test_features {biogram}

Permutation test for feature selection

Description

Performs a feature selection on positioned n-gram data using a Fisher's permutation test.

Usage

test_features(
  target,
  features,
  criterion = "ig",
  adjust = "BH",
  threshold = 1,
  quick = TRUE,
  times = 1e+05
)

Arguments

target

integer vector with target information (e.g. class labels).

features

integer matrix of features with number of rows equal to the length of the target vector.

criterion

criterion used in permutation test. See Details for the list of possible criterions.

adjust

name of p-value adjustment method. See p.adjust for the list of possible values. If NULL, p-values are not adjusted.

threshold

integer. Features that occur less than threshold and more often than nrow(features)-threshold are discarded from the permutation test.

quick

logical, if TRUE Quick Permutation Test (QuiPT) is used. If FALSE, normal permutation test is performed.

times

number of times procedure should be repeated. Ignored if quick is TRUE.

Details

Since the procedure involves multiple testing, it is advisable to use one of the avaible p-value adjustment methods. Such methods can be used directly by specifying the adjust parameter.

Available criterions:

ig

Information Gain: calc_ig.

kl

Kullback-Leibler divergence: calc_kl.

cs

Chi-squared-based measure: calc_cs.

Value

an object of class feature_test.

Note

Both target and features must be binary, i.e. contain only 0 and 1 values.

Features occuring too often and too rarely are considered not informative and may be removed using the threshold parameter.

References

Radivojac P, Obradovic Z, Dunker AK, Vucetic S, Feature selection filters based on the permutation test in Machine Learning: ECML 2004, 15th European Conference on Machine Learning, Springer, 2004.

See Also

binarize - binarizes input data.

calc_criterion - computes selected criterion.

distr_crit - distribution of criterion used in QuiPT.

summary.feature_test - summary of results.

cut.feature_test - aggregates test results in groups based on feature's p-value.

Examples

# significant feature
tar_feat1 <- create_feature_target(10, 390, 0, 600) 
# significant feature
tar_feat2 <- create_feature_target(9, 391, 1, 599)
# insignificant feature
tar_feat3 <- create_feature_target(198, 202, 300, 300)
test_res <- test_features(tar_feat1[, 1], cbind(tar_feat1[, 2], tar_feat2[, 2], 
                          tar_feat3[, 2]))
summary(test_res)
cut(test_res)

# real data example
# we will analyze only a subsample of a dataset to make analysis quicker
ids <- c(1L:100, 701L:800)
deg_seqs <- degenerate(human_cleave[ids, 1L:9], 
                       list(`a` = c(1, 6, 8, 10, 11, 18), 
                            `b` = c(2, 5, 13, 14, 16, 17, 19, 20), 
                            `c` = c(3, 4, 7, 9, 12, 15)))

# positioned n-grams example
bigrams_pos <- count_ngrams(deg_seqs, 2, letters[1L:3], pos = TRUE)
test_features(human_cleave[ids, 10], bigrams_pos)

# unpositioned n-grams example, binarization required
bigrams_notpos <- count_ngrams(deg_seqs, 2, letters[1L:3], pos = TRUE)
test_features(human_cleave[ids, 10], binarize(bigrams_notpos))