Empirical networks are sparse: Enhancing multiedge models with zero-inflation.

Area of Science:

• Network science
• Data science
• Statistical modeling

Background:

• Empirical networks exhibit significant sparsity, meaning most potential connections are absent.
• Traditional network models struggle to represent the high number of disconnected node pairs in real-world data.
• Existing models like configuration and stochastic block models fail to capture network sparsity accurately.

Purpose of the Study:

• To address the limitations of classical network models in representing network sparsity.
• To introduce and evaluate zero-inflated models for analyzing real-world networks.
• To improve the accuracy of network models by accounting for excess zeros (disconnected pairs).

Main Methods:

• Analysis of all datasets from the Sociopatterns repository.
• Implementation and comparison of zero-inflated network models against classical models.
• Evaluation of model performance in capturing sparsity and heavy-tailed edge distributions.

Main Results:

• Zero-inflated models provide a more accurate representation of real-world network sparsity.
• These models effectively capture the heavy-tailed distributions of edge counts observed in empirical data.
• Classical network models demonstrate significant biases due to their inability to account for excess zeros.

Conclusions:

• Zero-inflation is crucial for accurately modeling sparse real-world networks.
• Failure to incorporate zero-inflation leads to biased network models and inaccurate system dynamics.
• Zero-inflated models offer a superior framework for understanding complex systems and their interactions.