Efficient Spatio-Temporal Randomly Wired Neural Networks for Traffic Forecasting

Abstract

Designing neural architecture is very important for traffic forecasting. In the past decade, it mainly depends on human experts. Recently, neural architecture search (NAS) has been leveraged to automatically find the optimal candidates for diversity spatio-temporal forecasting applications. However, it still depends on experts' experiments to narrow down the search space to make a trade-off between accurate forecasting and the high computational cost. In this paper, we try to generate the neural architecture for spatio-temporal forecasting. We propose a novel and efficient spatio-temporal randomly wired neural network generation method, termed ST-RWNet. ST-RWNet is not a fixed neural architecture. It can produce a bunch of spatio-temporal neural networks. ST-RWNet includes a spatio randomly neural network generator, a temporal randomly neural network generator, and a randomly bipartite neural network generator. Those neural network generators are based on random graphs, which control the attributes of neural networks, e.g., the number of pathways, the length distribution of pathways. The spatio and temporal neural network generators are designed to capture the spatio and temporal dependencies respectively. And the randomly bipartite neural network generator is designed to fuse the complicated intermediate spatio-temporal features. Extensive experiments on real-world traffic forecasting datasets have demonstrated that the performance is competitive with state-of-the-art methods. And our method achieves up to $6\times$ speed up compared to other NAS-based methods.