In SomaLogic/SomaDataIO: Input/Output 'SomaScan' Data

library(dplyr)
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
options(width = 90)

Introduction

Most machine learning (ML) analyses require a random split of original data into training/test data sets, where the model is fit on the training data and performance is evaluated on the test data set. The split proportions can vary, though 80/20 training/test is common. It often depends on the number of available samples and the class distribution in the splits.

Among many alternatives, there are 3 common approaches, all are equally viable and depend on the analyst's weighing of pros/cons of each. I recommend one of these below:

1. base R data frame indexing with [sample()] and [
2. use dplyr::slice_sample() or dplyr::sample_frac() in combination with dplyr::anti_join()
3. use the rsample package (not demonstrated)

---

Original Raw Data

In most analyses, you typically start with a raw original data set that you must split randomly into training and test sets.

raw <- SomaDataIO::rn2col(head(mtcars, 10L), "CarModel") |>
  SomaDataIO::add_rowid("id") |> # set up identifier variable for the join()
  tibble::as_tibble()
raw

---

Option #1: sample()

n     <- nrow(raw)
idx   <- withr::with_seed(1, sample(1:n, floor(n / 2))) # sample random 50% (n = 5)
train <- raw[idx, ]
test  <- raw[-idx, ]
train

test

Option #2: anti_join()

# sample random 50% (n = 5)
train <- withr::with_seed(1, dplyr::slice_sample(raw, n = floor(n / 2)))

# or using `dplyr::sample_frac()`
# train <- withr::with_seed(1, dplyr::sample_frac(raw, size = 0.5))

# use anti_join() to get the sample setdiff
test <- dplyr::anti_join(raw, train, by = "id")
train

test

SomaLogic/SomaDataIO documentation built on Feb. 8, 2025, 12:19 p.m.