Effect of similarity measures in diffusion prediction on homogeneous and heterogeneous bibliographic network

Abstract

This paper evaluates the effect of similarity measures in predicting diffusion on homogeneous and heterogeneous bibliographic networks. The bibliographic network is analyzed within a homogeneous network and heterogeneous network, where a co-author relationship exists for the former, and multiple types of meta paths are considered for the latter. The supervised learning method is used to predict whether a node will be active with a topic or not. The features are extracted as the activation probability of a node, which represents the maximum of the activation probabilities of the neighbors of this node. In a homogeneous network, the activation probability from the activated node to the inactive node is measured based on one relationship co-author with basic similarity measures while it can be calculated based on diverse meta paths with dissimilar meta path-based similarity measures in the heterogeneous network. We performed our analysis on three different datasets. Our experimental results show that diffusion prediction in bibliographic networks provides better accuracy among heterogeneous networks than among homogeneous networks and that the Bayesian similarity measure provides the best efficiency.