quanteda (version 4.0.1)

dfm_trim: Trim a dfm using frequency threshold-based feature selection

Description

Returns a document by feature matrix reduced in size based on document and term frequency, usually in terms of a minimum frequency, but may also be in terms of maximum frequencies. Setting a combination of minimum and maximum frequencies will select features based on a range.

Feature selection is implemented by considering features across all documents, by summing them for term frequency, or counting the documents in which they occur for document frequency. Rank and quantile versions of these are also implemented, for taking the first \(n\) features in terms of descending order of overall global counts or document frequencies, or as a quantile of all frequencies.

Usage

dfm_trim(
  x,
  min_termfreq = NULL,
  max_termfreq = NULL,
  termfreq_type = c("count", "prop", "rank", "quantile"),
  min_docfreq = NULL,
  max_docfreq = NULL,
  docfreq_type = c("count", "prop", "rank", "quantile"),
  sparsity = NULL,
  verbose = quanteda_options("verbose"),
  ...
)

Value

A dfm reduced in features (with the same number of documents)

Arguments

x

a dfm object

min_termfreq, max_termfreq

minimum/maximum values of feature frequencies across all documents, below/above which features will be removed

termfreq_type

how min_termfreq and max_termfreq are interpreted. "count" sums the frequencies; "prop" divides the term frequencies by the total sum; "rank" is matched against the inverted ranking of features in terms of overall frequency, so that 1, 2, ... are the highest and second highest frequency features, and so on; "quantile" sets the cutoffs according to the quantiles (see quantile()) of term frequencies.

min_docfreq, max_docfreq

minimum/maximum values of a feature's document frequency, below/above which features will be removed

docfreq_type

specify how min_docfreq and max_docfreq are interpreted. "count" is the same as [docfreq](x, scheme = "count"); "prop" divides the document frequencies by the total sum; "rank" is matched against the inverted ranking of document frequency, so that 1, 2, ... are the features with the highest and second highest document frequencies, and so on; "quantile" sets the cutoffs according to the quantiles (see quantile()) of document frequencies.

sparsity

equivalent to 1 - min_docfreq, included for comparison with tm

verbose

print messages

...

not used

See Also

dfm_select(), dfm_sample()

Examples

dfmat <- dfm(tokens(data_corpus_inaugural))

# keep only words occurring >= 10 times and in >= 2 documents
dfm_trim(dfmat, min_termfreq = 10, min_docfreq = 2)

# keep only words occurring >= 10 times and in at least 0.4 of the documents
dfm_trim(dfmat, min_termfreq = 10, min_docfreq = 0.4)

# keep only words occurring <= 10 times and in <=2 documents
dfm_trim(dfmat, max_termfreq = 10, max_docfreq = 2)

# keep only words occurring <= 10 times and in at most 3/4 of the documents
dfm_trim(dfmat, max_termfreq = 10, max_docfreq = 0.75)

# keep only words occurring 5 times in 1000, and in 2 of 5 of documents
dfm_trim(dfmat, min_docfreq = 0.4, min_termfreq = 0.005, termfreq_type = "prop")

## quantiles
toks <- as.tokens(list(unlist(mapply(rep, letters[1:10], 10:1), use.names = FALSE)))
dfmat <- dfm(toks)
dfmat

# keep only the top 20th percentile or higher features

# keep only words above the 80th percentile
dfm_trim(dfmat, min_termfreq = 0.800001, termfreq_type = "quantile", verbose = TRUE)

# keep only words occurring frequently (top 20%) and in <=2 documents
dfm_trim(dfmat, min_termfreq = 0.2, max_docfreq = 2, termfreq_type = "quantile")

if (FALSE) {
# compare to removeSparseTerms from the tm package
(dfmattm <- convert(dfmat, "tm"))
tm::removeSparseTerms(dfmattm, 0.7)
dfm_trim(dfmat, min_docfreq = 0.3)
dfm_trim(dfmat, sparsity = 0.7)
}