AGNN论文笔记

2019-11-17  本文已影响0人  第一个读书笔记

AGNN: Attention-based Graph Neural Network for Semi-supervised learning (Mar 2018)

Background

These architectures alternate between a propagation layer that aggregates the hidden states of the local neighborhood and a fully connected layer.

AGNN:

  1. 通过移除intermediate fully connected layer,降低模型参数。在半监督学习任务中,标签数据比较少,这种方式可以为设计innovative的传播层留有更多余地;
  2. 在传播层中使用attention mechanisms,动态调节局部信息,以此获得更高的准度。

Contribution

  1. 使用a linear classifier of multinomial logistic regression,移除了intermediate non-linear activation layers, 只保留了图神经网络中的邻接线性传播,这种模型结果可以得到与最好的图模型媲美的结果,同时也表明了图上邻接信息聚合的重要性;

  2. 基于attention mechanisms:

    1. 降低模型复杂度,在每个intermediate layer,只有一个scalar parameter
    2. Discover dynamically and adaptively which nodes are relevant to the target node for clarification

Model

节点特征向量:X = [X_1, ...,X_n]X_i ∈ R^{d_x}
标签:Y_i
有标签的子集: L ⊂V
假设邻近节点更有可能具有相同的标签,损失:L(X,Y_L) = L_{label}(X_L, Y_L) + λL_G(X)

邻接矩阵:A ∈\{0,1\}^{n×n}
目标函数:Z = F(X,A) ∈ R^{n×d_y} ,来预测每个节点所属标签,其中:
Z_{ic}:节点i数据标签c的概率

Propagation Layer

第t层的隐层:H^t ∈ R^{n×d_h}
传播矩阵:P∈ R^{n×n}
传播层:\tilde H^t = PH^t,可以是局部平均或者是随机游走:

单层传播: H^{t+1} =σ(\tilde H^tW^t)

GCN

Is a special case of GNN which stacks two layers of specific propagation and perceptron:

H^1 = ReLU((PX))W^0)
Z = f(X,A) = softmax((PH^1)W^1) = softmax((PReLU((PX))W^0))W^1)
其中:
P = \tilde D^{-1/2}\tilde A\tilde D^{-1/2}
\tilde A = A+I
\tilde D = diag(\tilde A1)
损失:L = -\sum_{i∈L} \sum_{c=1}^{d_y}Y_{ic}lnZ_{ic}

GLN

将GCN的intermediate非线性激活移除,就是GLN:
Z = f(X,A) = softmax((P^2X)W^0W^1)
The two propagation layers simply take linear local average of the raw features weighted by their degrees, and at the output layer, a simple linear classifier( multinomial logistics regression) is applied.

AGCN

GCN的层与层的传播是不变(static)的,并不会考虑到节点的状态(adaptive propagation)。
比如:P_{i,j} = 1 /\sqrt[2]{|N(i)|N(j)},无法知道哪个邻接的节点与分类的节点更有关。
Embedding_layer: H^1 = ReLU(XW^0)
Attention-guided propagation layers: H^{t+1} = P^tH^t
Output row-vector of node i: H_i^{t+1} = \sum_{j∈N(i)} P_{i,j}^tH_j^t
其中:
P_i^t = softmax([β^tcos(H_i^t,H_j^t)]_{j∈N(i)})
cos(x,y) = x^Ty/||x||||y||, ||x||_{l^2}
Attention from node j to node I is: P_{ij}^t = (1/C)e^{β^tcos(H_i^t,H_j^t)}
Add self loop in propagation:
Z = f(X,A) = softmax(H^{l+1}W^1)

上一篇 下一篇

猜你喜欢

热点阅读