The von Mises Graphical Model:
Expectation Propagation for Inference
Narges Sharif Razavian*, Hetunandan Kamisetty, Christopher James Langmead**
Also appears as Computer Science Department
Technical Report CMU-CS-11-130
Keywords: Inference, Expectation Propagation, von Mises, Probabilistic Graphical Models, Proteins
The von Mises model encodes a multivariate circular distribution as an undirected probabilistic graphical model. Presently, the only algorithm for performing inference in the model is Gibbs sampling, which becomes inefficient for large graphs. To address this issue, we introduce an
Expectation Propagation based algorithm for performing inference in the von Mises graphical model. Our approach introduces a moment-matching technique for trigonometric functions to approximate the Expectation Propagation messages efficiently. We show that our algorithm has better speed of convergence and similar accuracy compared to Gibbs sampling, on synthetic data as well as real-world data from protein structures.