y_train = y[index_train, :]
x_test = x[index_test, :]
y_test = y[index_test, :]
# gp
model.optimize(x_train, y_train)
model.predict(x_test)
# evaluation
predictive_class_rbf = np.sign(model.ym)
ACC_rbf = valid.ACC(predictive_class_rbf, y_test)
## DIFFUSION Kernel
# compute kernel matrix and initalize GP with precomputed kernel
model = pyGPs.GPC()
M1, M2 = graphUtil.formKernelMatrix(Matrix, index_train, index_test)
k = pyGPs.cov.Pre(M1, M2)
model.setPrior(kernel=k)
# if you only use precomputed kernel matrix, there is no training data needed,
# but you still need to specify x_train (due to general structure of pyGPs)
# e.g. you can use the following:
n = len(index_train)
x_train = np.zeros((n, 1))
# gp
model.optimize(x_train, y_train)
model.predict(x_test)
# evaluation
predictive_class_diff = np.sign(model.ym)
print('number of kernel iterations =', t)
Matrix = K[:,:,t]
# normalize kernel matrix (not useful for MUTAG)
# Matrix = graphUtil.normalizeKernel(Matrix)
# start cross-validation for this t
for index_train, index_test in valid.k_fold_index(N, K=10):
y_train = graph_label[index_train,:]
y_test = graph_label[index_test,:]
n1 = len(index_train)
n2 = len(index_test)
model = pyGPs.GPC()
M1,M2 = graphUtil.formKernelMatrix(Matrix, index_train, index_test)
k = pyGPs.cov.Pre(M1,M2)
model.setPrior(kernel=k)
# gp
x_train = np.zeros((n1,1))
x_test = np.zeros((n2,1))
model.getPosterior(x_train, y_train)
model.predict(x_test)
predictive_class = np.sign(model.ym)
# evaluation
acc = valid.ACC(predictive_class, y_test)
ACC.append(acc)