clusterloco is a Python package for Cluster LOCO feature importance methods for clustering interpretability.
Cluster LOCO is a model-agnostic framework for quantifying feature importance in clustering. The package provides methods that evaluate how much removing a feature affects the generalizability and stability of a clustering solution, enabling feature-level interpretation for unsupervised learning workflows.
Our package can be installed directly from PyPI:
pip install clusterlocoor for additional experiment dependencies, use
pip install "clusterloco[experiments]"Clone the repository and install the package in editable mode:
git clone https://github.com/DataSlingers/ClusterLOCO.git
cd ClusterLOCO
pip install -e .To check that the package is correctly installed:
python -c "import clim; print(clim.__file__)"
python -c "from clim import ClusterLOCOMP; print('import ok')"To build the package:
python -m pip install build
python -m buildThis should create a source distribution and wheel in the dist/directory.
For experiment dependencies, such as anndata and scanpy, install:
pip install -e ".[experiments]"The core package requires Python 3.10 or higher. Core dependencies include:
numpy
scipy
pandas
scikit-learn
joblib
tqdm
matplotlib
seaborn
leidenalg
igraphOptional experiment dependencies include:
anndata
scanpyThe package offers Cluster LOCO via data splitting, via minipatches and with adaptive recursive trimming. Two example notebooks for running our models are available under the example folder with simulated data and a real application to PBMC 68k data. For the latter you will need to install anndata and scanpy.
Cluster LOCO Split is recommended for data with few features (less than 10 features).
from clim.data_splitting import Cluster_LOCO_Split
from clim.utils import hinge_errorBasic usage: for any sklearn clustering algorithm, default transfer classifier is RandomForestClassifier.
from sklearn.cluster import SpectralClustering
model = SpectralClustering(n_clusters=K)
feature_importance, feature_importance_se = Cluster_LOCO_Split(X_train, X_test, model=model, error_metric=hinge_error, use_proba=True, seed=42)Cluster LOCO-MP implements a minipatch ensemble version of Cluster LOCO. This approach is suited for large data.
from clim import ClusterLOCOMPBasic usage: for any sklearn clustering algorithm, first fit() the minipatch model, then compute the feature importance via score(). We recommend to use parallelization during model fitting but not during computing scores where the overhead can be consequential.
g = ClusterLOCOMP(base_clusterer = model, base_classifier = RandomForestClassifier(), K=3, B=500)
g.fit(X, standardize=False, alpha_N = 0.2, alpha_M = 0.2, parallel=par)
out = g.score(error_metric=hinge_error, agg='mean', proba_error=True, parallel_features=False) Cluster LOCO-RAMPART is a sped-up version of Cluster LOCO-MP based on adaptive recursive trimming of active feature set. We recommend using this with high-dimensional data.
Basic usage:
from clim import ClusterLOCO_RAMPART, RAMPART
from clim.utils import transform_scores_to_rankingRAMPART directly fits the model and computes the scores.
gen_fn = ClusterLOCO_RAMPART(base_clusterer=model, K=3, error_metric=hinge_error, parallel_MP=True,
parallel={"n_jobs_features": 3, "backend": "loky", "prefer": "processes", "verbose": 0},
standardize=False, alpha_N = 0.2, alpha_M = 0.2)
out = RAMPART(X, generalizability_fn=gen_fn, B=1000, ranking_fn=transform_scores_to_ranking, top_k=50)ClusterLOCO/
├── pyproject.toml
├── README.md
└── clim/
├── __init__.py
├── minipatches/
├── data_splitting/
├── models/
└── utils/
└── benchmarking/
└── simulations/
└── example/
└── paper_figures/
This repository additionally contains code to reproduce the results from our paper: the folder paper_figures contains the notebook to make the main figures from our paper.
If you use this package, please cite the corresponding Cluster LOCO paper.
@preprint{he2026clusterloco,
title={Cluster LOCO: Feature Importance for Interpreting Clusters},
author={He, Claire and Allen, Genevera},
url={https://arxiv.org/pdf/2606.14592},
year={2026}
}