A Unified Machine Learning Interface for R

Overview

unifiedml provides a (work in progress) consistent, sklearn-like interface for various (any) machine learning models in R.

It eliminates the need to remember different function signatures across packages by automatically detecting the appropriate interface (formula vs matrix) and task type (regression vs classification).

Key Features

For now:

Automatic Task Detection: Automatically detects regression vs classification based on response variable type (numeric → regression, factor → classification)
Universal Interface: Works seamlessly with glmnet, randomForest, e1071::svm, and other popular ML packages with formula or matrix interface
Built-in Cross-Validation: Consistent cross_val_score() function with automatic metric selection
Model Interpretability: Numerical derivatives and statistical significance testing via summary()
Partial Dependence Plots: Visualize feature effects with plot()
Method Chaining: Clean, pipeable syntax with R6 classes

Installation

From CRAN:
```
install.packages("unifiedml")
```

From Github (development version):

# Install from GitHub (development version, for now)
devtools::install_github("Techtonique/unifiedml")

Quick Start

Regression Example

library(unifiedml)
library(glmnet)
# Prepare data
data(mtcars)
X <- as.matrix(mtcars[, -1])
y <- mtcars$mpg  # numeric → automatic regression
# Fit model
mod <- Model$new(glmnet::glmnet)
mod$fit(X, y, alpha = 0, lambda = 0.1)
# Make predictions
predictions <- mod$predict(X)
# Visualize partial dependence
mod$plot(feature = 1)
# Cross-validation (automatically uses RMSE for regression)
cv_scores <- cross_val_score(mod, X, y, cv = 5)
cat("Mean RMSE:", mean(cv_scores), "\n")

Classification Example

library(randomForest)
# Prepare data
data(iris)
X <- as.matrix(iris[, 1:4])
y <- iris$Species  # factor → automatic classification
# Fit model
mod <- Model$new(randomForest::randomForest)
mod$fit(X, y, ntree = 100)
# Make predictions
predictions <- mod$predict(X)
# Cross-validation (automatically uses accuracy for classification)
cv_scores <- cross_val_score(mod, X, y, cv = 5)
cat("Mean Accuracy:", mean(cv_scores), "\n")

Core Functionality

The Model Class

The Model R6 class provides a unified interface for any machine learning function:

# Create a model wrapper
mod <- Model$new(model_function)
# Fit the model (task type auto-detected from y)
mod$fit(X, y, ...)
# Make predictions
predictions <- mod$predict(X_new)
# Get interpretable summary
mod$summary(h = 0.01, alpha = 0.05)
# Visualize feature effects
mod$plot(feature = 1)
# Print model info
mod$print()

Cross-Validation

The cross_val_score() function provides consistent k-fold cross-validation:

See the vignettes.

Model Interpretability

The summary() (work in progress) method uses numerical derivatives to assess feature importance:

mod$summary()
# Output:
# Model Summary - Numerical Derivatives
# ======================================
# Task: regression
# Samples: 150 | Features: 4
# 
# Feature         Mean_Derivative  Std_Error  t_value  p_value  Significance
# Sepal.Length    0.523           0.042       12.45    < 0.001  ***
# Sepal.Width    -0.234           0.038       -6.16    < 0.001  ***
# ...

Automatic Task Detection

The package automatically determines the task type:

# Regression (numeric y)
y_reg <- c(1.2, 3.4, 5.6, ...)
mod$fit(X, y_reg)  # → task = "regression"
# Classification (factor y)
y_class <- factor(c("A", "B", "A", ...))
mod$fit(X, y_class)  # → task = "classification"

Advanced Features

Partial Dependence Plots

Work in progress.

# Visualize how feature j affects predictions
mod$plot(feature = 3, n_points = 100)

Model Cloning

# Create independent copy for parallel processing
mod_copy <- mod$clone_model()

Examples

See the package vignettes for comprehensive examples.

Why unifiedml?

Traditional R modeling requires remembering different interfaces:

# Different interfaces = cognitive overhead
glmnet(x = X, y = y, ...)               # matrix interface
randomForest(y ~ ., data = df, ...)     # formula interface
svm(x = X, y = y, ...)                  # matrix interface

With unifiedml, it's always the same:

# One interface to rule them all
Model$new(glmnet)$fit(X, y, ...)
Model$new(randomForest)$fit(X, y, ...)
Model$new(svm)$fit(X, y, ...)

Contributing

Contributions are welcome, feel free to submit a Pull Request.

License

The Clear BSD License - see LICENSE file for details.

Citation

If you use this package in your research, please cite:

@Manual{unifiedml,
  title = {unifiedml: Unified Machine Learning Interface for R},
  author = {T. Moudiki},
  year = {2025},
  note = {R package version 0.1.0}
}