Link prediction on evolving social network using spectral analysis

This paper revisits the spectral based link prediction problem of evolutionary social networks reported in [9] and focuses on addressing two empirically observed issues which affect the prediction performance. First, the assumption that eigenvectors are constant over time is not valid for lower order eigenvectors and eigenvectors evolve over time as network evolves. A regression based method is proposed to predict evolving eigenvectors. Second, the spectral condition that higher order eigenvalues are greater than or equal to lower order eigenvalues may not be guaranteed by traditional curve fitting. Two smoothing methods are proposed to address this issue. From various experiments using two large datasets namely DBLP and Facebook, it is observed that proposed methods enhance prediction performance as compared to that of their counterparts.