Fig. 2

Adjusted Rand indices (ARIs) for 100 replicates of synthetic data under seven mixture scenarios with varying noise. The first column (panels x.1) shows the performance of medoidshift without a kernel function; the second column (panels x.2) show the performance of using the negative exponential kernel function; and the third column (panels x.3) is our new 2-stage medoidshift clustering method. Each row has increasing noise; the first row (panels 1.y) has no noise, the second row has σ=0.05 noise added, the third row has σ=0.1 noise added, the fourth row has σ=0.15 noise added, and the fifth row has σ=0.2 noise added