Link Prediction and Topological Feature Importance in Social Networks

The problem of link prediction describes how to account for the development of connection structure in a graph. There are many applications of link prediction, such as predicting missing links and future links in online social networks. Much of the literature has focused on limited characteristics of the graph topology or on node attributes, rather than a broad range of measures. There is a rich spectrum of topological features associated with a graph, such as neighbourhood similarity scores, node centrality measures, community structure and path-based distance measures. In this paper we formulate a supervised learning approach to link prediction using a feature set of graph measures chosen to capture a wide range of topological structure. This approach has the advantage that it can be applied to any graph where the connection structure is known. Random forest learning models are used for their high accuracy and measures of feature importance. The feature importance scores reveal the strength of contribution of the topological predictors for link prediction in a variety of synthetically generated network datasets, as well as three real world citation networks. We investigate both undirected and directed cases. Our results show that this approach can deliver very high model precision and recall performance in certain graphs, and good performance generally. Our models also consistently outperform a simpler comparison model we developed to resemble earlier work. In addition, our analysis of variable importance for each dataset reveals meaningful information regarding deep network properties.