[论文精读]Line Graph Neural Networks for Link Prediction

①For a graph $G=\left ( V, E \right )$ , $V=\left \{ v_1, v_2,...,v_n \right \}$ is the node set, $E\subseteq V \times V$ is the edge set, and adjacency matrix $A$

2.4.2. Overall Framework

①The original graph for link prediction in social networks (this method mainly focus on 1-hop neighbors and extracting subgraphs) (1-hop can be extended to h-hop).

extracted subgraph:

The common connections determine the relations between A and B

②Steps of link prediction: enclosing subgraph extraction, node labeling, feature learning and link prediction

③Overall framework:

where the double circles are the centered nodes, and the pure yellow circle inthe forth graph is the aimed link

2.4.3. Line Graph Neural Networks

（1）Line graph space transformation

①⭐我觉得它和传统线图不一样，传统的是无向→有向，但是它还是无向→无向。作者定义的是当原图中的两个边有共享顶点时，两条边在线图中变成俩顶点并存在一条连接

②Example:

③⭐The number of edges in L(G) is $\frac{1}{2}\sum_{i=1}^{m}d^2_i-n$ , where $d_i$ denotes the degree of node $i$ in original graph, $n$ is the edge number in the original graph

（2）Node label transformation

①Node label generation:

$l_{(v_1,v_2)}=\mathrm{concate}(\min(f_l(v_1),f_l(v_2)),\max(f_l(v_1),f_l(v_2)))$

where the $f$ denotes the node labeling function, $v_1$ and $v_2$ represent the end nodes of edge. In undirected graph, edge ( $v_1,v_2$ ) equals to edge ( $v_2,v_1$ )

②作者意思是①确实适用于无向图因为无论 $v_1$ 和 $v_2$ 顺序和编号咋样得到的 $l$ 都一样。但是如果加进去俩节点属性就会有顺序上的问题，所以节点属性还是应该取和：

$l_{(v_1,v_2)}=\mathrm{concate}(\min(f_l(v_1),f_l(v_2)),\max(f_l(v_1),f_l(v_2)),X_{v_1}+X_{v_2})$

where $X_{v_1}$ and $X_{v_2}$ are the node attributes of $v_1$ and $v_2$ . And this $l$ will be the node attribute of this edge in the line graph

（3）Feature Learning by Graph Neural Networks

①Updating of GNN:

$Z_{(v_i,v_j)}^{(k+1)}=\left(Z_{(v_i,v_j)}^{(k)}+\beta\sum_{d\in\mathcal{N}_{(v_i,v_j)}}Z_d^{(k)}\right)W^{(k)}$

where $\mathcal{N}_{(v_i,v_j)}$ is all the neighbors of node ${(v_i,v_j)}$ , $W$ is a weight matrix and $\beta$ denotes the normalization coefficient. $Z_{(v_{i},v_{j})}^{0}=l_{(v1,v2)}$

②Cross entropy loss:

$\mathcal{L}_{CE}=-\sum_{l\in L_{t}}(y_{l}\mathrm{log} (p_{l})+(1-y_{l})\mathrm{log} (1-p_{l}))$

where $L_t$ is the links which need to be predicted, $p_l$ is the probability that the link $l$ exists, $y_{l}\in\{0,1\}$ represents if the link exists

（4）Connection With Learning on Original Graphs

①Edge feature:

$f_e=g(f_l(v1),f_l(v2))$

where $g\left ( \cdot \right )$ denotes GNN

②They rewrite the updating function:

$\begin{aligned}Z_{(v_{i},v_{j})}^{1}&=(l_{(v1,v2)}+\beta\sum_{d_{1}\in\mathcal{N}_{v1}}\sum_{d_{2}\in\mathcal{N}_{d1}}l_{(d1,d2)}+\beta\sum_{d_{3}\in\mathcal{N}_{v2}}\sum_{d_{4}\in\mathcal{N}_{d3}}l_{(d3,d4)})W^{(0)}\end{aligned}$