Research map / Learning theory and algorithms / Prediction and decision losses
Optimizing ROC Curves with a Sort-Based Surrogate Loss for Binary Classification and Changepoint Detection
Jonathan Hillman, Toby Dylan Hocking
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AUM with FP/FN counts instead of rates (ablation) | AUM linear model · zip 0/1, 50% positive train | median test AUC: 0.9979 -> 0.9993 | Improved |
| Use AUM with FP/FN counts instead of rates (ablation) | AUM linear model · zip 0/1, 1% positive train | median test AUC: 0.9928 -> 0.9979 | Improved |
| Train with AUM instead of weighted logistic loss | linear model · zip 0/1, 1% positive train | median test AUC: 0.9950 -> 0.9979 | Improved |
| Train with AUM instead of squared hinge all pairs loss | linear model · zip 0/1, balanced train | median test AUC: 0.9991 -> 0.9993 | No clear effect |
| Minimize AUM on predicted values directly (chipseq, 68 train sets) | Initial min-label-error predictions · chipseq changepoint train sets | train AUC | Improved |
| Constrain predictions to a linear model when minimizing AUM | AUM gradient descent · chipseq H3K4me3_XJ_immune fold 4 (p=27, n=54) | train AUC | No clear effect |
| AUM gradient descent from squared hinge initialization, with early stopping | linear model · neuroblastoma changepoint (5 train sets) | test AUC | Mixed |
| Use AUM instead of squared hinge label-error loss | changepoint detection · chipseq changepoint | label error rate | Hurt |
| Use AUM O(n log n) gradient instead of all-pairs squared hinge | binary classification · speed test, 1 second budget | problem size computable: n=thousands -> n=millions | Improved |
| Use AUM instead of per-example squared hinge | changepoint detection · speed test, n about 5000 | gradient time: 1ms -> 10ms | Hurt |
Closest papers by meaning
- Interplay of ROC and Precision-Recall AUCs: Theoretical Limits and Practical Implications in Binary Classification (ICML 2024 · similarity 0.89)
- Large-scale Optimization of Partial AUC in a Range of False Positive Rates (NeurIPS 2022 · similarity 0.89)
- Estimating the Arc Length of the Optimal ROC Curve and Lower Bounding the Maximal AUC (NeurIPS 2022 · similarity 0.89)
- Positive-unlabeled AUC Maximization under Covariate Shift (ICML 2025 · similarity 0.87)
- AUC Maximization under Positive Distribution Shift (NeurIPS 2024 · similarity 0.87)
- Beyond the ROC Curve: Classification Trees Using Cost-Optimal Curves, with Application to Imbalanced Datasets (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
