DSets-DDC clustering

The paper proposes a new clustering algorithm DSets-DDC. The idea is to use the technique of cluster assignment in DDC for growing clusters. The cores of clusters are identified through Dominant Sets. Instead of preparing all clusters at once, a peeling - off strategy is adopted. Data model of graph representation and weighted edges of similarity is used. The similarity among objects is made free from effects of any control parameter using histogram equalization. Proposed method is an iterative method of detecting a dominant set, growing a cluster around it and peeling it off to have a reduced data set for next iteration. The proposal is for pattern recognition where irregular shaped clusters are common. Hence, experiments are performed over such popular datasets which have clusters of odd shapes and different sizes.