Research map / Time series / Long-term forecasting
MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification
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 MPTSNet, which uses FFT-detected main periods to segment multivariate time series into multiscale periodic segments and applies an Inception-style CNN local extractor followed by attention global capturer, reporting SOTA average accuracy on UEA benchmarks.
Problem
Multivariate time series classification (MTSC) methods often fail to jointly capture local intra-period features and global inter-period dependencies across multiple periodic scales, and variable correlations vary across those scales.
Approach
A PeriodicBlock that segments series by FFT-identified main periods, then sequentially applies a multi-kernel Inception-style convolutional Local Extractor and a multi-head attention Global Capturer per scale, aggregating scales with amplitude-based softmax weights; attention maps per scale give interpretability.
Key results
- Setting 1 (10 UEA datasets): MPTSNet reaches 75.4% average accuracy and average rank 2.4, versus 74.2% for the best baseline M.TCN (ModernTCN).
- Setting 2 (25 UEA datasets): MPTSNet reports 74.0% average accuracy and average rank 3.12, versus 72.6% for TodyNet.
- Ablation on Setting 1: removing multiscale periodic decomposition drops accuracy from 0.754 to 0.709, the largest drop among variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MPTSNet instead of ModernTCN | MPTSNet · UEA Setting 1, 10 datasets | average accuracy: 74.2 -> 75.4; 1.2% | Helped |
| use MPTSNet instead of TodyNet | MPTSNet · UEA Setting 2, 25 datasets | average accuracy: 72.6 -> 74.0 | Helped |
| add the Local Extractor (ablation) | MPTSNet · UEA Setting 1, 10 datasets | average accuracy: 0.737 -> 0.754 | Helped |
| add the Global Capturer (ablation) | MPTSNet · UEA Setting 1, 10 datasets | average accuracy: 0.731 -> 0.754 | Helped |
| add multiscale periodic decomposition (ablation) | MPTSNet · UEA Setting 1, 10 datasets | average accuracy: 0.709 -> 0.754 | Helped |
| set the number of periodic scales to 1 instead of 5 (ablation) | MPTSNet · UEA Setting 1, 10 datasets | average accuracy: 0.734 -> 0.754 | Hurt |
| set the number of periodic scales to 3 instead of 5 (ablation) | MPTSNet · UEA Setting 1, 10 datasets | average accuracy: 0.742 -> 0.754 | Hurt |
| set the number of periodic scales to 7 instead of 5 (ablation) | MPTSNet · UEA Setting 1, 10 datasets | average accuracy: 0.754 -> 0.751 | Hurt |
Surprising findings
- On several Setting 2 datasets (e.g., ArticularyWordRecognition, Cricket, Handwriting), MPTSNet is outperformed by multiple baselines, with 1-to-1 losses against several methods.
Limitations
- Baseline results are copied from original publications rather than rerun under identical training settings.
- Setting 2 shows MPTSNet losing to several baselines on many datasets, and the claim of top performance rests on averages.
- Hyperparameter k selection is left to manual tuning; adaptive k selection is noted as future work.
Experimental setup
Datasets: UEA multivariate time series archive
Benchmarks: UEA 10-dataset subset (Setting 1) UEA 25-dataset set (Setting 2)
Metrics: accuracy average accuracy average rank 1-to-1 wins/draws/losses
Baselines: ModernTCN Crossformer TimesNet MICN PatchTST DLinear SCINet FEDformer Flowformer LSSL LSTNet TodyNet
Builds on: TimesNet OS-CNN Inception
Closest papers by meaning
- MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting (AAAI 2024 · similarity 0.88)
- TimeMIL: Advancing Multivariate Time Series Classification via a Time-aware Multiple Instance Learning (ICML 2024 · similarity 0.88)
- TVNet: A Novel Time Series Analysis Method Based on Dynamic Convolution and 3D-Variation (ICLR 2025 · similarity 0.86)
- Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach (ICML 2024 · similarity 0.86)
- TSLANet: Rethinking Transformers for Time Series Representation Learning (ICML 2024 · similarity 0.86)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · 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.