def forward(self, node_ids, neighbor_ids, label):
node_embedding = self.node_emd(node_ids)
neighbor_node_embedding = self.node_emd(neighbor_ids)
bias = self.bias_vector.gather(0, neighbor_ids)
score = (node_embedding*neighbor_node_embedding).sum(dim=1)+bias
loss = (F.binary_cross_entropy_with_logits(score, label)).mean() + \
self.lambda_dis * (l2_loss(node_embedding)+l2_loss(neighbor_node_embedding)+l2_loss(bias))
return loss
def forward(self, node_ids, neighbor_ids, reward):
node_embedding = self.node_emd(node_ids)
neighbor_node_embedding = self.node_emd(neighbor_ids)
bias = self.bias_vector.gather(0, neighbor_ids)
score = (node_embedding*neighbor_node_embedding).sum(dim=1)+bias
prob = score.sigmoid().clamp(1e-5, 1)
loss = -(prob.log()*reward).mean() + self.lambda_gen * (l2_loss(node_embedding)+l2_loss(neighbor_node_embedding)+l2_loss(bias))
return loss
def forward_taxo(self, node_ids, category_ids, label, levels):
if self.transform:
transforms = F.normalize(self.transforms, p=2, dim=1)
# transforms = self.transforms
node_embedding = self.node_emd(node_ids).unsqueeze(
1) @ transforms[levels]
else:
node_embedding = self.node_emd(node_ids)
category_embedding = self.taxo_emd(category_ids)
bias = self.taxo_bias_vector.gather(0, category_ids)
score = (node_embedding.squeeze(1) *
category_embedding).sum(dim=1) + bias
loss = F.binary_cross_entropy_with_logits(
score, label) + self.lamb * (l2_loss(bias))
return loss * self.lambda_taxo
def forward_taxo(self, node_ids, category_ids, reward, levels):
if self.transform:
transforms = F.normalize(self.transforms, p=2, dim=1)
# transforms = self.transforms
node_embedding = self.node_emd(node_ids).unsqueeze(
1) @ transforms[levels]
else:
node_embedding = self.node_emd(node_ids)
category_embedding = self.taxo_emd(category_ids)
bias = self.taxo_bias_vector.gather(0, category_ids)
score = (node_embedding.squeeze(1) *
category_embedding).sum(dim=1) + bias
prob = score.sigmoid().clamp(1e-5, 1)
loss = -(prob.log() * reward).mean() + self.lamb * (l2_loss(bias))
return loss * self.lambda_taxo