PoBOC: an Overlapping Clustering Algorithm. Application to Rule-Based Classification and Textual Data
Résumé
This paper presents the clustering algorithm PoBOC (Pole-Based Overlapping Clustering). It has two main characteristics: the number of final clusters is unknown a priori and PoBOC allows an object to belong to one or several clusters. Given a similarity matrix over a set of objects, PoBOC builds small and homogeneous sets of objects (the poles), and then it assigns the objects to the poles. The clustering method is evaluated on two different research areas. First, on the Rule-Based Learning (RBL) task: classification rules are generated by organizing the instances of a class so that each cluster is covered with a single rule ; PoBOC is compared with different clustering methods and usual classifiers, on traditional datasets from the UCI repository. Otherwise, we observe the behaviour of PoBOC on the structuring of textual data in a semantic way. The efficiency of the proposed method on the two applications leads to conclude that PoBOC is also a general algorithm.