Link prediction in dynamic Weighted heterogeneous social Networks

Abstract

Forecasting possible future relationships between people in a social newlinenetwork requires a study of the evolution of their links. To capture network newlinedynamics and temporal variations in link strengths between various types of nodes newlinein a network, a dynamic weighted heterogeneous network is to be considered. newlineLink strength prediction in such networks is still an open problem. Moreover, a newlinestudy of variations in link strengths with respect to time has not yet been explored. newlineThe time granularity at which the weights of various links change remains to be newlinedelved into. To predict future link strengths in dynamic weighted heterogeneous newlinesocial networks, regular snapshots of a weighted bibliographic network are taken newlineand three new weighted statistical meta path-based features which capture the newlinetopological structure of the network are extracted for each of its snapshots. The newlineweighting scheme for the different types of links is based on the node and link newlinefrequencies. Features are forecast using Auto Regressive Integrated Moving newlineAverage (ARIMA) time series forecasting method suitable for long term forecasts, newlinewhich extrapolates the features for a future time interval. These forecast features newlineare fed to a neural network framework with a newly designed Beta kernel initializer newlinewhich enables faster convergence and makes the learning better. The proposed link newlinestrength prediction model is the first of its kind that predicts future relationships newlinebetween people, along with a measure of the strength of those relationships in a newlinenetwork modeled as dynamic, weighted and heterogeneous. newlineRanking of social links plays an important role in analyzing whom one newlineis largely influenced by, which has potential applications in viral marketing, newlinesentiment analysis, influence diffusion etc. newline