Loscar extends residual causal discovery to location-scale noise
Original title: Residual-Based Causal Discovery via Pretrained In-Context Quantile Regression
Pretrained in-context quantile regression beats feasible generalized least squares, especially when causes are multivariate.
The paper proposes Loscar, extending RESIT from additive-noise models to location-scale noise models by estimating conditional mean and scale with quantile regression. Pretrained tabular foundation models can then serve as plug-in residual estimators without parametric residual assumptions. On bivariate and multivariate benchmarks, in-context quantile regression substantially outperforms an FGLS variant.
- Existing LSNM methods use parametric assumptions that can misspecify residuals.
- RESIT was proposed for additive-noise models and finds sink nodes nonparametrically.
- The FGLS variant is consistent when suitable basis functions are used.
- Gains are largest for multivariate causes.
From the original article
1 min read at arxiv.orgResidual-based causal discovery assumes that each variable is generated from its parents and an independent noise source. It estimates this noise and tests its independence from candidate causes. For identifiable functions such as additive-noise models (ANMs) and the broader class of location-scale noise models (LSNMs), this principle identifies the causal direction, as only the causal model admits this independence.
By Daniel Klippert · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.