in directed networks has received less attention and still is not well understood. In this paper, we make use of classical link prediction indices for undirected networks, adapt them to directed version which could predict both the existence and direction of an arc between two nodes, and investigate their prediction ability on six real-world directed networks. Experimental results demonstrate that those modified indices perform quite well in directed networks. Compared with Bi-fan predictor, some of them can provide more accurate predictions.

Uncovering community structure in social networks by clique correlation

Community is tightly-connected group of agents in social networks and the discovery of such subgraphs has aroused considerable research interest in the past few years. Typically, a quantity function called modularity is used to guide the division of the network. By representing the network as a bipartite graph between its vertices and cliques, we show that community structure can be uncovered by the correlation coefficients derived from the bipartite graph through a suitable optimization procedure. We also show that the modularity can be seen as a special case of the quantity function built from the covariance of the vertices. Due the the heteroscedaticity, the modularity suffers a resolution limit problem. And the quantity function based on correlation proposed here exhibits higher resolution power. Experiments show that the proposed method can achieve promising results on synthesized and real world networks. It outperforms several state-of-the-art algorithms.