Hierarchical network models for exchangeable structured interaction processes.

Area of Science:

• Network Science
• Statistical Modeling
• Data Analysis

Background:

• Network data often involves complex, structured interactions between elements.
• Examples include email exchanges and co-authorship in scientific articles.
• Existing models may not fully capture the nuances of such structured interactions.

Purpose of the Study:

• To introduce a novel statistical model, the Pitman-Yor hierarchical vertex components model (PY-HVCM).
• To effectively model complex relational data with structured interactions.
• To analyze properties like sparsity and power law degree distribution in networks.

Main Methods:

• Development of the Pitman-Yor hierarchical vertex components model (PY-HVCM).
• Utilizing partial pooling of local information via a latent, shared population-level distribution.
• Derivation of a computationally tractable Gibbs sampling algorithm for inference.

Main Results:

• The PY-HVCM provides clear model interpretation and establishes global sparsity.
• The model demonstrates a power law degree distribution in complex networks.
• Successful application and validation on the Enron email and ArXiv datasets.

Conclusions:

• The PY-HVCM is a powerful tool for analyzing exchangeable structured interaction-labeled networks.
• The model effectively captures sparsity and power law properties.
• Posterior predictive validation confirms the model's goodness of fit for real-world network data.