学习论文 "Edge-Labeling Graph Neural Network for Few-shot Learning"笔记

个人笔记对模型数学上的解读部分很大程度上受到这篇博客的启发与参考

Notation

$T=S\cup Q$,support set and query set, support set $S$ in each episode serves as the labeled training set
$x_i$ and y i ∈ { C 1 , . . . , C N } = C T ⊂ C y_i \in \{C_1,...,C_N\}=C_T \subset C yi​∈{ C1​,...,CN​}=CT​⊂C: $i$ th input data and its label, $C$ is the set of all
classes of either training or test dataset. $C_{train}\cap C_{test}=\varphi$
$G=(V;\xi ;T)$: the graph constructed with samples from the task$T$. V：node set; E: edge set;
$y_{ij}$ ground-truth edge-label, defined by the ground-truth node labels
e i j = { e i j d } d = 1 2 ∈ [ 0 , 1 ] 2 \mathbf e_{ij}=\{e_{ijd}\}^2_{d=1} \in [0,1]^2 eij​={ eijd​}d=12​∈[0,1]2: edge feature, representing the (normalized) strengths of the intra- and inter-class relations of the two connected nodes.
${\tilde{e}}_{ijd}$ 归一化的edge feature
$f_v^l$ node feature transformation network
$f_e^{;}$ metric network
${\hat{y}}_{ij}$ probability that the two nodes $V^{}$