Research map / Probabilistic and scientific modeling / Causal discovery

PRIM: Meta-Learned Bayesian Root Cause Analysis

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Constraint-based structure learning with latents · 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.

Meta-learns a MACE transformer neural process on synthetic SCM priors to score root-cause nodes zero-shot from normal and anomalous data plus a symptom mask, with optional lightweight fine-tuning on domain simulators.

Problem

Root cause analysis in multivariate systems requires identifying which variables' mechanisms changed between normal and anomalous periods, but existing methods need a causal graph, costly causal discovery at test time, or statistical tests that scale poorly.

Approach

Frames RCA as amortised Bayesian inference: a prior-fitted transformer neural process (MACE-TNP) trained on synthetic causal models directly maps (observational data, interventional data, anomaly mask) to a distribution over root-cause nodes, implicitly marginalising graph and mechanism uncertainty and performing a distribution-shift comparison via mean-pooled differences.

Key results

• On CausRCA, PRIM-FT achieves average MAP@3 of 0.43 (full) and 0.77 (subsystem), the best among graph-not-given methods.
• On PetShop, PRIM-FT achieves average Recall@3 of 0.65, the highest among graph-not-given methods, versus 0.78 for Corr.
• PRIM inference takes about 17 ms on an A100 GPU for graphs up to 100 variables, constant across node sizes.

Surprising findings

• Zero-shot PRIM scores 0.00 Recall@3 on PetShop high-latency and zero on the CausRCA probe full setting, indicating prior misspecification.
• Graph-given methods such as Traversal and CIRCA beat PRIM-FT on several PetShop scenarios despite the PFN not using the graph.

Limitations

• Identifiability relies on acyclicity, causal sufficiency, I-faithfulness and known anomaly mask; violations weaken guarantees.
• Misspecified synthetic prior degrades accuracy on some setups (PetShop high-load, CausRCA probe).
• Cannot exploit a ground-truth causal graph when one is available.

Experimental setup

Datasets: PetShop CausRCA synthetic SCM prior microservice simulator (PetShop finetuning) synthetic factory simulator (CausRCA finetuning)

Benchmarks: PetShop CausRCA synthetic confounder/mediator/multi-root-cause/scalability setups

Metrics: Recall@k MAP@k inference time

Baselines: Traversal CIRCA Counterfactual attribution PageRank CausTR epsilon-diagnosis RCD Corr Baro Time Rank PC-Traversal PC-CIRCA

Builds on: MACE-TNP prior-fitted networks TabPFN CIRCA RCD

Closest papers by meaning

• A Meta-Learning Approach to Bayesian Causal Discovery (ICLR 2025 · similarity 0.88)
• Estimating Interventional Distributions with Uncertain Causal Graphs through Meta-Learning (NeurIPS 2025 · similarity 0.87)
• Root Cause Analysis of Outliers with Missing Structural Knowledge (NeurIPS 2025 · similarity 0.85)
• Identifying Causal Direction via Variational Bayesian Compression (ICML 2025 · similarity 0.85)
• CausalPFN: Amortized Causal Effect Estimation via In-Context Learning (NeurIPS 2025 · similarity 0.85)
• Learning Structural Causal Models from Ordering: Identifiable Flow Models (AAAI 2025 · similarity 0.85)