.png){kind=icon}
.png){kind=icon}
Return
Breaking multilayer perceptron limitations for traffic flow forecasting with structured patch learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2026.114339.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Transformer neural networks and multilayer perceptron (MLP) architectures are two dominant paradigms in multivariate spatiotemporal sequence modeling. While Transformers excel at capturing complex dependencies, their attention mechanisms incur high computational costs, limiting practical deployment. MLP methods are structurally simple and inference-efficient, but existing approaches typically rely on flattened modeling processes, making it difficult to capture hierarchical temporal patterns in traffic data. Moreover, the lack of structural decoupling between spatial and temporal components often leads to entangled representations, reducing overall modeling effectiveness. To address these challenges, we propose a structured, purely MLP model, termed the Structured Patch Multilayer Perceptron (SPMLP), for traffic flow forecasting. The model employs a temporal patch multilayer perceptron with overlapping sliding windows to extract hierarchical temporal features from local to global scales, and a graph-structured multilayer perceptron to capture node dependencies over a predefined spatial graph. A lightweight linear fusion integrates spatial and temporal representations into a unified spatiotemporal context. Extensive experiments on standard traffic flow benchmarks demonstrate that SPMLP outperforms strong baselines in both prediction accuracy and computational efficiency, highlighting the practicality of structured multilayer perceptron architectures for real-time traffic forecasting in intelligent transportation systems.
Keywords:
Structured Patch Learning
Multilayer Perceptron
Traffic Flow Forecasting
Spatiotemporal Modeling
Graph-Structured MLP
Journal
IF:
8
Papers:
5.4K
Citations:
3.5W
Organization
No organization information available