Integrating node centralities, similarity measures, and machine learning classifiers for link prediction

Link prediction is a widely studied topic in graph data analytics and finds numerous applications like friend recommendations in social networks and product recommendations in e-commerce. It refers to predicting new connections or edges that may arise in the near future amongst the nodes of the network. There exist various methods of link prediction generally based on either local, semi-local, or global features of networks and usually suffers from the problems of consistency in their performances over different and large size networks. In this paper, we intend to propose a generic and improved method of link prediction named as NSMLLP by integrating Node centralities, Similarity measures, and Machine Learning classifiers. We calculate popularity measures for every node and evaluate similarity measures for every pair of nodes in the network. The combined popularity and similarity measures form the features for every node pair in the network. The combined features of the nodes at the end of the edges, along with the positive or negative edge label, form a well-defined dataset for the task of link prediction. This dataset is then fed into machine learning classifiers like Random Forest classifier, AdaBoost classifier, and an ANN based classifier. The results obtained from these classifiers are then combined to make the final link prediction. We provide an information gain study aiming to quantify the improvement brought on by our proposed method. A feature importance study is also provided to comprehend better the relative importance of the various popularity and similarity measures used by us. The experimental results obtained on multiple real-life networks demonstrate that the proposed technique outperforms many existing popular methods of link predication based on several evaluation criteria.