Model Builder

Build, evaluate, and interpret statistical models - without writing a line of code

A guided wizard that fits 15 model types live in your R session - supervised and unsupervised. Validate with cross-validation, generate diagnostic plots, compare models side-by-side, and get AI interpretation. All from the AiR panel inside RStudio.

Get Started Free

Guided Workflow

Six steps from dataset to interpretation

The Model Builder walks you through every stage. Pick a dataset, select variables, configure your model, and get results - all in a structured flow that keeps you in control.

Select Dataset

Choose from any dataframe loaded in your R environment. AiR shows dimensions, column types, and a preview.

Variables & Interactions

Pick target and predictors (supervised) or just variables (unsupervised). Add interaction terms between any pair of predictors.

Choose Model

12 supervised models or 3 unsupervised methods. AiR filters by your task type and shows package availability.

Configure

Validation method, metrics, stepwise selection, class imbalance handling, diagnostic plots, and model-specific parameters.

Results & Compare

Metrics, coefficients, plots, confusion matrices. Compare multiple models side-by-side and highlight the best performer.

AI Interpretation

A plain-English summary that cites actual values from your model output. Key findings, significance, and next steps.

Model Types

15 models, from linear to ensemble to unsupervised

Cover the full spectrum of statistical and machine learning models. Each generates clean R code and structured JSON results.

Regression

Linear Regression

base R

The workhorse. Interpretable coefficients with p-values and R². Supports stepwise AIC/BIC.

Polynomial Regression

base R

Fits curved relationships using polynomial terms of your predictors.

Ridge Regression

glmnet

L2 regularization. Handles multicollinearity and prevents overfitting.

Lasso Regression

glmnet

L1 regularization. Built-in feature selection by shrinking coefficients to zero.

Elastic Net

glmnet

Combines L1 and L2 penalties. Configurable alpha for the mixing ratio.

Classification

Logistic Regression

base R

Binary classification with interpretable odds ratios and log-likelihood.

GLM (General)

base R

Gaussian, Poisson, Gamma, inverse Gaussian families with configurable link functions. Supports stepwise.

Naive Bayes

naivebayes

Fast probabilistic classifier. Works well with many features and small datasets.

K-Nearest Neighbors

class

Instance-based learning. Classifies by majority vote of K nearest data points.

Regression & Classification

Decision Tree

rpart

Recursive partitioning. Visual tree structure with pruning via complexity parameter.

Random Forest

randomForest

Ensemble of decision trees with OOB error, variable importance, and configurable ntree.

Support Vector Machine

e1071

Finds optimal separating hyperplane. Radial, linear, and polynomial kernels with tunable cost.

XGBoost

xgboost

Gradient boosted trees. Configurable rounds, depth, and learning rate. Top-tier predictive performance.

Unsupervised

PCA

base R

Principal component analysis with scree plot and biplot. Explore variance structure and reduce dimensions.

K-Means Clustering

base R

Partition data into K clusters. Elbow plot for optimal K and cluster visualization.

Hierarchical Clustering

base R

Agglomerative clustering with dendrogram. Multiple linkage methods and distance metrics.

Validation

Four ways to validate your model

Go beyond a single train/test split. Cross-validation gives you robust, reliable performance estimates with confidence intervals.

Train/Test Split

Default

Classic holdout. Split your data into training and test sets with a configurable ratio (50-90%). Fast and simple.

Best for: Quick iteration, large datasets, exploratory modeling.

K-Fold Cross-Validation

Recommended

Divide data into K folds, train on K-1, test on the held-out fold. Rotate K times. Reports mean ± SD across all folds.

Best for: Reliable estimates, moderate datasets. Default: 10 folds.

Repeated K-Fold CV

Robust

Run K-Fold CV multiple times with different random splits. Reduces variance in your performance estimates.

Best for: Final model evaluation, publication-grade results. Default: 3 repeats × 10 folds.

Leave-One-Out (LOOCV)

Exhaustive

Each observation is used as a test set exactly once. Maximum use of your data - ideal when every data point counts.

Best for: Small datasets (<200 rows). Warning shown for large datasets.

How K-Fold Cross-Validation Works

Fold 1

Test

Train

Fold 2

Train

Test

Train

Fold 3

Train

Test

Train

Fold 4

Train

Test

Train

Fold 5

Train

Test

Each fold takes a turn as the test set. Performance is averaged across all folds, giving you a mean \u00B1 SD estimate.

Advanced Features

Beyond fitting a model

Tools for the full modeling workflow - from feature engineering to model selection to diagnostics.

Model Comparison

Run multiple models on the same data and compare them in a side-by-side table. Best metric values are highlighted automatically.

Diagnostic Plots

Residuals vs Fitted, Q-Q plots, Predicted vs Actual, and ROC curves - rendered as inline images directly in the results panel.

Interaction Terms

Add interaction terms between any pair of predictors with a click. The formula builder handles backtick-escaping automatically.

Stepwise Selection

Forward, backward, or both - using AIC or BIC. Available for LM and GLM models. See which variables were kept or dropped.

Class Imbalance

Undersample majority or oversample minority class before training. Applied to training data only to prevent data leakage.

Unsupervised Mode

Toggle to unsupervised mode for PCA, K-Means, and hierarchical clustering. No target variable needed - just select numeric features.

Metrics

Choose exactly what to measure

Select the evaluation metrics that matter for your analysis. AiR calculates and displays only what you ask for.

Regression

R²

Proportion of variance explained. 1.0 = perfect fit, 0 = no better than the mean.

RMSE

Root mean squared error. In the same units as your target variable. Lower is better.

MAE

Mean absolute error. Less sensitive to outliers than RMSE. Lower is better.

Adjusted R²

R² penalized for number of predictors. Prevents overfitting from adding useless variables.

MAPE

Mean absolute percentage error. Intuitive percentage scale, but unstable near zero.

Classification

Accuracy

Percentage of correct predictions. Simple but can be misleading on imbalanced data.

Precision

Of all positive predictions, how many were actually positive. Matters when false positives are costly.

Recall

Of all actual positives, how many did the model catch. Matters when false negatives are costly.

F1 Score

Harmonic mean of precision and recall. A balanced single metric for classification.

AUC-ROC

Area under the ROC curve. Measures discrimination ability across all thresholds. Binary only.

Log Loss

Penalizes confident wrong predictions. Lower is better. Evaluates predicted probabilities.

Results

Structured output, not raw R console

Every model produces a clean results dashboard: metrics grid, coefficient tables, confusion matrices, and variable importance - formatted and readable.

Example: Regression Results (10-Fold CV)

Model fitted - lm on mtcars10-Fold CV

0.826

± 0.048

R²

3.214

± 0.891

RMSE

2.567

± 0.634

MAE

Example: Multi-Class Confusion Matrix

Predicted ↓

setosa

versicolor

virginica

setosa

versicolor

virginica

Multi-class confusion matrices render automatically for models with 3+ classes. Binary models show TP/FP/FN/TN.

AI Interpretation

Model Interpretation

Model Summary

Linear regression predicting mpg using cyl, hp, and wt on 32 observations with 10-fold cross-validation.

Key Findings

The model explains 82.6% \u00B1 4.8% of variance across folds. Weight has the strongest effect (-3.17 per 1000lbs, p<0.001). Horsepower is not significant (p=0.142) and may be removed.

Next Steps

Try removing hp (non-significant), compare with XGBoost or Random Forest using the comparison table, or add a wt:cyl interaction term.

The AI reads your actual model output - coefficients, p-values, metrics, variable importance - and produces a contextual interpretation. No generic summaries.

Resilience

When code fails, AiR fixes it automatically

Model code execution can fail for many reasons - missing packages, data type issues, convergence problems. AiR detects the failure, sends the code and error to AI for diagnosis, and retries with corrected code. You see results, not errors.

Generate R Code

Execute in R

Failed?

AI Diagnoses Error

Fix & Retry

Results

The auto-fix pass sends the full R code and complete console output to AI for diagnosis. It understands that datasets are pre-loaded in your R session and focuses on actual code bugs. If the fix succeeds, you see results with an “Auto-fixed” badge. If the issue is truly unfixable (e.g., missing data), you get a clear plain-English explanation instead of a cryptic R error.

Export

Every model produces reproducible R code

The exact code that trained your model is available as a clean, self-contained R script. Copy it, save it, share it - it runs identically every time.

model_lm_mtcars.RCopy to clipboard

# Model: Linear Regression on mtcars
# Target: mpg | Predictors: cyl, hp, wt
# Validation: 10-Fold Cross-Validation
# Generated by AiR

set.seed(42)
library(caret)

.air_data <- mtcars[, c("mpg", "cyl", "hp", "wt")]
.air_data <- na.omit(.air_data)

ctrl <- trainControl(method = "cv", number = 10,
                     savePredictions = "final")

model <- train(mpg ~ cyl + hp + wt,
               data = .air_data,
               method = "lm",
               trControl = ctrl)

# Results
print(model)
summary(model$finalModel)

Ready to build models in RStudio?

Install AiR, load a dataset, and start modeling in under a minute.

Get Started Free Back to Home