Differentially Private Undirected Graphical Model Learning Based on Collective Graphical Models

The study proposes a differentially private method for learning discrete undirected graphical models based on collective graphical models and the EM framework, releasing noisy sufficient statistics via the Laplace mechanism. A naive algorithm that directly uses these statistics outperforms generic differentially private learning algorithms, but ignores the data generation process, has uncertain theoretical grounding, and exhibits ill-conditioned behavior. On synthetic data and real human mobility data, the models learned by the method outperform comparison methods.