def infomax_loss(self, prop_emb_list, neg_prop_emb_list):
prop_result = np.concatenate(prop_emb_list, axis=1)
if self.svd:
prop_result = get_embedding_dense(prop_result, self.dim)
def sigmoid(x):
return 1.0 / (1 + np.exp(-x))
pos_glb = prop_result.mean(0)
pos_info = sigmoid(pos_glb.dot(prop_result.T))
pos_loss = np.mean(np.log(pos_info)).mean()
neg_loss = 0
neg_step = 1
for _ in range(neg_step):
neg_prop_result = np.concatenate(neg_prop_emb_list, axis=1)
if self.svd:
neg_prop_result = get_embedding_dense(neg_prop_result,
self.dim)
neg_info = sigmoid(pos_glb.dot(neg_prop_result.T))
neg_loss += np.mean(np.log(1 - neg_info)).mean()
random.shuffle(neg_prop_emb_list)
return -(pos_loss + neg_loss) / (1 + neg_step)
def __call__(self, emb, adj):
if len(self.filter_types) == 1 and self.filter_types[0] == "identity":
return emb
dim = emb.shape[1]
prop_result = []
for tp in self.filter_types:
prop_result.append(propagate(adj, emb, tp))
prop_result_emb = np.concatenate(prop_result, axis=1)
if self.svd:
prop_result_emb = get_embedding_dense(prop_result_emb, dim)
return prop_result_emb
def __call__(self, emb, adj):
self.init_data(emb, adj)
study = optuna.create_study()
study.optimize(self.target_func, n_jobs=self.n_workers, n_trials=self.max_evals)
best_params = study.best_params
best_result = self.prop(best_params)[0]
best_result = np.concatenate(best_result, axis=1)
print(f"best parameters: {best_params}")
if self.svd:
best_result = get_embedding_dense(best_result, self.dim)
return best_result
def _chebyshev_gaussian(self,
A,
a,
order=5,
mu=0.5,
s=0.2,
plus=False,
nn=False):
# NE Enhancement via Spectral Propagation
num_node = a.shape[0]
if order == 1:
return a
A = sp.eye(num_node) + A
DA = preprocessing.normalize(A, norm="l1")
L = sp.eye(num_node) - DA
M = L - mu * sp.eye(num_node)
Lx0 = a
Lx1 = M.dot(a)
Lx1 = 0.5 * M.dot(Lx1) - a
conv = iv(0, s) * Lx0
conv -= 2 * iv(1, s) * Lx1
for i in range(2, order):
Lx2 = M.dot(Lx1)
Lx2 = (M.dot(Lx2) - 2 * Lx1) - Lx0
# Lx2 = 2*L.dot(Lx1) - Lx0
if i % 2 == 0:
conv += 2 * iv(i, s) * Lx2
else:
conv -= 2 * iv(i, s) * Lx2
Lx0 = Lx1
Lx1 = Lx2
del Lx2
emb = mm = conv
if not plus:
mm = A.dot(a - conv)
if not nn:
emb = get_embedding_dense(mm, self.dimension)
return emb
