R/2_2_textTrainRegression.R
textTrainRegression.RdtextTrainRegression() trains word embeddings to a numeric or a factor variable.
textTrainRegression(
x,
y,
x_append = NULL,
append_first = FALSE,
cv_method = "validation_split",
outside_folds = 10,
inside_folds = 3/4,
strata = "y",
outside_strata = TRUE,
outside_breaks = 4,
inside_strata = TRUE,
inside_breaks = 4,
model = "regression",
eval_measure = "default",
save_aggregated_word_embedding = FALSE,
language_distribution = NULL,
language_distribution_min_words = 3,
preprocess_step_center = TRUE,
preprocess_step_scale = TRUE,
preprocess_PCA = NA,
penalty = 10^seq(-6, 6),
parameter_selection_method = "first",
mixture = c(0),
first_n_predictors = NA,
impute_missing = FALSE,
method_cor = "pearson",
model_description = "Consider writing a description of your model here",
multi_cores = "multi_cores_sys_default",
save_output = "all",
simulate.p.value = FALSE,
seed = 2020,
...
)Word embeddings from textEmbed (or textEmbedLayerAggregation). If several word embedding are provided in a list they will be concatenated.
Numeric variable to predict.
(optional) Variables to be appended after the word embeddings (x); if wanting to preappend them before the word embeddings use the option first = TRUE. If not wanting to train with word embeddings, set x = NULL (default = NULL).
(boolean) Option to add variables before or after all word embeddings (default = False).
(character) Cross-validation method to use within a pipeline of nested outer and inner loops of folds (see nested_cv in rsample). Default is using cv_folds in the outside folds and "validation_split" using rsample::validation_split in the inner loop to achieve a development and assessment set (note that for validation_split the inside_folds should be a proportion, e.g., inside_folds = 3/4); whereas "cv_folds" uses rsample::vfold_cv to achieve n-folds in both the outer and inner loops.
(numeric) Number of folds for the outer folds (default = 10).
(numeric) The proportion of data to be used for modeling/analysis; (default proportion = 3/4). For more information see validation_split in rsample.
(string or tibble; default "y") Variable to stratify according; if a string the variable needs to be in the training set - if you want to stratify according to another variable you can include it as a tibble (please note you can only add 1 variable to stratify according). Can set it to NULL.
(boolean) Whether to stratify the outside folds.
(numeric) The number of bins wanted to stratify a numeric stratification variable in the outer cross-validation loop (default = 4).
Whether to stratify the outside folds.
The number of bins wanted to stratify a numeric stratification variable in the inner cross-validation loop (default = 4).
Type of model. Default is "regression"; see also "logistic" and "multinomial" for classification.
(character) Type of evaluative measure to select models from. Default = "rmse" for regression and "bal_accuracy" for logistic. For regression use "rsq" or "rmse"; and for classification use "accuracy", "bal_accuracy", "sens", "spec", "precision", "kappa", "f_measure", or "roc_auc",(for more details see the yardstick package).
(boolean) If TRUE, the aggregated word embeddings (mean, min, and max) are saved for comparison with other language input when the model is applied to other types of data.
(Character column) If you provide the raw language data used for making the embeddings, the language distribution (i.e., a word and frequency table) will be saved to the model object. This enables calculating similarity scores when the model is being applied to new language domains. Note that this saves the individual words, which, if you are analyzing sensitive data, can be problematic from a privacy perspective; to some extent this can be mitigated by increasing the number of words needed to be saved.
(numeric) Minimum number a words need to occur in the data set to be saved to the language distribution.
(boolean) Normalizes dimensions to have a mean of zero; default is set to TRUE. For more info see (step_center in recipes).
(boolean) Normalize dimensions to have a standard deviation of one; default is set to TRUE. For more info see (step_scale in recipes).
Pre-processing threshold for PCA (to skip this step set it to NA). Can select amount of variance to retain (e.g., .90 or as a grid c(0.80, 0.90)); or number of components to select (e.g., 10). Default is "min_halving", which is a function that selects the number of PCA components based on number of participants and feature (word embedding dimensions) in the data. The formula is: preprocess_PCA = round(max(min(number_features/2), number_participants/2), min(50, number_features))).
(numeric) Hyper parameter that is tuned (default = 10^seq(-16,16)).
If several results are tied for different parameters (i.e., penalty or mixture), then select the "first" or the "median" order.
A number between 0 and 1 (inclusive) that reflects the proportion of L1 regularization (i.e. lasso) in the model (for more information see the linear_reg-function in the parsnip-package). When mixture = 1, it is a pure lasso model while mixture = 0 indicates that ridge regression is being used (specific engines only).
By default this setting is turned off (i.e., NA). To use this method, set it to the highest number of predictors you want to test. Then the X first dimensions are used in training, using a sequence from Kjell et al., 2019 paper in Psychological Methods. Adding 1, then multiplying by 1.3 and finally rounding to the nearest integer (e.g., 1, 3, 5, 8). This option is currently only possible for one embedding at the time.
Default FALSE (can be set to TRUE if something else than word_embeddings are trained).
Type of correlation used in evaluation (default "pearson"; can set to "spearman" or "kendall").
(character) Text to describe your model (optional; good when sharing the model with others).
If TRUE it enables the use of multiple cores if the computer system allows for it (i.e., only on unix, not windows). Hence it makes the analyses considerably faster to run. Default is "multi_cores_sys_default", where it automatically uses TRUE for Mac and Linux and FALSE for Windows.
(character) Option not to save all output; default = "all". see also "only_results" and "only_results_predictions".
(Boolean) From fisher.test: a logical indicating whether to compute p-values by Monte Carlo simulation, in larger than 2 × 2 tables.
(numeric) Set different seed (default = 2020).
For example settings in yardstick::accuracy to set event_level (e.g., event_level = "second").
A (one-sided) correlation test between predicted and observed values; tibble of predicted values (t-value, degree of freedom (df), p-value, alternative-hypothesis, confidence interval, correlation coefficient), as well as information about the model (preprossing_recipe, final_model and model_description).
By default, NAs are treated as follows: 1. rows with NAs in word embeddings are removed. 2. rows with NAs in y are removed 3. rows with NAs in x_append are removed; if impute_missing is set to TRUE, missing values will be imputed using k-nearest neighbours. When rows are omitted, the user will get a warning. The CV predictions will include NAs with the same length as the input.
# Examines how well the embeddings from the column "harmonytext" can
# predict the numerical values in the column "hilstotal".
if (FALSE) { # \dontrun{
trained_model <- textTrainRegression(
x = word_embeddings_4$texts$harmonytext,
y = Language_based_assessment_data_8$hilstotal,
multi_cores = FALSE # This is FALSE due to CRAN testing and Windows machines.
)
# Examine results (t-value, degree of freedom (df), p-value, alternative-hypothesis,
# confidence interval, correlation coefficient).
trained_model$results
} # }