How do I create a model for node regression?

[TSWG-test1]

Hello.
I'm trying node regression in social network.

For regression, What should be the final output of linear layer?
Is it best to have a final output of 1?

Please let me know if there is a better model

Below is the model.

print(data)
> Data(x=[58, 1], edge_index=[2, 770], edge_attr=[770], y=[58], train_mask=[58])

class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = GCNConv(58, 64)
        self.conv2 = GCNConv(64, 32)
        self.linear1 = Linear(32, 4)
        
    def forward(self, data):
        x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr
        x = self.conv1(x, edge_index, edge_attr)
        x = F.relu(x)
        x = self.conv2(x, edge_index, edge_attr)
        x = F.relu(x)
        x = self.linear1(x)
        
        return x

[rusty1s]

Yes, the output should be of feature dimension 1 if you are training against a single scalar per node.

[TSWG-test1]

Thank you for your reply.

I want to predict the features of all 58 nodes. Is Leave One Out appropriate for cross-validation?

The current average MSE is about 0.8, and the coefficient of determination between prediction and correct values is between 0.05and 0.1 . How can I improve it?

https://drive.google.com/drive/mobile/folders/1WgRehfAyzWHmIbaaanITQ_hYGGKDbsxz

The following is doing leave one out

def kyotangoGNN(row, col, node_feature, edge_feature, node):
  """
  row：始点ノード
  col：終点ノード
  node_feature：ノードの特徴量
  edge_feature：エッジの特徴量
  node：テストするノード
  """

  # エッジ形状
  edge_index = torch.tensor([row, col], dtype=torch.long)
  
  # ラベル
  y = torch.tensor(node_feature, dtype=torch.float)

  # ノードの特徴量
  x = torch.tensor([[i.item()] for i in y])

  # エッジの特徴量
  edge_attr = torch.tensor(edge_feature, dtype=torch.float)

  train_mask = torch.tensor([True for i in range(len(y))])
  test_mask = torch.tensor([False for i in range(len(y))])
  train_mask[node] = False
  test_mask[node] = True

  data = Data(x=x, y=y, edge_index=edge_index, edge_attr=edge_attr, train_mask=train_mask)
  return data, test_mask

row = row
col = col
node_feature = standard_procommunity_attitude      #standard_procommunity_attitude
edge_feature = edge_feature
data, test_mask = kyotangoGNN(row, col, node_feature, edge_feature, 0)

graph_loss = []
node_loss = []
true = [i.item() for i in data.y]
pred1 = []

for node in range(len(data.y)):
    # デバイス設定
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    # モデルのインスタンス生成
    model = Net()

    # モデルを訓練モードに設定
    model.train()

    # optimizer
    optimizer = torch.optim.SGD(model.parameters(), lr=0.01, weight_decay=5e-4)

    #standard_procommunity_attitude
    data, test_mask = kyotangoGNN(row, col, node_feature, edge_feature, node)
    train_mask = data.train_mask

    # learnig loop
    for epoch in range(100):
        optimizer.zero_grad()
        out, _ = model(data).max(dim=1)
        loss = F.mse_loss(out, data.y)
        loss.backward()
        optimizer.step()
        if epoch == 99:
            print()
            print("node", node, ',Epoch %d | Loss: %.4f' % (epoch, loss.item()))
            graph_loss.append(loss.item())
            loss = F.mse_loss(out[node], data.y[node])
            print("true {:.5f} | pred {:.5f} | loss {:.5f}".format(true[node], out[node].item(), loss.item()))
            pred1.append(out[node].item())
            node_loss.append(loss.item())

[rusty1s]

Aren't you currently also training on your test nodes? Where does train_mask gets applied?