Research map / Label-efficient and robust learning / Recognition and long-tail
Adaptive Meta-Learning Probabilistic Inference Framework for Long Sequence Prediction
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 AMPIF, a meta-learning probabilistic framework for long sequence prediction that decomposes sequences into seasonal/trend parts, builds meta-learning tasks via clustering-matching of learned global representations, and uses a dual-stream amortized network (ST-DAN) to generate task-specific parameters.
Problem
Long sequence time-series prediction under domain distribution shift (e.g., new stocks and new market indices), where meta-learning tasks are hard to define a priori.
Approach
Adaptively constructs meta-learning tasks by clustering support-set global representations (learned with a VAE plus mutual-information constraint over input and label) and matching query samples to clusters, then uses a VERSA-style dual-stream amortized network over separate seasonal and trend streams to generate task-specific parameters.
Key results
- On the A-share dataset, AMPIF reduced MSE by 35.8% and MAE by 24.5% on average relative to LaST.
- On Cross-Market, AMPIF reduced MSE by 46.0% and MAE by 36.1% on average relative to LaST.
- On ETT, AMPIF reduced MSE by an average of 15.8% and MAE by 8.5% relative to Autoformer when using Autoformer's autocorrelation as encoders.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AMPIF instead of LaST | long-sequence prediction · A-share, input 24, horizon 24 | MSE: 0.105 -> 0.034 | Helped |
| Use AMPIF instead of LaST | long-sequence prediction · Cross-Market, 528 trading days | MSE: 0.895 -> 0.418 | Helped |
| Use AMPIF instead of Autoformer | ETT long-sequence prediction · ETTh2, horizon 24 | MSE: 0.261 -> 0.230 | Helped |
| Add frequency decomposition (ablation) | AMPIF · Cross-Market, 22 trading days | MSE: 0.111 -> 0.096 | Helped |
| Add frequency decomposition (ablation) | AMPIF · Cross-Market, 22 trading days | MAE: 0.234 -> 0.214 | Helped |
| Add ST-DAN (ablation) | AMPIF · Cross-Market, 22 trading days | MSE: 0.199 -> 0.096 | Helped |
| Add ST-DAN (ablation) | AMPIF · Cross-Market, 22 trading days | MAE: 0.304 -> 0.214 | Helped |
| Add the seasonal component and frequency decomposition (ablation) | AMPIF · Cross-Market, 22 trading days | MSE: 0.114 -> 0.096 | Helped |
| Add the seasonal component and frequency decomposition (ablation) | AMPIF · Cross-Market, 22 trading days | MAE: 0.242 -> 0.214 | Helped |
| Add the trend component and frequency decomposition (ablation) | AMPIF · Cross-Market, 22 trading days | MSE: 0.104 -> 0.096 | Helped |
| Add the trend component and frequency decomposition (ablation) | AMPIF · Cross-Market, 22 trading days | MAE: 0.229 -> 0.214 | Helped |
Limitations
- Limitations are discussed only in an appendix not included in the provided text.
Experimental setup
Datasets: A-share Cross-Market ETT
Benchmarks: A-share Cross-Market ETT
Metrics: MSE MAE
Baselines: LSTM Autoformer FEDformer LaST Informer LogTrans Reformer LSTnet LSTMa
Builds on: VERSA VAE Autoformer LaST MAML
Closest papers by meaning
- CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns (NeurIPS 2024 · similarity 0.84)
- Introducing Spectral Attention for Long-Range Dependency in Time Series Forecasting (NeurIPS 2024 · similarity 0.84)
- SparseTSF: Modeling Long-term Time Series Forecasting with *1k* Parameters (ICML 2024 · similarity 0.84)
- Sequence Complementor: Complementing Transformers for Time Series Forecasting with Learnable Sequences (AAAI 2025 · similarity 0.84)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.84)
- Efficiently Enhancing Long-term Series Forecasting via Ultra-long Lookback Windows (AAAI 2025 · 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.