Research map / Probabilistic and scientific modeling / Causal discovery
Robust Root Cause Diagnosis using In-Distribution Interventions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes In-Distribution Interventions (IDI), a root cause diagnosis method that checks the fix condition by probing a fitted SCM only on in-distribution inputs, with theory bounding interventional vs counterfactual errors.
Problem
Root cause diagnosis (RCD) in interconnected cloud systems requires assessing the fix condition, but counterfactuals from an SCM fitted on usual data are unreliable because anomalous abduction falls outside the training distribution.
Approach
Combine an anomaly condition filter with a fix check that uses interventional estimates by sampling exogenous noise from in-distribution residuals rather than abducting anomalous noise; use Shapley values over candidate subsets to handle multiple root causes.
Key results
- On PetShop latency, IDI reaches Recall@1 of 0.90 (low), 0.90 (high), and 1.00 (temporal), beating all nine baselines at k=1.
- On synthetic non-linear non-invertible SCMs with unique root cause, IDI achieves Recall@1 of 0.70 versus 0.40 for the best baseline (smooth traversal) ... reported 0.60 for IDI in main table.
- Theory: counterfactual error bound scales with M times total variation distance between training and root-cause distributions, while interventional error is bounded by standard deviation of exogenous noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IDI's in-distribution interventions rather than counterfactuals in step 2 (ablation) | IDI · Synthetic RQ2 Non-Lin Invertible, multiple root causes | Recall@1: 0.77 -> 0.83 | Helped |
| Use IDI's in-distribution interventions rather than counterfactuals in step 2 (ablation) | IDI · Synthetic RQ3 Non-Lin Non-Invertible, multiple root causes | Recall@1: 0.12 -> 0.63 | Helped |
| Use IDI's in-distribution interventions rather than counterfactuals in step 2 (ablation) | IDI · Synthetic RQ1 Linear, unique root cause | Recall@1: 1.00 -> 1.00 | No clear effect |
| Use IDI's in-distribution interventions rather than counterfactuals in step 2 (ablation) | IDI · PetShop Latency, low | Recall@1: 0.70 -> 0.90 | Helped |
| Use IDI's in-distribution interventions rather than counterfactuals in step 2 (ablation) | IDI · PetShop Latency, high | Recall@1: 0.70 -> 0.90 | Helped |
| Use IDI's in-distribution interventions rather than counterfactuals in step 2 (ablation) | IDI · PetShop Latency, temporal | Recall@1: 1.00 -> 1.00 | No clear effect |
| Use IDI rather than the strongest baseline | RCD on PetShop Low latency · PetShop Low latency | Recall@1: 0.80 -> 0.90 | Helped |
| Use IDI rather than the strongest baseline | RCD on PetShop High latency · PetShop High latency | Recall@3: 1.00 -> 0.90 | Hurt |
| Use IDI rather than the strongest baseline | RCD on Synthetic RQ2 Non-Lin Invertible, unique root cause · Synthetic RQ2 Non-Lin Invertible, unique root cause | Recall@1: 0.60 -> 1.00 | Helped |
| Use IDI rather than the strongest baseline | RCD on Synthetic RQ3 Non-Lin Non-Invertible, multiple root causes · Synthetic RQ3 Non-Lin Non-Invertible, multiple root causes | Recall@3: 0.73 -> 0.83 | Helped |
Surprising findings
- CIRCA outperforms IDI at Recall@3 on high latency in PetShop (1.00 vs 0.90).
- Counterfactual-based IDI variant is strong only in linear SCMs; interventions are better for non-linear non-invertible cases.
Limitations
- Performance degrades when the at-most-one-root-cause-per-path assumption is violated.
- Interventional error plateaus at the standard deviation of exogenous variables even with large training data, while counterfactual estimates converge.
- Assumes additive noise for the theoretical bounds; the non-invertible case relies on empirical evidence.
Experimental setup
Datasets: PetShop RCD benchmark synthetic SCMs generated via NetworkX causal graphs toy four-variable additive noise dataset
Benchmarks: PetShop RCD (low, high, temporal latency) synthetic linear, non-linear invertible, non-linear non-invertible SCMs toy four-variable dataset
Metrics: Recall@1 Recall@3 validation error counterfactual error interventional error
Baselines: Random Walk Ranked Correlation epsilon-Diagnosis Circa Traversal Smooth Traversal HRCD TOCA CF Attribution
Builds on: structural causal model counterfactual root cause analysis (SAGE) CIRCA PetShop RCD library
Closest papers by meaning
- Root Cause Analysis of Outliers with Missing Structural Knowledge (NeurIPS 2025 · similarity 0.86)
- Root Cause Analysis of Measurement and Mechanistic Anomalies (NeurIPS 2026 · similarity 0.86)
- Where Root Cause Analysis Fails: A Retrieval-Reranking Decomposition (NeurIPS 2026 · similarity 0.85)
- PRIM: Meta-Learned Bayesian Root Cause Analysis (NeurIPS 2026 · similarity 0.83)
- Root Cause Analysis in Microservice Using Neural Granger Causal Discovery (AAAI 2024 · similarity 0.82)
- When and How Does In-Distribution Label Help Out-of-Distribution Detection? (ICML 2024 · similarity 0.79)
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.