Help for package theftdlc

Type:

Package

Title:

Analyse and Interpret Time Series Features

Version:

0.2.0

Date:

2025-07-10

Maintainer:

Trent Henderson <then6675@uni.sydney.edu.au>

Description:

Provides a suite of functions for analysing, interpreting, and visualising time-series features calculated from different feature sets from the 'theft' package. Implements statistical learning methodologies described in Henderson, T., Bryant, A., and Fulcher, B. (2023) <doi:10.48550/arXiv.2303.17809>.

BugReports:

https://github.com/hendersontrent/theftdlc/issues

License:

MIT + file LICENSE

Encoding:

UTF-8

Depends:

R (≥ 3.5.0), theft (≥ 0.8.1)

Imports:

rlang, stats, tibble, dplyr, ggplot2, tidyr, purrr, furrr, future, reshape2, scales, broom, Rtsne, e1071, janitor, umap, MASS, mclust, normaliseR, correctR, glmnet

Suggests:

lifecycle, cachem, bslib, knitr, rmarkdown, pkgdown, testthat

RoxygenNote:

7.3.2

VignetteBuilder:

knitr

URL:

https://hendersontrent.github.io/theftdlc/

NeedsCompilation:

Packaged:

2025-07-10 08:02:45 UTC; trenthenderson

Author:

Trent Henderson [cre, aut]

Repository:

CRAN

Date/Publication:

2025-07-10 08:20:02 UTC

Fit classifiers using time-series features using a resample-based approach and get a fast understanding of performance

Description

Fit classifiers using time-series features using a resample-based approach and get a fast understanding of performance

Usage

classify(
  data,
  classifier = NULL,
  train_size = 0.75,
  n_resamples = 30,
  by_set = TRUE,
  use_null = FALSE,
  seed = 123
)

tsfeature_classifier(
  data,
  classifier = NULL,
  train_size = 0.75,
  n_resamples = 30,
  by_set = TRUE,
  use_null = FALSE,
  seed = 123
)

Arguments

data

feature_calculations object containing the raw feature matrix produced by theft::calculate_features

classifier

function specifying the classifier to fit. Should be a function with 2 arguments: formula and data containing a classifier compatible with R's predict functionality. Please note that classify z-scores data prior to modelling using the train set's information so disabling default scaling if your function uses it is recommended. Defaults to NULL which means the following linear SVM is fit: classifier = function(formula, data){mod <- e1071::svm(formula, data = data, kernel = "linear", scale = FALSE, probability = TRUE)}

train_size

numeric denoting the proportion of samples to use in the training set. Defaults to 0.75

n_resamples

integer denoting the number of resamples to calculate. Defaults to 30

by_set

Boolean specifying whether to compute classifiers for each feature set. Defaults to TRUE. If FALSE, the function will instead find the best individually-performing features

use_null

Boolean whether to fit null models where class labels are shuffled in order to generate a null distribution that can be compared to performance on correct class labels. Defaults to FALSE

seed

integer to fix R's random number generator to ensure reproducibility. Defaults to 123

Value

list containing a named vector of train-test set sizes, and a data.frame of classification performance results

Author(s)

Trent Henderson

Examples


library(theft)

features <- theft::calculate_features(theft::simData,
  feature_set = "catch22")

classifiers <- classify(features,
  by_set = FALSE,
  n_resamples = 3)

Perform cluster analysis of time series using their feature vectors

Description

Perform cluster analysis of time series using their feature vectors

Usage

cluster(
  data,
  norm_method = c("zScore", "Sigmoid", "RobustSigmoid", "MinMax", "MaxAbs"),
  unit_int = FALSE,
  clust_method = c("kmeans", "hclust", "mclust"),
  k = 2,
  features = NULL,
  na_removal = c("feature", "sample"),
  seed = 123,
  ...
)

Arguments

data

feature_calculations object containing the raw feature matrix produced by theft::calculate_features

norm_method

character denoting the rescaling/normalising method to apply. Can be one of "zScore", "Sigmoid", "RobustSigmoid", "MinMax", or "MaxAbs". Defaults to "zScore"

unit_int

Boolean whether to rescale into unit interval [0,1] after applying normalisation method. Defaults to FALSE

clust_method

character specifying the clustering algorithm to use. Can be one of "kmeans" for k-means clustering, "hclust" for hierarchical clustering, or "mclust" for Gaussian mixture model clustering. Defaults to "kMeans"

k

integer denoting the number of clusters to extract. Defaults to 2

features

character vector denoting the names of time-series features to use in the clustering algorithm. Defaults to NULL for no feature filtering and usage of the entire feature matrix

na_removal

character defining the way to deal with NAs produced during feature calculation. Can be one of "feature" or "sample". "feature" removes all features that produced any NAs in any sample, keeping the number of samples the same. "sample" omits all samples that produced at least one NA. Defaults to "feature"

seed

integer to fix R's random number generator to ensure reproducibility. Defaults to 123

...

arguments to be passed to stats::kmeans or stats::hclust, or mclust::Mclust depending on selection in clust_method

Value

object of class feature_cluster containing the clustering algorithm and a tidy version of clusters joined to the input dataset ready for further analysis

Author(s)

Trent Henderson

Examples


library(theft)

features <- theft::calculate_features(theft::simData,
  feature_set = "catch22")

clusts <- cluster(features,
  k = 6)

Conduct statistical testing on time-series feature classification performance to identify top features or compare entire sets

Description

Conduct statistical testing on time-series feature classification performance to identify top features or compare entire sets

Usage

compare_features(
  data,
  metric = c("accuracy", "precision", "recall", "f1"),
  by_set = TRUE,
  hypothesis = c("null", "pairwise"),
  p_adj = c("none", "holm", "hochberg", "hommel", "bonferroni", "BH", "BY", "fdr"),
  n_workers = 1
)

Arguments

data

list object containing the classification outputs produce by tsfeature_classifier

metric

character denoting the classification performance metric to use in statistical testing. Can be one of "accuracy", "precision", "recall", "f1". Defaults to "accuracy"

by_set

Boolean specifying whether you want to compare feature sets (if TRUE) or individual features (if FALSE). Defaults to TRUE but this is contingent on whether you computed by set or not in tsfeature_classifier

hypothesis

character denoting whether p-values should be calculated for each feature set or feature (depending on by_set argument) individually relative to the null if use_null = TRUE in tsfeature_classifier through "null", or whether pairwise comparisons between each set or feature should be conducted on main model fits only through "pairwise". Defaults to "null"

p_adj

character denoting the adjustment made to p-values for multiple comparisons. Should be a valid argument to stats::p.adjust. Defaults to "none" for no adjustment. "holm" is recommended as a starting point for adjustments

n_workers

integer denoting the number of parallel processes to use. Defaults to 1 for serial processing

Value

data.frame containing the results

Author(s)

Trent Henderson

References

Henderson, T., Bryant, A. G., and Fulcher, B. D. Never a Dull Moment: Distributional Properties as a Baseline for Time-Series Classification. 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, (2023).

Examples


library(theft)

features <- theft::calculate_features(theft::simData,
  feature_set = NULL,
  features = list("mean" = mean, "sd" = sd))

classifiers <- classify(features,
                        by_set = FALSE,
                        n_resamples = 3)

compare_features(classifiers,
                 by_set = FALSE,
                 hypothesis = "pairwise")

Remove duplicate features that exist in multiple feature sets and retain a reproducible random selection of one of them

Description

Remove duplicate features that exist in multiple feature sets and retain a reproducible random selection of one of them

Usage

filter_duplicates(data, preference = NULL, seed = 123)

Arguments

data

feature_calculations object containing the raw feature matrix produced by calculate_features

preference

deprecated. Do not use

seed

integer denoting a fix for R's pseudo-random number generator to ensure selections are reproducible. Defaults to 123

Value

feature_calculations object containing filtered feature data

Author(s)

Trent Henderson

Filter resample data sets according to good feature list

Description

Filter resample data sets according to good feature list

Usage

filter_good_features(data, x, good_features)

Arguments

data

list of "Train" and "Test" data

x

integer denoting the resample index to operate on

good_features

character vector of good features to keep

Value

list of filtered train and test data

Author(s)

Trent Henderson

Helper function to find features in both train and test set that are "good"

Description

Helper function to find features in both train and test set that are "good"

Usage

find_good_features(data, x)

Arguments

data

list of "Train" and "Test" data

x

integer denoting the resample index to operate on

Value

character vector of "good" feature names

Author(s)

Trent Henderson

Fit classification model and compute key metrics

Description

Fit classification model and compute key metrics

Usage

fit_models(data, iter_data, row_id, is_null_run = FALSE, classifier)

Arguments

data

list containing train and test sets

iter_data

data.frame containing the values to iterate over for seed and either feature name or set name

row_id

integer denoting the row ID for iter_data to filter to

is_null_run

Boolean whether the calculation is for a null model. Defaults to FALSE

classifier

function specifying the classifier to fit. Should be a function with 2 arguments: formula and data. Please note that tsfeature_classifier z-scores data prior to modelling using the train set's information so disabling default scaling if your function uses it is recommended.

Value

data.frame of classification results

Author(s)

Trent Henderson

Calculate central tendency and spread values for all numeric columns in a dataset

Description

Calculate central tendency and spread values for all numeric columns in a dataset

Usage

get_rescale_vals(data)

Arguments

data

data.frame containing data to normalise

Value

list of central tendency and spread values

Author(s)

Trent Henderson

Calculate interval summaries with a measure of central tendency of classification results

Description

Calculate interval summaries with a measure of central tendency of classification results

Usage

interval(
  data,
  metric = c("accuracy", "precision", "recall", "f1"),
  by_set = TRUE,
  type = c("sd", "qt", "quantile"),
  interval = NULL,
  model_type = c("main", "null")
)

calculate_interval(
  data,
  metric = c("accuracy", "precision", "recall", "f1"),
  by_set = TRUE,
  type = c("sd", "qt", "quantile"),
  interval = NULL,
  model_type = c("main", "null")
)

Arguments

data

list object containing the classification outputs produce by tsfeature_classifier

metric

character denoting the classification performance metric to calculate intervals for. Can be one of "accuracy", "precision", "recall", "f1". Defaults to "accuracy"

by_set

Boolean specifying whether to compute intervals for each feature set. Defaults to TRUE. If FALSE, the function will instead calculate intervals for each feature

type

character denoting whether to calculate a +/- SD interval with "sd", confidence interval based off the t-distribution with "qt", or based on a quantile with "quantile". Defaults to "sd"

interval

numeric scalar denoting the width of the interval to calculate. Defaults to 1 if type = "sd" to produce a +/- 1 SD interval. Defaults to 0.95 if type = "qt" or type = "quantile" for a 95 per cent interval

model_type

character denoting whether to calculate intervals for main models with "main" or null models with "null" if the use_null argument when using tsfeature_classifier was use_null = TRUE. Defaults to "main"

Value

interval_calculations object which is a data frame containing the results

Author(s)

Trent Henderson

Examples


library(theft)

features <- theft::calculate_features(theft::simData,
  feature_set = NULL,
  features = list("mean" = mean, "sd" = sd))

classifiers <- classify(features,
  by_set = FALSE,
  n_resamples = 3)

interval(classifiers,
  by_set = FALSE,
  type = "sd",
  interval = 1)

Helper function for converting to title case

Description

Helper function for converting to title case

Usage

make_title(x)

Arguments

x

character vector

Value

character vector

Author(s)

Trent Henderson

Produce a plot for a feature_calculations object

Description

Produce a plot for a feature_calculations object

Usage

## S3 method for class 'feature_calculations'
plot(
  x,
  type = c("matrix", "cor", "violin", "box", "quality"),
  norm_method = c("zScore", "Sigmoid", "RobustSigmoid", "MinMax", "MaxAbs"),
  unit_int = FALSE,
  clust_method = c("average", "ward.D", "ward.D2", "single", "complete", "mcquitty",
    "median", "centroid"),
  cor_method = c("pearson", "spearman"),
  feature_names = NULL,
  ...
)

Arguments

x

feature_calculations object containing the raw feature matrix produced by theft::calculate_features

type

character specifying the type of plot to draw. Can be one of "matrix", "cor", "violin", "box", or "quality". Defaults to "matrix"

norm_method

character specifying a rescaling/normalising method to apply if type = "matrix" or if type = "cor". Can be one of "zScore", "Sigmoid", "RobustSigmoid", "MinMax", or "MaxAbs". Defaults to "zScore"

unit_int

Boolean whether to rescale into unit interval [0,1] after applying normalisation method. Defaults to FALSE

clust_method

character specifying the hierarchical clustering method to use if type = "matrix" or if type = "cor". Defaults to "average"

cor_method

character specifying the correlation method to use if type = "cor". Defaults to "pearson"

feature_names

character vector denoting the name of the features to plot if type = "violin". Defaults to NULL

...

Arguments to be passed to ggplot2::geom_bar if type = "quality", ggplot2::geom_raster if type = "matrix", ggplot2::geom_raster if type = "cor", or ggplot2::geom_point if type = "violin"

Value

object of class ggplot that contains the graphic

Author(s)

Trent Henderson

Produce a plot for a feature_projection object

Description

Produce a plot for a feature_projection object

Usage

## S3 method for class 'feature_projection'
plot(x, show_covariance = TRUE, ...)

Arguments

x

feature_projection object containing the two-dimensional embedding calculated by project

show_covariance

Boolean specifying whether covariance ellipses should be shown on the plot. Defaults to TRUE

...

Arguments to be passed to methods

Value

object of class ggplot that contains the graphic

Author(s)

Trent Henderson

Produce a plot for a interval_calculations object

Description

Produce a plot for a interval_calculations object

Usage

## S3 method for class 'interval_calculations'
plot(x, ...)

Arguments

x

interval_calculations object containing the summary calculated by interval

...

Arguments to be passed to methods

Value

object of class ggplot that contains the graphic

Author(s)

Trent Henderson

Project a feature matrix into a two-dimensional representation using PCA, MDS, t-SNE, or UMAP ready for plotting

Description

Project a feature matrix into a two-dimensional representation using PCA, MDS, t-SNE, or UMAP ready for plotting

Usage

project(
  data,
  norm_method = c("zScore", "Sigmoid", "RobustSigmoid", "MinMax", "MaxAbs"),
  unit_int = FALSE,
  low_dim_method = c("PCA", "tSNE", "ClassicalMDS", "KruskalMDS", "SammonMDS", "UMAP"),
  na_removal = c("feature", "sample"),
  seed = 123,
  ...
)

reduce_dims(
  data,
  norm_method = c("zScore", "Sigmoid", "RobustSigmoid", "MinMax", "MaxAbs"),
  unit_int = FALSE,
  low_dim_method = c("PCA", "tSNE", "ClassicalMDS", "KruskalMDS", "SammonMDS", "UMAP"),
  na_removal = c("feature", "sample"),
  seed = 123,
  ...
)

Arguments

data

feature_calculations object containing the raw feature matrix produced by theft::calculate_features

norm_method

character denoting the rescaling/normalising method to apply. Can be one of "zScore", "Sigmoid", "RobustSigmoid", "MinMax", or "MaxAbs". Defaults to "zScore"

unit_int

Boolean whether to rescale into unit interval [0,1] after applying normalisation method. Defaults to FALSE

low_dim_method

character specifying the low dimensional embedding method to use. Can be one of "PCA", "tSNE", "ClassicalMDS", "KruskalMDS", "SammonMDS", or "UMAP". Defaults to "PCA"

na_removal

seed

integer to fix R's random number generator to ensure reproducibility. Defaults to 123

...

arguments to be passed to stats::prcomp or Rtsne::Rtsne, stats::cmdscale, MASS::isoMDS, MASS::sammon, or umap::umap depending on selection in low_dim_method

Value

object of class feature_projection which is a named list containing the feature_calculations data supplied to the function, the wide matrix of filtered data, a tidy data.frame of the projected 2-D data, and the model fit object

Author(s)

Trent Henderson

Examples



library(theft)

features <- theft::calculate_features(theft::simData,
  feature_set = "catch22")

pca <- project(features,
  norm_method = "zScore",
  low_dim_method = "PCA")

Helper function to create a resampled dataset

Description

Helper function to create a resampled dataset

Usage

resample_data(data, train_rows, test_rows, train_groups, test_groups, seed)

Arguments

data

data.frame containing time-series features

train_rows

integer denoting the number of cases in the train set

test_rows

integer denoting the number of cases in the test set

train_groups

data.frame containing proportions of each class in original train split

test_groups

data.frame containing proportions of each class in original test split

seed

integer denoting fixed value for R's pseudorandom number generator

Value

list containing new train and test data

Author(s)

Trent Henderson

Calculate z-score for all columns in a dataset using train set central tendency and spread

Description

Calculate z-score for all columns in a dataset using train set central tendency and spread

Usage

rescale_zscore(data, rescalers)

Arguments

data

data.frame containing data to normalise

rescalers

list containing central tendency and spread values for the train set

Value

data.frame of rescaled data

Author(s)

Trent Henderson

Helper function to select only the relevant columns for statistical testing

Description

Helper function to select only the relevant columns for statistical testing

Usage

select_stat_cols(data, by_set, metric, hypothesis)

Arguments

data

data.frame of classification accuracy results

by_set

Boolean specifying whether you want to compare feature sets (if TRUE) or individual features (if FALSE).

metric

character denoting the classification performance metric to use in statistical testing. Can be one of "accuracy", "precision", "recall", "f1". Defaults to "accuracy"

hypothesis

Value

object of class data.frame

Author(s)

Trent Henderson

Use a cross validated penalized maximum likelihood generalized linear model to perform feature selection

Description

Use a cross validated penalized maximum likelihood generalized linear model to perform feature selection

Usage

shrink(data, threshold = c("one", "all"), plot = FALSE, ...)

Arguments

data

feature_calculations object containing the raw feature matrix produced by theft::calculate_features

threshold

character denoting whether to retain features that have at least one non-zero coefficient "one" across all group levels or features that have non-zero coefficients across all group levels "all". Applicable to multinomial case only. Defaults to "one" for less aggressive filtering

plot

Boolean whether to draw the misclassification error lambda plot for a cv.glmnet object. Defaults to FALSE

...

arguments to be passed to glmnet::cv.glmnet

Value

feature_calculations object containing a data frame of the reduced feature set

Author(s)

Trent Henderson

Examples


library(theft)

features <- theft::calculate_features(theft::simData,
  feature_set = "catch22")

best_features <- shrink(features)

Calculate p-values for feature sets or features relative to an empirical null or each other using resampled t-tests

Description

Calculate p-values for feature sets or features relative to an empirical null or each other using resampled t-tests

Usage

stat_test(
  data,
  iter_data,
  row_id,
  by_set = FALSE,
  hypothesis,
  metric,
  train_test_sizes,
  n_resamples
)

Arguments

data

data.frame of raw classification accuracy results

iter_data

data.frame containing the values to iterate over for seed and either feature name or set name

row_id

integer denoting the row ID for iter_data to filter to

by_set

Boolean specifying whether you want to compare feature sets (if TRUE) or individual features (if FALSE).

hypothesis

metric

character denoting the classification performance metric to use in statistical testing. Can be one of "accuracy", "precision", "recall", "f1". Defaults to "accuracy"

train_test_sizes

integer vector containing the train and test set sample sizes

n_resamples

integer denoting the number of resamples that were calculated

Value

object of class data.frame

Author(s)

Trent Henderson

Analyse and Interpret Time Series Features

Description

Analyse and Interpret Time Series Features