面向交通拥堵预测的可解释性时空分析方法

Download paper here

The invention discloses an interpretable spatio-temporal analysis method for traffic congestion prediction, which extracts the key features causing congestion events and the deep connection between roads from the interpretation. Traditional data mining methods often explore the correlation between traffic spatio-temporal data from the statistical point of view, and it is difficult to fully reveal the deep connection and key factors of traffic congestion. Therefore, the invention proposes a spatio-temporal interpretation generation model based on STGCN, utilizes the characteristic that neural networks are good at finding hidden features, and uses the interpretable technology of deep learning to extract key input features that neural networks pay attention to. The model uses perturbation-based interpretation method to generate mask, and gradient-based interpretation method to generate gradient map of mask. In view of the coarse granularity and poor pertinence of spatial mask, a step-by-step mask method is proposed to reduce the interpretation granularity. In this way, the effective extraction of hidden information is increased, thus obtaining more accurate and comprehensive congestion key information.

Recommended citation: 关佶红，孔令百，杨涵晨，李文根，张毅超. 面向交通拥堵预测的可解释性时空分析方法, 202211094177.3, 2022/09/08