Research map / Time series / Foundation models for time series
SimPSI: A Simple Strategy to Preserve Spectral Information in Time Series Data Augmentation
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.
Proposes SimPSI, which mixes original and augmented spectra under a frequency-wise preservation map (magnitude, saliency, or a learned spectrum-preservative map) so time-series augmentations keep core spectral information, improving accuracy across simulation, HAR, SleepEDF and Waveform benchmarks.
Problem
Time series data augmentations (jittering, permutation, time warping, frequency masking, etc.) distort frequency-domain information that is essential for the task, so their gains do not generalize across time series benchmarks.
Approach
Mix the original and augmented spectra per frequency using a preservation map in [0,1] that marks informative frequencies; three maps are proposed: magnitude-based, gradient saliency-based, and a learned spectrum-preservative map trained with a preservation contrastive loss that ranks original-preserving views above random and inverted ones.
Key results
- On HAR, SimPSI (magnitude spectrum) raised Scale-Shift-Jittering accuracy by 0.9 points (94.9 vs 94.0 baseline-free comparison in table), and Frequency masking improved by 1.0 with all SimPSI maps.
- On SleepEDF, Frequency masking with the spectrum-preservative map gained 1.0 points (81.7 vs 80.7 without augmentation).
- On Waveform, Scale-Shift-Jittering with the spectrum-preservative map improved accuracy from 92.0 to 95.2 and AUPRC from 62.6 to 64.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SimPSI spectrum-preservative map to Frequency masking | Frequency masking · Simulation (FSK32) | accuracy: +1.5 | Helped |
| add SimPSI spectrum-preservative map to Frequency masking | Frequency masking · SleepEDF | accuracy: +1.0 | Helped |
| add random preservation map (uniform U(0,1)) to Jittering (ablation) | Jittering · HAR | accuracy: -0.4 | Hurt |
| add SimPSI magnitude spectrum to Jittering | Jittering · HAR | accuracy: +0.2 | No clear effect |
| add SimPSI spectrum-preservative map (vs random map) to Frequency masking (ablation) | Frequency masking · HAR | accuracy: +1.0 (all maps equal) | No clear effect |
| remove preservation contrastive loss from SimPSI spectrum-preservative map (ablation) | SimPSI (Spectrum-preservative map) · Atrial Fibrillation (Waveform) | accuracy: 0.3 accuracy drop when removed (95.2 vs 94.9) | Helped |
| joint training of classifier and preservation map generator instead of separate training (ablation) | SimPSI (Spectrum-preservative map) · Atrial Fibrillation (Waveform) | accuracy: -0.7 | Hurt |
| add SimPSI (spectrum-preservative map) to Permutation | Permutation · Simulation (FSK8/FSK32, Table D.1) | accuracy: +1.4 (from -12.1 to +1.4 relative to baseline) | Helped |
| add SimPSI magnitude spectrum to Permutation (ablation) | Permutation · Simulation (Table D.1) | accuracy: 0 relative to baseline (94.8 vs 82.7 base) | No clear effect |
| add SimPSI saliency map to Dropout (ablation) | Dropout · Simulation (Table D.1) | accuracy: -0.4 vs Dropout alone (93.0 vs 93.4) | Hurt |
| add SimPSI spectrum-preservative map to Scale-Shift-Jittering | Scale-Shift-Jittering · Atrial Fibrillation (Waveform) | accuracy: +3.2 (92.0 to 95.2) | Helped |
| add SimPSI saliency map to Jittering (ablation) | Jittering · Atrial Fibrillation (Waveform) | accuracy: +0.4 (94.2 to 95.1 with saliency; +0.9 vs baseline) | Helped |
Surprising findings
- A random preservation map (uniform U(0,1) mixing) also improved several augmentations, showing that mixing itself helps in some settings.
- Magnitude-spectrum and saliency preservation maps frequently gave marginal or negative gains, with the benefit depending strongly on the data domain.
Limitations
- Magnitude warping showed no improvement on Waveform with any preservation map, and hyperparameters (beta1, beta2) were not carefully tuned.
- Saliency and spectrum-preservative maps add substantial training time compared to the magnitude spectrum.
- Gains are often within one standard deviation across seeds, and the method's benefit depends on the data domain.
Experimental setup
Datasets: Simulation (FSK8/FSK32 signal demodulation, synthetic) HAR SleepEDF Waveform (atrial fibrillation ECG)
Benchmarks: Simulation (FSK8/FSK32) HAR SleepEDF Waveform
Metrics: accuracy AUPRC training time
Baselines: Jittering Scale-Shift-Jittering Frequency masking Magnitude warping Dropout Time warping Permutation Random preservation map
Builds on: Frequency masking KeepAugment AugMix ResNet1D
Closest papers by meaning
- Spectral Transformer Neural Processes (NeurIPS 2026 · similarity 0.85)
- Introducing Spectral Attention for Long-Range Dependency in Time Series Forecasting (NeurIPS 2024 · similarity 0.84)
- Rethinking Fourier Transform from A Basis Functions Perspective for Long-term Time Series Forecasting (NeurIPS 2024 · similarity 0.83)
- Guiding Visual Autoregressive Models through Spectrum Weakening (NeurIPS 2026 · similarity 0.83)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.83)
- LSCD: Lomb--Scargle Conditioned Diffusion for Time series Imputation (ICML 2025 · similarity 0.83)
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.