A Clustering Approach using Weighted Similarity Majority Margins

We propose a meta-heuristic for clustering objects that are described on multiple incommensurable attributes of nominal, ordinal and/or cardinal type. Our approach makes use of an innovative bipolar-valued dual similarity-dissimilarity relation characterized by pairwise weighted majority margins of similar minus dissimilar attribute evaluations. The clustering is computed in two steps. First, an evolutionary algorithm searches for a suitable subset of maximal similarity cliques that will best serve as cluster cores. In a second step, we construct with a greedy heuristic, around these initial cluster cores, a corresponding final partition which best fits the given bipolar-valued similarity relation.