Traffic prediction is a crucial component of data management systems, leveraging historical data to learn spatio-temporal dynamics for forecasting future traffic and enabling efficient decision-making and resource allocation. Despite efforts to develop increasingly complex architectures, existing traffic prediction models often struggle to generalize across diverse datasets and contexts, limiting their adaptability in real-world applications. In contrast to existing traffic prediction models, large language models (LLMs) progress mainly through parameter expansion and extensive pre-training while maintaining their fundamental structures. In this paper, we propose ST-LLM+, the graph enhanced spatio-temporal large language models for traffic prediction. Through incorporating a proximity-based adjacency matrix derived from the traffic network into the calibrated LLMs, ST-LLM+ captures complex spatio-temporal dependencies within the traffic network. The Partially Frozen Graph Attention (PFGA) module is designed to retain global dependencies learned during LLMs pre-training while modeling localized dependencies specific to the traffic domain. To reduce computational overhead, ST-LLM+ adopts the LoRA-augmented training strategy, allowing attention layers to be fine-tuned with fewer learnable parameters. Comprehensive experiments on real-world traffic datasets demonstrate that ST-LLM+ outperforms state-of-the-art models. In particular, ST-LLM+ also exhibits robust performance in both few-shot and zero-shot prediction scenarios. Additionally, our case study demonstrates that ST-LLM+ captures global and localized dependencies between stations, verifying its effectiveness for traffic prediction tasks.
20 April 2021 in IEEE Transactions on Intelligent Transportation Systems
Vehicle trajectory collection usually faces challenges such as inaccurate and incomplete trajectory data, mainly due to missing trajectories caused by Global Navigation Satellite System (GNSS) outages. In this paper, a novel ensemble transfer regression framework is proposed for urban environments with transfer learning as the primary solution for constructing a fine-grained trajectory dataset during GNSS outages. First, GNSS and motion information are fused for the training process. Then, a regression-to-classification (R2C) process is employed to implement incremental training to adapt to dynamically changing environments. Third, to account for GNSS outages, transfer learning is integrated to construct a data filtering strategy that minimizes negative sample weights during the current scenario. Finally, a more accurate classification-type loss function for ensemble learning is designed to obtain the ensemble transfer regression model. We utilize real-world datasets to verify the accuracy of the comparative methods and the proposed framework in trajectory interpolation prediction. The experimental results show that our framework is significantly superior to the comparative methods.
Vehicles traveling through cities generate extensive vehicle trajectory collected by scalable sensors, providing excellent opportunities to address urban challenges such as traffic congestion and public safety. In this survey, we systematically review vehicle trajectory collection, preprocessing, analytics, and applications. First, we focus on the standard techniques for vehicle trajectory collection and corresponding datasets. Next, we introduce representative approaches for the latest advances in vehicle trajectory processing. We further discuss individual travel behavior and collective mobility analytics using vehicle trajectory. Since private cars constitute the majority of urban vehicles and form the basis for many recent research findings, we emphasize analytics based on private car trajectory data. We then compile vehicle trajectory-boosted applications from the perspective of computing vehicle trajectory. Finally, we go through unresolved problems with vehicle trajectory and outline potential future research directions.
29 December 2021 in IEEE Transactions on Network Science and Engineering
Accurate prediction of private car flows in urban regions is strategically vital for constructing smart cities. Private car flows are essentially reflected in the arrive-stay-leave (ASL) behaviors of car users. Specifically, to fulfill daily travel needs, private car users usually $arrive$ at specified locations, $stay$ for a certain period participating in their activities, then $leave$ to the next destination. In this paper, we strive to explore global spatiotemporal correlations of ASL behaviors to predict private car flow. Therefore, we formulate the dynamic distribution of ASL behaviors in urban regions through multiple graphs and propose the novel m ulti g raph d ense c onvolutional n etwork ( MGDCN ) to represent spatiotemporal correlations of ASL behaviors for private car flow prediction. The proposed MGDCN is an end-to-end framework consisting of three modules: $i$ ) multigraph dense convolutions, in which we introduce densely connected blocks to aggregate global spatial correlations effectively, $ii$ ) convolutional gated recurrent units to capture sequential temporal correlations of ASL behaviors in all regions, and $iii$ ) attention networks for learning the stay duration correlations in each region. Experiments based on large-scale private car trajectory data and region-of-interest (ROI) datasets verify that the proposed method achieves superior performance than the baselines in all five metrics.
The mining of private vehicle trajectories can be used to establish the regular travel behavior of private vehicles and provide novel solutions to address problems related to urban traffic. With the help of automotive sensors in private vehicles, we can obtain a considerable amount of trajectory data on connected private vehicles. We herein assessed the regular travel behavior from connected private vehicle trajectory data. First, we explored the travel features of private vehicles through an empirical evaluation of move-and-stay frequency, time, duration, and distance. Second, we acquired the fitting distribution of private vehicle move-and-stay features based on the sum of the squared errors. We then proposed a move-and-stay feature based detection (MSFD) method, i.e., a semi-supervised model with incomplete data labels, to identify regular travelers from a massive private vehicle trajectory dataset. Extensive experiments were conducted on the private vehicle trajectory dataset collected from more than 68,069 private vehicles in Shenzhen, China, from June 1, 2016 to August 31, 2016. We compared the proposed MSFD with the existing baselines. The quantitative experimental results demonstrated that the MSFD outperformed these baselines.