Propagation source identification of infectious diseases with graph convolutional networks

Area of Science:

• Network science
• Computational epidemiology
• Machine learning

Background:

• Accurate source identification is crucial for controlling infectious disease spread.
• Existing methods often struggle to balance accuracy and error distance.
• Network topology and infection rates significantly impact source identification challenges.

Purpose of the Study:

• To develop an effective method for identifying infection sources in networks.
• To improve both the accuracy and error distance of source identification.
• To leverage graph convolutional networks for epidemiological network analysis.

Main Methods:

• Proposed a label propagation framework incorporating both infected and uninfected nodes.
• Developed a novel Source Identification Graph Convolutional Network (SIGN) framework.
• Introduced a modified loss function, neighborhood loss, to minimize average error distance.

Main Results:

• SIGN demonstrated effective source identification across diverse network topologies and infection sizes.
• The proposed method achieved strong performance, particularly under large infection sizes.
• SIGN outperformed four mainstream approaches in extensive experimental evaluations.

Conclusions:

• The SIGN framework offers a robust solution for network source identification.
• The approach effectively addresses challenges posed by varying infection rates and network structures.
• This work advances the application of graph convolutional networks in disease propagation analysis.