rk.tuning: Advanced Hyperparameter Tuning and ML Workflows for RKWard

A Zero-Code GUI Wrapper for the tidymodels Ecosystem bringing Advanced Machine Learning to RKWard

rk.tuning bridges the gap between state-of-the-art machine learning engineering and practical Business Analytics. Designed with modern data science workflows in mind, this plugin provides a point-and-click interface to build strict, leak-proof predictive pipelines. Implement feature engineering, cross-validation, grid search tuning, and final model evaluation—all without writing a single line of R code.

Inspired by the philosophy of Business Analytics: Communicating with Numbers (Jaggia et al.), this tool ensures your models are statistically rigorous while keeping the outputs visual, interpretable, and business-ready.

🌟 Key Features

What's New in Version 0.0.1

⚙️ Prerequisites

You must have RKWard installed along with the following R packages. The plugin will automatically attempt to load them when required:

install.packages(c("tidymodels", "rsample", "recipes", "parsnip", "tune", "workflows", "yardstick", "ggplot2", "ranger", "xgboost", "kernlab", "themis"))

🚀 Installation

You can install this plugin directly from GitHub using remotes or devtools inside your RKWard console:

# Install the plugin
library("remotes")
# install_github("AlfCano/rk.tuning")
install.packages("rk.tuning", repos = c('https://alfcano.r-universe.dev', 'https://cloud.r-project.org'))

Once installed, restart RKWard to ensure the new menu entries and translations appear correctly.

🛠️ Usage Workflow (The 4 Modules)

This plugin breaks down the complex tidymodels framework into 4 digestible, logical steps:

1. Resample Data (Validation Strategy)

Define how you want to test your model's robustness. Create a 10-Fold Cross-Validation object (e.g., cv_folds) to ensure your model trains and tests on different slices of data, preventing overfitting.

2. Build Recipe (Feature Engineering)

Define your Target (Y) and Predictors (X). Add preprocessing steps sequentially: impute missing values, normalize numerical data, and balance classes. Save this pipeline as rec_obj.

3. Model Spec (Algorithms & Tuning)

Choose your algorithm (e.g., Random Forest). Check the boxes for the parameters you want the computer to test blindly (e.g., tune mtry and min_n), and define the grid size. Save as model_spec.

4. Execute & Finalize Model

The magic happens here. Drag and drop your original data, your cv_folds, your rec_obj, and your model_spec. Hit Submit. The plugin will search for the best combination, plot the results, and output a final_model ready for production.

📊 Real-World Case Study: Automated Optimization

Let's test our tuning pipeline using a classic dataset to see how it automatically finds the best configuration without human bias.

Step 1: Resampling

  1. Go to Analysis > Predictive Analytics > Advanced Tuning > 1. Resample.
  2. Select iris as your dataset.
  3. Choose K-Fold Cross-Validation (10 folds). Save as cv_folds. Hit Submit.

Step 2: The Recipe

  1. Go to 2. Build Recipe.
  2. Training Dataset: iris. Target Variable: Species.
  3. Check the box to Normalize all numeric predictors (crucial for good ML performance). Save as rec_obj. Hit Submit.

Step 3: The Model Spec

  1. Go to 3. Model Spec.
  2. Algorithm: Random Forest (ranger). Mode: Classification.
  3. Check the boxes to tune mtry (Randomly Selected Predictors) and min_n (Minimal Node Size). Set Grid Size to 10. Save as model_spec. Hit Submit.

Step 4: The Execution

  1. Go to 4. Execute & Finalize Model.
  2. Insert iris, cv_folds, rec_obj, and model_spec into their respective slots.
  3. Under Plot Options, title the plot "Random Forest Optimization".
  4. Hit Submit.

🗣️ Communicating with Numbers (Translating Output to Business)

True to the Jaggia philosophy, we don't hand management a report full of code. We use the tables and charts generated by RKWard to speak in terms of optimization and operational readiness.

Here is how you "pitch" the clean HTML reports RKWard just gave you:

1. Interpreting the Hyperparameter Grid Plot: Look at the generated scatterplot showing the different combinations tested by the algorithm.

"To ensure we don't overfit our model to past data, we ran a blind simulation of 10 different configurations using a 10-Fold Cross-Validation strategy. The visual grid shows that our algorithm peaks in accuracy when we restrict the tree to evaluating only 2 random variables at a time (mtry=2), and stopping the tree branching at groups of 35 (min_n=35)."

2. Interpreting the Accuracy & Brier Score: Look at the Top 5 Hyperparameter Combinations and the Workflow [trained] summary.

"By systematically tuning these parameters, our winning model achieved a verified Accuracy of 97% on unseen data. Furthermore, our Out-Of-Bag Brier Score is extremely low (0.034), meaning that when the algorithm predicts a category, it does so with immense statistical confidence, rarely guessing."

3. The Production Pipeline (The Recipe):

"Most importantly, we didn't just train a mathematical algorithm; we built a Production Workflow. If we feed new, raw customer data into this system tomorrow, the model automatically knows how to normalize and prepare the data before predicting. It is 100% ready for deployment."

🌍 Internationalization (i18n)

The graphical interface automatically adapts to your RKWard language settings. Currently supported languages:

📝 License and Author

Author: Alfonso Cano (@AlfCano)
Email: alfonso.cano@correo.buap.mx

This project is licensed under the GPL (>= 3) License.