PRISM: Asymmetric Precision-Recall Optimization for Million-Scale Root Cause Analysis

Multi-dimensional Root Cause Analysis (RCA) requires searching an exponential $\mathcal{O}(2^M)$ attribute space over millions of records within seconds to isolate the root-cause combination. Existing methods fail at this industrial scale due to a fundamental modeling oversight: employing symmetric objectives that treat anomalous and normal samples interchangeably, allowing the massive normal background to dominate computation. To overcome this, we propose \textbf{PRISM} (\underline{P}recision-\underline{R}ecall \underline{I}nformed \underline{S}ubspace \underline{M}ining), which reformulates RCA as an asymmetric Precision-Recall optimization. Precision is defined over anomaly-normal pairs and Recall over anomaly-anomaly pairs, eliminating uninformative normal-normal comparisons entirely. Coupled with tensor compression and a two-stage solver, comprising Bernoulli continuous relaxation and combinatorial verification, PRISM reduces pairwise complexity from a full-dataset quadratic scale to an anomaly-bounded $\mathcal{O}(|\mathcal{P}|^2)$. Evaluations on four public benchmarks and a new 5.13M-record industrial dataset show that PRISM achieves the highest average F1 (0.90 vs.\ 0.65 for the strongest baseline), while maintaining sub-second latency across a $2^{110}$ attribute space. We publicly release this dataset as \textbf{NFC}, establishing the first production-sourced, million-scale benchmark for multi-dimensional RCA. \href{https://anonymous.4open.science/status/PRISM-09FC}{[Code]}