Self-supervision Spatiotemporal Part-Whole Convolutional Neural Network for Traffic Prediction

Traffic is a relatively broad concept, including transportation, travel, trade, and internet networks. It is a kind of method to analyze, model and give predictive results for a given sequence with temporal and spatial relations. Traffic forecasting has always been a hot issue for researchers. It is a non-stationary time series with a high degree of nonlinearity, and it is very challenging to accurately forecast it. We propose a novel self-supervision Spatiotemporal Part-Whole Convolutional Neural Network (STPWNet), which simultaneously captures the temporal and spatial correlations of the traffic sequence to accurately predict the traffic data at the next moment. In order to improve the inference accuracy and speed of the deep network, we designed a lightweight convolutional network module with a part-whole structure to improve the accuracy and speed of network prediction. Compared with traditional neural networks, STPWNet has fewer parameters, faster inference speed, and can produce good prediction performance on a variety of traffic data sets. Experiments show that our proposed network uses only a small number of parameters compared with other networks, and can achieve quite good performance. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

Traffic is a broad concept that involves transportation, travel, trade, and internet networks. Forecasting traffic data accurately is a challenging issue due to its non-stationary and highly nonlinear nature. The proposed STPWNet model shows improved performance with fewer parameters and faster inference speed compared to traditional neural networks.