Abstract
ModWaveMLP is a streamlined, MLP-based framework for traffic forecasting that outperforms complex Transformer and Graph Neural Network (GNN) models. By integrating multi-level mode decomposition with wavelet-based denoising, it achieves state-of-the-art results on standard benchmarks while maintaining a simpler, more efficient architecture.
Introduction
Traffic forecasting—predicting future flow, speed, or demand—is a pillar of Intelligent Transportation Systems. Recently, the field has been swept up in the "Large Model" trend, with researchers deploying increasingly intricate architectures like Transformers and GNNs to capture spatial and temporal dependencies. However, traffic data is essentially sequential numerical information, which is lower-dimensional and simpler than natural language or high-resolution images.
Heavy models often lead to redundancy, high computational costs, and noise sensitivity. Sensor-collected traffic data is inherently noisy, and complex models can inadvertently overfit to this noise rather than learning the underlying traffic "modes" or patterns. ModWaveMLP addresses these challenges by shifting the focus from architectural complexity to signal processing principles.
Methodology
The ModWaveMLP architecture consists of four primary components designed to handle signal noise and periodicity efficiently:
- Wavelet Decomposition Learning Module: Instead of feeding raw, noisy data directly into the network, this module uses wavelet transforms (specifically Daubechies wavelets) to decompose the signal. This allows the model to filter out high-frequency noise and focus on meaningful underlying trends.
- Mode Decomposition Learning (MDL) Block: This is the core of the model. It uses a residual separation idea where stacked MLP layers learn to isolate different traffic patterns step-by-step. One branch predicts the current mode, while another subtracts it from the input to let the next layer focus on the remaining information.
- Time Gate Module: To handle the strong periodicity of traffic (e.g., daily rush hours), this module uses embeddings for "time of day" and "day of week." A multiplicative gate then decodes these temporal effects to restore or remove periodic disturbances as needed.
- Information Aggregation Module: This captures the dynamic history of nodes in the traffic network using node embeddings, allowing the model to learn spatial correlations without requiring a pre-defined graph topology.
Experimental Results
The model was evaluated on two gold-standard real-world traffic datasets: METR-LA (Los Angeles traffic speed) and PEMS-BAY (San Francisco Bay Area traffic speed). Both datasets are aggregated into 5-minute intervals, and the task is to predict the next 12 time steps (60 minutes) based on the previous 12.
ModWaveMLP achieved State-of-the-Art (SOTA) performance across both datasets, consistently outperforming top-tier GNNs (like DCRNN and Graph WaveNet) and Transformers (like GMAN and PDFormer). On the PEMS-BAY dataset, ModWaveMLP improved the Mean Absolute Error (MAE) for short-term (15-min) predictions by 20.97% and the Root Mean Squared Error (RMSE) by 27.31% compared to the next best baseline.
Key Insights
- The "Less is More" Philosophy: The paper reinforces that well-designed MLPs can often beat Transformers in time-series tasks. By focusing on how the signal is decomposed rather than how "attention" is calculated, the model achieves better performance with less overhead.
- End-to-End Denoising: Integrating wavelet transforms directly into the learning pipeline is a powerful way to handle sensor error. The model doesn't just see the noise; it learns to separate it from the signal.
- Implicit Topology: Unlike GNNs, which require a pre-defined map of the road network, ModWaveMLP learns these relationships through node embeddings, making it more generalizable.
Conclusion and Future Work
ModWaveMLP demonstrates that traffic forecasting is more about smart signal decomposition and less about building the largest possible neural network. While the model shows impressive robustness to noise, future work could explore different wavelet selection criteria and the computational scaling of the Information Aggregation Module for "mega-cities" involving thousands of sensors.
Summary updated