An Efficient and Fast Parzen-Window Density Based Clustering Method for Large Data Sets

Density based clustering technique like DBSCAN finds arbitrary shaped clusters along with noisy outliers. DBSCAN finds the density at a point by counting the number of points falling in a sphere of radius epsi and it has a time complexity of O(n²). Hence it is not suitable for large data sets. The proposed method in this paper is an efficient and fast Parzen-Window density based clustering method which uses (i) prototypes to reduce the computational burden, (ii) a smooth kernel function to estimate density at a point and hence the estimated density is also varies smoothly. Enriched prototypes are derived using counted leaders method. These are used with a special form of the Gaussian kernel function which is radially symmetrical and hence the function can be completely specified by a variance parameter only. The proposed method is experimentally compared with DBSCAN which shows that it is a suitable method for large data sets.