EM: EM algorithm

In activetext/activeR: a semi-supervised active learning algorithm for text classification.

EM algorithm

Description

Use EM algorithm to maximize the marginal posterior. The marginal posterior is the probability of parameters given both labeled and unlabeled documents and the labels for the labeled documents

Usage

EM(
  .D_train = NULL,
  .C_train = NULL,
  .D_test,
  .n_class = 2,
  .lambda = 0.1,
  .max_iter = 100,
  .alpha = 0.1,
  .lazy_eval = F,
  .counter_on = T,
  .active_iter = NULL,
  .maxactive_iter = NULL,
  .fixed_words = NULL,
  .supervise = T,
  .class_prob = NULL,
  .word_prob = NULL,
  .export_all = F
)

Arguments

.D_train	document term matrix of the labeled documents
.C_train	vector of class labels for the labeled documents
.D_test	document term matrix of the unlabeled documents
.n_class	number of classes
.lambda	vector of document weights
.max_iter	maximum number of iteration of the EM algorithm
.alpha	the threshold of the convergence. If the increase of the maximand becomes less than alpha, the iteration stops.
.lazy_eval	boolean object. If lazy_eval == T, convergence is measured by comparing changes in log likelihood across model iterations rather than directly computing maximand.
.counter_on	boolean object. If counter_on == T, displays the progress of the EM algorithm.
.active_iter	integer value that tells the EM algorithm which iteration of the active loop it is in.
.maxactive_iter	integer value that tells the EM algorithm the maximum allowed active iterations.
.fixed_words	matrix of fixed words with class probabilities, where ncol is the number of classes.
.supervise	T if supervised. F is unsupervised.
.class_prob	required if .supervise == T. Starting value of class probability (logged)
.word_prob	required if .supervise == T. Starting value of word probability (logged)
.export_all	If T, model parameters from each iteration of the EM algorithm are returned. If F, only model results from the last iteration are returned.

Details

The inputs must conform to the following specifications D_train: a matrix with dimension: the number of labeled documents * the number of unique words D_test: a matrix with dimension: the number of labeled documents * the number of unique words The column length of D_train and D_test must be the same. The elements of the D_train, D_test are integers (the counts of each unique word appeard in each document) C_train: vector of labels for the labeled documents. The length must be the same as the row length of D_test

Value

maximands is a vector of maximands in each iteration. Each element of the vector contains the log maximand in each step. pi is a vector of log class probabilities. (length = 2) eta is a matrix of log word probabilities (nrow = the number of all documents, ncol = 2)

References

Active EM in overleaf

activetext/activeR documentation built on May 31, 2024, 10:21 a.m.