irapy 0.1.2

Impact Range Assessment (IRA): An Interpretable Sensitivity Measure for Regression Modeling

irapy

irapy is a Python library for implementing single-execution Impact Range Assessment (IRA) and repeated IRA analyses.

Installation

pip install irapy

Requirements

irapy Requirements

irapy depends on the following Python libraries:

• NumPy
• pandas
• joblib

These dependencies are installed automatically when using pip install irapy.

irapy was developed using Python 3.12 (NumPy 2.2.5, pandas 2.3.3, joblib 1.5.3) and has been tested on Python environments including Python 3.9 (NumPy 1.26, pandas 2.0) and Python 3.12 (NumPy 2.2, pandas 2.3). Compatibility issues may occur when NumPy and pandas are installed in binary-incompatible combinations.

Usage

Run irapy

You can apply either a single-execution or a repeated IRA to a trained regression model using the corresponding dataset.

from irapy import single_ira, repeated_ira

# single-execution IRA
single_ira_result = single_ira(input_data=X, model=trained_model)
print(single_ira_result)

# repeated IRA
repeated_ira_result = repeated_ira(input_data=X, model=trained_model, n_repeats=50)
print(repeated_ira_result)

Arguments

• input_data: predictor pandas DataFrame (i.e., the training predictor matrix, often denoted as X_train; use the original, unscaled predictors when a scaler is applied); all predictors must be continuous numeric variables.
• model: trained model (object with '.predict()', tested with 'scikit-learn' models) or a callable function. A Pipeline object can also be used to combine a scaler and prediction model. When using a Pipeline that already includes a scaler, leave the scaler argument as None.
• scaler: fitted scaler (optional, e.g., 'StandardScaler' / 'MinMaxScaler' in Python).
• num_interpol: number of interpolated points (default: 100).
• num_background_samples: number of background observations (default: 200).
• random_state: seed for reproducibility (default: 42).
• sorted_output: whether to sort results by IRA value (default: False).
• n_repeats (repeated IRA only): number of repeated runs (default: 50).
• n_jobs (repeated IRA only): number of parallel jobs to run (default: 1). Parallel execution uses the default joblib backend. In notebook or limited-memory environments (e.g., Google Colab), large models or datasets may occasionally produce memory or temporary storage issues when n_jobs > 1.

Project Homepage

https://github.com/jyou2025/impact-range-assessment