On the use of valued action profiles for relational multi-criteria clustering

Abstract

Clustering techniques aim at eliciting hidden structures of a dataset by partitioning it into groups of similar elements. We will focus on the special case of relational clustering, where the similarity is based on relations that exist between elements rather than on their respective intrinsic features. As will be shown, particular attention has to be paid when applying a relational approach to the context of multi-criteria decision making. This paper introduces the concept of valued action profiles, a formalism for handling elements that are defined by pairwise valued outranking relations. For our experimental study, we then integrate these profiles into an adapted k-means algorithm that returns a relational partition. Results on both artificial and real datasets show that the use of the proposed method leads to meaningful relational partitions.