Abstract:
Modern cities heavily rely on complex transportation, making accurate traffic speed prediction crucial for traffic management authorities. Classical methods, including statistical techniques and traditional machine learning techniques, fail to capture complex relationships, while deep learning approaches may have weaknesses such as error accumulation, difficulty in handling long sequences, and overlooking spatial correlations. Graph neural networks (GNNs) have shown promise in extracting spatial features from non-Euclidean graph structures, but they usually initialize the adjacency matrix based on distance and may fail to detect hidden statistical correlations. The choice of correlation measure can have a significant impact on the resulting adjacency matrix and the effectiveness of graph-based models. This thesis proposes a novel approach for accurately forecasting traffic patterns by utilizing a multi-view spatio-temporal graph neural network that captures data from both realistic and statistical domains. Unlike traditional correlation measures such as Pearson correlation, copula models are utilized to extract hidden statistical correlations and construct multivariate distribution functions to obtain the correlation relationship among traffic nodes. A two-step approach is adopted, which involves selecting and testing different types of bivariate copulas to identify the ones that best fit the traffic data, and utilizing these copulas to create multi-weight adjacency matrices. The second step involves utilizing a graph convolutional network to extract spatial information and capturing temporal trends using dilated causal convolutions. The proposed ST-CopulaGNN model outperforms previous approaches such as DCRNN and Graph WaveNet, indicating the effectiveness of incorporating copulas in trafficforecasting. Experiments on theMETR-LA and PEMS-BAY datasets show that the proposed model outperforms previous approaches with a slight improvement.
