Feature Engineering for Supervised Link Prediction on Dynamic Social Networks
Link prediction is an important network science problem in many domains such as social networks, chem/bio-informatics, etc. Most of these networks are dynamic in nature with patterns evolving over time. In such cases, it is necessary to incorporate time in the mining process in a seamless manner to aid in better prediction performance. We propose a two-step solution strategy to the link prediction problem in dynamic networks in this work. The first step involves a novel yet simple feature construction approach using a combination of domain and topological attributes of the graph. In the second phase, we perform unconstrained edge selection to identify potential candidates for prediction by any generic two-class learner. We design various experiments on a real world collaboration network and show the effectiveness of our approach.