Machine Learning with Tidymodels

A comprehensive introduction to machine learning workflows using the tidymodels ecosystem in R, developed for the Machine Learning intensive course at LCGEJ (January 2024).

Overview

This project covers essential machine learning concepts and implementations using the tidymodels framework:

• Data preprocessing and tidying with dplyr and tidyr
• Train/test splitting and cross-validation strategies
• Feature engineering with recipes
• Regression and classification modeling
• Hyperparameter tuning (grid search and Bayesian optimization)
• Model evaluation and visualization
• Variable importance analysis

Data

The course utilizes multiple datasets for hands-on learning:

• Ames Housing Dataset: Housing prices in Ames, Iowa with ~2,900 observations and 80+ features including lot area, living area, year built, and neighborhood characteristics
• Microarray Expression Data: Gene expression dataset (microarray_expression_data.csv) with 5,537 genes and 40 experimental conditions for data tidying exercises
• Taxi Dataset: Used for classification tasks and demonstrating decision tree models

Course Materials

File	Description
00_package_installation.R	Installation script for required packages (tidyverse, tidymodels, easystats, etc.)
01_tidyng_data.Rmd	Data tidying fundamentals using dplyr and tidyr with gene expression data
03_test.Rmd	K-nearest neighbors example with hyperparameter tuning on Ames dataset
04_regression_ames.Rmd	Complete regression pipeline: linear models, workflows, cross-validation, and metrics
05_tidymodels.Rmd	Core tidymodels concepts: data splitting, stratification, and classification workflows

Key Concepts Covered

Data Preprocessing

• Data splitting with stratification
• Feature engineering with recipes
• Handling missing data and outliers
• Variable transformations (log, Yeo-Johnson, splines)
• Creating dummy variables

Model Building

• Linear regression with multiple predictors
• K-nearest neighbors (KNN) regression
• Decision trees for classification
• Model specification with parsnip
• Workflow creation combining recipes and models

Model Evaluation

• Cross-validation (v-fold, stratified)
• Performance metrics: RMSE, MAE, R²
• Coefficient visualization with dotwhisker plots
• Variable importance analysis
• QQ plots and diagnostic visualizations

Hyperparameter Tuning

• Grid search with tune_grid()
• Bayesian optimization with tune_bayes()
• Parameter space exploration
• Model selection strategies

Dependencies

# Core packages
install.packages(c(
  "tidyverse",      # Data manipulation and visualization
  "tidymodels",     # ML framework
  "easystats",      # Statistical analysis
  "performance"     # Model diagnostics
))

# Additional packages
install.packages(c(
  "rmdformats",     # R Markdown templates
  "dotwhisker",     # Coefficient plots
  "vip",            # Variable importance
  "DataExplorer",   # Automated EDA
  "kknn",           # KNN implementation
  "AmesHousing",    # Ames dataset
  "embed",          # Feature engineering
  "skimr",          # Data summaries
  "assertr"         # Data validation
))

Alternatively, run the installation script:

source("00_package_installation.R")

Usage

Running Individual Modules

Each R Markdown file can be executed independently:

# Data tidying tutorial
rmarkdown::render("01_tidyng_data.Rmd")

# Regression modeling
rmarkdown::render("04_regression_ames.Rmd")

# Tidymodels fundamentals
rmarkdown::render("05_tidymodels.Rmd")

Example Workflow

library(tidymodels)
library(tidyverse)

# 1. Load and split data
data(ames)
set.seed(123)
ames_split <- initial_split(ames, prop = 0.8, strata = Sale_Price)
ames_train <- training(ames_split)
ames_test <- testing(ames_split)

# 2. Create recipe and model
lm_rec <- recipe(Sale_Price ~ Lot_Area + Year_Built + Gr_Liv_Area, 
                 data = ames_train) %>%
  step_log(Sale_Price, base = 10) %>%
  step_normalize(all_numeric_predictors())

lm_spec <- linear_reg() %>% 
  set_engine("lm")

# 3. Build workflow and fit
ames_wf <- workflow() %>%
  add_recipe(lm_rec) %>%
  add_model(lm_spec)

ames_fit <- fit(ames_wf, data = ames_train)

# 4. Evaluate with cross-validation
ames_cv <- vfold_cv(ames_train, v = 10)
cv_results <- fit_resamples(ames_wf, resamples = ames_cv)
collect_metrics(cv_results)

Output

Each tutorial generates:

• HTML reports: Self-contained documents with code, visualizations, and explanations
• Model objects: Fitted models with estimated parameters
• Performance metrics: RMSE, MAE, R² for model evaluation
• Visualizations:
  • Coefficient plots with confidence intervals
  • Variable importance plots
  • Cross-validation performance
  • Model diagnostic plots

Author

Axel Rodriguez Perez
Genomic Sciences, UNAM · January 2024