The investigation of node similarity is an important component in link prediction and community detection.In this paper, four kinds of algorithms including common neighbor (CN), resource allocation (RA), Adamic-Adar (AA) and Sorenson are introduced into various kinds of real networks and two kinds of simulation networks comprised of small world network and scale free network.The Area Under the Curve (AUC) is computed to compare their predictive accuracy.It's found that RA performs much better than the other three kinds of algorithms.Then four algorithms are adopted in functional connectivity networks that characterize electroencephalograph (EEG) recordings from eight patients with generalized epilepsy.It's demonstrated that RA performs best from the point of prediction accuracy.According to RA technique, clusters could be determined from nodes that own maximum similarity which provides an objective index for quantifying brain condition, and this might be applied for clinical auxiliary diagnosis in the future.