Explaining outliers - Isolation forest with shapley interactions (SHAP-IQ)

This repository contains the code for the experiments for the paper: Explaining Outliers using Isolation Forest and Shapley Interactions (Visser et al. 2025). This paper has been accepted for publication, and will be presented, at ESANN 2025.

Use method

Method has been added to the SHAP-IQ package which can be installed through pip:

pip install shapiq

Below you can see a basic usage example to generate explanations and visualize them. Check the shapiq documentation for more details.

from sklearn.ensemble import IsolationForest
import shapiq

# Train isolation forest
clf = IsolationForest(n_estimators=200, contamination='auto', random_state=None, bootstrap=False)
clf.fit(X_train)

# Create shapiq treeexplainer for model
explainer = shapiq.TreeExplainer(model=clf, index="k-SII", min_order=1, max_order=3)

# Generate feature interaction explanations
x = X_test[0]
interaction_values = explainer.explain(x)

# Visualize explanations
shapiq.network_plot(
    first_order_values=interaction_values.get_n_order_values(1),
    second_order_values=interaction_values.get_n_order_values(2),
    feature_names=column_names # Where column names are a list of column names used for printing
)

Experiments

In order to run the paper experiments follow the instructions below.

Setup, explanation generation, and evaluation

• install requirements (full environments available in requirements_full.txt and requirements_full_conda.txt)
• run experiments for synthetic data (see below) and real-world dataset (experiment_realworld.py)
• run evaluation scripts for synthetic (synthetic_evaluation.py) and real-world (glass_evaluation.py) data experiments.

Run & create synthetic experiment

• python experiment.py diagonal_feature --data_type "diagonal_feature" --n_trials 10 --results_folder "./results"
• python experiment.py single_outlier_quandrant --data_type "single_outlier_quandrant" --n_trials 10 --results_folder "./results"