Information Localization in Growing Networks

This paper introduces the concept of information localization in growing network models, demonstrating that information about model parameters is often contained within small subgraphs. This has significant implications for inference, allowing for the use of graph neural networks (GNNs) with limited receptive fields to approximate the posterior distribution of model parameters. The work provides a theoretical justification for analyzing local subgraphs and using GNNs for likelihood-free inference, which is crucial for complex network models where the likelihood is intractable. The paper's findings are important because they offer a computationally efficient way to perform inference on growing network models, which are used to model a wide range of real-world phenomena.

• •Information about model parameters in growing network models is often localized.
• •This localization allows for efficient inference using GNNs with limited receptive fields.
• •The approach is applicable even for non-localized models, offering a computationally efficient alternative to model-specific inference methods.
• •The findings justify the analysis of local subgraphs and the use of GNNs for likelihood-free inference.