Bayesian Generative Network

Data collection biases pose a persistent challenge in social network analysis, particularly in animal social network studies where observations are uneven, censored, and incomplete. These biases can lead to distorted network inference and incorrect conclusions about social behaviour. We present a generative Bayesian framework based on the Social Relationship Model (SRM) that jointly estimates latent social structure while explicitly accounting for sampling and node-level censoring biases. Simulation experiments reflecting realistic observational scenarios show that this approach reliably recovers true social connections, even when key individuals are intermittently unobserved, and outperforms commonly used methods such as permutation-based tests and linear regression models. To support the practical application of such models, we also introduce Bayesian Inference (BI), a cross-platform Bayesian modeling software available in Python, R, and Julia. BI combines an intuitive model specification syntax with the flexibility of low-level Bayesian programming and leverages GPU acceleration to enable efficient inference for high-dimensional models, achieving substantial speed gains over comparable Stan implementations. Together, these contributions improve both the robustness and accessibility of Bayesian social network analysis under realistic data collection constraints.