Research map / Time series / Long-term forecasting
MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting
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 MSGNet, which uses FFT-detected periodic time scales and per-scale adaptive mixhop graph convolution plus multi-head self-attention to model multi-scale inter-series correlations for multivariate forecasting.
Problem
Existing deep forecasting models fail to capture how inter-series correlations among multiple time series vary across different time scales, and fixed or dynamic graphs ignore scale-linked stability.
Approach
Use FFT-detected periods as time scales, reshape the series per scale, learn a distinct adaptive adjacency matrix per scale with mixhop graph convolution, apply multi-head attention within each scale, and aggregate scales by FFT amplitude-based softmax weights.
Key results
- On Flight (7:1:2 split), MSGNet reports average MSE 0.208 and MAE 0.321 versus TimesNet 0.265 and 0.372, a 21.5% MSE reduction as stated in the text.
- MSGNet achieves the best average rank (1.813) across the eight datasets, ahead of TimesNet (2.750) and NLinear (2.813).
- Under the 4:4:2 COVID-affected partition, MSGNet's Flight MSE is 0.252 versus 0.335 for TimesNet, with a 21.29% degradation versus TimesNet's 26.47%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MSGNet rather than TimesNet | Multivariate forecasting · Flight (average over prediction lengths) | MSE: 0.265 to 0.208 | Helped |
| Add adaptive graph convolution (graph learning) (ablation) | MSGNet · Flight, Weather, and ETTm2 (average over prediction lengths 96, 336) | MSE: 0.302 -> 0.195; 0.232 -> 0.218; 0.253 -> 0.245 | Helped |
| Use multi-scale graph convolution instead of a shared graph convolution (ablation) | MSGNet · Flight, Weather, and ETTm2 (average over prediction lengths 96, 336) | MSE: 0.213 -> 0.195; 0.226 -> 0.218; 0.250 -> 0.245 | Helped |
| Add multi-head self-attention (ablation) | MSGNet · Flight, Weather, and ETTm2 (average over prediction lengths 96, 336) | MSE: 0.198 -> 0.195; 0.224 -> 0.218; 0.247 -> 0.245 | Mixed |
| Use mix-hop convolution instead of traditional convolution (ablation) | MSGNet · Flight, Weather, and ETTm2 (average over prediction lengths 96, 336) | MSE: 0.202 -> 0.195; 0.224 -> 0.218; 0.247 -> 0.245 | Helped |
Surprising findings
- Ablation shows removing the adaptive graph layer (w/o-AdapG) raises Flight MSE from 0.195 to 0.302, while removing multi-head attention changes results only marginally.
Limitations
- Relies on FFT periodicity as the only source of time scales, which may not capture non-periodic correlation structure.
- Flight dataset is newly constructed and not stated as released, limiting reproducibility of the main claims.
- Computationally heavier than linear models such as DLinear.
Experimental setup
Datasets: Flight Weather ETTh1 ETTh2 ETTm1 ETTm2 Electricity Exchange
Benchmarks: Flight Weather ETTh1 ETTh2 ETTm1 ETTm2 Electricity Exchange
Metrics: MSE MAE average rank
Baselines: TimesNet DLinear NLinear MTGnn Autoformer Informer
Builds on: TimesNet MTGNN Mixhop Informer
Closest papers by meaning
- Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling (ICML 2024 · similarity 0.88)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.88)
- MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification (AAAI 2025 · similarity 0.88)
- FilterTS: Comprehensive Frequency Filtering for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.88)
- Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting (AAAI 2025 · similarity 0.87)
- Time-MMD: Multi-Domain Multimodal Dataset for Time Series Analysis (NeurIPS 2024 · similarity 0.87)
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.