Research map / Probabilistic and scientific modeling / Causal discovery
On the Identifiability and Estimation of Causal Location-Scale Noise Models
Alexander Immer, Christoph Schultheiss, Julia E Vogt, Bernhard Schölkopf, Peter Bühlmann, Alexander Marx
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace additive noise model estimator with location-scale noise model estimator (same feature maps/architecture) (ablation) | maximum likelihood and HSIC estimators · AN, ANs, LS, LSs, MNU synthetic datasets | accuracy: large gains on LS, LSs, MNU | Improved |
| use neural network instead of spline feature maps as estimator (ablation) | LOCI maximum likelihood · LSs vs synthetic LSNM benchmarks | accuracy: LSs 100 (NN) vs 86 (feature maps) at Table 2 for LOCI_M? Feature-map LOCI_M LSs=94 vs NN 100 | Improved |
| add HSIC independence test after residual estimation (RESIT-style) on top of LSNM estimator (ablation) | feature-map LOCI on real-world benchmarks · Tuebingen, Net, Multi | accuracy: Tuebingen 52 to 56, Net 75 to 84, Multi 66 to 66 (LOCI_M vs LOCI_H) | Mixed |
| add HSIC independence test to LSNM estimator (ablation) | feature-map LOCI on synthetic LSNM benchmarks · LS and LSs | accuracy: LS 94 to 85; LSs 94 to 53 | Hurt |
| use maximum likelihood instead of HSIC for decision (neural network) (ablation) | NN-LOCI on synthetic benchmarks · LS, LSs, MNU, AN, ANs | accuracy: NN-LOCI_M 100 on all vs NN-LOCI_H 95 on LS, 89 on LSs | Mixed |
| use neural network LSNM estimator with ML decision (NN-LOCI_M) vs baselines on the real-world benchmarks | cause-effect benchmarks · Tuebingen, SIM, Cha, Net, Multi (accuracy) | accuracy: NN-LOCI_M underperforms GRCI on Tuebingen (57 vs 82) and Cha (43 vs 70) while beating on Net-ish? Net 76 vs 85 | Mixed |
| switch from ANM to LSNM likelihood estimator in the ML decision (ablation) | ML-based cause-effect inference · LS dataset | accuracy: ANM_M ~3 points better than LOCI_M on LS | No clear effect |
| use concave natural-parameter Gaussian likelihood instead of IFGLS for heteroscedastic regression (ablation) | spline feature-map estimator · simulated sinusoidal heteroscedastic regression (small T) | KL divergence to true density | Improved |
Closest papers by meaning
- Causal Discovery via Transformed Low-Rank Quantile Surfaces (NeurIPS 2026 · similarity 0.86)
- Strong and Weak Identifiability of Optimization-based Causal Discovery in Non-linear Additive Noise Models (ICML 2025 · similarity 0.85)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.85)
- A Skewness-Based Criterion for Addressing Heteroscedastic Noise in Causal Discovery (ICLR 2025 · similarity 0.84)
- Consistency of Neural Causal Partial Identification (NeurIPS 2024 · similarity 0.84)
- It’s Hard to Be Normal: The Impact of Noise on Structure-agnostic Estimation (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
