mpmm.cgs: Collapsed Gibbs sampler for the marginal product mixture...

In mpmm: Marginal Product Mixture Model for Relational Event Data

Description

Performs collapsed Gibbs sampling for the marginal-product mixture model (DuBois 2010). This function takes a dataset of observed relational events, perform inference, and return point estimates of the latent parameters using the state at the last iteration of Gibbs sampling.

The model is appropriate for analyzing event data in 1, 2, or 3-mode graphs. The main assumption is that each event belongs to a latent class. We use MCMC to explore the posterior distribution of these assignments given observed data. The resulting class assignments can be useful for exploratory data analysis and predicting missing or future data.

At the moment, the number of latent classes must be chosen a priori, though a DP version is in the works.

With the current implementation, each dimension is treated separately, e.g. for a given latent class, there will be a distribution over likely senders and a separate distribution over likely recipients, even if this involves the same set of actors.

Usage

mpmm.cgs(dataset, C, dims, priors, niter = 10, z.init = NULL)

Arguments

dataset	A Tx2 or Tx3 matrix of integers, each row corresponding to an event, each integer identifying the participating vertices.
C	Number of desired latent classes.
dims	Maximum value for each categorical random variable.
priors	pi the value for the symmetric Dirichlet prior on pi (called alpha in the paper) phi a list where each element is the value for the symmetric Dirichlet prior for the corresponding phi vector
niter	Number of Gibbs iterations to perform.
z.init	vector of initial latent class assignments for all observations.

Value

assignments	Latent class assignments for the T observations from the last iteration of the Gibbs sampler.
params	pi MAP estimate of the pi vector phi MAP estimate for the phi vectors
counts	A list where each element is a matrix counting the number of times values of the corresponding dimension were observed and assigned to a particular latent class. (These matrices are used for estimates of phi.)
dims	Maximum value for each categorical random variable.

Author(s)

Christopher DuBois (<email: [e-mail]>)

References

Christopher DuBois and Padhraic Smyth. Modeling Relational Events via Latent Classes. Proceedings of the 16th ACM SIGKDD, 2010.

Examples

## See demo(synthetic) and demo(enron).

mpmm documentation built on May 2, 2019, 4:55 p.m.