Research map / Time series / Foundation models for time series
Temporal Streaming Batch Principal Component Analysis for Time Series Classification (Student Abstract)
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 TSBPCA, a streaming, time-block-updated PCA compression step that reduces long multivariate time series before classification, improving accuracy and speed of LSTM, Transformer, Informer, iTransformer and TimesNet.
Problem
Sequence classification models struggle with long multivariate time series, where extended length and many variables hurt accuracy and efficiency.
Approach
Extends streaming/histPCA-style batch power iteration with temporal ordering: each time block's projection matrix is updated using the previous block's eigenvectors and eigenvalues, so the compact representation retains temporal dependencies and is used as a preprocessing reduction for downstream sequence models.
Key results
- TSBPCA + LSTM improves average accuracy by 4.23% and time efficiency by 62.15% across the five datasets (Table 1).
- On SR2 and EC, accuracy increases by up to 18% and time efficiency improves by up to 80%.
- Averaged over datasets, TSBPCA raises Transformer accuracy from 66.7 to 68.7 and reduces its time from 39.48 to 16.59.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TSBPCA (ablation) | LSTM · AVERAGE | Acc: 59.1 -> 61.6 | Helped |
| Add TSBPCA (ablation) | LSTM · AVERAGE | Time: 79.81 -> 30.21 | Helped |
| Add TSBPCA (ablation) | Transformer · AVERAGE | Acc: 66.7 -> 68.7 | Helped |
| Add TSBPCA (ablation) | Transformer · AVERAGE | Time: 39.48 -> 16.59 | Helped |
| Add TSBPCA (ablation) | Informer · AVERAGE | Acc: 67.6 -> 67.8 | Helped |
| Add TSBPCA (ablation) | Informer · AVERAGE | Time: 19.69 -> 18.12 | Helped |
| Add TSBPCA (ablation) | iTransformer · AVERAGE | Acc: 65.1 -> 65.5 | Helped |
| Add TSBPCA (ablation) | iTransformer · AVERAGE | Time: 16.49 -> 15.97 | Helped |
| Add TSBPCA (ablation) | TimesNet · AVERAGE | Acc: 66.7 -> 65.5 | Hurt |
| Add TSBPCA (ablation) | TimesNet · AVERAGE | Time: 37.83 -> 35.75 | Helped |
| Add TSBPCA (ablation) | LSTM · SR2 | Acc: 53.9 -> 54.8 | Helped |
| Add TSBPCA (ablation) | LSTM · SR2 | Time: 85.28 -> 32.29 | Helped |
Surprising findings
- Gains from TSBPCA are inconsistent across models other than LSTM, varying by dataset, and for some model-dataset pairs accuracy drops (e.g., iTransformer on SR1 and UWave-like cases in Table 1).
Limitations
- Dual-iteration reconstruction may struggle with highly complex time series or datasets with extreme variability (stated in conclusion).
- Improvements are inconsistent across non-LSTM models and datasets.
Experimental setup
Datasets: UWaveGestureLibrary Heartbeat SelfRegulationSCP1 SelfRegulationSCP2 EthanolConcentration
Benchmarks: UWaveGestureLibrary Heartbeat SelfRegulationSCP1 SelfRegulationSCP2 EthanolConcentration
Metrics: classification accuracy running time
Baselines: LSTM Transformer Informer iTransformer TimesNet
Builds on: streaming PCA (Oja) hist-PCA
Closest papers by meaning
- MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification (AAAI 2025 · similarity 0.83)
- T-Rep: Representation Learning for Time Series using Time-Embeddings (ICLR 2024 · similarity 0.82)
- TimeMIL: Advancing Multivariate Time Series Classification via a Time-aware Multiple Instance Learning (ICML 2024 · similarity 0.82)
- TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis (ICLR 2025 · similarity 0.82)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.82)
- TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation (ICML 2025 · similarity 0.82)
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.