Research map / Time series / Foundation models for time series
Learning under Temporal Label Noise
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Formalizes temporal label noise in sequential time-series classification and proposes forward-correction methods (Continuous, Discontinuous, Plug-In) that jointly or separately estimate a time-varying noise transition matrix, outperforming static noise-robust baselines on four healthcare/activity datasets and a stress-detection demo.
Problem
Label noise in time series classification often varies over time (temporal label noise), but existing noise-robust methods assume a static noise transition matrix, leading to misspecification and poor performance.
Approach
Model the label noise as a time-dependent transition matrix Q(t) parameterized by a neural network, learn it jointly with a sequential classifier via a forward-corrected loss and an augmented Lagrangian with a minimum-volume simplex constraint; also offers discontinuous and anchor-point plug-in variants.
Key results
- Under Mixed 30% temporal noise, the Continuous method reduces test error on moving to 4.2% versus 14.9% for the best static baseline (VolMinNet).
- On the stress detection demo, Continuous reduces approximation error to 9.7% vs 16.4% for static correction and test error to 25.5% vs 29.5%.
- Under Periodic noise, Continuous achieves lowest test error on all four datasets, e.g., 1.7% on moving vs 13.5% for VolMinNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| model temporal label noise with Continuous instead of static correction | GRU time-series classifier · moving, Mixed noise | Test Error: 14.9 $\pm$ 2.7\% -> 4.2 $\pm$ 2.2\% | Helped |
| model temporal label noise with Continuous instead of static correction | GRU time-series classifier · moving, Periodic noise | Test Error: 13.5 $\pm$ 6.0\% -> 1.7 $\pm$ 0.6\% | Helped |
| impose continuity across time steps rather than estimate each time step independently (ablation) | Temporal noise correction for the GRU · Stress detection | Test Error: 26.0 $\pm$ 2.5\% -> 25.5 $\pm$ 1.7\% | Helped |
| add temporal noise correction instead of ignoring label noise (ablation) | Sequential classifier · Stress detection | Test Error: 31.1 $\pm$ 4.3\% -> 25.5 $\pm$ 1.7\% | Helped |
| use temporal rather than static noise correction | Time-series classifiers · Real-world healthcare datasets with Mixed and Periodic noise | Test Error | Helped |
| use Continuous rather than Plug-In or Discontinuous estimation | Temporal noise estimation · Real-world healthcare datasets with Mixed and Periodic noise | Test Error | Mixed |
Surprising findings
- Static noise-correction methods can perform no better than or worse than ignoring noise when noise is temporal, e.g., Anchor is worse than Ignore on some settings.
Limitations
- Assumes all sequences share the same temporal noise function, ignoring annotator-specific noise.
- Assumes stationarity of time series.
- Few public time series datasets contain both noisy and clean labels, making validation hard.
Experimental setup
Datasets: moving senior blinking sleeping synthetic HMM stress detection physiological/self-report dataset
Benchmarks: moving senior blinking sleeping stress detection
Metrics: clean test error approximation error of noise matrix
Baselines: Ignore Anchor VolMinNet Plug-In Discontinuous
Builds on: forward loss correction anchor points VolMinNet
Closest papers by meaning
- Learning with Feature-Dependent Label Noise: A Progressive Approach (ICLR 2021 · similarity 0.88)
- Safeguarded Dynamic Label Regression for Noisy Supervision (AAAI 2019 · similarity 0.86)
- An Unsupervised Approach for Periodic Source Detection in Time Series (ICML 2024 · similarity 0.85)
- Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.85)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.84)
- Learning Discriminative Dynamics with Label Corruption for Noisy Label Detection (CVPR 2024 · similarity 0.84)
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.