DeCoR: decoupled multiplex causal ranking for microservice root cause analysis

Failures in microservice systems can produce correlated symptoms across services, motivating the study of structural information for root-cause localization alongside anomaly-based approaches. We study service-level localization from multivariate monitoring metrics under known fault-onset and metric-to-service mapping assumptions. Existing temporal and multilayer ranking methods rarely use relation-estimation reliability to control node-level information exchange between contemporaneous and lagged evidence. We propose Decoupled Multiplex Causal Ranking (DeCoR) and its core algorithm, Regime-aware Innovation Causal Estimation with Confidence-aware Multiplex Coupling (RICE-CM). RICE-CM derives both evidence types from a shared dynamics model fitted to the pre-fault normal segment. It aggregates edge confidence scores into node reliability, applies nonuniform diagonal coupling between same-metric replicas, and ranks services through reverse PageRank on a supra-adjacency matrix. We evaluated RICE-CM on a 300-case subset drawn from three RCAEval datasets: Sock Shop 1 (SS1), Sock Shop 2 (SS2), and Online Boutique (OB). With dataset-specific coupling settings selected on the same benchmark cases, it achieved Avg@5 scores of 0.6000, 0.7040, and 0.4864 on SS1, SS2, and OB, respectively. It achieved the highest Avg@5 among the evaluated methods on all three datasets and exceeded RUN on OB by 0.0288, while requiring approximately 1/33 of RUN’s runtime per case. Compared with RICE-only, RICE-CM improved Avg@5 by 0.0640, 0.0704, and 0.0624, respectively. Paired tests supported these gains on SS1 and SS2 after Holm correction, but not on OB. Against fixed-coupling RICE-MX, the gains were 0.0720, 0.0400, and 0.0336, respectively, with a Holm-adjusted p value of 0.0456 on each dataset. Positive observed gains over RICE-only occurred across all 12 dataset-fault-type combinations, whereas the coupling-parameter sweep showed dataset-dependent sensitivity. These results support node-reliability-driven nonuniform coupling over fixed coupling within the evaluated benchmarks and task assumptions.