Evaluating performance of link prediction in scale-free evolving networks and a Facebook community

Networks exhibit different behaviors, which govern patterns of connectivity. Predicting new links is a fundamental and challenging problem. This work examines the link prediction problem in different scale-free networks and in a Facebook community. Feature-based methods, which can be used on large-scale networks, are used to predict novel links in the graphs. Results show that predictive accuracy is largely dependent on the underlying dynamic process and connectivity pattern. A simulation engine and testing paradigm are described as resources for testing novel link prediction methods in a natural dynamic context, and guide for the selection of appropriate technique for applications.