A Return to Biased Nets: New Specifications and Approximate Bayesian Inference.

Area of Science:

• Network analysis and statistical modeling
• Social network analysis
• Computational social science

Background:

• The biased net paradigm offers a framework for understanding complex network dependencies beyond random graph models.
• Previous specifications required approximations, limiting their empirical tractability and potentially introducing inconsistencies.
• Local specifications of biased nets aimed to improve upon earlier tracing-based methods.

Purpose of the Study:

• To address inconsistencies found in existing local specifications of biased nets.
• To develop a more robust and extensible framework for network parameterization.
• To introduce and evaluate the utility of inhibitory bias events and approximate Bayesian inference methods.

Main Methods:

• Development of a Markovian specification for biased net models, resolving prior inconsistencies.
• Introduction of inhibitory bias events, exemplified by satiation, to prevent model degeneracies.
• Application of approximate Bayesian inference using random forest prevision due to the lack of a computable likelihood.

Main Results:

• The proposed Markovian specification successfully evades inconsistencies and allows for the incorporation of new effects.
• Inhibitory bias events effectively mitigate degeneracies arising from closure bias terms.
• The approach was successfully demonstrated on a college student friendship network, validating prior findings on sibling bias and tie strength.

Conclusions:

• The refined Markovian biased net specification provides a more consistent and flexible tool for network analysis.
• Inhibitory bias events represent a valuable addition for modeling complex network structures.
• Random forest prevision offers a viable strategy for approximate Bayesian inference in network models where direct computation is infeasible.