Transductive Rewiring

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 2

Transductive Rewiring

Diffusive rewiring
tends to work better in homophilic graphs

Curvature
Cheeger constant

h G = min h S ,

squashing

Ric(i, j) =
S⊆V
hS =

If we consider edge curvature, we find that:

# △ (i, j)

#
i

(i, j)

zero curvature
cut(S, S̄)

λ2

2
¯
min(vol(S), vol(S))

Hard to calculate, but we can use spectral bounds:

≤ h G ≤ √ 2λ 2

and of course we can also bound the Fiedler eigenvalue with the Cheeger constant:

2h G ≤ λ 2 ≤
h

2
2
G
¯
cut(S, S) =

- Triangles based at edge (i, j) contribute to positive curvature

- Maximal number of 4-caycles based at (i, j) traversing a common node is


γ max (i, j)

also useful
We wish to avoid edges with very negative curvature - tree shape will lead to over-

A better metric for curvature is needed, that takes all this into account

Balanced Forman curvature

di
+
2

dj
− 2 + 2

Stochastic Discrete Ricci Flow (SDRF)

**Algorithm**
Ric(i, j) = 0

|# △ (i, j)|

max{d i , d j }
+
if min{d i , d j } = 1

|# △ (i, j)|

min{d i , d j }
+ ∣
- Neighbors of i forming a 4-cycle based on (i, j) without diagonals contribute to

(γ max )
−1

max{d i , d j }
i
¯
{(u, v) : u ∈ S, v ∈ S}

( # ■ (i, j) + #
j


(i, j) )
**Input:** graph $G$, temperatue $\tau > 0$, max number of iterations, optional Ric upper-
bound $C^+$
**Repeat**
1) For edge $i \sim j$ with minimal Ricci curvature $Ric(i,j)$:
+
Calculate vector $\mathbf{x}$ where $x_{kl} = Ric_{kl}(i,j) - Ric(i,j)$, the improvement to
$Ric(i,j)$ from adding edge $k \sim l$ where $k \in B_1(i), l\in B_1(j)$; Sample index $k, l$
with probability $softmax(\tau\mathbf{x})_{kl}$ and add edge $k \sim l$ to $G$. 2) Remove
edge $i \sim j$ with maximal Ricci curvature $Ric(i,j)$ if $Ric(i,j) > C^+$
**Until**
convergence, or max iterations reached;

You might also like