The Impact of Network Indices Integration on Traffic Flow Imputation Accuracy: A Machine Learning Approach

Traffic flow imputation aims to estimate missing flow values within a traffic network. While machine learning methods have been widely applied to this challenge and outperformed conventional approaches, the current literature must explore the potential benefits of integrating network indices into traffic flow imputation. Network indices encompass attributes associated with nodes and links, such as centrality measures, which provide valuable insights into the network’s structure. This study proposes adopting network indices in addition to link features to enhance the accuracy of traffic flow imputation. Our proposed feature set incorporates network indices from the original graph and its line graph transformation, alongside inherent link features. Subsequently, we employ a correlation matrix analysis to discern relationships between features, thus excluding dependent features. We estimate missing traffic flow based on the proposed feature set by utilizing five state-of-the-art machine learning and deep learning methods, including KNN, Random Forest, XGBoost, Graph Convolutional Network (GCN), and Graph Attention Network (GAT). The results of testing our proposed method across various traffic networks with varying sizes (Sioux Falls, Anaheim, and Chicago) and under different missing rates (0.1-0.9) consistently demonstrate that each machine learning method experiences an improvement in imputation accuracy when augmented with network indices. More specifically, on the Anaheim dataset, our proposed approach improves by 8.74% with KNN, 3.26% with Random Forest, 2.69% with XGBoost, 4.88% with GCN, and 4.85% with GAT, compared to using only link properties as the feature set, in terms of PCC. Our study also includes further investigations into feature importance and computational cost analysis of network indices.