Prototype-based Clustering for Relational Data using Barycentric Coordinates

Data clustering is a very important and challenging task in Artificial Intelligence (AI) field with many applications such as bio-informatics, medical, enhancing recommendation engines or fraud detection. Among the different families of clustering algorithms, one of the most widely used is the prototype-based clustering, because of its simplicity and reasonable computational time. Prototype-based algorithms compute a compact model of the data structure in the form of a set of prototypes described in the same vectorial space as the data, each prototype representing a cluster. However, in real world applications we often face non-vectorial data that cannot be treated with most prototype-based approaches. Such non-vectorial data are usually represented in the form of relational data, i.e. data defined by their relation to each other (their dissimilarities). In this study, we propose a prototype-based clustering algorithm for relational data based on the Barycentric Coordinates formalism. We compared experimentally the quality of the proposed approach on artificial and real data-sets. The experiments show the high quality of the algorithm in terms of clustering results. We also showed that our approach is a significant improvement in terms of computational and memory complexity compared to the state-of-the-art approaches. We consider that these results are encouraging and pave the road to numerous applications in data clustering.