lda.svi (version 0.1.0)

lda_svi: Fit a Latent Dirichlet Allocation model to a text corpus

Description

Fit a Latent Dirichlet Allocation model to a text corpus

Usage

lda_svi(dtm, passes = 10, batchsize = 256, maxiter = 100, K,
  eta = 1/K, alpha = 1/K, kappa = 0.7, tau_0 = 1024,
  tidy_output = TRUE)

Arguments

dtm

This must be a DocumentTermMatrix (with term frequency weighting) from the tm package.

passes

The number of passes over the whole corpus - how many times we update the local variables for each document.

batchsize

The size of the minibatches.

maxiter

The maximum iterations for the "E step" for each document (the updating of the per-document parameters within each minibatch). The default of 100 follows the reference implementation in python by the authors.

K

The number of topics

eta

Dirichlet prior hyperparameter for the document-specific topic proportions.

alpha

Dirichlet prior hyperparameter for the topic-specific term proportions.

kappa

learning rate parameter. Lower values give greater weight to later iterations. For guaranteed convergence to a local optimum, kappa must lie in the interval (0.5,1].

tau_0

learning rate parameter. Higher values reduce the influence of early iterations.

tidy_output

if true, the parameter estimates are returned as 'long' data frames; otherwise they are returned as matrices.

Value

A named list of length two. The element named 'beta' gives the proportions for the terms within the topics, while the element named 'theta' gives the proportions for the topics within the documents. If the tidy_output argument is true these are data frames in 'long' format; otherwise they are matrices.

Details

The implementation here is based on the python implementation by Matthew D. Hoffman accompanying the paper

References

Hoffman, M., Bach, FM., and Blei, DM. (2010) 'Online Learning for Latent Dirichlet Allocation', _Conference and Workshop on Neural Information Processing Systems_

Hoffman, M., Blei, DM., Wang, C, and Paisley, J. (2013) 'Stochastic Variational Inference', _Journal of Machine Learning Research_. Preprint: https://arxiv.org/abs/1206.7051_

Examples

# NOT RUN {
library(topicmodels)
data(AssociatedPress)
ap_lda_fit <- lda_svi(AssociatedPress,passes=1,K=50)
#I use a single pass because CRAN requires examples to run quickly; 
#generally one would use more. 20 often seems to be sufficient as a rule of thumb,
#but it might be worth experimenting with more or fewer
# }