Research map / Graph learning / Graph algorithms and clustering

Where Root Cause Analysis Fails: A Retrieval-Reranking Decomposition

NeurIPS 2026 · Graph learning › Graph algorithms and clustering · approach: Continuous DAG structure learning · 0 citations (OpenAlex)

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.

Decomposes root cause analysis evaluation into retrieval and reranking failures, audits four benchmarks, and builds a multi-signal retriever plus single-call LLM reranker (gpt-oss-120b) that matches or beats statistical and graph baselines on top@1.

Problem

Top@k accuracy for root cause analysis conflates failures to retrieve the true cause into a candidate set with failures to rank it highly, hiding where methods break, especially on propagating-fault industrial CPS benchmarks.

Approach

A retrieval-reranking decomposition with two metrics (Retrieval@K and Rerank@k) that separately measure candidate coverage and ordering; a multi-signal retriever (magnitude, earliest onset, discrete state-change) feeds a single-call LLM reranker over per-candidate evidence with optional domain-knowledge document.

Key results

• On WADI the LLM reranker reaches top@1 0.333±0.082 vs BARO 0.214, and on RE1-TT 0.653 without DK and 0.685 with DK vs BARO 0.560.
• Magnitude-only retrieval caps Retrieval@15 at 0.35-0.64 on propagating-fault CPS benchmarks; the multi-signal retriever lifts it to 0.63-0.67.
• Under retrieval-controlled pools with domain knowledge, the LLM leads the best same-pool baseline by +6.7 to +18.1 pp on every benchmark.

Surprising findings

• Graph-based methods never clearly beat the best statistical baseline on industrial CPS benchmarks, even when graphs are learned on candidate pools guaranteed to contain the cause.
• Domain knowledge lowers end-to-end top@1 on SWaT (-6.5 pp) while raising top@1 under controlled retrieval, because its benefit falls on causes the retriever misses.

Limitations

• Propagating-fault benchmarks are small (n=14 WADI, 36 SWaT, 48 HVAC) and cover only water and building systems.
• CPS gains assume detection timestamps accurate to better than 5 minutes.
• The LLM reranker is nondeterministic and retriever settings were tuned on the same benchmarks.

Experimental setup

Datasets: WADI SWaT HVAC (LBNL FDD ORNL RTU) RCAEval RE1

Benchmarks: WADI SWaT HVAC RCAEval RE1-OB RCAEval RE1-SS RCAEval RE1-TT

Metrics: top@1 top@3 top@5 Avg@5 Retrieval@K Rerank@k

Baselines: BARO RCD ε-Diagnosis CIRCA (PC) CIRCA (FCI) PageRank (PC) PageRank (FCI) RandomWalk (PC) RandomWalk (FCI)

Builds on: RankGPT PC algorithm FCI CIRCA BARO RCD PageRank

Closest papers by meaning

• Root Cause Analysis of Outliers with Missing Structural Knowledge (NeurIPS 2025 · similarity 0.86)
• Robust Root Cause Diagnosis using In-Distribution Interventions (ICLR 2025 · similarity 0.85)
• Root Cause Analysis of Measurement and Mechanistic Anomalies (NeurIPS 2026 · similarity 0.84)
• PRIM: Meta-Learned Bayesian Root Cause Analysis (NeurIPS 2026 · similarity 0.82)
• DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules (NeurIPS 2026 · similarity 0.82)
• Root Cause Analysis in Microservice Using Neural Granger Causal Discovery (AAAI 2024 · similarity 0.82)