Sensitivity Analysis of SDOclust Hyperparameters

Context and methodology

This artifact was created for the paper "Limits of Clustering Models Based on Distance Similarity: A Sensitivity Study of SDOclust" (Khazari, Zseby & Iglesias, SISAP 2026). It studies under which conditions SDOclust — a lightweight, model-based, noise-robust clustering method — remains reliable, by analyzing the sensitivity of its main hyperparameters (model size k₀, connectivity rank χ, minimum representativeness e, and global-local weighting ζ) across controlled scenarios of dataset size, dimensionality, number of clusters, cluster overlap/density, non-globular shapes, and outlier proportion/dispersion. The artifact provides full reproducibility of this analysis.

Please cite our work as:

@inproceedings{KhazariZsebyIglesias2026SDOclustSensitivity,
  author    = {Khazari, Sabina and Zseby, Tanja and Iglesias, F{\'e}lix},
  title     = {Limits of Clustering Models Based on Distance Similarity: A Sensitivity Study of SDOclust},
  booktitle = {Proceedings of the 19th International Conference on Similarity Search and Applications (SISAP 2026)},
  year      = {2026},
  address   = {Brno, Czechia},
  month     = oct,
  note      = {To appear},
}

Technical details

The artifact provides code and results for the full sensitivity analysis:

• results.zip — results per analysis type (outs, clus, inspread, outspread, dims, factorial, size), each run producing ALL_combinations_LONG.csv (ARI, execution time, effective parameters) and args.json.
• sensitivity_analysis.py + configs/ — main runner and JSON configs defining each analysis type; run.sh executes the full suite.
• docker-reproducible-env.tar (recommended) plus Docker build files, for a ready-to-use reproducible environment.

Full setup and usage instructions are in the included README.md.