GCN原理回顾论文导读

Cora_dataset description

Cora数据集是一个常用的学术文献用网络数据集，用于研究学术文献分类和图网络分析等任务。
该数据集由机器学习领域的博士论文摘要组成，共计2708篇论文，涵盖了7个不同的学科领域。每篇论文都有一个唯一的ID，并且被分为以下7个类别之一：Case_Based、Genetic_Algorithms、Neural_Networks、Probabilistic_Methods、Reinforcement_Learning、Rule_Learning和Theory。

除了论文之间的引用关系外，Cora数据集还包含了每篇论文的词袋表示，即将每篇论文表示为一个词频向量(0-1嵌入，每行有多个1，非one-hot vector，feature of node)。这些词频向量表示了论文中出现的单词及其在该论文中的出现频率。

Cora数据集常用于图神经网络的研究和评估，可以用于学术文献分类、引文网络分析、节点嵌入等任务。

print cora

dataset = Planetoid("./tmp/Cora", name="Cora", transform=T.NormalizeFeatures())
num_nodes = dataset.data.num_nodes
# For num. edges see:
# - https://github.com/pyg-team/pytorch_geometric/issues/343
# - https://github.com/pyg-team/pytorch_geometric/issues/852num_edges = dataset.data.num_edges // 2
train_len = dataset[0].train_mask.sum()
val_len = dataset[0].val_mask.sum()
test_len = dataset[0].test_mask.sum()
other_len = num_nodes - train_len - val_len - test_len
print(f"Dataset: {dataset.name}")
print(f"Num. nodes: {num_nodes} (train={train_len}, val={val_len}, test={test_len}, other={other_len})")
print(f"Num. edges: {num_edges}")
print(f"Num. node features: {dataset.num_node_features}")
print(f"Num. classes: {dataset.num_classes}")
print(f"Dataset len.: {dataset.len()}")

在这里插入图片描述

GCN原理与实现

卷积公式： $f*g=F^{-1}(F(f)·F(g))$
给定一个图信号x和一个卷积核, $x*g=U(U^Tx\odot U^Tg)=U(U^Tx\odot g_{\theta})=\widetilde D^{-0.5}\widetilde A\widetilde D^{-0.5}X\Theta$
其中A为图的邻接矩阵，D为图的度数矩阵，
$\widetilde D=D+\gamma I,\widetilde A=A+\gamma I$ ，添加自环以缩小 $\lambda$ (Laplace matrix)

1.computation of $\widetilde D^{-0.5}\widetilde A\widetilde D^{-0.5}$

def gcn_norm(edge_index, edge_weight=None, num_nodes=None,
add_self_loops=True, flow="source_to_target", dtype=None):
fill_value = 1.
num_nodes = maybe_num_nodes(edge_index, num_nodes)
if add_self_loops: #添加自环
edge_index, edge_weight = add_remaining_self_loops(
edge_index, edge_weight, fill_value, num_nodes)
edge_weight = torch.ones((edge_index.size(1), ), dtype=dtype,
device=edge_index.device)
row, col = edge_index[0], edge_index[1]
idx = col
deg = scatter(edge_weight, idx, dim=0, dim_size=num_nodes, reduce='sum')
deg_inv_sqrt = deg.pow_(-0.5)
deg_inv_sqrt.masked_fill_(deg_inv_sqrt == float('inf'), 0)
edge_weight = deg_inv_sqrt[row] * edge_weight * deg_inv_sqrt[col]
return edge_index, edge_weight

代码解释

edge_index, edge_weight = add_remaining_self_loops(edge_index, edge_weight,fill_value, num_nodes)：
$\widetilde D=D+\gamma I,\widetilde A=A+\gamma I$ ;
deg = scatter(edge_weight, idx, dim=0, dim_size=num_nodes, reduce='sum')：
根据edge_weight和idx=edge_index[1]得到度数矩阵，deg=D

explantation：edge_weight是要放入的对角阵，

deg_inv_sqrt = deg.pow_(-0.5)：require $D^{-0.5}$
deg_inv_sqrt.masked_fill_(deg_inv_sqrt == float('inf'), 0)：
由于D非对角元=0，其-0.5次幂=∞，需要转化为0，
edge_weight = deg_inv_sqrt[row] * edge_weight * deg_inv_sqrt[col]：
输出归一化后的edge_index

2. PairNorm

在这里插入图片描述

3.GCNConv的实现如下(删改自torch_geometric.nn.GCNConv)

class myGCNConv2(MessagePassing):def __init__(self, in_channels: int, out_channels: int,add_self_loops: bool = True,bias: bool = True):super().__init__()self.in_channels = in_channelsself.out_channels = out_channelsself.add_self_loops = add_self_loopsself.lin = Linear(in_channels, out_channels, bias=False,weight_initializer='glorot')if bias:self.bias = Parameter(torch.Tensor(out_channels))else:self.register_parameter('bias', None)self.reset_parameters()def reset_parameters(self):super().reset_parameters()self.lin.reset_parameters()         #卷积层zeros(self.bias)                    #偏置层def forward(self, x: Tensor, edge_index: Adj,edge_weight: OptTensor = None) -> Tensor:edge_index, edge_weight = gcn_norm(  # yapf: disableedge_index, edge_weight, x.size(self.node_dim),self.add_self_loops, self.flow, x.dtype)x = self.lin(x)# propagate_type: (x: Tensor, edge_weight: OptTensor)out = self.propagate(edge_index, x=x, edge_weight=edge_weight,size=None)if self.bias is not None:out = out + self.biasreturn outdef message(self, x_j: Tensor, edge_weight: OptTensor) -> Tensor:return x_j if edge_weight is None else edge_weight.view(-1, 1) * x_jdef message_and_aggregate(self, adj_t: SparseTensor, x: Tensor) -> Tensor:return spmm(adj_t, x, reduce=self.aggr)

代码解释

x = self.lin(x)： $X'=X\Theta，X\in R^{n*d1}，\Theta \in R^{d1*d2}$ ，对X降维
out = self.propagate(edge_index, x=x, edge_weight=edge_weight,size=None)：
out= $A'X'=\widetilde D^{-\frac 1 2}\widetilde A \widetilde D^{-\frac 1 2 } X \Theta$
Converge {x1’,…,xn’} ，each of which be a sampled vector，into target form.

message&message_and_aggregate为MessagePassing.propagate的相关函数，
经测试，删除后，val acc下降，故予以保留

4.Net(GCN)的实现

class GCN(torch.nn.Module):
def __init__(
self,
num_node_features: int,
num_classes: int,
hidden_dim: int = 16,
dropout_rate: float = 0.5,
) -> None:
super().__init__()
self.dropout1 = torch.nn.Dropout(dropout_rate)
self.conv1 = myGCNConv2(num_node_features,
hidden_dim,add_self_loops=True)
self.relu = torch.nn.ReLU(inplace=True)
self.dropout2 = torch.nn.Dropout(dropout_rate)
self.conv2 = myGCNConv2(hidden_dim, num_classes,add_self_loops=True)
self.pn=PairNorm()
def forward(self, x: Tensor, edge_index: Tensor) -> torch.Tensor:
x = self.pn(x)
x = self.dropout1(x)
x = self.conv1(x, edge_index)
x = self.relu(x)
x = self.dropout2(x)
x = self.conv2(x, edge_index)
return x

代码解释

x = self.pn(x)：对x作PairNorm处理，之后xi~N(0,s2)，各节点特征范数大小平衡，作用不明显；
采用2层GCN卷积层，中间用relu激活，dropout避免过拟合

DropEdge Realization的手动实现

idea

首先把有向图的边，转化为无向图的边，保存在single_edge_index中，实现时先用single_edge字
典存储每条无向边(key-value 任意)1次，再把single_edge转化成无向图的边集索引(2-dim tensor
array)

#single_edge_index
single_edge={}
for i in range(len(dataset.data.edge_index[0])):if(((dataset.data.edge_index[0][i],dataset.data.edge_index[1][i]) not in single_edge.items()) and ((dataset.data.edge_index[1][i],dataset.data.edge_index[0][i]) not in single_edge.items())):single_edge[dataset.data.edge_index[0][i]]=dataset.data.edge_index[1][i]single_edge_index=[[],[]]for key,value in single_edge.items():single_edge_index[0].append(key)single_edge_index[1].append(value)        single_edge_index=torch.tensor(single_edge_index)

再把无向边集舍去dropout_rate比例的部分，之后转成有向边集索引

def drop_edge(single_edge_index, dropout_rate):# 计算需要丢弃的边数num_edges = single_edge_index.shape[1]num_drop = int(num_edges * dropout_rate)# 随机选择要丢弃的边remain_indices = torch.randperm(num_edges)[num_drop:]remain_single_edges = single_edge_index[:, remain_indices]reverse_edges = torch.stack([remain_single_edges[1],remain_single_edges[0]],dim=0)remain_edges=torch.cat([remain_single_edges,reverse_edges],dim=1)return remain_edges