USL10 Paper Portfolio
USL-05 Discovery · Transport AAAI-27 abstract registered · forum linked public contract: drafting

Causal Discovery under Unit Mixtures: Composition Shift or Mechanism Change?

When a distribution changes across environments, did the unit composition change or did unit-conditioned mechanisms change?

venue AAAI-27 Main Track topic ML: Causal Learning state abstract registered / writing owner gong · CausaClaw × DiscoSeed OpenReview forum ↗

基本研究问题

这篇论文独立可证伪的问题是什么；它与共享 ontology 中其他九篇不重叠的部分。

When a distribution changes across environments, did the unit composition change or did unit-conditioned mechanisms change?

核心 claim
Environment-conditioned Population laws $P(U\in du\mid E=e)$ can create, erase or reverse pooled dependencies while separately declared fixed-$u$ causal kernels remain unchanged.
如果成立会改变什么
Discovery under nonstationarity needs unit-aligned comparison before interpreting distribution shift as mechanism change.
最强 reviewer 反对 / kill signal

反对：This is standard latent-mixture confounding or Simpson's paradox.

Kill signal：kill if alignment information does not change any discovery decision, if soft $Q_\phi$ is silently treated as true grouping, or if distinguishability requires observed unit labels throughout.

基本思路

论文如何回答这个问题：形式化对象、论证 spine 与所需证据。

USL-01 type contract

The Unit Selection Variable $U\sim\Pi$ performs individual selection under the task-declared Population law $\Pi$. The event $E=e$ conditions the same Population world, so composition is written directly as $P(U\in du\mid E=e)$. This unit-Population conditional is not automatically the observed-row mixture: inclusion and per-individual observation intensity can induce an event-weighted law. Empirical claims must declare that process, repeated-event dependence and whether the target is unit-weighted or event-weighted. A fixed-individual causal or observational mechanism must be typed separately, for example $\kappa_{j,e}(dv_j\mid pa_j;u)$, and can remain stochastic after $u$ is fixed. Equality of mechanisms for distinct $u$ and $v$ does not make them the same individual. Observed attribution means the protocol reliably reveals $u$; inferred alignment means evidence $\mathcal O^{\mathrm{align}}$ conditions the world law $P(U\in du\mid E=e,\mathcal O^{\mathrm{align}})$ and the learner forms $Q_\phi(du\mid E=e,\mathcal O^{\mathrm{align}}) \approx P(U\in du\mid E=e,\mathcal O^{\mathrm{align}})$. Soft beliefs are not oracle groups. Cross-environment same-individual linkage and repeated-observation dependence require an explicit assignment/coupling law rather than an IID default.

论证 spine

  1. Why environment shift is not synonymous with mechanism shift.
  2. Environment-conditioned Population law $P(U\in du\mid E=e)$ and separately typed fixed-$u$ mechanism kernels.
  3. Composition, mechanism and intervention shift operators.
  4. Identifiability with observed attribution, alignment-conditioned world law $P(U\in du\mid E=e,\mathcal O^{\mathrm{align}})$, learner approximation $Q_\phi(du\mid E=e,\mathcal O^{\mathrm{align}})\approx P(U\in du\mid E=e,\mathcal O^{\mathrm{align}})$ and absent alignment.
  5. Cross-environment assignment/coupling, edge-confidence and false-change diagnostics.
  6. Factorial SCM benchmark and discovery baselines.
  7. Failure cases under ambiguous individuals and granularity mismatch.

所需证据

  • At least one pooled edge creation/reversal with unchanged within-unit mechanisms.
  • Formal comparison with ICP and nonstationary discovery assumptions.
  • Ablation removing unit alignment.
  • Separate oracle-attribution and learner $Q_\phi(du\mid E=e,\mathcal O^{\mathrm{align}})$ evaluations against $P(U\in du\mid E=e,\mathcal O^{\mathrm{align}})$; a soft alignment belief must not be scored or used as a true group without justification.
  • An explicit cross-environment individual-assignment and repeated-observation law, plus a negative control where different individuals have identical mechanisms.
已注册 abstract（点击展开）

We distinguish distribution changes caused by different compositions of actual units from genuine changes in their causal mechanisms. Unit mixtures can create, remove, or reverse pooled dependencies even when every unit-level mechanism remains unchanged.

当前进展

状态只记录可验证 delta：claim、SOTA opponent、theorem、experiment、manuscript 或 owner decision。

当前状态

AAAI-27 abstract registered; full paper drafting.

下一写作门

state the environment-conditioned Population laws $P(U\in du\mid E=e)$ directly, together with fixed-$u$ mechanism kernels, observed-versus-inferred alignment, cross-environment assignment/coupling and the invariance/identifiability assumptions for each shift type.

下一证据门

produce the minimal three-factor benchmark and at least one edge reversal.

48h 最小实验

factorially vary $P(U\in\cdot\mid E=e)$ across environments, fixed-$u$ mechanisms and interventions in a controlled SCM; compare oracle observed attribution, soft inferred alignment and absent-alignment regimes against pooled and environment-aware discovery methods.

Closest SOTA opponents

invariant causal prediction; CD-NOD/nonstationary causal discovery; mixture and latent-class causal models.

canonical 文件

research-questions/USL05-causal-discovery-under-unit-mixtures/seed.md
papers/USL05-causal-discovery-under-unit-mixtures/paper.md

进展日志

2026-07-29 · AAAI-27 abstract registration 由 direct OpenReview forum link 验证；full paper drafting 启动；本页面建立为持续 review 面。

后续每次可验证 delta 追加在此；canonical 状态以 seed.md / paper.md / submission-ledger 为准，本页由 build_paper_pages.py 重新生成。