GraphEdgeNormalization {SMRUCC.genomics.Analysis.GEARS.Graph}

.NET clr documentation

GraphEdgeNormalization

Description

图卷积的边权重归一化方式。

三种方式只改变"每跳聚合的幅度"，边与边之间的相对强弱完全一致：

• GraphEdgeNormalization.Mean：除以自环与入边权重之和，得到凸组合平均；梯度最稳，

但单点扰动经过 L 跳后幅度大致按 (1/度)^L 衰减。

• GraphEdgeNormalization.Sqrt：除以平方根的同样总和，幅度按 (1/√度)^L 衰减；

在入度 20~70 的真实先验图上相当于每跳放大 5~8 倍。

• GraphEdgeNormalization.Sum：完全不归一化，幅度不减，但高度数节点的激活更容易饱和。

Declare

            
# namespace SMRUCC.genomics.Analysis.GEARS.Graph
export class GraphEdgeNormalization extends Enum {
   # 凸组合平均（GCN 行归一化；GEARS 与旧模型的默认行为）
   Mean: GraphEdgeNormalization = 0;

   # 平方根归一化（GPS 默认，保留更多跨基因传播幅度）
   Sqrt: GraphEdgeNormalization = 1;

   # 求和聚合（GIN 式，传播幅度最强）
   Sum: GraphEdgeNormalization = 2;

}

        

.NET clr type reference tree

1. this class extends from Enum class: Enum
2. use by field member Mean: GraphEdgeNormalization
3. use by field member Sqrt: GraphEdgeNormalization
4. use by field member Sum: GraphEdgeNormalization

---