当前位置:网站首页>【GCN】《Adaptive Propagation Graph Convolutional Network》(TNNLS 2020)
【GCN】《Adaptive Propagation Graph Convolutional Network》(TNNLS 2020)
2022-07-25 11:11:00 【chad_lee】
《Adaptive Propagation Graph Convolutional Network》(TNNLS 2020)
为每个节点赋予一个停止单元,该单元输出一个值控制Propagation是否应该继续进行下一跳。聚合时停止单元的输出值就是聚合每跳的权重。可以理解成为每个节点找到自己的感受野。
首先节点的特征先经过一个MLP变成embedding,这就是Propagation的起点,然后开始递归的Propagation:
z i 0 = z i z i 1 = propagate ( { z j 0 ∣ j ∈ N i } ) z i 2 = propagate ( { z j 1 ∣ j ∈ N i } ) . . . . . . \begin{aligned} &\mathbf{z}_{i}^{0}=\mathbf{z}_{i}\\ &\mathbf{z}_{i}^{1}=\operatorname{propagate}\left(\left\{\mathbf{z}_{j}^{0} \mid j \in \mathcal{N}_{i}\right\}\right)\\ &\mathbf{z}_{i}^{2}=\operatorname{propagate}\left(\left\{\mathbf{z}_{j}^{1} \mid j \in \mathcal{N}_{i}\right\}\right) \\ &...... \end{aligned} zi0=zizi1=propagate({ zj0∣j∈Ni})zi2=propagate({ zj1∣j∈Ni})......
传播的步数应该由每个节点自身决定的,因此给每个节点附加一个线性二分类器作为传播过程的“停止单元”。经过k次迭代传播后的输出:
h i k = σ ( Q z i k + q ) h_{i}^{k}=\sigma\left(\mathbf{Q} \mathbf{z}_{i}^{k}+q\right) hik=σ(Qzik+q)
其中 Q , q Q,q Q,q 是训练参数, h i k h_{i}^{k} hik 是该节点当前迭代应该停止的概率(0~1)。为了确保传播步数合理,有两个技巧:规定一个最大步数 T T T;用halting values的加和来定义Propagation的边界:
K i = min { k ′ : ∑ k = 1 k ′ h i k > = 1 − ϵ } K_{i}=\min \left\{k^{\prime}: \sum_{k=1}^{k^{\prime}} h_{i}^{k}>=1-\epsilon\right\} Ki=min⎩⎨⎧k′:k=1∑k′hik>=1−ϵ⎭⎬⎫
其中 ϵ \epsilon ϵ 是通常设置为一个很小的值0.01,保证传播一次之后也可以终止。对于节点 i 的第 k 轮迭代,当 k = K i k=K_{i} k=Ki 时Propagation停止。
节点 i 每一次迭代的停止概率为:
p i k = { R i = 1 − ∑ k = 1 K i − 1 h i k , if k = K i or k = T ∑ k = 1 K i h i k , otherwise. p_{i}^{k}= \begin{cases}R_{i}=1-\sum_{k=1}^{K_{i}-1} h_{i}^{k}, & \text { if } k=K_{i} \text { or } k=T \\ \sum_{k=1}^{K_{i}} h_{i}^{k}, & \text { otherwise. }\end{cases} pik={ Ri=1−∑k=1Ki−1hik,∑k=1Kihik, if k=Ki or k=T otherwise.
很自然的可以用它来作为节点聚合每层embedding的权重:
z ^ i = 1 K i ∑ k = 1 K i p i k z i k + ( 1 − p i k ) z i k − 1 \widehat{\mathbf{z}}_{i}=\frac{1}{K_{i}} \sum_{k=1}^{K_{i}} p_{i}^{k} \mathbf{z}_{i}^{k}+\left(1-p_{i}^{k}\right) \mathbf{z}_{i}^{k-1} zi=Ki1k=1∑Kipikzik+(1−pik)zik−1
还定义了节点 i 的 propagation cost:
S i = K i + R i \mathcal{S}_{i}=K_{i}+R_{i} Si=Ki+Ri
最终loss有监督信号和惩罚正则化项构成:
L ^ = L + α ∑ i ∈ V S i \widehat{\mathcal{L}}=\mathcal{L}+\alpha \sum_{i \in \mathcal{V}} \mathcal{S}_{i} L=L+αi∈V∑Si
这个惩罚项控制了信息在图上传播的“难以程度”。每5个step优化一次惩罚项。
AP-GCN学出的停止步数分布。看起来符合直觉:稀疏的图感受野一般更大,稠密的图一般只聚合1~2阶邻居。
边栏推荐
- Plot ==pyqt5
- 30 sets of Chinese style ppt/ creative ppt templates
- Miidock Brief
- Learning to Pre-train Graph Neural Networks(图预训练与微调差异)
- 擎创科技加入龙蜥社区,共建智能运维平台新生态
- W5500 adjusts the brightness of LED light band through upper computer control
- PHP one server sends pictures to another. Curl post file_ get_ Contents save pictures
- JS中的数组
- Pycharm connects to the remote server SSH -u reports an error: no such file or directory
- 奉劝那些刚参加工作的学弟学妹们:要想进大厂,这些并发编程知识是你必须要掌握的!完整学习路线!!(建议收藏)
猜你喜欢

创新突破!亚信科技助力中国移动某省完成核心账务数据库自主可控改造

Management of software defects

什么是全局事件总线?
![[MySQL learning 08]](/img/9e/6e5f0c4c956ca8dc31d82560262013.png)
[MySQL learning 08]

Review in the middle of 2022 | understand the latest progress of pre training model
![[imx6ull notes] - a preliminary exploration of the underlying driver of the kernel](/img/0f/a0139be99c61fde08e73a5be6d6b4c.png)
[imx6ull notes] - a preliminary exploration of the underlying driver of the kernel
![[USB device design] - composite device, dual hid high-speed (64BYTE and 1024byte)](/img/ce/534834c53c72a53fd62ff72a1d3b39.png)
[USB device design] - composite device, dual hid high-speed (64BYTE and 1024byte)

Intelligent information retrieval(智能信息检索综述)

Differences in usage between tostring() and new string()

What is the global event bus?
随机推荐
PHP curl post length required error setting header header
JS中的函数
Teach you how to configure S2E as the working mode of TCP client through MCU
软件缺陷的管理
Innovation and breakthrough! AsiaInfo technology helped a province of China Mobile complete the independent and controllable transformation of its core accounting database
[high concurrency] I summarized the best learning route of concurrent programming with 10 diagrams!! (recommended Collection)
图神经网络用于推荐系统问题(IMP-GCN,LR-GCN)
Hardware peripherals =maixpy3
油猴脚本链接
Experimental reproduction of image classification (reasoning only) based on caffe resnet-50 network
[MySQL learning 08]
W5500 adjusts the brightness of LED light band through upper computer control
【6篇文章串讲ScalableGNN】围绕WWW 2022 best paper《PaSca》
Classification parameter stack of JS common built-in object data types
擎创科技加入龙蜥社区,共建智能运维平台新生态
【GCN-RS】Region or Global? A Principle for Negative Sampling in Graph-based Recommendation (TKDE‘22)
【USB设备设计】--复合设备,双HID高速(64Byte 和 1024Byte)
session和cookie有什么区别??小白来告诉你
创新突破!亚信科技助力中国移动某省完成核心账务数据库自主可控改造
Plot ==pyqt5