Research map / Probabilistic and scientific modeling / Causal discovery
Root Cause Analysis in Microservice Using Neural Granger Causal Discovery
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 RUN, a root cause analysis method for microservices that pre-trains a DLinear encoder with contrastive learning on same-timestamp positive pairs (SimSiam-style, no negatives), learns a Granger causal graph via per-variable forecasters with a learnable attention matrix, and ranks root causes with personalized PageRank.
Problem
Root cause analysis in microservice systems from multivariate KPI time series; prior causal-discovery methods (PC-algorithm based) ignore temporal order, and prior neural Granger methods do not exploit contextual information in time series.
Approach
Contrastive pre-training that treats only identical timestamps with different contexts as positives (no negatives, to avoid pushing apart periodically similar timestamps), followed by neural Granger causal discovery via a learnable causal attention matrix and PageRank with a personalization vector and modified access-distance tie-breaking.
Key results
- On sock-shop CPU hog, RUN achieves average HR@1 of 0.40 vs 0.28 for the best baseline (epsilon-Diagnosis), and HR@5 of 0.52 vs 0.32.
- On sock-shop memory leak, RUN average HR@3 is 0.32 vs 0.20 for epsilon-Diagnosis; RUN HR@1 is 0.12, second best behind epsilon-Diagnosis (0.20).
- Ablation shows removing the pre-training (contextual contrastive) stage notably reduces performance, while adding negative pairs did not substantially change it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RUN instead of ε-Diagnosis | Sock-shop CPU hog root cause analysis · Sock-shop, CPU hog | HR@1: 0.28 -> 0.40 | Helped |
| Use RUN instead of ε-Diagnosis | Sock-shop Memory leak root cause analysis · Sock-shop, Memory leak | HR@1: 0.20 -> 0.12 | Hurt |
| Use RUN instead of ε-Diagnosis | Sock-shop CPU hog root cause analysis · Sock-shop, CPU hog | HR@3: 0.32 -> 0.48 | Helped |
| Use RUN instead of ε-Diagnosis | Sock-shop Memory leak root cause analysis · Sock-shop, Memory leak | HR@3: 0.20 -> 0.32 | Helped |
| Use RUN instead of ε-Diagnosis | Sock-shop CPU hog root cause analysis · Sock-shop, CPU hog | HR@5: 0.32 -> 0.52 | Helped |
| Use RUN instead of ε-Diagnosis | Sock-shop Memory leak root cause analysis · Sock-shop, Memory leak | HR@5: 0.20 -> 0.36 | Helped |
| Use RUN instead of ε-Diagnosis | Sock-shop CPU hog root cause analysis · Sock-shop, CPU hog | MRR: 0.304 -> 0.484 | Helped |
| Use RUN instead of ε-Diagnosis | Sock-shop Memory leak root cause analysis · Sock-shop, Memory leak | MRR: 0.227 -> 0.248 | Helped |
| remove the pre-training stage (ablation) | RUN · Sock-shop | root cause analysis performance | Hurt |
| incorporate negative pairs in contrastive learning (ablation) | RUN · Sock-shop | root cause analysis performance | No clear effect |
Surprising findings
- Adding negative pairs to the contrastive objective did not substantially decrease performance, which the authors attribute to the short temporal extent of sock-shop data.
- RUN performs less well on synthetic data than on sock-shop, and the authors state the model is less suitable for the synthetic dataset.
Limitations
- Scalability to larger datasets is not yet addressed (stated as future work).
- Evaluated on a single small sock-shop dataset (38 series, 600 timestamps) and synthetic data; no public real-world RCA benchmark is used.
- Performance is weaker on memory leak HR@1 and on synthetic data relative to some baselines.
Experimental setup
Datasets: Synthetic data generated via RCD procedure with PyAgrum CPTs Sock-shop microservice dataset
Benchmarks: Sock-shop (CPU hog, memory leak) Synthetic data (10-50 nodes)
Metrics: HR@1 HR@3 HR@5 MRR
Baselines: epsilon-Diagnosis AutoMAP RCD Psi-PC CausalRCA
Builds on: Granger causality DLinear SimSiam GrootRank TS2Vec
Closest papers by meaning
- Federated Granger Causality Learning For Interdependent Clients With State Space Representation (ICLR 2025 · similarity 0.86)
- A Simple yet Scalable Granger Causal Structural Learning Approach for Topological Event Sequences (NeurIPS 2024 · similarity 0.84)
- PRIM: Meta-Learned Bayesian Root Cause Analysis (NeurIPS 2026 · similarity 0.84)
- Root Cause Analysis of Outliers with Missing Structural Knowledge (NeurIPS 2025 · similarity 0.82)
- Robust Root Cause Diagnosis using In-Distribution Interventions (ICLR 2025 · similarity 0.82)
- GCAD: Anomaly Detection in Multivariate Time Series from the Perspective of Granger Causality (AAAI 2025 · similarity 0.82)
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.