Non-Homophilic Graph Pre-Training and Prompt Learning
KDD25
#paper/⭐#
目的:对异配图进行prompt
方法
邻居节点的综合嵌入
s v = 1 ∣ V ( S v ) ∣ ∑ u ∈ V ( S v ) h u ⋅ s i m ( h u , h v ) , \mathbf{s}_{v}=\frac{1}{|V(S_{v})|}\sum_{u\in V(S_{v})}\mathbf{h}_{u}\cdot\mathrm{sim}(\mathbf{h}_{u},\mathbf{h}_{v}), sv=∣V(Sv)∣1∑u∈V(Sv)hu⋅sim(hu,hv),
在实践中,作者使用了2hop来表示邻域信息
邻域信息映射
p t , v = C o n d N e t ( s v ; ϕ t ) \mathbf{p}_{t,v}=\mathrm{CondNet}(s_{v};\phi_{t}) pt,v=CondNet(sv;ϕt),这里,condnet是一层MLP
h ~ t , v = p t , v ⊙ h v , \tilde{\mathbf{h}}_{t,v}=\mathbf{p}_{t,v}\odot\mathbf{h}_{v}, h~t,v=pt,v⊙hv,
与类prototype做prompt
L d o w n ( ϕ t ) = − ∑ ( x i , y i ) ∈ D t ln exp ( 1 τ sin ( h ~ t , x i , h ˉ t , y i ) ) ∑ c ∈ Y exp ( 1 τ sin ( h ~ t , x i , h ˉ t , c ) ) \mathcal{L}_{\mathrm{down}}(\phi_{t})=-\sum_{(x_{i},y_{i})\in\mathcal{D}_{t}}\ln\frac{\exp\left(\frac{1}{\tau}\sin(\tilde{\mathbf{h}}_{t,x_{i}},\bar{\mathbf{h}}_{t,y_{i}})\right)}{\sum_{c\in Y}\exp\left(\frac{1}{\tau}\sin(\tilde{\mathbf{h}}_{t,x_{i}},\bar{\mathbf{h}}_{t,c})\right)} Ldown(ϕt)=−∑(xi,yi)∈Dtln∑c∈Yexp(τ1sin(h~t,xi,hˉt,c))exp(τ1sin(h~t,xi,hˉt,yi))
其中, h t , c h_{t,c} ht,c代表属于类C的平均值。
结果: