Entering factors for classification

[Nathaniel-Haines]

I have realized that when factors are entered into easy_glmnet as the outcome variable, the "Replicating Metrics" part of the model fitting just sits at 0% complete. That said, if the factors are converted to 0 and 1 (i.e. as.integer(outcome_variable)), then things work fine. It may make things easier for users if this conversion is done automatically?

[paulhendricks]

Here's a minrex for this issue as of 4573549. We can use this to write a unit test and submit a patch.

library(easyml) # https://github.com/CCS-Lab/easyml

# Load data
data("cocaine_dependence", package = "easyml")

cocaine_dependence$diagnosis <- factor(cocaine_dependence$diagnosis)

results <- easy_support_vector_machine(cocaine_dependence, "diagnosis",
                                       family = "binomial", preprocess = preprocess_scale,
                                       exclude_variables = c("subject"),
                                       categorical_variables = c("male"),
                                       n_samples = .n_samples, n_divisions = .n_divisions,
                                       n_iterations = .n_iterations, random_state = 12345, n_core = .n_core)

[1] "Generating predictions for a single train test split:"
  |==================================================| 100% elapsed =00s, remaining ~00s
 Show Traceback
 
 Rerun with Debug
 Error in stats::cor(y_true, y_pred) : 'x' must be numeric 
> 

[paulhendricks]

Proposal for solution within core:

  # Set dependent variable
  y <- set_dependent_variable(.data, dependent_variable)
  
  # Process dependent variable
  if (family == "binomial") {
    # Check that dependent variable has two classes
    if (length(unique(y)) != 2) {
      stop("Error! Dependent variable must have two classes!")
    }
    
    # Check if dependent variable is a factor 
    if (is.factor(y)) {
      y <- as.numeric(y) - 1
    }
  }
  
  # Capture dependent variable
  object[["y"]] <- y

Will test and merge if it works.

[paulhendricks]

See c9c9184.