Category: A1; Team name: louisvl; Dataset: GraphUniverse

[LouisVanLangendonck]

Description

GraphUniverse [0] is a synthetic, generative graph dataset framework that creates families of graphs with controllable structural properties. Unlike traditional static datasets, GraphUniverse takes generation parameters as input to produce entire families of graphs with consistent semantic communities across different graph instances. This enables systematic evaluation of inductive generalization - how models perform on completely unseen graphs rather than just unseen parts of the same graph.

The framework generates graphs using an extended Degree-Corrected Stochastic Block Model (DC-SBM) with three hierarchical levels: Universe (global community properties), Family (generation constraints), and Graph (individual instances). Users specify parameters like homophily ranges, degree distributions, node counts, and community structures to generate graph families on-demand.

Two task configurations included:

• Community Detection (node classification)
• Triangle Counting (graph-level regression)

[0] Anonymous (2025). GraphUniverse: Enabling Systematic Evaluation of Inductive Generalization. In Submitted to The Fourteenth International Conference on Learning Representations.

Caveat

This PR includes modifications to TopoBench's splitting utilities because GraphUniverse community-detection task is an inductive, node classification tasks that were not previously supported.

Changes made:

• Modified split_utils.py to detect and handle inductive node classification
• Fixed data loading to properly batch multiple graphs with node-level targets
• Updated dataset configurations to specify task_level: node for proper splitting