maxwailab 1.0.4

Bootstrap-based model stability and supervised binning toolkit

Bootstrap ML Diagnostics

A lightweight toolkit for statistically robust model diagnostics using bootstrap resampling, with utilities for:

• supervised tree binning
• bootstrap-based feature selection
• model stability analysis
• hyperparameter sensitivity analysis

The library focuses on reducing overfitting and improving model interpretability through bootstrap distributions rather than single-point estimates.

---

Installation

pip install maxwailab 

or

pip install git+https://github.com/MaxWienandts/maxailab.git

---

Core Philosophy

Most ML workflows rely on single train/validation splits.

This library instead uses bootstrap resampling to estimate:

• performance distributions
• feature selection stability
• hyperparameter robustness

Benefits:

• reduces variance from a single split
• identifies unstable variables
• provides confidence intervals for model performance

---

Workflow Overview

Typical modeling workflow using this library:

1️⃣ Supervised binning (optional)

tree_supervised_binning
bootstrap_tree_binning_auc_analysis


2️⃣ Feature selection

bootstrap_lightgbm_forward_selection


3️⃣ Diagnostics

performance_forward_selection_boxplot
variable_frequency_forward_selection


4️⃣ Extract best variables

top_k_forward_selection_variables
top_k_variables_by_forward_selection_boxplot


5️⃣ Hyperparameter analysis

lightgbm_hyperparameter_auc_curve_bootstrap

---

Example Workflow

import maxwailab

# --------------------------------
# Forward Selection with Bootstrap
# --------------------------------

result_bootstrap = maxwailab.bootstrap_lightgbm_forward_selection(
    df=data,
    target="target",
    n_bootstrap=30,
    n_max_variables=15,
    metric_to_optimize="auc_roc",
    hyperparameters=lgb_params
)

---

Analyze performance stability

bml.performance_forward_selection_boxplot(
    result_bootstrap["auc_roc"],
    "AUC"
)

This visualizes how performance behaves as variables are added.

---

Variable selection stability

maxwailab.variable_frequency_forward_selection(
    result_bootstrap["variables"],
    n_bootstraps=30
)

Heatmap showing how frequently variables appear in models of different sizes.

---

Extract best variables

Based on selection frequency

maxwailab.top_k_forward_selection_variables(
    result_bootstrap["variables"],
    n_bootstraps=30,
    k=10
)

Based on best model performance

variables, auc = maxwailab.top_k_variables_by_forward_selection_boxplot(
    result_bootstrap,
    k=6,
    metric="auc_roc"
)

---

Tree-based Supervised Binning

Supervised binning using decision trees.

from maxwailab import tree_supervised_binning

tree_supervised_binning(
    df=data,
    feature="age",
    target="target",
    max_leaf_nodes=5
)

---

Bootstrap Binning Stability

bootstrap_tree_binning_auc_analysis(
    df_train,
    df_val,
    feature="age",
    target="target"
)

Evaluates how binning performance varies across bootstrap samples.

---

Hyperparameter Sensitivity Analysis

Evaluate how model performance reacts to hyperparameter changes.

lightgbm_hyperparameter_auc_curve_bootstrap(
    X_train,
    y_train,
    X_val,
    y_val,
    hyperparameters=lgb_params,
    hyperparameter_name="num_leaves",
    hyperparameter_values=[5,10,20,40],
    n_bootstrap=50
)

Bootstrap is applied only to the training set while keeping validation fixed (out-of-time).

---

Example Output

The library produces:

• performance distributions
• boxplots
• stability heatmaps
• hyperparameter sensitivity curves

These diagnostics help detect:

• overfitting
• unstable features
• fragile hyperparameters

---

Module Structure

maxwailab
│
├── binning
│   ├── tree_supervised_binning
│   └── bootstrap_tree_binning_auc_analysis
│
├── feature_selection
│   ├── bootstrap_lightgbm_forward_selection
│   ├── performance_forward_selection_boxplot
│   ├── variable_frequency_forward_selection
│   ├── top_k_forward_selection_variables
│   └── top_k_variables_by_forward_selection_boxplot
│
└── hyperparameter_analysis
    └── lightgbm_hyperparameter_auc_curve_bootstrap

---

When to Use This Library

This library is particularly useful for:

• credit risk models
• tabular ML problems
• high-stakes predictive modeling
• interpretable ML workflows

---

License

MIT License