Research map / Time series / Long-term forecasting
Addressing Spatial-Temporal Heterogeneity: General Mixed Time Series Analysis via Latent Continuity Recovery and Alignment
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 MiTSformer, which recovers latent continuous variables behind discretized variables via dilated-CNN recovery adversarially aligned with continuous variables, then models intra- and cross-modality dependencies with attention, evaluated on five mixed time series tasks.
Problem
Mixed time series containing both continuous and discrete variables suffer spatial-temporal heterogeneity: discrete variables have different temporal patterns and distributions from continuous ones, leading to biased, imbalanced modeling.
Approach
Treat discrete variables as discretizations of latent continuous variables (LCVs); recover LCVs with a residual dilated CNN regularized by a temporal smoothness loss and adversarial variable-modality discrimination against continuous variables, then apply intra-modality self-attention and cross-modality cross-attention.
Key results
- Averages 71.9% accuracy on 10 UEA classification datasets, above the best baseline at 70.1%.
- Achieves average MAE 0.524 on 10 UCR extrinsic regression datasets, outperforming all baselines.
- Reports SOTA F1 on five anomaly detection benchmarks and best CV forecasting on 76% of settings per Table 16.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove variable-modality discrimination loss (LDis) (ablation) | MiTSformer · UEA classification average | accuracy: 71.9 -> 70.1 | Hurt |
| Remove temporal smoothness constraint (Lsmooth) (ablation) | MiTSformer · UEA classification average | accuracy: 71.9 -> 70.7 | Hurt |
| Remove reconstruction loss (LRec) (ablation) | MiTSformer · UEA classification average | accuracy: 71.9 -> 69.7 | Hurt |
| Remove LCV recovery losses and LCVs (ablation) | MiTSformer · UEA classification average | accuracy: 71.9 -> 69.3 | Hurt |
| Remove cross-attention blocks (ablation) | MiTSformer · UEA classification average | accuracy: 71.9 -> 70.2 | Hurt |
| Use MiTSformer instead of the strongest listed baseline | mixed time series imputation · ETTm2 | MAE: 0.145 -> 0.116 | Helped |
| Use MiTSformer instead of the strongest listed baseline | mixed time series imputation · ETTm1 | MAE: 0.135 -> 0.156 | Hurt |
| Use MiTSformer instead of the strongest listed baseline | mixed time series long-term forecasting · ETTm1 | MAE: 0.379 -> 0.376 | Helped |
| Use MiTSformer instead of the strongest listed baseline | mixed time series long-term forecasting · ETTh2 | MAE: 0.442 -> 0.430 | Helped |
Surprising findings
- Frequency-based methods FiLM and FEDformer perform poorly on mixed data, attributed to discrete variables corrupting frequency estimation.
- DLinear performs notably worse on extrinsic regression than attention-based models.
Limitations
- Smoothness loss may not suit discrete variables with inherent sudden changes.
- Memory footprint grows on high-dimensional datasets such as Electricity; complexity quadratic in number of discrete variables.
- Mixed datasets are synthetically created by discretizing existing continuous benchmarks rather than real mixed data.
Experimental setup
Datasets: UEA UCR ETT Electricity Weather SMD MSL SMAP SWaT PSM Traffic Exchange
Benchmarks: UEA UCR ETT Electricity Weather SMD MSL SMAP SWaT PSM Traffic Exchange
Metrics: Accuracy MAE RMSE MSE Precision Recall F1-score
Baselines: ModernTCN TimesNet MICN iTransformer PatchTST Crossformer FEDformer Pyraformer LightTS DLinear FiLM
Builds on: TimesNet iTransformer PatchTST Gradient reversal layer
Closest papers by meaning
- Explaining Time Series via Contrastive and Locally Sparse Perturbations (ICLR 2024 · similarity 0.87)
- TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis (ICLR 2025 · similarity 0.86)
- TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (ICLR 2024 · similarity 0.86)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.86)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.86)
- MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting (AAAI 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.