Inferencing hourly traffic volume using data-driven machine learning and graph theory

Abstract

Traffic volume is a critical piece of information in many applications, such as transportation long-range planning and traffic operation analysis. Effectively capturing traffic volumes on a network scale is beneficial to Transportation Systems Management & Operations (TSM&O). Yet it is impractical to install sensors to cover a large road network. To address this issue, spatial prediction techniques are widely performed to estimate traffic volumes at sites without sensors. In retrospect, most relevant studies resort to machine learning methods and treat each prediction location independently during the training process, ignoring the potential spatial dependency among them.This paper presents an innovative spatial prediction method of hourly traffic volume on a network scale. To achieve this, we applied a state-of-the-art tree ensemble model - extreme gradient boosting tree (XGBoost) - to handle the large-scale features and hourly traffic volume samples, due to the model's powerful scalability. Moreover, spatial dependency among road segments is taken into account in the proposed model using graph theory. Specifically, we created a traffic network graph leveraging probe trajectory data, and implemented a graph-based approach - breadth first search (BFS) - to search neighboring sites in this graph for computing spatial dependency. The proposed spatial dependency feature is subsequently incorporated as a new feature fed into XGBoost. The proposed model is tested on the road network in the state of Utah. Numerical results not only indicate high computational efficiency of the proposed model, but also demonstrate significant improvement in prediction accuracy of hourly traffic volume comparing with the benchmarked models.