Optimised link prediction techniques using tailored artificial neural networks Pundhir Sandhya

Abstract

Link Prediction (LP) is about inferring the link among unobserved links that would newlineemerge in time to come or unobserved links among pairs of nodes in complex networks. newlineThe performance of LP gets improved with the application of machine learning (ML) newlinemethods. An enormous scope of automation for LP by using machine learning is there. newlineAn Artificial Neural Network (ANN) is one of the pervasively utilized ML techniques. newlineDeep Learning (DL) is one of the driving force of current ML success that has attained newlinequotstate of the artquot performance in many fields. In this work, simple ANN, Multiple Layer newlinePerceptron (MLP), Deep Neural Network (DNN) and customized ANN have been newlineexperimented with various standard datasets. newlineMany techniques have been instigated for LP in divergent areas. Many endeavors are newlinethere to addressthe problems of LP through multiple approaches. The primary approach newlineis to estimate the closeness or similarity of verges or nodes in divergent interactions or newlineinformation exchange. Likeness changes over time because of changes in the behavior newlineof systems or networks. Current static relevance methods cannot stand with speedily newlinetransforming networks. Therefore, they are not a good choice for LP, and a need for newlinesome more accurate methods arises. newlineIn this work, an attempt is made to instigate and generalize techniques for efficient LP, newlinewhich can work with multiple types of dataset or network. We applied and evaluated newline40 link prediction algorithms (10 traditional link prediction methods and 30 link newlineprediction using different ANN configurations). This Ph.D. thesis comprises divergent newlinework lines, all of them closely related to optimized LP using a customized artificial newlineneural network. Not only a tailored ANN used for prediction but also compact newline(Subgraph based), robust (quotAvgNewquot based), unbiased (legitANN based), data ... newline