RandomWalk(Diffusion) Covariance Kernel for GPR

#1

Would it be possible to implement a Random Walk Covariance Kernel (a.k.a Diffusion Kernel) in the pymc3.gp.cov module? I have inferred a network among my predictors (X) and would like to implement something similar to, e.g. Equation 1 in Urry and Sollich, 2013 for my covariance function (Random Walk Kernels and Learning Curves for Gaussian Process Regression on Random Graphs).

This takes the form: Cov_func = a x exp(M x l), where M is a NxN covariance matrix of X (txN), with diagonal elements equal to the negative sum of each respective row: M_{ii} = -1* \sum_{j\neq i} M_{ij}. a and l are then hyperparameters of the covariance function.

#2

I ended up adding my own function to my pymc3/gp/cov.py file for this, in case anybody is interested. You will have to supply your own network matrix M (adjacency matrix of weights). Something like:

M = np.cov(X, rowvar=False, bias=True)
M = M - np.eye(len(M))*M.diagonal()
M = M - np.nansum(M,axis=0)*np.eye(len(M))

and then assign it as a variable when you call the function, along with hyperparameters sigmaf (signal variance) and ls (characteristic length-scale).

Σprior = pm.gp.cov.Diffusion(input_dim=X.shape[1],active_dims=np.arange(X.shape[1]),M=M,sigf=σf,ls=ℓ)

For the pymc3/gp/cov.py file:

class Diffusion(Covariance):
    R"""
    Random Walk (Diffusion Kernel) based on linear covariance.
    W. Gregory
    """
    def __init__(self, input_dim, M, sigf=None, ls=None, ls_inv=None, active_dims=None):
        super().__init__(input_dim, active_dims)
        self.ls = tt.as_tensor_variable(ls)
        self.sigf = tt.as_tensor_variable(sigf)
        self.M = tt.as_tensor_variable(M)
        
        
    def full(self, X, Xs=None):
        X, Xs = self._slice(X, Xs)
        if Xs is None:
            K = tt.dot(tt.dot(X,(self.sigf*tt.exp(self.M*self.ls))),tt.transpose(X)) #K(X,X) / K(X_*,X_*) - nxn(n*xn*)
        else:
            K = tt.dot(tt.dot(X,(self.sigf*tt.exp(self.M*self.ls))),tt.transpose(Xs)) #K(X,X_*) - nxn*
        
        K2 = tt.patternbroadcast(K, (False, False))
        return K2
3 Likes
#3

There is a mistake above. The exponential should be a matrix exponential:

i.e. tt.slinalg.expm(self.M*self.ls)

#4

Thanks for follow up on your own post! Would you be interested to check this into PyMC3?

#5

Sure

1 Like