PROJECT TITLE :

Learning Dynamics and Heterogeneity of Spatial-Temporal Graph Data for Traffic Forecasting

ABSTRACT:

It is essential to have accurate traffic forecasting in order to improve the safety, stability, and overall effectiveness of intelligent transportation systems. Despite years of research, accurate traffic prediction still has a number of obstacles to overcome. These obstacles include modeling the dynamics of traffic data along both temporal and spatial dimensions, as well as capturing the periodicity and the spatial heterogeneity of traffic data. The problem is even more challenging for long-term forecasting. In this article, we propose a traffic forecasting system called an Attention based Spatial-Temporal Graph Neural Network, or ASTGNN. Specifically, in the temporal dimension, we design a brand new self-attention mechanism that is capable of making use of the local context. The local context is tailored specifically for the transformation of numerical sequence representations. It makes it possible for our prediction model to capture the temporal dynamics of the traffic data and to take advantage of global receptive fields, both of which are helpful for making long-term forecasts. In the spatial dimension, we develop a dynamic graph convolution module using self-attention to capture the spatial correlations in a dynamic way. This allows us to better understand the relationship between space and time. In addition to this, we model the periodicity explicitly, and we capture the spatial heterogeneity by embedding modules. Experiments carried out on five different real-world traffic flow datasets have shown that ASTGNN is superior to the state-of-the-art baselines in terms of its performance.