Stratified Features via Clustering

Create cluster-specific supervised models for heterogeneous data

Usage

tl_stratified_models(
  data,
  formula,
  cluster_method = "kmeans",
  k = 3,
  supervised_method = "tree",
  ...,
  cluster_args = list()
)

Arguments

data

A data frame

formula

Model formula

cluster_method

Clustering method: "kmeans" (default), "pam", "clara" or "hclust", whose tree is cut at k. Only k-means can assign new rows, so the others predict their training data alone.

k

Number of clusters

supervised_method

Supervised learning method (default: "tree", which handles both regression and classification). "linear" needs a numeric response and refuses a factor.

...

Additional arguments for the supervised models

cluster_args

A named list of arguments for the clustering step, such as list(nstart = 5) for k-means. Pass k as k.

Value

A list with class "tidylearn_stratified" containing:

cluster_model

The fitted clustering model.

clusters

The training rows' cluster assignments.

supervised_models

Named list of tidylearn models, one per cluster that holds more than one class.

single_class_clusters

Named character vector giving, for each cluster whose rows all hold one class, that class.

formula

The model formula.

data

The original training data.

Details

The rows are clustered on the predictors the formula names, and a model is fitted to each cluster. A cluster whose rows all hold one class has nothing for a classifier to separate, so it gets no model and its rows are predicted as that class.

Examples

# \donttest{
models <- tl_stratified_models(mtcars, mpg ~ ., cluster_method = "kmeans",
                                k = 3, supervised_method = "linear")
#> Note: Response 'mpg' has 6 unique numeric values. Treating as regression. Convert to factor for classification.
# }