Semi-Supervised Learning via Clustering — tl

Train a supervised model with limited labels by first clustering the data and propagating labels within clusters.

Usage

tl_semisupervised(
  data,
  formula,
  labeled_indices,
  cluster_method = "kmeans",
  supervised_method = "tree",
  ...,
  cluster_args = list()
)

Arguments

data: A data frame
formula: Model formula. The response must be categorical: a factor, character or logical column, or an expression that computes a factor or character vector from a column, such as factor(am) or factor(mpg > 20). For an expression, each row's propagated label is written to that column as its value in a labelled row of the same class. An expression that then computes other labels, such as cut(mpg, 2), whose breaks follow the column's range, is refused.
labeled_indices: Indices of labeled observations
cluster_method: Clustering method for label propagation: "kmeans" (default), "pam", "clara" or "hclust", whose tree is cut at k
supervised_method: Supervised learning method for the final model (default: "tree", which handles any number of classes). "logistic" is binary-only and errors on a response with more than two levels.
...: Additional arguments for the supervised model
cluster_args: A named list of arguments for the clustering step, such as list(nstart = 5) for k-means. k is set from the labelled classes and cannot be given here.

Value

A tidylearn model object with additional class "tidylearn_semisupervised", trained on pseudo-labeled data. The model includes a semisupervised_info element with labeled_indices, cluster_model, label_mapping, and n_unlabelled_dropped, the number of rows left out for having no label to train on.

Details

Labels are propagated by majority vote within each cluster, so the response must be categorical. The rows are clustered on the predictors the formula names, into as many clusters as the labelled rows hold classes. A labelled row whose label is missing takes no part in the vote. Rows in a cluster where no labelled observation carries a label have no label to take, and labelled rows whose own label is missing have none either; both are left out of training, with a warning giving the counts. The pseudo-labels keep the response's level order, so the second level stays the positive class.

Examples

# \donttest{
# Use only 10% of labels
labeled_idx <- sample(nrow(iris), size = 15)
model <- tl_semisupervised(iris, Species ~ ., labeled_indices = labeled_idx,
  cluster_method = "kmeans",
  supervised_method = "tree"
)
# }