The contextual reasoning and generalization capabilities of large language models (LLMs) are crucial for next-generation intelligent transportation systems to achieve accurate traffic flow prediction. However, existing schemes that integrate LLMs with spatiotemporal traffic forecasting face challenges such as spatiotemporal coupling characteristics, unified representation of multi-source heterogeneous data, and the need for cross-scenario generalization. To address these challenges, this paper proposes a scene-adaptive prediction framework based on a spatiotemporal large language foundation model. Firstly, a spatiotemporal large language foundation model for transportation is constructed through structured spatiotemporal prompt engineering and full-parameter fine-tuning, achieving domain adaptation of the general-purpose LLM. Secondly, a scene-adaptive prediction framework is designed. This framework innovatively proposes a tree structure-based spatial embedding layer to capture the hierarchical spatial dependencies of traffic networks, combines it with a temporal embedding layer to learn cyclical patterns, and employs parameter-efficient fine-tuning methods such as dynamically rank-allocated Low-Rank Adaptation and bottleneck Adapters to achieve rapid adaptation to different scenarios under limited computational resources. Experiments on multiple real-world traffic datasets demonstrate that the proposed framework achieves certain accuracy in terms of MAE and RMSE, showing certain generalization capability and robustness. It achieves an average prediction error reduction rate of 6.62% compared to existing representative advanced baseline models. Open-source links: https://huggingface.co/ubinet/LightTrafficLLM-FM; https://github.com/williamZQ/LightTrafficLLM.

Urban mobile crowdsensing captures urban states through the routine trajectories of mobile participants. Its central challenge is not only to generate a feasible plan, but also to enable stable and constraint-valid improvement over existing worker trajectories. Existing task allocation methods mainly target initial planning and lack reusable operation-level refinement interfaces. Directly asking a local large language model to generate complete plans or freely edit trajectories can also violate hard constraints on budget, time, and reachability. To address this interface gap, this paper proposes ACIC-Core, which reformulates urban mobile sensing plan improvement as operation-level refinement under contract constraints. In ACIC-Core, the local large language model no longer generates worker trajectories directly. Instead, it proposes local modification intentions within explicit operation contracts, while deterministic modules convert these intentions into executable and verifiable trajectory modifications. This design converts open-ended language planning into worker-level modifications within a finite operation space. As a result, each refinement step is executable, verifiable, and traceable. Experiments on T-Drive and Porto show that ACIC-Core consistently improves plan quality across task scales, initial planners, and local model settings. The framework also maintains high path feasibility across these settings. Ablation studies and trace analysis further show that the gains mainly come from explicit action contracts, programmatic execution, and verification-based adjudication. They do not depend on the free-form planning ability of the large language model.

Efficient urban traffic management and intelligent urban planning are pivotal drivers for the development of modern cities. Among them, accurate prediction of taxi flow, as a core component of the urban transportation system, has significant implications for traffic management, resource allocation, and the daily lives of citizens. However, practical taxi flow prediction faces numerous challenges, particularly in fully utilizing periodic features. To address the issue of capturing periodic characteristics in urban taxi flow prediction research, this study proposes a prediction model named Period-based Spatial-Temporal Graph Convolution Network (PSTGCN). In the PSTGCN, graph convolutional techniques are employed in the spatial convolution module to capture both local and global spatial correlations of taxi flow, effectively integrating the two. This further enhances the understanding of flow interaction between different areas. Furthermore, in the temporal convolution module, PSTGCN transforms the one-dimensional time series data into two-dimensional data, enabling the simultaneous analysis of both the within-period changes in taxi flow and the patterns of variation between different periods. This significantly strengthens the capability to capture flow periodic features. Lastly, this study extensively validates the predictive performance of the PSTGCN model. Experimental results demonstrate its significant advantages over various existing benchmark models in terms of prediction effectiveness. Moreover, the model's excellent generalization ability is showcased through validation using multiple real-world datasets.

The Intelligent Transport Systems represent a pivotal research area within the broader context of smart city con struction. It constitutes a vital component of the contemporary urban transport system, with the potential to facilitate high-quality development. The prediction of traffic flow represents a significant research area within the field of ITS. It offers a valuable opportunity to develop a robust data foundation for the planning and op timisation of urban traffic road networks. The majority of studies in this field currently employ static graphs and graph neural networks to complete the traffic flow prediction task. The use of static graphs for traffic flow prediction is inadequate for capturing the dynamic spatial and temporal characteristics of the traffic network structure. Furthermore, graph neural networks are only capable of performing local spatial characteristic analy sis. To address the issue of global feature analysis of traffic network topology, multi-layer graph neural networks are required for iterative computation. The number of layers of graph neural networks increases in line with the number of nodes in the traffic network. To address the aforementioned issues, this study proposes a neural network architecture that employs a tree structure for attention computation, namely the Spatio-temporal Tree Attention Network (STTAT). In particular, this study proposes a tree-structured representation of traffic network topology and a tree-structured attention computation method for learning global features of traffic network topol ogy. The proposed model has been evaluated on several real-world traffic datasets, and its performance has been compared with that of several baseline models. The results demonstrate that the proposed model significantly outperforms the baseline models in terms of prediction accuracy.

Traffic flow forecasting is a key technical component of Intelligent Transportation Systems (ITS). It overcomes the limitations of traditional passive traffic management and shifts the paradigm from "reactive response" to "proactive forecasting." Existing studies have primarily focused on modeling nonlinear traffic flow using models with massive parameters. However, long computing time is always a problem happened on these models. While maintaining high accuracy, ways to reduce computing time have rarely been explored in depth. Additionally, current research seldom leverages implicit nonlinear trends by proposing a framework to preprocess multi-source factors influencing traffic flow. The balance between accuracy and efficiency is critical to determining the practicality and reliability of ITS. The "accuracy-efficiency" trade-off in nonlinear traffic flow forecasting is addressed by proposing a Fast Deep Spatial-Temporal Perception (FDSTP) model, which integrates Tucker decomposition and multi-source factors for efficient forecasting. Experiments on four real-world datasets demonstrate that FDSTP reduces mean average error, root mean square error, combined accuracy, and computing time by approximately 35.68%, 29.88%, 31.99%, and 50.12%, respectively, compared to the baselines. Besides, FDSTP increases coefficient of determination (R2) by about 15.43%.

Traffic prediction is a foundational element of intelligent transportation systems. Contemporary spatio-temporal graph learning methodologies have achieved encouraging outcomes. However, their efficacy is frequently constrained by the necessity of task-specific architectures, which impedes their generalizability and transferability. Recently, large language models (LLM) have emerged as a unifying sequence modeling paradigm. Nevertheless, the direct application of these methods to traffic prediction remains challenging. This is due to the continuous-valued nature of traffic observations and the complex spatio-temporal dependencies that are inherent in traffic networks. To address this issue, we hereby propose SPSTLLM, an advanced statistical spatio-temporal large language model for traffic prediction. Specifically, statistical priors, including trends, dispersion, distributional characteristics, and autocorrelation, are derived from historical traffic observations. The aim is to construct multi-perspective spatio-temporal representations. These representations are then aligned with the LLM’s hidden space and combined with task instructions for prompt-based prediction. In addition, low-rank adaptation (LoRA) is introduced to facilitate parameter-efficient adaptation, and a unified optimization objective is designed to jointly account for numerical regression, language supervision, and spatial consistency. A series of experiments conducted on four PeMS benchmark datasets demonstrated that SPSTLLM exhibits superior performance in comparison to both representative spatio-temporal prediction models and LLM-based baselines. Ablation studies were employed to verify the effectiveness of its key components, while few-shot results further demonstrated its strong adaptability and generalization ability under data-scarce scenarios.

Missing traffic flow data poses a significant challenge to intelligent transportation systems (ITS). It often results from power interruptions, sensor failures, or network congestion, which undermines data reliability for real-time decision-making. While existing imputation methods achieve high accuracy through complex, parameter-heavy frameworks, they suffer from long computing times. It means that they fail to balance accuracy and efficiency effectively. Urban traffic networks are constantly evolving. This means that traffic flow imputation models must be regularly retrained and maintained to ensure the long-term, smooth operation of ITS. Designing efficient models can reduce system downtime and maintenance time. Furthermore, efficient model design strategies enable ITS to reduce their reliance on high-cost hardware, thereby achieving low-cost traffic management. Therefore, Normalized Spatial-Temporal Factor Autoregression (NSTFA) model is proposed for efficient missing traffic flow imputation. NSTFA begins by representing incomplete traffic data as a third-order tensor and applying Tucker decomposition to extract a low-rank core tensor and spatial-temporal factor matrices. Normalization is then applied to reduce computational complexity and prevent overfitting. A spatial-temporal autoregressive process is integrated to model unidirectional temporal dependencies and multidirectional spatial effects directly on the factor matrices. Finally, inverse Tucker decomposition reconstructs the complete tensor. NSTFA is versatile for grid-based Euclidean data and, via Multiple Delay Embedding Transform (MDT), for non-Euclidean graph-structured data. Experiments on four real-world Beijing datasets show NSTFA reduces mean absolute error by 11.32%, root mean square error by 10.65%, and computing time by 95.15% compared to baselines, with stable performance at missing rates up to 75%.

Shared mobility is critical to urban transportation, yet its complex spatiotemporal dynamics challenge traditional prediction methods. We propose the Tucker Decomposition-based Spatiotemporal Tensor Autoregressive Model (T-STAR), which leverages tensor-structured data modeling and Tucker decomposition to efficiently capture multi-dimensional dependencies. Unlike conventional methods, T-STAR preserves high-dimensional structures by decomposing raw spatiotemporal data into a low-rank core tensor and mode-specific factor matrices, reducing complexity and enhancing interpretability by decoupling spatial, temporal, and modal interactions. Experimental results on three benchmark datasets demonstrate T-STAR's strong performance. On the Beijing Taxi Trajectory Dataset (TaxiBJ), T-STAR achieves Mean Absolute Error (MAE) of 23.53 and Root Mean Square Error (RMSE) of 37.71, improving performance by 18.5 % and 21.2 % over baseline averages. On the New York City Taxi Dataset (NYCtaxi), it records MAE of 18.18 and RMSE of 46.87, reducing errors by 22.7 % and 15.4 %. In the sparse-demand New York City Bike-Sharing Dataset (NYCbike), it maintains robust accuracy with MAE of 7.95 and RMSE of 14.32, outperforming baselines by 14.1 % and 17.9 %, respectively. Most notably, T-STAR achieves these results at high speed: on TaxiBJ, it completes a prediction in just 0.35 seconds-87 % faster than the Adaptive Graph Convolutional Recurrent Network (AGCRN) and 99.8 % faster than the Diffusion Convolutional Recurrent Neural Network (DCRNN). By retaining over 95 % of key spatiotemporal correlations through Tucker compression, T-STAR reduces prediction error by 20-30 % while delivering real-time performance, offering a scalable framework for urban traffic prediction and shared vehicle scheduling. Code and data are both available at yanhongyu0/TSTAR (github.com)

Pedestrian trajectory prediction is of great significance in fields such as autonomous driving, intelligent security, and robot navigation. Its accuracy directly affects the safety and reliability of these systems. Traditional models often rely solely on fixed information like pedestrian speed, direction, and distance. They cannot effectively capture complex pedestrian interaction relationships and fail to deeply integrate scene and social features, limiting the improvement of prediction accuracy. In this study, the Dynamic Scene Trajectory Fusion Adaptive Graph Convolution Trajectory Prediction Model (DSTFAGCN) is proposed. This model innovatively introduces a dynamic weight adjustment mechanism and a hierarchical graph structure to achieve more accurate dynamic social interaction modeling. It also uses a temporal scene feature propagation mechanism combined with a scene attention mechanism to realize more effective spatio-temporal scene feature fusion. Meanwhile, a Transformer-based graph convolution network (TGCN) is adopted to significantly enhance the graph convolution effect. Experiments on the ETH-UCY dataset show that compared with existing models, the overall prediction error of the DSTFAGN model is reduced by 12.3%, remarkably improving the accuracy and reliability of pedestrian trajectory prediction.

As the urbanization process accelerates, the problem of traffic congestion is becoming increasingly severe. Efficient traffic flow prediction is crucial for traffic management and urban planning. This work proposes a spatiotemporal analysis model based on multi-order sampling neighbor aggregation, aiming to overcome the over-smoothing problem inherent in the traditional GCN multi-layer stacking. In the model construction, this work designs a spatial analysis module of multi-order sampling neighbor aggregation, which effectively solves the over smoothing problem and reduces the consumption of computing resources. In addition, this work also designs a time analysis module based on residual gated units, which uses a residual structure to complete the temporal analysis task of traffic data. The model is evaluated on real open-source datasets, and the results show that this model is superior to other comparative models, with a comprehensive error reduction of 6.34

Accurate causal discovery in telecommunication alarm event sequences is crucial for reliable root cause analysis, but presents significant challenges due to complex topological dependencies and inherent alarm redundancy. To address these challenges, we propose the topological-temporal convolution transformer Hawkes process (TTCTH), a novel framework for learning hidden causal relationships in alarm event sequences. To start with, a topological-temporal graph convolution is designed to initialize a causal graph for mainly exploring local dependencies among event types in the topological and temporal domains. We next design three separable transformer variants of topological transformer (TopoT), causal transformer and temporal transformer, which are sequentially stacked to globally capture causal dependencies among event types. In particular, the TopoT fuses prior topological information into alarm sequence representation. The causal transformer models causal dependencies with self-attention mechanism to capture spatially causal relations among event types. The temporal transformer is developed to model long-range causal dependencies across multiple events. We then leverage the high-dimensional hidden state representation, encoded by the stacked transformers, to construct feature evolution with the constraint-based Hawkes process. In order to observe the performance of TTCTH, we conduct extensive experiments on real-world alarm datasets. It demonstrates TTCTH's excellence in comparison with ten baselines.

In the field of Intelligent Transportation Systems (ITS), nonlinear sparse traffic flow prediction faces significant challenges of high sparsity, nonlinear spatial-temporal heterogeneity, and "accuracy-speed" balance. Statistical learning methods and traditional machine learning models usually find it difficult to capture the spatial-temporal continuous evolution of nonlinear traffic flow dynamics systems. The code executing time of deep spatial-temporal models with many parameters is usually long, and researchers have not deeply addressed the computing efficiency problem. To address the above problems, a novel hybrid model (Fast-TrafficNet) is proposed, which integrates Spatial-Temporal Graph Neural Control Differential Equations (STG-NCDE) and Block-Term Decomposition (BTD) for efficient prediction of nonlinear spatial-temporal dynamical systems of sparse traffic flow. STG-NCDE captures the non-Euclidean nonlinear spatial dependence features. BTD extracts global patterns and local heterogeneous features through efficient tensor low-rank mapping, which in turn effectively compresses the high-dimensional sparse data and suppresses the noise interference to reduce the model complexity. Experiments show that under four types of real-world datasets and 10 %-75 % sparse data, Fast-TrafficNet reduces mean average error and root mean square error by an average of 17.3 % and 13.6 % compared with the optimal baseline, and consumes only 4.22 % of the code executing time. This study provides a new paradigm for modeling nonlinear sparse traffic flows, and the proposed Fast-TrafficNet can be extended to other nonlinear sparse traffic dynamics system prediction tasks, which contributes to the construction and sustainable development of ITS and smart cities. Open-source link: https://github.com/qdu318/Fast-TrafficNet.

This study proposes an innovative framework that combines spatiotemporal long- and short-term memory networks (ST-LSTM) with hierarchical Tucker decomposition (HTD), aiming at efficiently processing and predicting complex spatiotemporal data, such as the demand for shared trips. The framework compresses the original tensor data into low-rank cores and factor matrices through a recursive hierarchical decomposition strategy, which not only significantly reduces the storage and computation overheads, but also improves the data processing efficiency, especially in sparse data scenarios showing superior performance. In addition, ST-LSTM achieves accurate modeling of multi-scale features through a lightweight spatio-temporal gating mechanism, capturing the long-term and short-term dependencies in time series.HT-STNet also proposes a dynamic feature selection and gradient masking mechanism, which effectively solves the problem of localized sparsity in traffic data, and avoids redundant computation of invalid information and zero-valued elements through sparsity-aware decomposition of rank adjustment. The experimental results show that HT-STNet outperforms multiple mainstream baseline models in terms of prediction accuracy, computational efficiency, and sparse data processing capability, especially in multi-scale feature extraction and dimensionality reduction. The method provides an efficient and robust solution for high-dimensional spatio-temporal data modeling, which is especially suitable for complex travel demand prediction tasks and breaks through the bottleneck of traditional models in characterizing complex spatio-temporal relationships.

Spatio-temporal forecasting has wide applications across various domains, particularly in intelligent transportation systems, where it plays a crucial role. Traffic flow prediction, a typical spatio-temporal forecasting task, involves complex dependencies across both time and space dimensions. Current research predominantly relies on graph neural networks (GNNs) for modeling. However, deep GNN architectures often face the issue of over-smoothing. To address this challenge, recent studies have explored integrating residual connections or neural ordinary differential equations (ODEs) with GNNs. Nonetheless, existing graph ODE methods have limitations in initializing latent feature representations for time series data and capturing higher order spatio-temporal dependencies. Additionally, they struggle to extract multi-scale temporal dependencies. In this paper, we propose a framework called the Multiple Second-order Continuous Graph Neural Network. The framework utilizes a second-order continuous GNN, and experiments on four real-world datasets demonstrate that it outperforms mainstream baseline models, thereby confirming the effectiveness of the proposed method.

With the development of urban traffic, road networks are becoming increasingly complex and traffic congestion is getting worse. Accurately predicting traffic data is one of the most important means to keep urban roads smooth and is one of the key technologies to improve urban traffic management capabilities. Graph Convolutional Networks (GCNs) are a type of deep learning method used to analyze spatial information in traffic networks. There are two problems with traditional GCNs. First, GCNs can only solve first-order models that only consider the influence between adjacent nodes’ temporal data, and stacking multiple layers of graph neural networks can easily lead to over-smoothing phenomena. Second, as the number of layers increases, the number of trainable parameters in the model increases, so does leading to model overfitting. Therefore, this paper proposes a traffic flow prediction model based on multi-order Markov chains. The paper estimates the parameters of this model and refers to the urban road traffic operation evaluation index system. Through prediction experiments, it is proven that the prediction accuracy of this model is higher than other baseline models.

Recent years, the Transformer has achieved remarkable results in scene text spotting. This paper uses the classic encoderdecoder architecture to propose a Text Spotting model based on Transformer with bidirectional Explicit Points sampling (TSEP). As a sequence, the text contains rich semantic information in its forward and backward features. We model each text instance through bidirectional explicit point sampling. After decoding by the decoder, the positional and semantic information of the text is integrated into the explicit points. Therefore, a basic prediction head is capable of producing the boundary of the text region, the text content, and the corresponding confidence scores. Additionally, we propose a reference point feature enhancement module constructed using one-dimensional convolutions and MLP to address the spatial inductive bias of non-local self-attention in Transformers. Experimental results across various public datasets indicate that our model outperforms numerous other leading models.

The graph neural network model has a wide range of application value in the intelligent transportation system. However, the data design pattern based on graph structure cannot solve the directional and hierarchical problems of node information transmission. This makes the deep models represented by graph convolutional networks lack a certain predictive ability in node distribution scenarios. In this study, a tree spatial-temporal model with tree structure as the sample space is designed for the traffic node distribution scenarios. Firstly, road nodes and spatial relationships are abstracted according to the graph structure, so as to realize the preliminary spatial distribution relationship of nodes. Secondly, different nodes are used as the root nodes of the tree to construct the plane tree structure and plane tree matrix to complete the conversion process from the graph structure to the tree structure. Finally, the plane tree matrices of all nodes are fused into a spatial tree matrix representing the spatial global relationship of the nodes. This study designs the deep tree traffic forecast model based on tree structure, which converts the graph structure of small-scale aggregated nodes into tree structure. The deep tree traffic forecast model realizes the mining and prediction tasks of various traffic measurements based on the spatial tree convolution module and the temporal convolution module. This study demonstrates the excellent predictive ability of the deep tree traffic forecast model in traffic node distribution scenarios by comparing with multiple existing baselines on real datasets.