Do We Really Need GCNs in Traffic Forecasting? A Graph-Less Pure-MLP Architecture

Traffic flow forecasting, as a fundamental task of intelligent transportation systems driven by the Web, remains challenging due to the complex spatial and temporal dependencies between different roads. Considering the inherent graph structure of road networks, recent works capture the spatial dependencies by utilizing graph convolutional networks (GCNs) and have devoted great efforts to enhancing GCNs. However, graph convolution, as the core of GCNs, has quadratic time and space complexity due to the internal matrix multiplication operation on the adjacency matrix, which prevents existing works from being deployed on huge road networks. In this paper, we doubt the necessity of graph convolution in capturing spatial dependencies and find that the aggregating operation in graph convolution can be replaced by Multi-Layer Perceptron (MLP) with layer normalization in traffic forecasting. Inspired by the emerging MLP models on other deep learning tasks, we further propose a pure-MLP traffic forecasting model without the dependence on graph convolution. Experiments on several real-world traffic datasets show that our pure-MLP model is more efficient and has competitive performances compared to state-of-the-art works.