Propagation source identification of infectious diseases with graph convolutional networks

Abstract

Source identification in networks has drawn considerable interest to understand and control the infectious disease propagation processes. It is usually difficult to achieve both high accuracy and short error distance when we try to solve the problem. This study attempts to introduce the graph convolutional network for the problem of source identification in a given network with different infection rates. First, we put forward a label propagation framework, which can locate the infection source based on both infected and uninfected nodes. Then, a novel Source Identification Graph Convolutional Network (SIGN) framework is proposed inspired by label propagation. Third, we modify the classical cross-entropy loss function and presented neighborhood loss to optimize the average error distance. Finally, extensive experiments are performed on eight datasets with different topologies and varying infection sizes to demonstrate the effectiveness of the proposed algorithms. We compare the proposed method with four mainstream approaches, and our method shows strong performances especially under the large infection size.

Keywords: Graph convolutional network; Infectious disease; Label propagation; Source identification.