Localized persistent homology features can make graph neural networks more expressive and slightly more accurate, but the state-of-the-art claim is not uniformly supported.
A limit theorem for the $1$st Betti number of layer-$1$ subgraphs in random graphs
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
We initiate the study of local topology of random graphs. The high level goal is to characterize local "motifs" in graphs. In this paper, we consider what we call the layer-$r$ subgraphs for an input graph $G = (V,E)$: Specifically, the layer-$r$ subgraph at vertex $u \in V$, denoted by $G_{u; r}$, is the induced subgraph of $G$ over vertex set $\Delta_{u}^{r}:= \left\{v \in V: d_G(u,v) = r \right\}$, where $d_G$ is shortest-path distance in $G$. Viewing a graph as a 1-dimensional simplicial complex, we then aim to study the $1$st Betti number of such subgraphs. Our main result is that the $1$st Betti number of layer-$1$ subgraphs in Erd\H{o}s--R\'enyi random graphs $G(n,p)$ satisfies a central limit theorem.
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Enhancing Graph Representation Learning with Localized Topological Features
Localized persistent homology features can make graph neural networks more expressive and slightly more accurate, but the state-of-the-art claim is not uniformly supported.