Exploring to predict the tipping points in traffic flow: A lightweight spatio-temporal information-enhanced neural point process approach

Accurate prediction of traffic flow status is a fundamental and crucial task in intelligent transportation systems construction. To better serve various urban traffic scenarios, predicting traffic flow status requires understanding the overall evolving trends as well as capturing tipping points. In this paper, tipping points in traffic flow are defined as moments when significant local maxima or minima occur in traffic flow. They represent the starting points of traffic flow changes (transition from congestion to smooth flow or smooth flow to congestion), which facilitate early traffic scheduling and decision-making. Currently, complex deep learning models based on spatio-temporal graph neural networks, spatio-temporal attention mechanisms, etc., have made breakthrough progress in short-term overall trend prediction for traffic flow. However, accurately predicting tipping points in traffic flow remains a challenging task due to the sparsity of these points and the complex inherent spatio-temporal correlations. To address this issue, in this paper, we exploratively propose a Spatio-Temporal Information-Enhanced Neural Point Process (STENPP) approach for predicting tipping points in traffic flow. We first model tipping points in traffic flow as event sequences based on self-exciting point process, then learn the hidden representation of event sequences through a lightweight architecture based on MLP-Mixer and parameterize the intensity function calculation in the point process. Simultaneously, we utilize the temporal patch network and learnable spatio-temporal embeddings in conjunction with multiple layers of MLP networks for the spatio-temporal representation learning of long-term traffic flow. Finally, by fusing the learned event representation and spatio-temporal representation, we optimize the overall model by maximizing the likelihood of event sequences and prediction absolute errors. Our proposed model achieves a minimum improvement of 5.92% and a maximum improvement of 11.45% over baseline models on two real-world datasets, paving the way for further research in accurately predicting tipping points in traffic flow.