Headline
Fault-type macro F1 on grouped 5-fold cross-validation across all three systems, grouped by case so that windows of one case never sit in both train and test. Mean ± std across folds, with min–max.
Per-class results
Precision, recall and F1 for each fault type, pooled over the held-out folds of the grouped cross-validation.
| Fault | Precision | Recall | F1 | Rows |
|---|
Confusion matrix: rows are the true fault, columns the predicted one. Shaded cells hold more windows, and every cell shows its count.
Detection delay
Macro F1 when the model sees only the first seconds of metrics after the alert. Mean ± std across folds, with min–max.
| Seconds | Macro F1 | Min–max |
|---|
Generalizing to an unseen system
Leave-one-system-out trains on two of the RCAEval RE2 systems and tests on the third. This is the situation of a team running FaultTrack on its own cluster.
| Held-out system | Macro F1 |
|---|
Sample held-out cases
A few cases from the held-out folds. The misses are kept. The Model column marks the fault type (from the model); the Rule column marks the suspect service (from the rule).
| Case | True fault | Model says | Probability | Model | Injected service | Rule suspects | Rule |
|---|