Exploiting optimised communities in directed weighted graphs for link prediction

The most developing issue in analysing complex networks and graph mining is link prediction, which can be studied for both content and structural-based analysis in a social network. Link prediction deals with the prediction of missing links by determining whether a link can be created between two nodes in a future snapshot of a given directed weighted graph. Existing link prediction methods are only studied for unsigned graphs and work on principles of the common neighbourhood. However, the link prediction problem can also be studied for signed graphs where signed links can give an interesting insight into user associations. Obstruction of studies in this domain is caused by imbalance of class, i.e., positive links are frequent than negative ones, and forbearance of hidden communities. A signed network is a combination of dense and hidden communities. A hidden community structure is overlooked by majority of existing applications, taking dense community structure, i.e., one whole graph as input for developing a link prediction model. Hence, complete network information is required by majority of existing approaches, which seems unrealistic in modern social network analytics. In this article, we exploit hidden network communities to address link prediction problem in the signed network, focusing on negative links. A number of observation were made regarding negative links and a principle ensemble framework, i.e., � NeLp, is proposed, having two phases, i.e, network embedding and classifier prediction. Using a probabilistic embedding framework, network representation of hidden signed communities is learned, which were then passed to a learning classifier to predict negative links, keeping intact the ensemble framework. Despite the limited availability of signed network datasets, an extensive experimental study was performed to evaluate � NeLp pertinency, robustness, and scalability. The performance result shows that � NeLp can be a promising consideration for addressing link prediction tasks in signed networks and gives encouraging results.