Limits of Clustering Models Based on Distance Similarity: A Sensitivity Study of SDOclust

Abstract

Parameter tuning is a major bottleneck in machine learning, particularly in large-scale applications where it becomes costly or impractical. In clustering, SDOclust stands out for its ability to run in a wide variety of settings with its default configuration. It additionally shows a set of properties that are rarely achieved simultaneously: lightweight, model-based, robust to noise, capable of capturing non-convex structures, updatable, and amenable to batch-wise processing. The model is a low-density data abstraction, which makes it naturally interpretable.

This versatility motivates the following question: under which conditions does such a model remain reliable? We study the impact of hyperparameter adjustment in SDOclust with scenarios of variable dataset size, dimensionality, number of clusters, cluster overlap, cluster densities, non-globular shapes, outlier proportion, and outlier dispersion. The studied parameters are: model size, model connectivity rank, minimum cluster representativeness, and global-local weighting.

The analysis reveals that SDOclust is generally robust, particularly with regard to size, dimensionality, density differences, and noise, but it has specific and predictable failure points that can be mitigated with external adjustment. In particular, the number of clusters becomes critical at high granularity, this being the main condition for overriding defaults in favor of external adjustments. Overall, SDOclust shows that statistical distance-based models can operate with robust defaults in a broad envelope of data scenarios, while performance degrades predictably at the boundaries of that envelope. This work is a clear map of when SDOclust requires adjustment and how to configure it, providing grounded instructions for real deployment.