Comparing model explainability across algorithms — Random Forest, Logistic Regression, and XGBoost — using modelStudio and DALEX.
Training a model is the easy part. Explaining why it made a specific prediction — and whether that reasoning holds across different algorithms — is where real insight lives. This project builds 4 models across regression and classification tasks, then uses interactive explainability dashboards to compare how each one arrives at its predictions.
| # | Script | Model | Task | Dataset |
|---|---|---|---|---|
| 1 | 1, random forest - regression.R |
Random Forest | Regression | World Happiness |
| 2 | 2, logistic regression - classification.R |
Logistic Regression | Classification | Titanic |
| 3 | 3, XGBoost - regression.R |
XGBoost | Regression | mpg |
| 4 | 4, XGBoost - classification.R |
XGBoost | Classification | Bank Churners |
Each script follows the same flow: fit model → create DALEX explainer → launch interactive dashboard. The consistency makes it easy to compare how different algorithms explain the same type of problem.
- Break Down — feature contributions to individual predictions (local explainability)
- Shapley Values — global feature importance (which features matter most overall)
- Ceteris Paribus — how changing one feature affects the outcome (what-if analysis)
- Partial Dependence — marginal effect of each feature across the dataset
- Random Forest distributes importance more evenly across features — no single feature dominates
- XGBoost concentrates importance in fewer features — more aggressive feature selection
- Logistic Regression provides coefficient-based explanations — easy to interpret but less flexible with non-linear relationships
git clone https://github.com/wsamuelw/model-studio.git
cd model-studioinstall.packages(c("DALEX", "modelStudio", "ranger", "tidymodels", "tidyverse"))
source("1, random forest - regression.R")Each script launches an interactive dashboard in your browser.
| Dataset | Source | Records | Task |
|---|---|---|---|
| World Happiness | Kaggle | 150+ countries | Predict happiness score |
| Titanic | DALEX::titanic_imputed |
2,207 | Predict survival |
mpg |
ggplot2::mpg (built-in) |
234 | Predict highway fuel economy |
| Bank Churners | data/bank_churners.csv (included) |
10,127 | Predict customer churn |
- DALEX — model-agnostic explainability framework
- modelStudio — interactive explainability dashboards
- ranger — fast Random Forest implementation
- tidymodels — unified modelling framework (used for XGBoost)
- tidyverse — data wrangling and visualisation
Explainer — wraps any model into a standardised format for explainability:
explainer <- DALEX::explain(
model = my_model,
data = my_data,
y = my_data$target,
label = "My Model"
)Per-instance explanations — pass specific observations to see how the model reasons about individual cases:
modelStudio(explainer, new_observations = my_data[1:3, ])- modelStudio documentation
- DALEX: Explainers for Complex Models
- DrWhy AI — the Explainable AI family of R packages
MIT