emperor-mew/voidly-forecast-asn-gnn-v1

Voidly Atlas ASN-GNN v1 (GraphSAGE-2layer)

Version: v1 | Trained: 2026-05-22T01:41:08.267598 | License: CC BY 4.0

GraphSAGE-2layer, hidden_dim=16, dropout 0.5, 60 epochs. Trained on the May 2026 CAIDA serial-2 AS-AS peering graph (7,060 nodes, 841K edges). 58 labeled ASNs (40 positive). Predicts had_shutdown_next_7d per ASN.

Eval (leave-one-out across 6 tier-1 ASNs)

Metric	Value
loocv_auc	0.8000
loocv_accuracy_at_0.5	0.8333
graph_n_nodes	7060
graph_n_edges	841064
hidden_dim	16
epochs	60
architecture	GraphSAGE-2layer-h64

Features

• n_evidence_30d
• n_evidence_180d
• block_rate_30d
• block_rate_180d
• n_unique_dates
• pct_dns_block
• pct_tcp_reset
• pct_outage
• pct_interference
• pct_block
• country_risk_tier
• log_degree
• has_evidence

Honest caveats

• Only 6 tier-1 ASNs have density >= 100 evidence rows AND >= 30 unique days. Leave-one-out CV across 6 folds is barely statistical — every miss is 16% accuracy.
• Training set is the OTHER 52 labeled ASNs (most have < 30 unique days). Label is had_shutdown_next_7d from the same evidence table — there's leakage potential via the country-risk and topological neighbors.
• Per-ASN forecasting is data-sparsity-bound, not model-architecture-bound. Even a perfect GNN is limited by how few ASNs we have ground truth for.

Citation

@misc{voidly_voidly_forecast_asn_gnn_v1,
  title  = {Voidly Atlas: voidly-forecast-asn-gnn-v1 (v1)},
  author = {Voidly},
  year   = {2026},
  url    = {https://huggingface.co/emperor-mew/voidly-forecast-asn-gnn-v1},
  note   = {Open censorship-research ML stack. CC BY 4.0.}
}

Method foundation: Hamilton et al. 2017 — Inductive Representation Learning on Large Graphs (GraphSAGE)