Normalized RandomWalk returning k(G,G) = -1 about grakel HOT 2 CLOSED

You should try a smaller lambda.
Theoretically geometric random walk converges if 1/lamda is bigger than the maximum eigenvalue of the kronecker product of the adjacency matrices of g1, g2. Only if the kernel converges it is psd. In another case the kernel value is not of a valid kernel and some other implementation set its value to zero.

A rule of thumb as noted by Viswanathan is setting lamda to the largest power of 10 which is smaller than 1/d^2, where d being the largest degree in the dataset.

B = [[0, 1, 1, 1, 1],
     [1, 0, 1, 1, 1],
     [1, 1, 0, 1, 1],
     [1, 1, 1, 0, 1],
     [1, 1, 1, 1, 0]]

rw_kernel_bl = RandomWalk(method_type="baseline", lamda=0.01)
rw_kernel_f = RandomWalk(lamda=0.01)

print(rw_kernel_bl.fit_transform([[B]])) # [[29.76190476]] -> [[1.0]] after normalization
print(rw_kernel_f.fit_transform([[B]])) # [[29.76190476]] -> [[1.0]] after normalization

Please install the latest grakel as I spotted an incosistency with the baseline.
Thanks a lot for taking our library into consideration 👍

from grakel.

Got it. New version installed and example working. The only difference to your example is that the result of print(rw_kernel_f.fit_transform([[B]])) is [[0.04761905]].

Thank you very much for your support and for the library!

from grakel.