Research map / Architectures and efficiency / RNNs and normalization
ModWaveMLP: MLP-Based Mode Decomposition and Wavelet Denoising Model to Defeat Complex Structures in Traffic 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 ModWaveMLP, a pure MLP traffic forecaster that uses residual mode decomposition blocks and a wavelet-denoising input branch, achieving SOTA on METR-LA and PEMS-BAY without graph or attention structures.
Problem
Traffic speed forecasting typically relies on complex GNN or transformer architectures with hand-designed features and is sensitive to sensor noise; the paper asks whether such elaborate structures are necessary.
Approach
An MLP stack in which each block predicts a traffic mode via a backcast/forecast residual split (subtracting the backcast from the input and summing forecasts), combined with a multi-wavelet-denoised input representation and a learned time gate, plus a node-embedding-based information aggregation module instead of an explicit graph.
Key results
- On PEMS-BAY, ModWaveMLP reaches average MAE 1.29, RMSE 2.96, MAPE 2.89% over horizons 3/6/12, versus the best baseline MegaCRN-level results of about 1.59 MAE.
- On METR-LA, ModWaveMLP achieves average MAE 2.67, RMSE 5.33, MAPE 7.23%, outperforming D2STGNN (2.94 MAE) and other GNN/transformer baselines.
- Under Gaussian noise injection, ModWaveMLP's maximum MAE change is 0.6 versus 1.2 for FC-GAGA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ModWaveMLP instead of D2STGNN | Traffic forecasting · METR-LA, Average | MAE: 2.94 -> 2.67* | Helped |
| use ModWaveMLP instead of D2STGNN | Traffic forecasting · PEMS-BAY, Average | MAE: 1.55 -> 1.29* | Helped |
| add day-cycle gating and encoding (ablation) | ModWaveMLP · METR-LA, Average | MAE: 2.78 -> 2.71 | Helped |
| add week-cycle gating and encoding (ablation) | ModWaveMLP · METR-LA, Average | MAE: 2.73 -> 2.71 | Helped |
| add information enrichment from mode decomposition (ablation) | ModWaveMLP · METR-LA, Average | MAE: 2.79 -> 2.71 | Helped |
| add wavelet decomposition and coding (ablation) | ModWaveMLP · METR-LA, Average | MAE: 3.06 -> 2.71 | Helped |
| replace MLPResidual with MDL blocks (ablation) | ModWaveMLP · METR-LA, Average | MAE: 2.96 -> 2.71 | Helped |
Surprising findings
- A pure MLP without graph structure or attention outperforms GNN and transformer-based traffic models on METR-LA and PEMS-BAY.
- Replacing MDL blocks with TiDE's MLPResidual block degrades performance, which the authors attribute to normalization and dropout being ill-suited to this setting.
Limitations
- Robustness advantage shown only with synthetic Gaussian noise; reliability declines at higher noise variance.
- Efficiency claims rely on an appendix not included in the main text.
- Evaluation limited to two speed datasets (METR-LA, PEMS-BAY).
Experimental setup
Datasets: METR-LA PEMS-BAY
Benchmarks: METR-LA PEMS-BAY
Metrics: MAE RMSE MAPE
Baselines: HA VAR SVR FC-LSTM DCRNN STGCN Graph WaveNet MTGNN MegaCRN D2STGNN GMAN ASTGCN
Builds on: N-BEATS STID TiDE Graph WaveNet
Closest papers by meaning
- Decomposed Spatio-Temporal Mamba for Long-Term Traffic Prediction (AAAI 2025 · similarity 0.83)
- WardropNet: Traffic Flow Predictions via Equilibrium-Augmented Learning (ICLR 2025 · similarity 0.83)
- SPD-DDPM: Denoising Diffusion Probabilistic Models in the Symmetric Positive Definite Space (AAAI 2024 · similarity 0.82)
- SSL-STMFormer Self-Supervised Learning Spatio-Temporal Entanglement Transformer for Traffic Flow Prediction (AAAI 2025 · similarity 0.82)
- WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.82)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.81)
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.