摘要：

本文提出了利用邻节点及其与中心节点邻边相乘的方法，提取节点间交互信息；同时设计了新颖的采样策略加速训练

方法关键：

在原有简单属性相加的基础上

h(i+1)(u)=σ(∑v∈N(u)W[h(i)(v)f(eu,v)])=σ(W1∑v∈N(u)h(i)(v)+W2∑v∈N(u)f(eu,v))\begin{aligned} h^{(i+1)}(u) &=\sigma\left(\sum_{v \in N(u)} \mathbf{W}\left[\begin{array}{c}{h^{(i)}(v)} \\ {f\left(e_{u, v}\right)}\end{array}\right]\right) \\ &=\sigma\left(\mathbf{W}_{1} \sum_{v \in N(u)} h^{(i)}(v)+\mathbf{W}_{2} \sum_{v \in N(u)} f\left(e_{u, v}\right)\right) \end{aligned}

可以存在异构的问题

本文提出让邻节点及其相应的边相乘的方法

f((v,eu,v)):=f(v)⊗f(eu,v)f\left(\left(v, e_{u, v}\right)\right):=f(v) \otimes f\left(e_{u, v}\right)

细节上，本文进一步解决了high dimensionalities and heavy redundancies的问题

LASE原理

LASE can be divided into three common modules, namely a gate, an amplifier and an aggregator

门：the gate evaluates v’s influence in u’s neighborhood

放大器：The amplifier amplifies the node attributes using link information

控制器：The aggregator sums up neighbor embeddings and combines them with the central node embedding using various strategies

Aggregators proposed in [Hamilton et al.,2017] may also be used in LASE.

实验

数据集

Reddit：节点为post，边为两条post的共同评论用户在不同社区的平均评论分布

dblp：文章为节点，以文章提取出来的tf-idf向量为节点属性，边属性为one-hot embeddings of the common authors。pca降维到200维

email和fmobile，节点为联系人，边为时间切片后节点之间的联系

结果分析

应用数据集做节点分类

汇总邻居属性的能力上，GCNs模型强于proximity-based models（LINE DeepWalk）

在利用边属性的能力上，LASE强于其他

LASE-RW和LASE-SAGE强于原始的简单拼接LASE-concat

Although there is no original features in two temporal networks, LASE still outperforms pre-trained features by exploring edge attributes, while GCN and GraphSAGE do not capture these additional information and struggles in over-fitting the proximity-based features.

贡献：

LASE provides a ubiquitous solution to a wider class of graph data by incorporating link attributes;

数据集适用广泛

LASE outperforms strong baselines and naive concatenating implementations by adequately leveraging the information in the link attributes;

充分利用边的属性信息，胜过原始方法

LASE adopt a more explainable approach in determining the neural architecture and thus enjoys better explainability

神经网络结构中采用了一种更具解释性的方法

不足：

  • Sampling setup不够好，在超大图上很笨拙
  • 模型为保持通用性而不够简洁优雅。应用到具体领域时可以进行优化

Sampling: To control batch scales, we leverage the Monte Carlo method to estimate the summed neighborhood information by sampling a fixed number of neighbors.