Computes consensus feature importance

Computes the consensus feature importance from features weight obtained with different integration methods (considering features importance for one latent component per integration method), or for different latent dimensions constructed by an integration method.

Usage

compute_consensus_importance(
  output_list,
  latent_dimensions,
  metric = "geometric",
  include_missing_features = FALSE
)

Arguments

output_list

List of integration methods output, each generated via the get_output() function, or a single integration method output (from get_output()).

latent_dimensions

Named list (if output_list is a list), where each element is a character giving the latent dimension to retain in the corresponding element of output_list (1 value). If output_list is a single output object, needs to be instead a character vector giving the latent dimensions to retain.

metric

Character, one of the metrics to use to compute the consensus score. Can be one of 'min', 'max', 'average', 'product', 'l2' (for L2-norm), 'geometric' (for geometric mean) or 'harmonic' (for harmonic mean). Default value is 'geometric'. Names must match those of output_list.

include_missing_features

Logical, whether features missing in some of the output should be included in the calculation (see Details). Default value is FALSE.

Value

A tibble giving the consensus importance of each feature.

Details

If include_missing_features is FALSE (default behaviour), and some features are present in the output of one integration method but not the other (e.g. because a different pre-filtering was applied to the input data of the two methods), these features will be ignored. This does not mean that features that were selected by one method but not the other are discarded; in that case the feature will be assigned a weight of 0 for the method that did not select it. This is the recommended behaviour, should only be changed in specific scenarios (e.g. to check whether using all features in a dataset vs doing a variance-based preselection affect which features are deemed most important). If include_missing_features is TRUE, missing features will be assigned a weight of 0. Note that the geometric and harmonic means only work for strictly positive values. Therefore, all importance scores of 0 are replaced with an offset when computing these metrics. The offset is calculated per dataset, and corresponds to the minimum non-null importance score observed across all features in the dataset (and across all latent dimensions), divided by 2. The calculation of the offset is done before removing missing features (if include_missing_features = FALSE) so that results are consistent between the two options for include_missing_features.