Experimental Setup
Evaluation Metrics
- Precision@K (PR@K) \( PR@K = \frac{1}{|\mathbb{A}|}\sum_{a \in \mathbb{A}}\frac{\sum_{i < k}R_a(i)\in V_a}{\min (K, |v_a|)} \), where \(\mathbb{A}\) is the set of system faults, \(a\) is one fault in \(\mathbb{A}\), \(V_a\) is the real root causes of \(a\), \(R_a\) is the predicted root causes of \(a\), and \(i\) is the \(i\)-th predicted cause of \(R_a\)
- Mean Average Precision@K (MAP@K) \( MAP@K = \frac{1}{K|\mathbb{A}|} \sum_{a \in \mathbb{A}} \sum_{i\leq j\leq K} PR@j \)
- Mean Reciprocal Rank (MRR) \( MRR@K = \frac{1}{|\mathbb{A}|}\sum_{a \in \mathbb{A}}\frac{1}{rank_{R_a}}\), where \(rank_{R_a}\) is the rank number of the first correctly predicted root cause for system fault \(a\).
Baseline Methods
| Method | Main Technique | Online/Offline | Time |
|---|---|---|---|
| PC | Constrain-based independence test | Offline | 2003 |
| DyNotears | Dynamic Bayesian network | Offline | 2020 |
| C-LSTM | Nonlinear Granger causality | Offline | 2022 |
| GOLEM | Relaxation of Notears | Offline | 2020 |
| REASON | Interdependent graph neural networks | Offline | 2023 |
| Nezha | Multi-modal anomaly detection | Offline | 2023 |
| MULAN | Multi-modal causal structure learning | Offline | 2024 |
| CORAL | Incremental disentagled causal graph learning | Online | 2023 |
For detailed experimental results, please refer to the experiment section in our paper.