Table 1 DRAGON Method

 1. Given a sample set χ _s (at the beginning χ _s  = χ), compute likelihood L (Eq. (12)).

 2. Until L ^* > L, remove one sample \( \widehat{\mathbf{x}}\in {\chi}_s \) (Eq. (18)), compute new likelihood L ^*, update χ _s and L.

 3. Find centroid μ _s  = E[χ _s ] and d _x  = δ(x, μ _s ) ∀ x ∈ χ.

 4. Partition {d _x } into two groups, for example using k-means algorithm (or divide into two groups based on their values). One of these groups will have lower d _x values (representing closeness to μ _s ) whereas the other will have higher d _x values (representing distance from μ _s ). Update χ _s by replacing it with the samples with the lower d _x values.

 5. If required repeat steps 3 and 4. Take out the cluster χ _s from χ. Update χ accordingly (the updated χ would contain all the samples except χ _s ; i.e. χ ∩ χ _s  = Φ).

 6. Repeat all the steps until all the possible clusters (or desired number of clusters) are obtained.