A list whose length is equal to the number of documents, D. Each
element of documents is an integer matrix with two rows. Each
column of documents[[i]] (i.e., document i) represents a
word occurring in the document.
documents[[i]][1, j] is a
0-indexed word identifier for the jth word in document i.
documents[[i]][2,j] is an integer specifying the number of times that word appears in
the document.
mlabels
The training labels for the documents.
ntopics
Number of topics to be used in modeling the corpus.
initial_c, lambda, rho
These are positive-valued regularization constants.
Default values are initial_c=0.5, lambda=0.1, rho=0.01
delta_ell
The parameter for the svm cost function, i.e., 0/(delta ell) loss.
Only positive values are allowed. Default value is 3600.
supervised
If the value is TRUE, the model is a supervised MedSTC; if FALSE, the model
is the unsupervised STC.
primal_svm
Only works when "supervised" is set at 1. If the value is 1, uses the
loss-augmented prediction (i.e., sub-gradient) to update document codes;
otherwise it uses the gradient with Lagrangian multipliers to update document codes.
var_max_iter
The maximum number of iterations of coordinate descent for a single document.
convergence
The convergence criteria for coordinate descent. Stop if
(objective_old - objective) / abs(objective_old) is less than this value (or
after the maximum number of iterations). Note that "objective" is the objective
value for a single document.
em_max_iter
The maximum number of iterations of hierarchical sparse coding, dictionary learning,
and svm training (for supervised MedSTC).
em_convergence
The convergence criteria for coordinate descent. Stop if (objective_old -
objective) / abs(objective_old) is less than this value (or after the
maximum number of iterations). Note that "objective" is the objective value
for the whole corpus.
svm_alg_type
If set to 0 then the n-slack multi-class SVM is used. If set to 2,
then the 1-slack multi-class SVM is used. In our testing, the 1-slack
SVM is faster.
output_dir
A directory for writing intermediate results.
Directory is removed after the calculation is done, but is needed during the run.
Value
model
A model object of the medSTC class, which has a state list with five elements:
The first two list elements are for storing the model parameter state after the model completed training.
The third list element is the LogProbabilityOfWordsForTopics, which can be used for
word assignments to topics.
The fourth and fifth model state list elements are Eta and Mu. (refer to paper)
The model also stores the original paramater values.