LINK PREDICTION IN SOCIAL NETWORKS

Abstract

In today's society, social networks play a significant role, with applications ranging from creating a more connected world to finding critical relationships in biological systems. The significant growth in the use of social networks has increased the need of recognizing node-to-node relationships even before they are formed. Several approaches for the task of link prediction utilizing various indices have been developed in the past. There has been a lot of work put into combining multiple indices utilizing machine learning techniques and analogies to the Law of Gravitation, with similarity measures serving as proxies for distance and popularity measures serving as proxies for mass. Merging different indices can improve overall link prediction efficacy, although only a few techniques have been proposed in the past. After integrating three popularity and four similarity metrics, we suggest the usage of a "Histogram based Gradient Boosting Regression Tree" for the task of link prediction in this work. Nature Inspired Approach using CC-CD, has also been proposed which makes use of node embeddings and closeness centrality. Node Embeddings is a way of representing the high dimensional vector representation of graphs to a low dimensional vector. We have used the cosine distances of node embeddings as a proxy of distances and Closeness Centrality as a proxy of masses in Newton’s Gravitational Law for prediction of new links.