Spectral-based graph convolutional neural network

这里写自定义目录标题

  • Spectral-based graph convolutional neural network
    • 1. 来自对《[A Comprehensive Survey on Graph Neural Networks](https://arxiv.org/abs/1901.00596?context=cs)》的spectral-based GCN部分的翻译
      • 1.1 图卷积的背景(background)
      • 1.2 图卷积的方法
        • 1.2.1 Spectral CNN
        • 1.2.2 Chebyshev Spectral CNN (chebNet)
        • 1.2.3 First order of chebNet (1stChebNet)
    • Relate Papers
    • 参考文献

Spectral-based graph convolutional neural network

1. 来自对《A Comprehensive Survey on Graph Neural Networks》的spectral-based GCN部分的翻译

1.1 图卷积的背景(background)

Spectral-based方法在图信号(graph signal processing)处理中已经有了一个非常好的基础。在Spectral-based的模型中,图通常被假定为无向图。对无向图比较鲁棒的表示方法是归一化的拉普拉斯矩阵(Normalized graph Laplacian matrix),归一化的拉普拉斯矩阵L被定义为:
L = I n − D − 1 / 2 A D − 1 / 2 , L=I_{n}-D^{-1/2}AD^{-1/2}, L=InD1/2AD1/2,
其中D是图的对角度数矩阵,A为图的邻接矩阵,所以有:
D i i = ∑ j ( A i , j ) . D_{ii}=\sum_{j}(A_{i,j}). Dii=j(Ai,j).
归一化的图的拉普拉斯矩阵L是一个实对称的半正定矩阵。根据这个性质,可以将拉普拉斯矩阵L分解为:
L = U Λ U T , 其 中 U = [ u 0 , u 1 , u 2 , . . . , u n − 1 ] ∈ R N × N L=U\Lambda U^{T}, 其中U=[u_{0},u_{1},u_{2},...,u_{n-1}] \in R^{N \times N} L=UΛUT,U=[u0,u1,u2,...,un1]RN×N
上述的表达式中,U是奇异值向量的矩阵。 Λ \Lambda Λ是奇异值的对角矩阵, Λ \Lambda Λ对角线上的每一个元素是一个奇异值 Λ i i = λ i \Lambda_{ii}=\lambda_{i} Λii=λi。U中的奇异值向量构成了一个正交的空间,因而有性质 U T U = I U^{T}U=I UTU=I。在图信号处理的过程当中。一个图信号 x ∈ R N x \in R^{N} xRN是一个特征向量,其中每一个元素 x i x_{i} xi是图中第i个结点的值。

有了上述的定义之后,可以定义图的傅里叶变换(graph Fourier transform):
x ‾ = F ( x ) = U T x \overline{x} = \mathscr{F}(x)=U^{T}x x=F(x)=UTx
图的傅里叶逆变换可以被定义为:
F − 1 ( x ‾ ) = U x ‾ \mathscr{F}^{-1}(\overline{x}) = U\overline{x} F1(x)=Ux
从图的傅里叶变换的定义中我们可以理解:

  1. 图的傅里叶变换是将输入的图信号x映射到由U中的奇异值向量所构成为的正交空间中去。
  2. 傅里叶变换的结果 x ‾ \overline{x} x是图信号在新的空间(由奇异值向量所构成的空间)对应的坐标。
  3. 据此,我们可以写出输入的图信号x在新的空间中的表示:
    x = ∑ i x ‾ i u i x =\sum_{i}\overline{x}_{i}u_{i} x=ixiui

根据图的傅里叶变换,可以定义图卷积,假设图的 输入信号为x,滤波器为 g ∈ R N g \in R^{N} gRN,图卷积可以被定义为:
x ∗ G g = F − 1 ( F ( x ) ⊙ F ( g ) ) = U ( U T x ⊙ U T g ) x \ast_{G} g = \mathscr{F^{-1}}( \mathscr{F}(x) \odot \mathscr{F}(g)) = U(U^{T}x \odot U^{T}g) xGg=F1(F(x)F(g))=U(UTxUTg)
其中 ⊙ \odot 表示Hadamard product. 如果换一种表示方法,把 g θ = d i a g ( U T g ) g_{\theta} = diag(U^{T}g) gθ=diag(UTg)表示为滤波器,图卷积可以被简化为:
x ∗ G g θ = U g θ U T x x \ast_{G} g_{\theta} = Ug_{\theta}U^{T}x xGgθ=UgθUTx

以上就是spectral-based的图卷积网络的定义。不同的网络结构在于对滤波器 g θ g_{\theta} gθ的选择。

1.2 图卷积的方法

1.2.1 Spectral CNN

Bruna 1大佬提出了第一代的图卷积神经网络(spectral convolutional neural network), 简称Spectral CNN. 论文中假设滤波器为 g θ = θ i , j k g_{\theta} = \theta^{k}_{i,j} gθ=θi,jk θ i , j k \theta^{k}_{i,j} θi,jk是一系列的可训练的参数,可以定义图卷积神经网络的一个layer为:
X : , j k + 1 = σ ( ∑ i = 1 f k − 1 U θ i , j k U T X : , i k ) ( j = 1 , 2 , . . . , f k ) X_{:,j}^{k+1} = \sigma( \sum_{i=1}^{f_{k-1}} U \theta^{k}_{i,j} U^{T}X_{:,i}^{k}) (j =1,2,...,f_{k}) X:,jk+1=σ(i=1fk1Uθi,jkUTX:,ik)(j=1,2,...,fk)
上述的表达式中:

  1. X k ∈ R N × f k − 1 X^{k} \in R^{N \times f_{k-1}} XkRN×fk1是当前层的输入
  2. N是图中结点的个数
  3. f k − 1 f_{k-1} fk1是输入channel的个数$f_{k}是输出channel的个数。
  4. θ i , j k \theta^{k}_{i,j} θi,jk是对角矩阵,矩阵对角线上的每一个值都是可学习的参数
  5. σ ( ) \sigma() σ()是激活非线性函数

1.2.2 Chebyshev Spectral CNN (chebNet)

Defferrard 2 大佬提出ChebNet. 在论文中,定义滤波器 g θ g_{\theta} gθ为一系列的切比雪夫(Chebyshev polynomials)多项式。例如:
g θ = ∑ i = 0 K − 1 θ i T i ( Λ ‾ ) g_{\theta}=\sum_{i=0}^{K-1}\theta_{i}T_{i}(\overline\Lambda) gθ=i=0K1θiTi(Λ)
其中,
Λ ‾ = 2 Λ / λ m a x − I N \overline\Lambda = 2\Lambda/\lambda_{max} - I_{N} Λ=2Λ/λmaxIN
切比雪夫多项式可以通过递归地定义为:
T k ( x ) = 2 x T k − 1 ( x ) − T k − 2 ( x ) T_{k}(x) = 2xT_{k-1}(x)-T_{k-2}(x) Tk(x)=2xTk1(x)Tk2(x) T 0 ( x ) = 1 T_{0}(x) = 1 T0(x)=1 T 1 ( x ) = x T_{1}(x) = x T1(x)=x
通过上述对滤波器 g θ g_{\theta} gθ的定义,可以得到基于切比雪夫多项式的图卷积的表示:
x ∗ G g θ = U ( ∑ i = 0 K − 1 θ i T i ( Λ ‾ ) ) U T x x \ast_{G} g_{\theta} = U(\sum_{i =0}^{K-1} \theta_{i}T_{i}(\overline\Lambda))U^{T}x xGgθ=U(i=0K1θiTi(Λ))UTx = ∑ i K − 1 θ i T i ( L ‾ ) x =\sum_{i}^{K-1}\theta_{i}T_{i}(\overline L)x =iK1θiTi(L)x
其中 L ‾ = 2 L m a x − I N \overline L = 2L_{max}-I_{N} L=2LmaxIN

通过基于切比雪夫多项式的图卷积的表示,我们可以发现,chebNet有一个优点,这个优点就是能够避免计算图傅里叶变换的基,从而将图卷积的计算复杂度从 O ( N ) O(N) O(N)减少为 O ( K M ) O(KM) O(KM)

1.2.3 First order of chebNet (1stChebNet)

Kipf 3 大佬提出了first-order approximation of ChebNet. 在Chebyshev Spectral CNN的基础上,假设 K = 1 K=1 K=1 λ m a x = 2 \lambda_{max}=2 λmax=2,从而基于切比雪夫多项式的图卷积可以被简化为:
x ∗ G g θ = θ 0 x − θ 1 D − 1 / 2 A D − 1 / 2 x x \ast_{G} g_{\theta} = \theta_{0}x - \theta_{1}D^{-1/2}AD^{-1/2}x xGgθ=θ0xθ1D1/2AD1/2x
进一步地,为了减少需要训练的参数,可以最进一步的假设 θ = θ 0 = θ 1 \theta = \theta_{0} = \theta_{1} θ=θ0=θ1,这样,卷积可以被进一步地化简为:
x ∗ G g θ = θ ( I θ + D − 1 / 2 A D − 1 / 2 ) x x \ast_{G} g_{\theta} = \theta(I_{\theta} +D^{-1/2}AD^{-1/2})x xGgθ=θ(Iθ+D1/2AD1/2)x
有了上述的表达式之后,可以根据First order of chebNet来定义图卷积神经网络中的一层:
X k + 1 = A ‾ X k Θ X^{k+1} =\overline A X^{k}\Theta Xk+1=AXkΘ
其中 A ‾ = I N + D − 1 / 2 A D − 1 / 2 \overline A=I_{N}+D^{-1/2}AD^{-1/2} A=IN+D1/2AD1/2


Relate Papers

  • M. Defferrard, X. Bresson, and P. Vandergheynst, “Convolutional neural networks on graphs with fast localized spectral filtering,” in Advances in Neural Information Processing Systems, 2016, pp. 3844–3852.
  • T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proceedings of the International Conference on Learning Representations, 2017.
  • M. Henaff, J. Bruna, and Y. LeCun, “Deep convolutional networks on graph-structured data,” arXiv preprint arXiv:1506.05163, 2015.
  • R. Li, S. Wang, F. Zhu, and J. Huang, “Adaptive graph convolutional neural networks,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2018, pp. 3546–3553.
  • R. Levie, F. Monti, X. Bresson, and M. M. Bronstein, “Cayleynets: Graph convolutional neural networks with complex rational spectral filters,” IEEE Transactions on Signal Processing, vol. 67, no. 1, pp. 97–109, 2017.

参考文献


  1. J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun, “Spectral networks and locally connected networks on graphs,” in Proceedings of International Conference on Learning Representations, 2014. ↩︎

  2. M. Defferrard, X. Bresson, and P. Vandergheynst, “Convolutional neural networks on graphs with fast localized spectral filtering,” in Advances in Neural Information Processing Systems, 2016, pp. 3844–3852. ↩︎

  3. T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proceedings of the International Conference on Learning Representations, 2017. ↩︎

你可能感兴趣的:(Spectral-based graph convolutional neural network)