Implementation-induced Inconsistency and Nondeterminism in Deterministic Clustering Algorithms

A deterministic clustering algorithm is designed to always produce the same clustering solution on a given input. Therefore, users of clustering implementations (toolkits) naturally assume that implementations of a deterministic clustering algorithm A have deterministic behavior, that is: (1) two different implementations I
₁
and I
₂
of A are interchangeable, producing the same clustering on a given input D, and (2) an implementation produces the same clustering solution when run repeatedly on D. We challenge these assumptions. Specifically, we analyze clustering behavior on 528 datasets, three deterministic algorithms (Affinity Propagation, DBSCAN, Hierarchical Agglomerative Clustering) and the deterministic portion of a fourth (K-means), as implemented in various toolkits; in total, we examined 13 algorithm-toolkit combinations. We found that different implementations of deterministic clustering algorithms make different choices, e.g., default parameter settings, noise insertion, input dataset characteristics. As a result, clustering solutions for a fixed algorithm-dataset combination can differ across runs (nondeterminism) and across toolkits (inconsistency). We expose several root causes of such behavior. We show that remedying these root causes improves determinism, increases consistency, and can even improve efficiency. Our approach and findings can benefit developers, testers, and users of clustering algorithms.