Robustness with inaccurate confusion matrices

Analyze the scheduling framework when the available confusion matrix is inaccurate, in order to determine the effect of errors in estimating classifier performance on the resulting robust scheduling guarantees.

Background

The scheduling framework assumes that the classifier’s confusion matrix is known exactly in advance. The paper notes that practical methods can estimate confusion matrices from unlabeled data, but such estimates may be inaccurate. The impact of errors in the confusion matrix on the algorithms and robustness guarantees is not analyzed in the paper and is explicitly deferred to future work.

References

In this paper, we propose to evaluate the gain that a perfect ``unsupervised performance estimation'' technique would permit on scheduling algorithms, and we leave the analysis of inaccurate confusion matrices for future work.

— Robust Non-Clairvoyant Scheduling with Classification Models  (2610.01343 - Dugois et al., 1 Oct 2026) in Section 2, “Knowledge model”