Task-AdaptiveFew-shot Node Classification

推荐指数： #paper/⭐
发表于：KDD22
分为节点级，类级别和任务级

节点级

$$\mathrm{H}={\mathrm{GNN}}_{\varphi }(\mathcal{V},\mathcal{E},\mathrm{X}),$$
$${\mathbf{h}}_{c_i}=\mathrm{MEAN}({\mathbf{h}}_v|v\in {\mathcal{S}}_i),$$

类级

α i = M L P α ( M E A N ( { h v ∣ v ∈ S i } ) ) , β i = M L P β ( M E A N ( { h v ∣ v ∈ S i } ) ) , \alpha_{i}=\mathrm{MLP}_{\alpha}\left(\mathrm{MEAN}\left(\{\mathrm{h}_{v}|v\in\mathcal{S}_{i}\}\right)\right),\\\beta_{i}=\mathrm{MLP}_{\beta}\left(\mathrm{MEAN}\left(\{\mathrm{h}_{v}|v\in\mathcal{S}_{i}\}\right)\right), αi​=MLPα​(MEAN({hv​∣v∈Si​})),βi​=MLPβ​(MEAN({hv​∣v∈Si​})),
$${\theta }_i=({\alpha }_i+1)\circ \theta +{\beta }_i,$$
$${\mathbf{s}}_i=\mathrm{Centroid}\left({\mathrm{GNN}}_{{\theta }_i}({\mathcal{V}}_i,{\mathcal{E}}_i,{\mathrm{X}}_i)\right),$$

任务级

$${\mathcal{L}}_N=-\sum _{i=1}^Q\log \frac{\exp ({\mathbf{q}}_i\cdot {\mathbf{s}}_i^{\prime }/{\tau }_i^{\prime })}{{\sum }_{j=1}^N\exp ({\mathbf{q}}_i\cdot {\mathbf{s}}_j/{\tau }_j)}.$$

最终损失

$$\begin{gathered} {\mathbf{p}}_i=\mathrm{Softmax}\left(\mathrm{MLP}({\mathbf{h}}_i)\right), \\ {\mathcal{L}}_{CE}=-\sum _{i=1}^Q\sum _{j=1}^{|C_b|}{y}_{i,j}\log p_{i,j}, \end{gathered}$$
$$\mathcal{L}={\mathcal{L}}_N+\gamma {\mathcal{L}}_{CE},$$