[{"source":{"position":1,"lines":["",""],"changePosition":null},"target":{"position":1,"lines":[],"changePosition":null},"type":"DELETE"},{"source":{"position":18,"lines":["$$\begin{gathered}h=\\\\{\\vec\\{h_1\\},\\vec\\{h_2\\},\\vec\\{h_3\\},..,\\vec\\{h_n\\}\\\\} \\tag\\{1.1\\}\end{gathered}$$"],"changePosition":null},"target":{"position":16,"lines":["$$\begin{gathered}","h=\\{\\vec{h_1},\\vec{h_2},\\vec{h_3},..,\\vec{h_n}\\} \\tag{1.1}","\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":20,"lines":["其中$\\vec\\{h_i\\} \\in R\\^F$,$N$表示节点邻居节点的个数，$F$表示相应的特征向量的维度，就是每个节点的特征。每一层的输出是一个新的节点特征向量集。"],"changePosition":null},"target":{"position":20,"lines":["其中$\\vec{h_i} \\in R^F$,$N$表示节点邻居节点的个数，$F$表示相应的特征向量的维度，就是每个节点的特征。每一层的输出是一个新的节点特征向量集。"],"changePosition":null},"type":"CHANGE"},{"source":{"position":22,"lines":["$$\begin{gathered}h'=\\\\{\\\\{\\vec\\{h_1'\\},\\vec\\{h_2'\\},\\vec\\{h_3'\\},..,\\vec\\{h_n'\\}\\\\} \\tag\\{1.2\\}\\\\}\end{gathered}$$"],"changePosition":null},"target":{"position":22,"lines":["$$\begin{gathered}","h'=\\{\\vec{h_1'},\\vec{h_2'},\\vec{h_3'},..,\\vec{h_n'}\\} \\tag{1.2}","\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":30,"lines":["$$\begin{gathered}e_\\{ij\\}=a(W\\vec\\{h_i\\},W\\vec\\{h_j\\}) \\tag\\{1.3\\}\end{gathered}$$"],"changePosition":null},"target":{"position":32,"lines":["$$\begin{gathered}e_{ij}=a(W\\vec{h_i},W\\vec{h_j}) \\tag{1.3}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":32,"lines":["其中$a$表示一个$R\\^\\{F'\\} \\times R\\^\\{F'\\} \\rightarrow R$ 的映射。$W \\in R\\^\\{F' \\times F\\}$是一个权值矩阵，一般来讲self-attention将注意力分配到图中所有节点，这种做法是会丢掉信息的。本身使用了masked attention的方式将注意力分配到邻居节点上。$j\\in N_i$."],"changePosition":null},"target":{"position":34,"lines":["其中$a$表示一个$R^{F'} \\times R^{F'} \\rightarrow R$ 的映射。$W \\in R^{F' \\times F}$是一个权值矩阵，一般来讲self-attention将注意力分配到图中所有节点，这种做法是会丢掉信息的。本身使用了masked attention的方式将注意力分配到邻居节点上。$j\\in N_i$."],"changePosition":null},"type":"CHANGE"},{"source":{"position":34,"lines":["$$\begin{gathered}a_\\{ij\\}=softmax(e_\\{ij\\})=\\frac\\{exp(e_\\{ij\\})\\}\\{\\sum_\\{k \\in N_i\\}exp(e_\\{ik\\})\\} \\tag\\{1.4\\}\end{gathered}$$"],"changePosition":null},"target":{"position":36,"lines":["$$\begin{gathered}a_{ij}=softmax(e_{ij})=\\frac{exp(e_{ij})}{\\sum_{k \\in N_i}exp(e_{ik})} \\tag{1.4}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":40,"lines":["$$\begin{gathered}a_\\{ij\\}=\\frac\\{exp(LeakyRelu(\\vec\\{a\\}\\^T[Wh_i || Wh_j]))\\}\\{\\sum_\\{k \\ inN_i\\}exp(LeakyRelu(\\vec\\{a\\}\\^T [wh_i||Wh_k]))\\} \\tag\\{1.5\\}\end{gathered}$$"],"changePosition":null},"target":{"position":42,"lines":["$$\begin{gathered}a_{ij}=\\frac{exp(LeakyRelu(\\vec{a}^T[Wh_i || Wh_j]))}{\\sum_{k \\ inN_i}exp(LeakyRelu(\\vec{a}^T [wh_i||Wh_k]))} \\tag{1.5}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":42,"lines":["其中$\\vec\\{a\\}\\^T \\in R\\^\\{2F'\\}$为前馈神经网络的$a$的参数。"],"changePosition":null},"target":{"position":44,"lines":["其中$\\vec{a}^T \\in R^{2F'}$为前馈神经网络的$a$的参数。"],"changePosition":null},"type":"CHANGE"},{"source":{"position":44,"lines":["$$\begin{gathered}h_i'=\\sigma(\\sum_\\{j \\in N_i\\}a_\\{ij\\}\\^\\{k\\}W\\^kh_j) \\tag\\{1.6\\}\end{gathered}$$"],"changePosition":null},"target":{"position":46,"lines":["$$\begin{gathered}h_i'=\\sigma(\\sum_{j \\in N_i}a_{ij}^{k}W^kh_j) \\tag{1.6}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":48,"lines":["math"],"changePosition":null},"target":{"position":50,"lines":[""],"changePosition":null},"type":"CHANGE"},{"source":{"position":52,"lines":[""],"changePosition":null},"target":{"position":54,"lines":[],"changePosition":null},"type":"DELETE"},{"source":{"position":54,"lines":["其中$||$表示连接，$a_\\{ij\\}\\^\\{k\\}$和$W\\^K$表示k个抽头的计算结果。由于$W\\^k \\in R\\^\\{F' \\times F\\}$,因此这里的$\\vec\\{h_i\\}' \\in R\\^\\{KF'\\}$。同样可以采取求和的方式得到$\\vec\\{h_i\\}'$"],"changePosition":null},"target":{"position":55,"lines":["其中$||$表示连接，$a_{ij}^{k}$和$W^K$表示k个抽头的计算结果。由于$W^k \\in R^{F' \\times F}$,因此这里的$\\vec{h_i}' \\in R^{KF'}$。同样可以采取求和的方式得到$\\vec{h_i}'$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":56,"lines":["math"],"changePosition":null},"target":{"position":57,"lines":[""],"changePosition":null},"type":"CHANGE"},{"source":{"position":58,"lines":["h_i'=\\sigma(\\frac{1}{K} \\sum^{K}_{k=1}\\sum_{j \\in N_i}a_{ij}^{k}W^kh_j) tag{1.8}"],"changePosition":null},"target":{"position":59,"lines":["h_i'=\\sigma(\\frac{1}{K} \\sum^{K}_{k=1}\\sum_{j \\in N_i}a_{ij}^{k}W^kh_j) \\tag{1.8}"],"changePosition":null},"type":"CHANGE"},{"source":{"position":60,"lines":[""],"changePosition":null},"target":{"position":61,"lines":[""],"changePosition":null},"type":"CHANGE"}]