Path-scan: A novel clustering algorithm based on core points and connexity
Résumé
A new clustering algorithm Path-scan aiming at discovering natural partitions is proposed. It is based on the idea that a (k,ɛ) coreset of the original data base represented by core and support patterns can be path-connected via a density differential approach. The Path-scan algorithm is structured in two main parts producing a connectivity matrix where partitions can be extracted at different levels of granularity. The first one aims to identify and select core and support points while the second one extracts connected components of core points and clusters with the help of support points. A simulation experiment based on synthetic and real world data sets was conducted to show the effectiveness of the proposed method.