Constrained overlapping clusters: minimizing the negative effects of bridge-nodes

This paper presents a new approach to forming overlapping clusters of objects by balancing the effects of incompleteness, impurity and overlap. Incompleteness results from similar objects separated into different clusters while impurity arises when a cluster contains dissimilar objects. Overlap is caused by nodes that appear in more than one cluster. The key to balancing these effects is the identification of bridge-nodes. We show the limitations of traditional clustering algorithms in handling bridge nodes and demonstrate the intractability of minimizing all three effects. Approximation algorithms based on graph mincut and genetic algorithm are proposed to minimize these effects. Our results with real data sets show significant improvement over traditional methods with regard to incompleteness, impurity and overlap.