\ub525\ub7ec\ub2dd\uc758 \ubc1c\uc804\uc73c\ub85c \uc778\ud574, \uc774\ubbf8\uc9c0\ub098 \ud14d\uc2a4\ud2b8\uc640 \uac19\uc740 \uc804\ud1b5\uc801\uc778 \ub370\uc774\ud130 \uc678\uc5d0\ub3c4 \uadf8\ub798\ud504 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc5f0\uad6c\uac00 \ud65c\ubc1c\ud574\uc84c\uc2b5\ub2c8\ub2e4. \uadf8\ub798\ud504 \ud569\uc131\uacf1 \ub124\ud2b8\uc6cc\ud06c(Graph Convolutional Network, GCN)\ub294 \uc774\ub7ec\ud55c \uadf8\ub798\ud504 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\uae30 \uc704\ud55c \uac15\ub825\ud55c \ub3c4\uad6c\uc785\ub2c8\ub2e4. \uc774 \uac15\uc88c\uc5d0\uc11c\ub294 GCN\uc758 \uc774\ub860\uc801 \ubc30\uacbd\uacfc PyTorch\ub97c \uc774\uc6a9\ud55c \uc2e4\uc81c \uad6c\ud604\uc744 \ub2e4\ub8e8\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uadf8\ub798\ud504\ub294 \ub178\ub4dc(\uc815\uc810)\uc640 \uc5e3\uc9c0(\uac04\uc120)\ub85c \uad6c\uc131\ub41c \ub370\uc774\ud130 \uad6c\uc870\uc785\ub2c8\ub2e4. \ub178\ub4dc\ub294 \uac1c\uccb4\ub97c \ub098\ud0c0\ub0b4\uace0, \uc5e3\uc9c0\ub294 \ub178\ub4dc \uac04\uc758 \uad00\uacc4\ub97c \ud45c\ud604\ud569\ub2c8\ub2e4. \uadf8\ub798\ud504\ub294 \uc18c\uc15c \ub124\ud2b8\uc6cc\ud06c, \ucd94\ucc9c \uc2dc\uc2a4\ud15c, \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n
GCN\uc740 \uadf8\ub798\ud504 \ub370\uc774\ud130\uc5d0\uc11c \ub178\ub4dc\uc758 \ud45c\ud604\uc744 \ud559\uc2b5\ud558\uae30 \uc704\ud574 \uace0\uc548\ub41c \uc2e0\uacbd\ub9dd \uad6c\uc870\uc785\ub2c8\ub2e4. GCN\uc740 \uc804\ud1b5\uc801\uc778 \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd(CNN)\uc758 \uac1c\ub150\uc744 \uadf8\ub798\ud504\uc5d0 \uc801\uc6a9\ud55c \ud615\ud0dc\ub85c, \ub178\ub4dc\uc758 \ud2b9\uc131\uacfc \uad6c\uc870\ub97c \uace0\ub824\ud558\uc5ec \uc815\ubcf4\ub97c \uc804\ud30c\ud569\ub2c8\ub2e4.<\/p>\n
GCN\uc758 \uae30\ubcf8 \uc544\uc774\ub514\uc5b4\ub294 \ub178\ub4dc\uc758 \ud2b9\uc9d5\uc744 \uc8fc\ubcc0 \ub178\ub4dc\uc640 \ud1b5\ud569\ud558\uc5ec \uc5c5\ub370\uc774\ud2b8\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uac01 \ub808\uc774\uc5b4\uc5d0\uc11c \ub2e4\uc74c\uacfc \uac19\uc740 \uc218\uc2dd\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4:<\/p>\n
H^{(l+1)} = \u03c3(A' H^{(l)} W^{(l)})<\/code><\/pre>\n\n- H^{(l)}: l\ubc88\uc9f8 \ub808\uc774\uc5b4\uc758 \ub178\ub4dc \ud2b9\uc9d5 \ud589\ub82c<\/li>\n
- A’: \ub178\ub4dc \uac04\uc758 \uc5f0\uacb0 \uc815\ubcf4\ub97c \ub098\ud0c0\ub0b4\ub294 \uc778\uc811 \ud589\ub82c<\/li>\n
- W^{(l)}: l\ubc88\uc9f8 \ub808\uc774\uc5b4\uc758 \uac00\uc911\uce58 \ud589\ub82c<\/li>\n
- \u03c3: \ud65c\uc131\ud654 \ud568\uc218 (\uc608: ReLU)<\/li>\n<\/ul>\n
2.2. GCN\uc758 \uc8fc\uc694 \ud2b9\uc9d5<\/h3>\n\n- \uc804\uc774 \ud559\uc2b5: GCN\uc740 \uadf8\ub798\ud504 \uad6c\uc870\ub97c \ud1b5\ud574 \ub178\ub4dc \uac04\uc758 \uc815\ubcf4\ub97c \uc804\uc774\ud569\ub2c8\ub2e4.<\/li>\n
- \uacb0\uacfc \ud574\uc11d \uac00\ub2a5\uc131: \ub178\ub4dc \uac04\uc758 \uc0c1\ud638\uc791\uc6a9\uc744 \uc2dc\uac01\uc801\uc73c\ub85c \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n
- \uc77c\ubc18\ud654 \ub2a5\ub825: \ub2e4\uc591\ud55c \uadf8\ub798\ud504 \uad6c\uc870\uc5d0 \uc801\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<\/ul>\n
3. PyTorch\ub85c GCN \uad6c\ud604\ud558\uae30<\/h2>\n
\uc774\uc81c \uc2e4\uc81c\ub85c PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec GCN\uc744 \uad6c\ud604\ud558\uaca0\uc2b5\ub2c8\ub2e4. PyTorch\ub294 \ub3d9\uc801 \ucef4\ud4e8 \uadf8\ub798\ud504\ub85c \uc720\uba85\ud558\uc5ec, \ubcf5\uc7a1\ud55c \ubaa8\ub378\uc744 \uc27d\uac8c \uad6c\ucd95\ud558\uace0 \ub514\ubc84\uae45\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
3.1. \ud658\uacbd \uc124\uc815<\/h3>\n
\uba3c\uc800 \ud544\uc694\ud55c \ud328\ud0a4\uc9c0\ub97c \uc124\uce58\ud569\ub2c8\ub2e4.<\/p>\n
!pip install torch torch-geometric<\/code><\/pre>\n3.2. \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\uc774\ubc88 \uc608\uc81c\uc5d0\uc11c\ub294 Cora<\/code> \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. Cora\ub294 \uac01 \ub178\ub4dc\uac00 \ub17c\ubb38\uc744 \ub098\ud0c0\ub0b4\uba70, \uc5e3\uc9c0\ub294 \ub17c\ubb38 \uac04\uc758 \uc778\uc6a9 \uad00\uacc4\ub97c \ub098\ud0c0\ub0c5\ub2c8\ub2e4.<\/p>\nimport torch\nfrom torch_geometric.datasets import Planetoid\n\ndataset = Planetoid(root='\/tmp\/Cora', name='Cora')\ndata = dataset[0]<\/code><\/pre>\n3.3. GCN \ubaa8\ub378 \uc815\uc758<\/h3>\n
GCN \ubaa8\ub378\uc744 \uc815\uc758\ud569\ub2c8\ub2e4. PyTorch\uc5d0\uc11c\ub294 \ud074\ub798\uc2a4 \uae30\ubc18\uc73c\ub85c \ubaa8\ub378\uc744 \uc815\uc758\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
import torch.nn.functional as F\nfrom torch.nn import Linear\nfrom torch_geometric.nn import GCNConv\n\nclass GCN(torch.nn.Module):\n def __init__(self, num_features, num_classes):\n super(GCN, self).__init__()\n self.conv1 = GCNConv(num_features, 16)\n self.conv2 = GCNConv(16, num_classes)\n\n def forward(self, data):\n x, edge_index = data.x, data.edge_index\n x = self.conv1(x, edge_index)\n x = F.relu(x)\n x = F.dropout(x, training=self.training)\n x = self.conv2(x, edge_index)\n return F.log_softmax(x, dim=1)<\/code><\/pre>\n3.4. \ubaa8\ub378 \ud559\uc2b5<\/h3>\n
\ubaa8\ub378\uc744 \ud559\uc2b5\ud558\uae30 \uc704\ud574 \uc190\uc2e4 \ud568\uc218\uc640 \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998\uc744 \uc124\uc815\ud569\ub2c8\ub2e4. \uc774\ubc88 \uc608\uc81c\uc5d0\uc11c\ub294 \ud06c\ub85c\uc2a4 \uc5d4\ud2b8\ub85c\ud53c \uc190\uc2e4\uacfc Adam \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
model = GCN(num_features=dataset.num_node_features, num_classes=dataset.num_classes)\noptimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=5e-4)\ncriterion = torch.nn.CrossEntropyLoss()\n\ndef train():\n model.train()\n optimizer.zero_grad()\n out = model(data)\n loss = criterion(out[data.train_mask], data.y[data.train_mask])\n loss.backward()\n optimizer.step()\n return loss.item()<\/code><\/pre>\n3.5. \ud559\uc2b5 \ubc0f \ud3c9\uac00<\/h3>\n
\uc774\uc81c \ubaa8\ub378\uc744 \ud559\uc2b5\ud558\uace0 \uc131\ub2a5\uc744 \ud3c9\uac00\ud569\ub2c8\ub2e4.<\/p>\n
for epoch in range(200):\n loss = train()\n if epoch % 10 == 0:\n print(f'Epoch {epoch}, Loss: {loss:.4f}')\n\n# \ubaa8\ub378 \ud3c9\uac00\nmodel.eval()\nout = model(data)\npred = out.argmax(dim=1)\ncorrect = (pred[data.test_mask] == data.y[data.test_mask]).sum()\nacc = int(correct) \/ int(data.test_mask.sum())\nprint(f'Accuracy: {acc:.4f}')<\/code><\/pre>\n4. GCN \ubaa8\ub378\uc758 \uc751\uc6a9<\/h2>\n
GCN\uc740 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \uc751\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uc18c\uc15c \ub124\ud2b8\uc6cc\ud06c\uc758 \uc0ac\uc6a9\uc790 \uae30\uc0ac \ucd94\ucc9c, \uadf8\ub798\ud504 \uae30\ubc18\uc758 \ud074\ub7ec\uc2a4\ud130\ub9c1 \ubc0f \ub178\ub4dc \ubd84\ub958 \ub4f1\uc5d0 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubaa8\ub378\uc758 \ud50c\ub809\uc2dc\ube14\ud55c \uc751\uc6a9\uc740 GCN\uc758 \ud070 \uc7a5\uc810 \uc911 \ud558\ub098\uc785\ub2c8\ub2e4.<\/p>\n
4.1. \uadf8\ub798\ud504 \ub370\uc774\ud130 \uc804\ucc98\ub9ac<\/h3>\n
\ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ub192\uc774\uae30 \uc704\ud574 \uadf8\ub798\ud504 \ub370\uc774\ud130\ub97c \uc804\ucc98\ub9ac\ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4. \ub370\uc774\ud130\uc758 \ud2b9\uc131\uc5d0 \ub530\ub77c \ub178\ub4dc \ud2b9\uc9d5\uc744 \uc815\uaddc\ud654\ud558\uace0, \uc5e3\uc9c0\uc758 \uac00\uc911\uce58\ub97c \uc870\uc815\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
4.2. \ub2e4\uc591\ud55c GCN \ubcc0\ud615\ub4e4<\/h3>\n
GCN \ud6c4\uc18d \uc5f0\uad6c\ub85c \uc5ec\ub7ec \ubcc0\ud615 \ubaa8\ub378\uc774 \uac1c\ubc1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, Graph Attention Networks (GAT)\ub294 \ub178\ub4dc\uc758 \uc911\uc694\ub3c4\ub97c \ud559\uc2b5\ud558\uc5ec \uac00\uc911\uce58\uac00 \ubc18\uc601\ub41c \ud569\uc131\uc744 \uc218\ud589\ud569\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubcc0\ud615\ub4e4\uc740 \ud2b9\uc815 \ubb38\uc81c\uc5d0 \ub300\ud574 \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n
5. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \uadf8\ub798\ud504 \ud569\uc131\uacf1 \ub124\ud2b8\uc6cc\ud06c(GCN)\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc PyTorch\ub97c \uc774\uc6a9\ud55c \uc2e4\uc81c \uad6c\ud604 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. GCN\uc740 \uadf8\ub798\ud504 \ub370\uc774\ud130\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \ucc98\ub9ac\ud560 \uc218 \uc788\ub294 \uac15\ub825\ud55c \ub3c4\uad6c\ub85c, \ub2e4\uc591\ud55c \ub3c4\uba54\uc778\uc5d0 \uc751\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c GCN \ubc0f \ub2e4\ub978 \uadf8\ub798\ud504 \uae30\ubc18 \ubaa8\ub378\uc5d0 \ub300\ud55c \uc5f0\uad6c\uac00 \ub354\uc6b1 \ud65c\ubc1c\ud574\uc84c\uc73c\uba74 \uc88b\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"