Traffic prediction method based on multi-graph convolution and Manhattan non-negative matrix factorization

Traffic sensing systems in the urban areas are susceptible to being offline, experiencing loss in network packets, video occlusions, and abnormal samples while operating. The aforementioned parameters make up for missing data in traffic and reduce the reliability of predictive models. To enhance the accuracy and robustness of traffic prediction when dealing with missing data, this study proposes an approach using multi-graph convolution and Manhattan non-negative matrix factorization. This method constructs a neural network to optimize the Manhattan non-negative matrix factorization completion module. It also integrates a multi-graph convolutional network that combines a spatial adjacency graph, a time similarity graph, a feature similarity graph, and a dynamic adaptive graph. This approach achieves collaborative modeling for missing data completion, traffic pattern extraction, and multi-step prediction. The results indicate that at a 30% missing rate, the completion performance evaluates to 18.64 in Mean Absolute Error, 30.27 in Root Mean Square Error, and 13.35% in Mean Absolute Percentage Error. As for the comprehensive prediction problem, the aforementioned model can reduce the above-mentioned three errors into 14.78, 25.19, and 10.46%, respectively, while the congestion recognition accuracy becomes 93.84%. It is clearly shown in this paper that this approach is capable of improving the efficiency of traffic prediction under conditions of lacking traffic data.