An empirical study to determine the optimal k in Ek-NNclus method

Yiru Zhang (1) , Tassadit Bouadi (2) , Arnaud Martin (3)

1 IMSIA - UMR 9219 - Institut des Sciences de la mécanique et Applications industrielles
2 IRISA - Institut de Recherche en Informatique et Systèmes Aléatoires
3 DRUID - Declarative & Reliable management of Uncertain, user-generated Interlinked Data

Ek-NNclus is a clustering algorithm based on the evidential k-nearest-neighbor rule. It has the advantage that the number of clusters can be detected unlike a c-means for example. However, the parameter k has crucial influence on the clustering results, especially for the number of clusters and clustering quality. Thus, the determination of k is an important issue to optimize the use of the Ek-NNclus algorithm. The authors of Ek-NNclus only give a large interval of k, which is not precise enough for real applications. In traditional clustering algorithms such as c-means and c-medoïd, the determination of c is a real issue and some methods have been proposed in the literature and proved to be efficient. In this paper, we borrow some methods from c determination solutions and propose a k determination strategy based on an empirical study.

Mots clés: Ek-NNclus; k determination; clustering

Domaines: Intelligence artificielle [cs.AI]; Système multi-agents [cs.MA]; Réseaux sociaux et d'information [cs.SI]

Yiru Zhang, Tassadit Bouadi, Arnaud Martin. An empirical study to determine the optimal k in Ek-NNclus method. 5th International Conference on Beleif Functions (BELIEF2018), BFAS, Sep 2018, Compiègne, France. ⟨hal-01880400⟩