Inferring network structure with unobservable nodes from time series data.

Area of Science:

• Network science
• Computational systems biology
• Data science

Background:

• Network structures are fundamental to various systems but often incompletely observed.
• Existing methods struggle with inferring network structures from partial node or connection information.
• Understanding complex dynamics requires complete network information.

Purpose of the Study:

• To develop a novel method for inferring complete network structures from incomplete data.
• To address the challenge of unobservable nodes and connections in real-world networks.
• To leverage time series data from network dynamics for network inference.

Main Methods:

• Proposed a data-driven deep learning model named Gumbel-softmax Inference for Network (GIN).
• GIN framework comprises a dynamics learner, network generator, and initial state generator.
• Treated network inference as minimizing prediction errors for observable node states.

Main Results:

• GIN achieved up to 90% accuracy in inferring unknown network structures and initial states.
• Accuracy showed a linear decline with increasing fractions of unobservable nodes.
• Experiments were conducted on artificial and empirical social networks with diverse dynamics.

Conclusions:

• The GIN framework effectively infers network structures and initial states under incomplete information.
• The method demonstrates significant potential for applications where network data is scarce.
• GIN offers a robust solution for network reconstruction using rich time series data.