README.md

metadata

license: mit
tags:
  - icml2026-repro
  - paper-YhEWC2HiJl
  - federated-graph-learning
  - graph-condensation
  - text-attributed-graphs

DANCE reproduction bundle

Reproduction of DANCE: Dynamic, Available, Neighbor-gated Condensation for Federated Text-Attributed Graphs (ICML 2026, OpenReview YhEWC2HiJl, arXiv 2601.16519).

Scope: local, single-machine (Apple M1 Pro). HF Jobs was credit-blocked (402), so the federated 8-dataset + LLM-surrogate setting (Table 1) is out of scope; we verify the mechanism, theory, and complexity claims, plus a single-dataset (Cora) proxy.

Contents

• scripts/dance_core.py — DANCE mechanisms (node condensation, budgeted neighbor gating, sparse self-expressive topology, hard-budget projection).
• scripts/dance_model.py — DANCE-lite GCN + chunk gating + fusion classifier (Sec 4.3-4.4).
• scripts/claim4_theorems.py — Theorem 5.4 & 5.6 numerical verification.
• scripts/claim5_complexity.py — O(Kq) edges, time scaling, frozen-vs-LLM cost.
• scripts/claim23_cora.py — Claim 2 architecture invariants + Claim 3 deletion/insertion evidence tests on Cora.
• scripts/claim1_condensation.py — Claim 1 accuracy/token proxy on Cora.
• scripts/make_figures.py — figures.
• outputs/ — result JSONs + figures.

Rerun

uv venv --python 3.11 .venv && source .venv/bin/activate
uv pip install torch numpy scipy scikit-learn matplotlib torch_geometric
python scripts/claim4_theorems.py      # Claim 4
python scripts/claim5_complexity.py    # Claim 5
python scripts/claim23_cora.py         # Claims 2 & 3 (downloads Cora via PyG)
python scripts/claim1_condensation.py  # Claim 1
python scripts/make_figures.py