Abstract
Traffic prediction is of great importance for the development of intelligent transportation system (ITS), including travel time estimation (TTE), route recovery and road condition analysis. Most previous works model both road network and vehicle trajectory by learn ing their spatio-temporal characteristics to conduct traffic prediction. Among them, graph neural networ (GNN) has become a popular option for graph-structured traffic data. How ever, there are still two key issues for trustworthy distributed traffic prediction in real-world location-based services, i.e., data sparsity and privacy protection. In this thesis, we aim to solve the above two issues through GNN. For the first issue of data sparsity, various complex factors such as network communi cation and energy constraints of mobile devices, make multiple vehicle trajectory collected at a low sampling rate in the scenes of Internet of Things (IoT). But existing efforts need to label the exact travel time and route between two consecutively sampled GPS points, and train the prediction model based on them. Thus these commonly sparse trajectories provides unreliable performance and leads to a decrease of accuracy. To overcome this, we formulate this problem as an inexact supervision in which the training data has coarsely grained labels, and propose an EM algorithm to jointly solve the tasks of both TTE and route recovery. More specifically, both two tasks are complementary to each other in the model-learning p ter inference for routes (Time → Route), in turn, resulting in a more accurate time estimation (Route → Time). In addition, the road condition can be estimated based on spatio-temporal embedding architecture that integrates the relational road graph and temporal correlations. Especially, when it comes to the extremely sparse scenarios, the origin-destination (OD) type of data, such as NYC taxi data, is more accessible. Many traditional or deep learning methods have been proposed to estimate the OD trips. These existing approaches mainly focus on independent OD input and route search-based algorithms. But the real-time road condition closely associated with both travel time and route is long ignored. Thus we pro pose a multi-task weakly supervised learning framework to infer transition probability be tween road segments, and the travel time of both road segments and intersections simulta neously. Technically, we design the stacked dual-graph architecture to generate the travel time distribution, and multi-layer perception to learn the transition probability. Moreover, an iterative update strategy is also used to update them during the training process. Then for the second issue of privacy protection, building a deep learning model for such data-driven tasks of the ITS requires massive trajectory data. But personal trajectory data sits in local devices, and is less likely to be shared due to the privacy concerns. Without
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
"Trustworthy Distributed Traffic Prediction through GNN", Technical Disclosure Commons, (April 29, 2025)
https://www.tdcommons.org/dpubs_series/8047