Explaining AutoClustering: Uncovering Meta-Feature Contribution in AutoML for Clustering

Matheus Camilo da Silva; Leonardo Arrighi; Ana Carolina Lorena; Sylvio Barbon Junior

cs.LGFeb 20, 2026

Explaining AutoClustering: Uncovering Meta-Feature Contribution in AutoML for Clustering

Authors: Matheus Camilo da Silva, Leonardo Arrighi, Ana Carolina Lorena, Sylvio Barbon Junior

Organizations: Department of Mathematics, Computer Science and Geosciences, University of Trieste, Piazzale Europa, 1, 34127 Trieste TS, Italy. · Division of Computer Science, Aeronautics Institute of Technology, Praça Marechal Eduardo Gomes, 50, 12228-900 São José dos Campos SP, Brazil.

Abstract

AutoClustering methods aim to automate unsupervised learning tasks, including algorithm selection (AS), hyperparameter optimization (HPO), and pipeline synthesis (PS), by often leveraging meta-learning over dataset meta-features. While these systems often achieve strong performance, their recommendations are often difficult to justify: the influence of dataset meta-features on algorithm and hyperparameter choices is typically not exposed, limiting reliability, bias diagnostics, and efficient meta-feature engineering. This limits reliability and diagnostic insight for further improvements. In this work, we investigate the explainability of the meta-models in AutoClustering. We first review 22 existing methods and organize their meta-features into a structured taxonomy. We then apply a global explainability technique (i.e., Decision Predicate Graphs) to assess feature importance within meta-models from selected frameworks. Finally, we use local explainability tools such as SHAP (SHapley Additive exPlanations) to analyse specific clustering decisions. Our findings highlight consistent patterns in meta-feature relevance, identify structural weaknesses in current meta-learning strategies that can distort recommendations, and provide actionable guidance for more interpretable Automated Machine Learning (AutoML) design. This study therefore offers a practical foundation for increasing decision transparency in unsupervised learning automation.

Figures & tables

Title	Reference
Ranking and selecting clustering algorithms using a meta-learning approach	de Souto et al. (2008)
Mining rules for the automatic selection process of clustering methods applied to cancer gene expression data	Nascimento et al. (2009)
An analysis of meta-learning techniques for ranking clustering algorithms applied to artificial data	Soares et al. (2009)
Clustering algorithm recommendation: a meta-learning approach	Ferrari and de Castro (2012)
Dealing with overlapping clustering: a constraint-based approach to algorithm selection	Adam and Blockeel (2015)
Clustering algorithm selection by meta-learning systems: a new distance-based problem characterization and ranking combination methods	Ferrari and De Castro (2015)

Table 1: Overview of 22 AutoClustering frameworks reviewed in this study.

Framework	Group	#Datasets	#Samples	#Dimensions	#Clusters
de Souto et al. (2008)	A	32	76.58	1383.38	3.00
Nascimento et al. (2009)	A	35	76.58	1383.38	3.00
Soares et al. (2009)	B	160	309.40	75.00	7.50
Ferrari and de Castro (2012)	C	30	1636.23	36.20	7.50
Adam and Blockeel (2015) (1)	C	14	522.93	11.07	3.92
Adam and Blockeel (2015) (2)	–	22	–	–	–

Table 2: Summary of dataset usage across AutoClustering frameworks. Dataset groups: A = Cancer gene expression microarray data, B = Synthetic datasets, C = General-purpose benchmark repositories, D = Clustering benchmarking collections. Note: Numerical columns for samples, dimensions, and clusters report average values per framework across the datasets used. Parentheses indicate different dataset settings within the same publication.

Figure 1: Heatmap showing the distribution of meta-feature families across various AutoClustering frameworks. The color intensity corresponds to the number of meta-features in each family for each framework. The frameworks are sorted ascending by publication date.

Panel of Figure 2 — Figure 2: Comparison of top and bottom meta-feature predicates per AutoClustering framework ( AutoClust , AutoCluster , ML2DAC , PoAC ) based on LRC. Top predicates (a) highlight features most embedded in the information flow, while bottom predicates (b) reflect less influential or potentially redundant aspects. Bars are color-coded by meta-feature category.

Figure 3: SHAP summary plot for ML2DAC showing the average contribution of each meta-feature to the model’s CVI recommendation. Colors denote the predicted CVI (CH, DBCV, SIL, COP).

Figure 4: Cohort-level SHAP summary plot analysing PoAC’s recommended clustering pipelines across 25 datasets. Each point represents a dataset/pipeline meta-feature value (color: red=high, blue=low) and its impact on the prediction (x-axis: SHAP value). Meta-features are ranked by mean absolute SHAP value.

Panel of Figure 5 — Figure 5: SHAP explanations for two PoAC-predicted pipelines on the thy dataset, representing low and high clustering quality.

Figure 6: Ablation experiment comparing the meta-feature extraction time across three different feature sets: (i) poac_full the complete set of 39 meta-features used by Poac; (ii) poac_dpg a reduced set of 24 meta-features with a minimum LRC value; (iii) poac_core a minimal set of the 10 most central meta-features based on LRC.

Meta-Feature Set	RMSE ↓	MAE ↓	R 2 ↑	Extraction Time
PoAC Full	0.0735	0.0521	0.8644	4h 54m
PoAC DPG	0.0782	0.0564	0.8465	4h 49m
PoAC CORE	0.0825	0.0600	0.8291	0h 7m

Table 3: Cross-validated performance of Random Forest regressors using different feature subsets. Lower values are better for RMSE and MAE; higher values are better for R 2 . The last column shows the total time required to extract each meta-feature set across all 6,000 datasets.

Observation	Why it matters
Decision influence is concentrated in a small subset of meta-features (DPG).	Large meta-feature sets often introduce redundancy and do not necessarily improve generalisation beyond a stable core set.
Internal-validation signals strongly shape recommendations (SHAP), indicating systematic preference effects.	Recommendations may inherit assumptions encoded by CVIs and landmarkers, which can diverge from user intent or domain semantics in unsupervised settings.
Global and local explanations provide complementary evidence (DPG vs. SHAP).	Global structure can highlight dominant decision logic, while local attributions justify individual recommendations and reveal regime-dependent feature reliance.
Explainability-guided pruning yields large efficiency gains with modest accuracy loss (ablation).	Meta-feature extraction can dominate end-to-end runtime; pruning based on relevance provides a principled route to faster, more deployable systems.

Table 4: Summary of findings for AutoClustering design.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Framework	Simple	Statistical	Info.-Theoretical	Complexity
Souto (2008)	3	4	0	0
Nascimento (2009)	3	4	0	0
Soares (2009)	0	9	0	0
Ferrari (2012)	5	3	2	0
Ferrari (2015)	0	19	0	0
CBOvalue (2015)	0	0	0	1

Appendix

Table 5: Meta-feature family usage across AutoClustering frameworks (2008–2024).

Framework	Predicate	LRC	LRC normalized	MF Category
autoclust	SIL > -0.24	33.0543	1.000	Landmarkers
autoclust	DBI <= 1.38	30.2646	0.7678	Landmarkers
autoclust	CH <= -12.63	29.9616	0.7426	Landmarkers
autoclust	CH > -12.63	29.1075	0.6715	Landmarkers
autoclust	COP > 0.36	27.9701	0.5768	Landmarkers
autoclust	DI > -0.44	27.6708	0.5519	Landmarkers

Appendix

Table 6: Top meta-feature predicates per AutoClustering framework based on original and normalized LRC.

Category	No.	Meta-feature
Simple	1	Number of instances
	2	Log number of instances
	3	Number of features
	4	Log number of features
	5	Dataset ratio
	6	Log dataset ratio

Appendix

Table 7: The implemented meta-features in AutoCluster.

Automated Machine Learning

Automated machine learning (AutoML) builds and tunes data-to-model pipelines with minimal manual effort, balancing predictive quality, compute, and deployment constraints.

See all 95 papers

Explainable cluster analysis: a bagging approach
Mar 20, 2026Federico Maria Quetti, Elena Ballante, Silvia Figini +1Clustering Unsupervised Clustering
Department of Mathematics, University of Pavia, Pavia, 27100, Italy. · Department of Social and Political Sciences, University of Pavia, Pavia, 27100, Italy. · Department of Economics and Management, University of Pavia, Pavia, 27100, Italy.
Interpretable Clustering: A Survey
Sep 1, 2024Lianyu Hu, Mudi Jiang, Junjie Dong +2Clustering Interpretable ML
College of Information Science and Engineering, Henan University of Technology, China · School of Software, Dalian University of Technology, China · Xinchang Power Supply Company, State Grid Corporation of China, China
Cluster LOCO: Feature Importance For Interpreting Clusters
Jun 12, 2026Claire M. He, Genevera I. AllenClustering
Department of Statistics · Columbia University · New York, NY 10027, USA

Mar 20, 2026J/K move · Enter open · S save

Explainable cluster analysis: a bagging approach

Federico Maria Quetti, Elena Ballante, Silvia Figini, Paolo Giudici

Department of Mathematics, University of Pavia, Pavia, 27100, Italy. · Department of Social and Political Sciences, University of Pavia, Pavia, 27100, Italy. · Department of Economics and Management, University of Pavia, Pavia, 27100, Italy.

Clustering Unsupervised Clustering

Open paper AskPDF arXiv

A major limitation of clustering approaches is their lack of explainability: methods rarely provide insight into which features drive the grouping of similar observations. To address this limitation, we propose an ensemble-based clustering framework that integrates bagging and feature dropout to generate feature importance scores, in analogy with feature importance mechanisms in supervised random forests. By leveraging multiple bootstrap resampling schemes and aggregating the resulting partitions, the method improves stability and robustness of the cluster definition, particularly in small-sample or noisy settings. Feature importance is assessed through an information-theoretic approach: at each step, the mutual information between each feature and the estimated cluster labels is computed and weighted by a measure of clustering validity to emphasize well-formed partitions, before being aggregated into a final score. The method outputs both a consensus partition and a corresponding measure of feature importance, enabling a unified interpretation of clustering structure and variable relevance. Its effectiveness is demonstrated on multiple simulated and real-world datasets.

Abstract

Automated Machine Learning

Explore similar work

Explainable cluster analysis: a bagging approach

Interpretable Clustering: A Survey

Cluster LOCO: Feature Importance For Interpreting Clusters