Rank features by their association with a target variable

For supervised problems, analyze_target() measures how strongly each column is associated with a chosen target column and returns a ranked table. The measure adapts to the variable types so numeric and categorical predictors are comparable on a common 0-1 scale:

Usage

analyze_target(data, target, types = NULL, max_levels = 50)

Arguments

data

A data frame.

target

Name of the target column (a single string).

types

Optional named vector of column types (as from infer_column_types()); computed if not supplied.

max_levels

Categorical variables with more distinct levels than this are skipped (reported with NA strength). Default 50.

Value

A data frame, sorted by descending strength, with columns feature, type, method, strength (0-1), p_value and n (the number of complete observations used).

Details

• numeric target and numeric feature - absolute Pearson correlation;

• numeric with categorical (either role) - the correlation ratio (eta), the square root of the ANOVA between-group variance proportion;

• categorical target and categorical feature - Cramer's V.

This gives a quick, model-free view of which variables are most predictive before any modelling.

See also

correlation_analysis(), categorical_association()

Examples

df <- data.frame(
  y = c(1, 2, 3, 4, 5, 6),
  x_num = c(1.1, 1.9, 3.2, 3.8, 5.1, 6.2),
  x_cat = factor(c("a", "a", "b", "b", "c", "c"))
)
analyze_target(df, "y")
#>   feature        type  method  strength      p_value n
#> 1   x_num     numeric pearson 0.9965033 1.831904e-05 6
#> 2   x_cat categorical     eta 0.9561829 2.509457e-02 6