The number of rejections for different graph kernels assuming a cut-off of α = 0.01.

The RW kernel discount was set to c ≈ 10−9 (varied a little between tests). The propagation used w = 0.0001 and tmax = 6. The number of WL iterations was set to h = 2. The WWL used a discount of λ = 0.1. The pyramid kernel used L = 10 and d = 6. Note that other hyperparameters were tested as well, giving very similar results. The SP kernel used a discretization of the continuous edge weights and node attributed, using rounding of 3 digits and 1 digit respectively. The MONK estimator used Q = 5 partitions.