Abstract:Predicting spatio-temporal traffic flow presents significant challenges due to complex interactions between spatial and temporal factors. Existing approaches often address these dimensions in isolation, neglecting their critical interdependencies. In this paper, we introduce the Spatio-Temporal Unitized Model (STUM), a unified framework designed to capture both spatial and temporal dependencies while addressing spatio-temporal heterogeneity through techniques such as distribution alignment and feature fusion. It also ensures both predictive accuracy and computational efficiency. Central to STUM is the Adaptive Spatio-temporal Unitized Cell (ASTUC), which utilizes low-rank matrices to seamlessly store, update, and interact with space, time, as well as their correlations. Our framework is also modular, allowing it to integrate with various spatio-temporal graph neural networks through components such as backbone models, feature extractors, residual fusion blocks, and predictive modules to collectively enhance forecasting outcomes. Experimental results across multiple real-world datasets demonstrate that STUM consistently improves prediction performance with minimal computational cost. These findings are further supported by hyperparameter optimization, pre-training analysis, and result visualization. We provide our source code for reproducibility at <a class="link-external link-https" href="https://anonymous.4open.science/r/STUM-E4F0" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

This paper attempts to solve two main problems in traffic flow prediction: 1. **Separation of Spatial and Temporal Modules**: - Existing methods usually handle the spatial and temporal dimensions separately, ignoring the crucial interdependence between them. This separation limits the effectiveness and efficiency of spatio - temporal representation learning. For example, Figure 1(c) shows how spatio - temporal relationships affect regional prediction. If the prediction module separates spatial and temporal processing, it cannot efficiently propagate regional relationships across time intervals. 2. **Heterogeneity of Spatio - Temporal Data**: - The heterogeneity of spatio - temporal data leads to large pattern changes at different spatial and temporal scales. For example, Figure 1(a) shows the sensor distribution in a region in the PEMS dataset, where traffic flow shows significant differences between different regions. Figure 1(b) shows the traffic flow waveforms in different time periods within the same region, indicating that even within the same region, different time periods also show different traffic dynamics. To solve these problems, the author proposes a new framework - **Spatio - Temporal Unitized Model (STUM)**. The core components of STUM include: - **Adaptive Spatio - temporal Unitized Cell (ASTUC)**: Through low - rank matrix decomposition, ASTUC can calculate, update, and store spatial, temporal, and correlation information in a single cell, thus overcoming the problem of separation of spatial and temporal modules in existing methods. - **Multi - layer Residual Fusion Block (MLRF)**: Utilizing the characteristics of ASTUC, MLRF better captures complex non - linear spatio - temporal dependencies, thus overcoming data heterogeneity and improving computational efficiency and performance. Through these innovations, STUM aims to uniformly handle spatio - temporal dependencies and integrate with various spatio - temporal graph neural networks through a modular design to enhance prediction results. Experimental results show that STUM significantly improves prediction performance on multiple real - world datasets while maintaining a low computational cost.

Cross Space and Time: A Spatio-Temporal Unitized Model for Traffic Flow Forecasting

Space-Time Adaptive Network for Origin-Destination Passenger Demand Prediction

A Foundation Model for Unified Urban Spatio-Temporal Flow Prediction

STSD: Modeling Spatial Temporal Staticity and Dynamicity in Traffic Forecasting

Space Meets Time: Local Spacetime Neural Network For Traffic Flow Forecasting

Spatiotemporal Fusion Transformer for large-scale traffic forecasting

Dynamic Spatial–Temporal Self-Attention Network for Traffic Flow Prediction

Efficient Adaptive Spatial-Temporal Attention Network for Traffic Flow Forecasting

A Lightweight and Accurate Spatial-Temporal Transformer for Traffic Forecasting

Adaptive and Interactive Multi-Level Spatio-Temporal Network for Traffic Forecasting

Location and time embedded feature representation for spatiotemporal traffic prediction

Unified Spatial-Temporal Neighbor Attention Network for Dynamic Traffic Prediction

Dynamic Spatial-Temporal Memory Augmentation Network for Traffic Prediction

Attention-Based Spatial–Temporal Convolution Gated Recurrent Unit for Traffic Flow Forecasting

Attention Based Spatial-Temporal Dynamic Interact Network for Traffic Flow Forecasting.

Deep Spatio-temporal Adaptive 3D Convolutional Neural Networks for Traffic Flow Prediction

Learning to Effectively Model Spatial-Temporal Heterogeneity for Traffic Flow Forecasting

Traffic Flow Prediction Based on Spatio-temporal Attention Block

Dynamic Spatiotemporal Straight-Flow Network for Efficient Learning and Accurate Forecasting in Traffic

Deep spatio-temporal neural network based on interactive attention for traffic flow prediction