Visualization of Gibbs Sampling in Social Networks

Christopher Hundt ( chundt at cs )

This is a visualization, made with the prefuse framework, of Gibbs sampling to infer stochastic blockmodel structures on social networks. It is based on Estimation and Prediction for Stochastic Blockstructures, by Krzysztof Nowicki and Tom Snijders, and Discovering latent classes in relational data, by Kemp, Griffiths, and Tenenbaum. You can see the Gibbs sampling in action, and manipulate the graph a little to gain more information.

Note: requires Java 1.5

The applet below gives the visualization. You can choose between two models and three datasets.

Don't forget to click "Update" after choosing a model and dataset. Then slide the "Speed" slider above 0 to start the animation.

The visualization works like this: the nodes don't care where they are, except that they wish to be close to other nodes that they are likely to be in a group with, and far from other nodes that they are not likely to be in a group in. The more sure they are, the harder they will push/pull. Once Gibbs sampling starts, the sampler will inform the nodes how likely they are to be grouped with other nodes.

• Mouse over a node to highlight its relationships.
• The color radio button shows or hides a proposed grouping.
• The Speed slider controls how fast Gibbs sampling is done. Slide it above 0 to start the animation.
• The parameter sliders control hyperparameters used for hyperparameters for the Gibbs sampler. If "sample" is available you can choose it to sample the hyperparameters along with the others.
• Count displays the number of iterations of Gibbs sampling so far
• Reset starts the Gibbs sampling over (but doesn't rearrange the nodes)