深度学习模型的可解释性为其预测提供了人类能够了解的推理。如果不解释预测背地的起因，深度学习算法就像黑匣子，对于一些场景说是无奈被信赖的。不提供预测的起因也会阻止深度学习算法在波及跨域偏心、隐衷和平安的要害应用程序中应用。

深度学习模型的可解释性有助于减少对模型预测的信赖， 进步模型对与偏心、隐衷和其余平安挑战相干的要害决策应用程序的透明度，并且能够让咱们理解网络特色，以便在将模型部署到事实世界之前辨认和纠正模型所犯错误的零碎模式。

图在事实世界中无处不在，代表社交网络、援用网络、化学分子、金融数据等。图神经网络 (GNN) 是一个弱小的框架，用于对图相干数据进行机器学习，例如节点分类、图分类、和链接预测。

所以本文探讨以下5方面

  • GNN 须要可解释性
  • 解释 GNN 预测的挑战
  • 不同的 GNN 解释方
  • GNNExplainer的直观解释
  • 应用 GNNExplainer 解释节点分类和图分类的实现

图卷积神经网络 (GCN)、GraphSAGE 和图留神网络 (GAT) 等 GNN 通过沿输出图的边缘递归传递神经音讯，将节点特色信息与图构造相结合。

同时联合图构造和特色信息会导致简单的模型；因而，解释 GNN 的预测是具备挑战性的。

  • 图数据不如图像和文本直观，这使得对图深度学习模型的人类能够了解的解释具备挑战性。
  • 图像和文本应用网格状数据；然而在拓扑图中，信息是应用特色矩阵和邻接矩阵来示意的，每个节点都有不同的街坊。因而图像和文本的可解释性办法不适宜取得对图的高质量解释。
  • 图节点和边对 GNN 的最终预测有显着奉献；因而GNN 的可解释性须要思考这些交互。
  • 节点分类工作通过执行来自其街坊的音讯遍从来预测节点的类别。音讯游走能够更好地理解 GNN 做出预测的起因，但这与图像和文本相比更具备挑战性。

GNN 解释办法

图可解释性须要答复以下问题：

  • 哪些输出边更要害，对预测奉献最大？
  • 哪些输出节点更重要？
  • 哪些节点特色更重要？
  • 什么图模式将最大化某个类的预测？

解释 GNN 的办法依据它们提供的解释类型分为两个分支。这些图解释办法侧重于图模型的不同方面，并提供不同的视图来了解 GNN 模型。

实例级办法：给定一个输出图，实例级办法通过辨认用于预测的重要输出特色来解释深度模型。

模型级办法提供了个别见解和高级了解来解释深度图模型。模型级办法专门钻研哪些输出图模式能够导致 GNN 进行某种预测。

上图为解释 GNN 的不同办法

实例级办法依据重要性分数的取得形式进行辨别，能够将它们分为四个不同的分支。

  • Gradients/Feature-based 办法应用梯度或暗藏特色图来示意不同输出特色的重要性，其中梯度或特征值越高示意重要性越高。基于梯度/特色的可解释性办法宽泛用于图像和文本工作。SA、Guided back propagation、CAM 和 Grad-CAM 是基于梯度/特色的可解释性办法的示例。
  • 基于扰动的办法监测不同输出扰动的输入变动变动。当保留重要的输出信息时，预测应该与原始预测类似。GNN 能够通过应用不同的掩码生成算法来取得不同类型的掩码来进行特色重要性的判断，如 GNNExplainer、PGExplainer、ZORRO、GraphMask、Causal Screening 和 SubgraphX。
  • 合成办法通过将原始模型预测合成为若干项来掂量输出特色的重要性，这些项被视为相应输出特色的重要性分数。
  • 代理办法采纳简略且可解释的代理模型来近似简单深度模型对输出示例的相邻区域的预测。代理办法包含 GraphLime、RelEx 和 PGM Explainer。

GNNExplainer

GNNExplainer 是一种与模型无关的基于扰动的办法，能够为任何基于图的机器学习工作上的任何基于 GNN 的模型的预测提供可解释的报告。

GNNExplainer 学习边和节点特色的软掩码，而后通过掩码的优化来解释预测。

GNNExplainer 会获取输出图并辨认紧凑的子图构造和在预测中起关键作用的一小部分节点特色。

GNNExplainer通过生成传递要害语义的掩码来捕捉重要的输出特色，从而产生与原始预测类似的预测。它学习边缘和节点特色的软掩码，通过掩码优化来解释预测。

以不同形式为输出图取得掩码能够取得重要的输出特色。还依据预测工作的类型生成不同的掩码，例如节点掩码、边掩码和节点特色掩码。

生成的掩码与输出图相结合，通过逐元素乘法取得蕴含重要输出信息的新图。最初，将新图输出经过训练的 GNN 以评估掩码并更新掩码生成算法。

GNNExplainer 示例

explain_node() 学习并返回一个节点特色掩码和一个边缘掩码，它们在解释 GNN 对节点分类所做的预测中起着至关重要的作用。

#Import Libraryimport numpy as npimport pandas as pdimport osimport torch_geometric.transforms as Tfrom torch_geometric.datasets import Planetoidimport matplotlib.pyplot as pltimport torchimport torch.nn.functional as Ffrom torch_geometric.nn import GCNConv, GNNExplainerimport torch_geometricfrom torch_geometric.loader import NeighborLoaderfrom torch_geometric.utils import to_networkx#Load the Planetoid datasetdataset = Planetoid(root='.', name="Pubmed")data = dataset[0]#Set the device dynamicallydevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')# Create batches with neighbor samplingtrain_loader = NeighborLoader(    data,    num_neighbors=[5, 10],    batch_size=16,    input_nodes=data.train_mask,)# Define the GCN modelclass Net(torch.nn.Module):    def __init__(self):        super().__init__()                self.conv1 = GCNConv(dataset.num_features, 16, normalize=False)        self.conv2 = GCNConv(16, dataset.num_classes, normalize=False)        self.optimizer = torch.optim.Adam(self.parameters(), lr=0.02, weight_decay=5e-4)    def forward(self, x, edge_index):        x = F.relu(self.conv1(x, edge_index))        x = F.dropout(x, training=self.training)        x = self.conv2(x, edge_index)        return F.log_softmax(x, dim=1)model = Net().to(device)def accuracy(pred_y, y):    """Calculate accuracy."""    return ((pred_y == y).sum() / len(y)).item()  # define the function to Train the modeldef train_nn(model, x,edge_index,epochs):    criterion = torch.nn.CrossEntropyLoss()    optimizer = model.optimizer    model.train()    for epoch in range(epochs+1):      total_loss = 0      acc = 0      val_loss = 0      val_acc = 0      # Train on batches      for batch in train_loader:        optimizer.zero_grad()                 out = model(batch.x, batch.edge_index)        loss = criterion(out[batch.train_mask], batch.y[batch.train_mask])                total_loss += loss        acc += accuracy(out[batch.train_mask].argmax(dim=1),                         batch.y[batch.train_mask])        loss.backward()        optimizer.step()        # Validation        val_loss += criterion(out[batch.val_mask], batch.y[batch.val_mask])        val_acc += accuracy(out[batch.val_mask].argmax(dim=1),                             batch.y[batch.val_mask])      # Print metrics every 10 epochs      if(epoch % 10 == 0):          print(f'Epoch {epoch:>3} | Train Loss: {total_loss/len(train_loader):.3f} '                f'| Train Acc: {acc/len(train_loader)*100:>6.2f}% | Val Loss: '                f'{val_loss/len(train_loader):.2f} | Val Acc: '                f'{val_acc/len(train_loader)*100:.2f}%')# define the function to Test the modeldef test(model, data):    """Evaluate the model on test set and print the accuracy score."""    model.eval()    out = model(data.x, data.edge_index)    acc = accuracy(out.argmax(dim=1)[data.test_mask], data.y[data.test_mask])    return acc  # Train the Modeltrain_nn(model, data.x, data.edge_index, 200)# Testprint(f'\nGCN test accuracy: {test(model, data)*100:.2f}%\n')# Explain the GCN for node node_idx = 20x, edge_index = data.x, data.edge_index# Pass the model to explain to GNNExplainerexplainer = GNNExplainer(model, epochs=100,return_type='log_prob')#returns a node feature mask and an edge mask that play a crucial role to explain the prediction made by the GNN for node 20node_feat_mask, edge_mask = explainer.explain_node(node_idx, x, edge_index)ax, G = explainer.visualize_subgraph(node_idx, edge_index, edge_mask, y=data.y)plt.show()print("Ground Truth label for node: ",node_idx, " is ", data.y.numpy()[node_idx])out = torch.softmax(model(data.x, data.edge_index), dim=1).argmax(dim=1)print("Prediction for node ",node_idx, "is " ,out[node_idx].cpu().detach().numpy().squeeze())

上图中所有色彩类似的节点都属于同一个类。可视化有助于解释哪些节点对预测奉献最大。

Explain_graph()用于图分类;它学习并返回一个节点特色掩码和一个边缘掩码，这两个掩码在解释GNN对一个图的预测时起着至关重要的作用

# Import libarariesimport numpy as npimport pandas as pdimport osimport torch_geometric.transforms as Tfrom torch_geometric.datasets import Planetoidimport matplotlib.pyplot as pltimport torchimport torch.nn.functional as Ffrom torch_geometric.nn import GraphConvimport torch_geometricfrom torch.nn import Parameterfrom torch_geometric.nn.conv import MessagePassingimport urllib.requestimport tarfilefrom torch.nn import Linearimport torch.nn.functional as Ffrom torch_geometric.nn import GCNConv, GNNExplainerfrom torch_geometric.nn import global_mean_poolfrom torch_geometric.datasets import TUDatasetfrom torch_geometric.loader import DataLoader# Load the datasetdataset = TUDataset(root='data/TUDataset', name='MUTAG')# print details about the graphprint(f'Dataset: {dataset}:')print("Number of Graphs: ",len(dataset))print("Number of Freatures: ", dataset.num_features)print("Number of Classes: ", dataset.num_classes)data= dataset[0]print(data)print("No. of nodes: ", data.num_nodes)print("No. of Edges: ", data.num_edges)print(f'Average node degree: {data.num_edges / data.num_nodes:.2f}')print(f'Has isolated nodes: {data.has_isolated_nodes()}')print(f'Has self-loops: {data.has_self_loops()}')print(f'Is undirected: {data.is_undirected()}')# Create train and test datasettorch.manual_seed(12345)dataset = dataset.shuffle()train_dataset = dataset[:50]test_dataset = dataset[50:]print(f'Number of training graphs: {len(train_dataset)}')print(f'Number of test graphs: {len(test_dataset)}')'''graphs in graph classification datasets are usually small, a good idea is to batch the graphs before inputting them into a Graph Neural Network to guarantee full GPU utilization___In pytorch Geometric adjacency matrices are stacked in a diagonal fashion (creating a giant graph that holds multiple isolated subgraphs), and node and target features are simply concatenated in the node dimension:'''train_loader = DataLoader(train_dataset,  batch_size=64, shuffle= True)test_loader= DataLoader(test_dataset,  batch_size=64, shuffle= False)for step, data in enumerate(train_loader):    print(f'Step {step + 1}:')    print('=======')    print(f'Number of graphs in the current batch: {data.num_graphs}')    print(data)    print()# Build the modelclass GNN(torch.nn.Module):    def __init__(self, hidden_channels):        super(GNN, self).__init__()        torch.manual_seed(12345)        self.conv1 = GraphConv(dataset.num_node_features, hidden_channels)        self.conv2 = GraphConv(hidden_channels, hidden_channels)        self.conv3 = GraphConv(hidden_channels, hidden_channels )        self.lin = Linear(hidden_channels, dataset.num_classes)    def forward(self, x, edge_index, batch):        x = self.conv1(x, edge_index)        x = x.relu()        x = self.conv2(x, edge_index)        x = x.relu()        x = self.conv3(x, edge_index)        x = global_mean_pool(x, batch)                x = F.dropout(x, p=0.5, training=self.training)        x = self.lin(x)                return xmodel = GNN(hidden_channels=64)print(model)# set the optimizeroptimizer = torch.optim.Adam(model.parameters(), lr=0.02)# set the loss functioncriterion = torch.nn.CrossEntropyLoss()# Creating the function to train the modeldef train():    model.train()    for data in train_loader:  # Iterate in batches over the training dataset.         out = model(data.x, data.edge_index, data.batch)  # Perform a single forward pass.         loss = criterion(out, data.y)  # Compute the loss.         loss.backward()  # Derive gradients.         optimizer.step()  # Update parameters based on gradients.         optimizer.zero_grad()  # Clear gradients.# function to test the modeldef test(loader):     model.eval()     correct = 0     for data in loader:  # Iterate in batches over the training/test dataset.         out = model(data.x, data.edge_index, data.batch)           pred = out.argmax(dim=1)  # Use the class with highest probability.         correct += int((pred == data.y).sum())  # Check against ground-truth labels.     return correct / len(loader.dataset)  # Derive ratio of correct predictions.# Train the model for 150 epochsfor epoch in range(1, 160):    train()    train_acc = test(train_loader)    test_acc = test(test_loader)    if(epoch % 10 == 0):          '''print(f'Epoch {epoch:>3} | Train Loss: {total_loss/len(train_loader):.3f} '                f'| Train Acc: {acc/len(train_loader)*100:>6.2f}% | Val Loss: '                f'{val_loss/len(train_loader):.2f} | Val Acc: '                f'{val_acc/len(train_loader)*100:.2f}%')          '''          print(f'Epoch: {epoch:03d}, Train Acc: {train_acc:.4f}, Test Acc: {test_acc:.4f}')#Explain the Graphexplainer = GNNExplainer(model, epochs=100,return_type='log_prob')data = dataset[0]node_feat_mask, edge_mask = explainer.explain_graph(data.x, data.edge_index)ax, G = explainer.visualize_subgraph(-1,data.edge_index, edge_mask, data.y)plt.show()

当可视化visualize_subgraph时，须要将node_idx设置为-1，因为这意味着一个图分类工作;否则会报错。

本文应用的是pytorch-geometric实现的GNNExplainer作为示例，有趣味理解的话能够查看其官网文档

https://avoid.overfit.cn/post/3a01457fe6094941a2bca2961f742dce

作者：Renu Khandelwal