Voidly Atlas Contagion-Chain Classifier v1
Version: v1.0 | Trained: 2026-05-21 | License: CC BY 4.0
Given a trigger event in country A, predict P(follower country B has an event within N days). Pairwise temporal XGBoost classifier — substitutes for full parametric Hawkes (data too sparse for parametric kernels).
Eval
| Metric | Value |
|---|---|
auc_3d |
0.6457 |
auc_7d |
0.6697 |
auc_14d |
0.6920 |
n_train |
17622 |
n_test |
9306 |
n_significant_pairs |
92 |
n_trigger_countries |
34 |
Feature importance
horizon_days(importance 0.1292)month_sin(importance 0.1107)global_30d_events(importance 0.1047)month_cos(importance 0.0977)follower_base_rate(importance 0.0954)follower_90d_events(importance 0.0922)follower_enc(importance 0.0868)days_since_follower_last(importance 0.0799)follower_30d_events(importance 0.0645)same_auth_cluster(importance 0.0357)pair_coevents_365d(importance 0.0326)trigger_enc(importance 0.0274)trigger_30d_events(importance 0.0273)same_region(importance 0.0161)
Honest caveats
- Pairwise temporal classifier substitutes for full Hawkes (data too sparse for parametric kernels)
- did_follow=1 is statistical co-event, not causal — use /v1/sentinel/attribute for causality
- Top-pair P~0.99 estimates have sparse support (n=2 held-out triggers)
Reproducibility
python3 scripts/build-pairwise-contagion-features.py
Citation
@misc{voidly_voidly_contagion_chain_v1,
title = {Voidly Atlas: voidly-contagion-chain-v1 (v1)},
author = {Voidly},
year = {2026},
url = {https://huggingface.co/emperor-mew/voidly-contagion-chain-v1},
note = {Open censorship-research ML stack. CC BY 4.0.}
}