Pre-training and online updating method for multi-view dynamic trajectory representation learning

Abstract

The invention discloses a pre-training and online updating method for multi-view dynamic trajectory representation learning, which comprises the following steps of: acquiring GPS (Global Positioning System) streaming data of a moving object, modeling a road network into a directed graph, and constructing multi-view joint representation by combining POI (Point of Interest) semantic features of a free space grid; road topology, regional semantics and multi-granularity time features are fused based on a space-time hybrid expert model, and universal track embedding is generated through span mask pre-training and triple contrast learning; dividing a model encoder into a frozen layer and a hot update layer according to dynamic evolution of trajectory data, and dynamically adjusting a parameter update proportion to balance historical mode retention and new feature learning; whenever the new trajectory data triggers the updating of the model, from the time-space fusion representation, the downstream task demand is adaptively matched through the gating weight, the incremental trajectory representation vector is generated, and the multi-task analysis report is performed, so that more accurate, general and efficient representation is provided for multiple downstream tasks of the trajectory data.

Description

Technical Field

[0001] This invention belongs to the field of trajectory data mining and representation learning technology, specifically relating to a pre-training and online update method for multi-view dynamic trajectory representation learning. Background Technology

[0002] With the widespread application of global location services and GPS (Global Positioning System) devices, an increasing amount of trajectory data is being collected. Trajectory data, as a crucial carrier for recording the spatiotemporal behavior of moving objects, has been widely applied in intelligent transportation, location services, trajectory similarity search, traffic pattern recognition, and travel time estimation. Trajectory Representation Learning (TRL) is a technique that transforms trajectory data into low-dimensional vectors, enabling these data to be effectively used by downstream tasks. However, traditional trajectory representation learning methods often suffer from several problems. For example, they can only model trajectory data from a single perspective, resulting in an inability to capture the multidimensional spatiotemporal features of the trajectory data. Furthermore, these methods are often task-specific, lacking cross-task versatility and flexibility. In addition, trajectory data is dynamic, with data patterns changing over time; therefore, existing methods often cannot update the model in a timely manner to adapt to these changes.

[0003] Through research and retrieval, we have summarized that currently popular trajectory representation learning methods are mainly divided into two categories: free space-based grid partitioning methods and road network topology-based graph embedding methods. The former captures point sequence patterns through recurrent neural networks, such as the TremBR (a trajectory embedding representation learning framework based on road networks) model, see the literature [Fu, Tao-Yang, Wang-Chien Lee. Trembr: Exploring road networks for trajectory representation learning. ACM Intelligent Systems & Technologies, Vol. 11, No. 1, 2020: pp. 1-25]. The latter uses graph attention networks to learn road segment embeddings, such as the START (Spatiotemporal Attention Road Network Transformer) algorithm, see the literature [Jiang, J, Pan, D, Ren, H, Jiang, X, Li, C, Wang, J. Self-supervised trajectory representation learning with temporal regularity and travel semantics. Proceedings of the 39th International Conference on Data Engineering, IEEE, 2023, pp. 843-855].

[0004] However, these two types of trajectory representation learning methods have the following key drawbacks: First, single-view modeling leads to incomplete feature capture. Existing methods are limited to a single spatial perspective (free space or road network view), making it difficult to collaboratively extract multi-dimensional spatiotemporal semantics of trajectories. While road network views can capture topological structures, they ignore regional commercial features. Although free space grid division reflects macroscopic movement trends, it lacks fine-grained features at the road segment level. This single-view modeling limits the accuracy of downstream tasks (such as semantically aware trajectory similarity calculation).

[0005] Secondly, task-specific training limits the model's versatility. Mainstream methods use fixed feature combinations for pre-training, making it difficult to dynamically adapt to the different spatiotemporal feature requirements of various tasks; existing best models only support 3 to 4 analysis tasks, and each task requires separate model training, resulting in a waste of computational resources.

[0006] Furthermore, the lack of dynamic update capabilities restricts practical applications. Urban trajectory data has a continuously evolving nature, and existing methods lack online update mechanisms, making it impossible to capture temporal changes in traffic flow patterns. Especially in scenarios involving sudden road conditions (such as traffic accidents), static models can exhibit timeliness biases, leading to a decrease in the accuracy of tasks such as trajectory prediction and time of arrival estimation.

[0007] More seriously, traditional methods employ simple location encoding and time segmentation strategies, failing to effectively integrate multi-source features such as road attributes (speed limits, number of lanes) and regional points of interest (POI) distribution. Furthermore, they lack interpretable design for attention weights, making it difficult to pinpoint the impact of key road segments on matching results.

[0008] In summary, designing a trajectory representation learning framework that supports multi-view feature fusion, dynamic parameter updates, and multi-task generalization has become a key challenge in improving the efficiency of urban mobile data mining. Summary of the Invention

[0009] In view of the above, this invention provides a pre-training and online update method for multi-view dynamic trajectory representation learning. Utilizing a novel trajectory representation learning framework, GTR (General, Multi-view, and Dynamic Trajectory Representation Learning Framework), it combines multi-view learning, spatiotemporal feature fusion, and dynamic model update strategies. This enables more accurate, general, and efficient representations for various downstream tasks of trajectory data. The framework is applicable to scenarios such as traffic flow prediction, mobile object behavior analysis, and travel pattern classification. It can transform raw trajectory data into low-dimensional semantic vectors, supporting collaborative optimization of various downstream analysis tasks.

[0010] A pre-training and online update method for multi-view dynamic trajectory representation learning includes the following steps: (1) Obtain GPS streaming data of the moving object. Each data point contains longitude, latitude and timestamp to form the original trajectory sequence; (2) The road network is modeled as a directed graph, and the POI semantic features of the free space grid are combined to generate a multi-view joint trajectory embedding representation by fusing information from the road network perspective and the free space perspective. (3) Based on the spatiotemporal hybrid expert model, road topology, regional semantics and multi-granular temporal features are dynamically fused, and a general trajectory embedding is generated through span mask pre-training and triplet comparison learning; (4) Based on the dynamic evolution of trajectory data, the model encoder is divided into a frozen layer and a hot update layer, and the parameter update ratio is dynamically adjusted; (5) Whenever new trajectory data triggers a model update, starting from the spatiotemporal fusion representation, the incremental trajectory representation vector is generated and a multi-task analysis report is generated by adaptively matching the downstream task requirements through gating weights.

[0011] Furthermore, the specific implementation of step (2) is as follows: S21: Based on graph neural network (GNN), embedding learning is performed on road segments. The graph convolutional network (GCN) of the Transformer model is used to capture the spatiotemporal dependencies in the road network. The static features (road type, speed limit, etc.) of the road segments are encoded based on graph attention network (GAT) to generate the embedding vector of each road segment. S22: The entire trajectory space is divided into multiple grid cells. Each point in the trajectory is mapped to the corresponding grid according to its geographical coordinates. The spatiotemporal dependencies of the trajectory are represented by location embedding and time embedding. S23: The four different features of road embedding, grid embedding, location embedding and time embedding are fused together and combined using a weighted sum to obtain the final trajectory embedding representation.

[0012] Furthermore, in step S21, the road network graph is input into a graph convolutional network. Each road segment in the road network graph is used as a node, and the connection relationship between road segments is used as an edge. Through the hierarchical propagation of the graph convolutional network, the spatiotemporal features of the road network graph can be learned. The information of different road segments is weighted through an attention mechanism, and the attention coefficient is calculated using a graph attention network to determine the strength of the relationship between different road segments, thereby generating an embedding vector for each road segment.

[0013] Furthermore, the specific implementation of step (3) is as follows: S31: The multi-head attention mechanism of the Transformer model is used to extract the global association between the road segment topology and regional semantics in the trajectory embedding representation, so as to obtain features in the spatial and temporal dimensions; S32: Model the dynamic temporal dependencies between trajectory points using gated cyclic units, and calculate the weights of spatial and temporal features; S33: Generate gating weights based on task IDs, dynamically adjust the fusion ratio of spatial and temporal features, and fuse spatial and temporal features through a weighted fusion strategy; S34: In the pre-training stage of spatiotemporal feature fusion, the contextual information of the trajectory is captured by the masking language model. Some trajectory information is randomly masked. The task of this model is to predict the masked part, thereby learning the spatiotemporal dependency of the trajectory. Then, triple loss is used to enhance the discriminativeness of the trajectory representation, so that the representations of similar trajectories are closer in the feature space, while the representations of different trajectories are further apart.

[0014] Furthermore, the specific implementation of step (4) is as follows: S41: The Transformer model in the multi-head attention mechanism is divided into frozen layers and hot-update layers. Frozen layers are layers in the model that have learned enough knowledge and are highly adaptable to new data. These layers remain unchanged during training. Hot-update layers are layers in the model that need to adapt to new data in a timely manner. These layers are updated according to the latest trajectory data, so that the model can continuously learn new trajectory patterns. S42: New trajectory data is filtered based on a sliding time window, and incremental training is performed only on the hot update layer; S43: By using comparative loss and regularization techniques to constrain the update process, we ensure the consistency between new features and historical patterns.

[0015] Furthermore, the specific implementation of step (5) is as follows: S51: For newly added trajectory data, activate the corresponding spatiotemporal hybrid expert branch based on the task ID to generate task-specific trajectory embeddings. S52: Lightweight fine-tuning modules are designed for different downstream tasks. For trajectory classification tasks, this module takes the trajectory representation learned in the pre-training stage as input and classifies the trajectory through a fully connected layer. For travel time estimation tasks, this module uses a regression model to predict the travel time of the trajectory. For trajectory generation tasks, this module uses a generative adversarial network to generate new trajectories. S53: Visualize the impact of key road segments on task results through attention weights and generate explanatory reports.

[0016] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described pre-training and online update method for learning dynamic trajectory representations for multiple views.

[0017] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described pre-training and online update method for learning dynamic trajectory representations for multiple views.

[0018] The inventiveness of the GTR framework design in this invention is mainly reflected in the following aspects: First, traditional trajectory learning methods typically focus on a single perspective, such as free space or road network, while neglecting the multidimensional information in trajectory data. Trajectory data includes not only spatial coordinates but also temporal information, road network information, and regional semantic information. Therefore, integrating information from multiple perspectives becomes a significant challenge.

[0019] Secondly, many existing trajectory representation learning methods are designed for specific tasks and lack generality. For example, some methods may only be used for trajectory similarity calculation or trajectory simplification, which limits the applicability and generalization ability of the model in other tasks. Therefore, a key innovation of the GTR framework is that it proposes a general representation learning framework that supports multiple trajectory analysis tasks, avoiding the high cost and complexity of training models separately for each task.

[0020] Finally, trajectory data is dynamic, especially in complex environments such as cities, where trajectory data is constantly changing, and the movement patterns of trajectories also change with traffic conditions. Therefore, how to efficiently update the trajectory representation learning model to maintain its adaptability to the latest data is an important issue. To address this, the GTR framework proposes an online freeze-hot update strategy, which can maintain the stability of the model and update its representation capabilities in a timely manner while new data is continuously flowing in.

[0021] To address this, GTR employs a multi-view encoder to integrate free space and road network perspectives, ensuring the capture of sufficient spatiotemporal features. It also introduces a spatiotemporal fusion pre-training method, using an "expert hybrid" model to dynamically adjust the fusion of spatial and temporal features across different tasks, thereby supporting a variety of downstream tasks. Furthermore, GTR proposes an online freeze-hot update strategy, which allows for continuous model updates based on the influx of new data without affecting the performance of existing tasks.

[0022] The core contribution of the GTR framework in this invention lies in its ability to overcome the limitations of existing trajectory representation learning methods through multi-view encoding, multi-task adaptation, and dynamic update strategies, providing a general and flexible trajectory representation learning framework. GTR not only achieves excellent performance in multiple trajectory analysis tasks but also ensures that the model can adapt in real time and maintain high performance when facing constantly changing data through an innovative online update strategy. These innovations make GTR a powerful tool that can be widely applied in various fields, including traffic management, urban planning, and intelligent transportation. The GTR framework, through its innovative multi-view learning, spatiotemporal feature fusion, and dynamic update strategies, solves some key problems in trajectory representation learning, providing a more general, flexible, and efficient solution. In the future, GTR can be further extended to combine with other types of machine learning methods, such as large-scale pre-trained models, to improve its performance in more complex scenarios.

[0023] Extensive experiments on two real-world trajectory datasets (trajectory data from Porto, Portugal and Beijing, China) demonstrate that the proposed GTR outperforms existing state-of-the-art methods in multiple tasks, including trajectory similarity calculation, trajectory classification, trajectory generation, and travel time estimation. Particularly in travel time estimation, GTR not only improves performance but also demonstrates its advantage in model updates. Through this dynamic update mechanism, GTR can adapt to changes in trajectory data in real time, ensuring high analytical capabilities and accuracy even in dynamic environments. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the pre-training and online update method for multi-view dynamic trajectory representation learning according to the present invention. Detailed Implementation

[0025] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] This implementation provides a novel trajectory representation learning framework, GTR, whose design and implementation mainly consist of the following four parts: 1. Design and implementation of Multi-View Encoder (MVE): MVE extracts rich spatiotemporal features by combining free space perspective and road network perspective, thereby capturing multi-dimensional information in trajectory data. This includes the implementation of technologies such as road embedding, grid embedding, location embedding and time embedding, and how each embedding technology will help build a more accurate trajectory representation.

[0027] 2. Implementation of the Spatiotemporal Feature Fusion Pre-training (STP) Mechanism: STP includes dynamically fusing spatiotemporal features to adapt to different downstream tasks. It adaptively fuses spatial and temporal features through expert mixing, further improving the flexibility and accuracy of the model when dealing with complex trajectory tasks.

[0028] 3. Implementation of Online Freeze-Hot Update (OFU) Strategy: The OFU strategy enables online updates of the model, maintaining its adaptability to new data without losing knowledge of historical data. In dynamic data environments, the OFU strategy allows GTR to perform incremental updates efficiently without affecting the performance of existing models.

[0029] 4. Model Training Process and Optimization: This includes the specific implementation details of the pre-training and fine-tuning stages, using appropriate loss functions to optimize the model, fine-tuning for different tasks to obtain the best results, and optimizing model performance through training data augmentation, network design, and other methods.

[0030] This implementation method, based on the GTR framework, presents a pre-training and online update method for multi-view dynamic trajectory representation learning. The specific implementation process is as follows: Figure 1 As shown: Step S11: Obtain GPS streaming data of the moving object. Each data point contains longitude, latitude, and timestamp, forming the original trajectory sequence.

[0031] Step S12: Model the road network as a directed graph (nodes are road segments, edges are connections), and construct a multi-view joint representation by combining the POI semantic features of the free space grid.

[0032] The complexity of trajectory data requires us to model from multiple perspectives. Traditional methods usually rely on a single perspective, such as the free space perspective or the road network perspective. The free space perspective usually divides the space into different regions by grid partitioning, while the road network perspective focuses on the topology and connectivity of roads.

[0033] In GTR, MVE can simultaneously process free space and road network perspectives, generating an embedded representation of the trajectory by fusing information from these two perspectives. Specifically, MVE maps trajectory data onto free space and road networks respectively, and uses graph convolutional networks to capture the spatiotemporal dependencies in the road network, while capturing regional features through grid partitioning. MVE can effectively combine these two types of information to generate richer and more accurate trajectory representations.

[0034] In GTR, MVE is one of the core modules. Its design aims to enhance the model's ability to represent trajectories by extracting complex spatiotemporal features from trajectory data from multiple perspectives. Trajectory data itself has multidimensional characteristics, including not only spatial coordinates but also temporal information, road network structure, and environment-related semantic features. Traditional trajectory representation learning methods typically model from only a single perspective—either free space or road network—making it difficult to capture the multidimensional information within the trajectory data. GTR, through its multi-view encoder design, effectively fuses information from different perspectives, enabling the model to more comprehensively understand the trajectory data and thus obtain a more accurate trajectory representation.

[0035] The design philosophy of the multi-view encoder is to encode the trajectory from multiple perspectives in order to capture the spatiotemporal characteristics of the trajectory. Specifically, GTR represents trajectory data from two perspectives: a road network perspective and a free space perspective, and extracts information from each perspective through different encoding methods. In this way, GTR can integrate information from different sources within the same framework, thereby effectively improving the accuracy and robustness of trajectory representation.

[0036] First, in the design of the multi-view encoder, the trajectory input data typically comes from two main sources: road network information and free space information. For road network information, GTR utilizes graph neural networks to learn the embedding of road segments, capturing the spatiotemporal dependencies within the road network through graph convolutional networks. By embedding road segments, GTR can effectively leverage the topological structure of the road network to construct a trajectory representation based on the road network. For the free space perspective, GTR employs a grid partitioning method, dividing the entire space into several grids, each representing a specific region, with trajectory points mapped to the corresponding grid. In this way, the regional information of free space helps GTR better capture the spatial features of the trajectory, thereby improving the model's trajectory representation capability.

[0037] In GTR's multi-view encoder, the road network perspective is encoded through map matching of the trajectory. Specifically, a map matching algorithm is first used to align GPS trajectory points with the actual road network, ensuring the connectivity of the trajectory points within the network. The key to this process is mapping each GPS point in the trajectory to a specific road segment within the road network. During map matching, GTR employs modern graph neural network methods, utilizing graph convolutional networks to learn features from road segments.

[0038] Graph Convolutional Networks (GCNs) are neural network models designed for graph-structured data. They capture relationships between nodes by propagating information across the graph. In GTR, the input to the GCN is a road network graph, where each road segment (or feature of a road segment) is a node, and the connections between road segments are edges. Through hierarchical propagation of the GCN, GTR effectively learns the spatiotemporal features of the road network and weights the information of different road segments through an attention mechanism to generate more accurate road segment embeddings. Specifically, GTR uses a graph attention network to calculate attention coefficients to determine the strength of relationships between different road segments, thereby generating embedding vectors for each road segment. These embedding vectors not only contain the geographical information of the road segment but also capture the topological information of the road network.

[0039] The core advantage of graph convolutional networks lies in their ability to effectively process spatiotemporal data with graph structures, making them particularly suitable for trajectory data learning. Through GAT, GTR can capture complex spatial dependencies within the context of road networks, thus providing more fine-grained road information for trajectory representation. This method not only considers the geographical location of trajectory points but also captures the dynamic changes and spatiotemporal dependencies between road segments.

[0040] Unlike the road network perspective, the encoding of the free space perspective is based on spatial grid partitioning. In GTR, the entire trajectory space is divided into multiple grids, each grid representing a spatial region. Each point in the trajectory is mapped to the corresponding grid according to its geographic coordinates. In this way, GTR can generate a grid coordinate for each trajectory point, thereby effectively capturing the spatial features of the trajectory.

[0041] In the free-space perspective encoding, GTR employs positional embedding and temporal embedding to represent the spatiotemporal dependencies of the trajectory. Positional embedding uses sine and cosine functions to generate positional codes, capturing the sequential information of trajectory points, while temporal embedding captures the temporal features of the trajectory using different time scales (such as minutes, weeks, and years). These embedding vectors, along with grid embedding, are input into the multi-view encoder for further processing.

[0042] A key innovation of GTR's multi-view encoder lies in how it fuses information from both the road network view and the free space view. In the multi-view encoder, GTR generates a unified trajectory representation by fusing the embedding vectors from the two views. Specifically, GTR fuses four different features: road embedding, grid embedding, location embedding, and temporal embedding, combining these information using a weighted sum to obtain the final trajectory representation.

[0043] The advantage of this fusion method is that it can simultaneously consider the spatial and temporal information of the trajectory, capturing both the local spatial dependence of the trajectory and the overall spatiotemporal trend of the trajectory. This multi-view fusion design enables GTR to achieve good performance in multiple downstream tasks, such as trajectory similarity calculation, travel time estimation, and trajectory classification.

[0044] GTR's multi-view encoder not only encodes trajectory information from multiple perspectives but also further improves the accuracy of trajectory representation through optimization strategies. During model training, GTR employs a self-supervised learning approach, continuously optimizing trajectory representations through pre-training and fine-tuning strategies. Specifically, GTR uses a transformer-based self-supervised task, leveraging a masked language model and triplet loss to optimize the quality of trajectory representations.

[0045] During the pre-training phase, GTR masks a portion of the trajectory information, training the model to predict the masked portion and thus learning the trajectory's contextual information. Simultaneously, triplet loss is used to optimize the trajectory similarity representation, bringing similar trajectories closer together in the embedding space and distancing different trajectories further apart. In this way, GTR can further enhance the representational capabilities of the multi-view encoder through self-supervised learning, resulting in a more accurate and robust trajectory representation across multiple downstream tasks.

[0046] The multi-view encoder is a key module in the GTR framework. By encoding trajectory data from multiple perspectives, it effectively captures the spatiotemporal features of trajectories. Through the fusion of road network and free-space perspectives, GTR generates trajectory representations rich in spatiotemporal information, giving it a significant advantage in handling various trajectory analysis tasks. GTR demonstrates innovation and efficiency in both spatial trajectory encoding and spatiotemporal feature fusion, enabling its widespread application in tasks such as trajectory similarity calculation, travel time estimation, and trajectory classification.

[0047] Step S13: Based on the spatiotemporal hybrid expert model (ST-MoE), road topology, regional semantics and multi-granular temporal features are dynamically fused, and a general trajectory embedding is generated through span mask pre-training and triple contrastive learning.

[0048] To enhance the multi-task adaptability of trajectory representation, GTR introduces a spatiotemporal fusion pre-training method. Traditional pre-training methods often employ a uniform strategy to train the model, but different downstream tasks may have different requirements for spatiotemporal features, thus necessitating a more flexible feature fusion mechanism. In GTR, we designed a spatiotemporal expert hybrid module, which includes two experts: a spatial expert and a temporal expert. By dynamically adjusting the weights of spatial and temporal features, ST-MoE can automatically select the most suitable feature combination according to the needs of different tasks. In this way, GTR can demonstrate good adaptability and generalization ability across multiple downstream tasks.

[0049] In GTR, the spatiotemporal feature fusion pre-training mechanism is a key component. It dynamically fuses the spatiotemporal features of the trajectory to provide a more comprehensive and accurate trajectory representation for subsequent downstream tasks. The core objective of STP is to enable the model to adaptively learn multi-level features of the trajectory from multiple spatiotemporal dimensions, thereby improving the generalization ability of the trajectory representation, especially across different task types. To achieve this goal, STP introduces a spatiotemporal expert hybrid module, which dynamically fuses spatial and temporal features using a hybrid expert mechanism, thus improving performance in multi-task learning.

[0050] GTR's STP mechanism not only enhances the model's expressive power through the synthesis of traditional temporal and spatial features, but also enables the model to balance and switch between different tasks through pre-training and fine-tuning. This mechanism not only effectively solves the limitations of traditional methods in spatiotemporal feature fusion, but also allows GTR to exhibit stronger adaptive capabilities when processing complex and dynamic trajectory data.

[0051] Trajectory data is essentially sequential data containing spatiotemporal information. Each trajectory point consists of spatial coordinates (such as longitude and latitude) and a timestamp (time information). In traditional trajectory learning methods, the fusion of spatiotemporal features is usually static, meaning that spatial and temporal features are fixedly combined, and these methods typically can only handle a single type of task. Specifically, traditional trajectory representation learning methods often process spatial and temporal features in a simplified way, making it difficult to extract sufficient features from complex spatiotemporal relationships, thus affecting the model's performance.

[0052] The STP mechanism proposed by GTR addresses this problem by introducing a spatiotemporal expert hybrid module, which enables dynamic fusion of spatiotemporal features. This means that the model can flexibly adjust the weights of spatial and temporal features under the requirements of different tasks to optimize trajectory representation and thus improve the model's performance in various trajectory analysis tasks.

[0053] In the STP mechanism, the ST-MoE module is one of the core innovations, which mainly solves the problem of how to dynamically adjust the fusion of spatial and temporal features. The ST-MoE module divides the spatiotemporal features of trajectory data into two sub-tasks: spatial features and temporal features. For each sub-task, an independent expert network (i.e., a spatial expert and a temporal expert) is designed. Each expert network processes the corresponding features and learns the spatiotemporal features through a specialized neural network structure.

[0054] Spatial experts primarily focus on the extraction and encoding of spatial features, including the spatial location of the trajectory, its relative spatial location, and gridded spatial information. Temporal experts, on the other hand, focus on the temporal features of the trajectory, particularly its changing trends across different time scales, such as minutes, hours, weeks, and months. Through the design of these two expert networks, GTR can effectively capture the interdependencies between spatial and temporal features.

[0055] In practice, the ST-MoE module first feeds spatial and temporal features into their respective expert networks for processing. Spatial and temporal experts extract important features from the spatial and temporal dimensions using a multi-head self-attention mechanism. These features are then combined using a weighted fusion strategy. Specifically, the ST-MoE module assigns a weight to each task, flexibly adjusting the fusion ratio of spatial and temporal features according to the different needs of the task.

[0056] The STP mechanism in GTR integrates spatial and temporal features adaptively. This adaptation relies primarily on learning task weights, specifically using task IDs to determine the integration ratio of spatial and temporal features. During training, the model calculates the weights of spatial and temporal features through a gating network based on different task requirements. This process adjusts the importance of spatial and temporal features in a specific task based on its characteristics.

[0057] In the ST-MoE module, the gated network encodes the task ID to obtain a vector representing the task features. This task feature vector is fed into a fully connected layer, and then the Softmax function is used to calculate the weights of spatial and temporal features. These weights determine the proportion of spatial and temporal features in the final fusion, thereby achieving adaptive adjustment for different tasks.

[0058] In this way, the STP mechanism can effectively fuse spatiotemporal features without losing task characteristics. This allows GTR to adapt to various complex trajectory analysis tasks and achieve good performance in these tasks. For example, in trajectory classification tasks, more spatial features may be needed to capture the spatial patterns of trajectories; while in travel time estimation tasks, stronger temporal features may be needed to capture the temporal dependencies of trajectories. In these cases, the STP mechanism can help GTR achieve optimal performance in each task by dynamically adjusting the feature fusion ratio.

[0059] The STP mechanism is designed not only to handle the fusion of spatiotemporal features, but also to be closely integrated with the GTR's pre-training and fine-tuning process. During the pre-training phase, GTR trains the spatiotemporal feature encoder through self-supervised learning on trajectory data. The core task of the pre-training process is to capture the contextual information of the trajectory by masking the language model. In this stage, some trajectory information is randomly masked, and the model's task is to predict the masked portion, thereby learning the spatiotemporal dependencies of the trajectory.

[0060] In addition to masking the language model, STP introduces triplet loss to enhance the discriminativeness of trajectory representations. In this process, STP learns to make representations of similar trajectories closer together in the feature space, while representations of different trajectories are further apart. Triplet loss helps GTR generate more discriminative trajectory embeddings, resulting in better performance in downstream tasks.

[0061] During the fine-tuning phase, the STP mechanism further adjusts the spatiotemporal feature fusion method according to different downstream tasks. Each downstream task may focus on different aspects of trajectory data (such as trajectory similarity, travel time, trajectory classification, etc.), thus requiring different feature combinations. Through task-specific fine-tuning, GTR can further optimize the spatiotemporal feature fusion method to improve performance on specific tasks.

[0062] GTR's STP mechanism employs a self-supervised learning approach, enabling the model to improve its representational capabilities by learning the inherent patterns of trajectories without human labels. Through self-supervised tasks such as masked language modeling and triple loss, GTR can automatically learn meaningful spatiotemporal features from trajectory data and generate general trajectory representations using these features. This self-supervised learning method not only reduces reliance on labeled data but also allows the model to exhibit stronger adaptability when handling different types of trajectory tasks.

[0063] Task adaptation is another important feature of the STP mechanism. GTR achieves task adaptation through fine-tuning for different tasks. In the pre-training phase, GTR learns a wide range of features from trajectory data to obtain a general trajectory representation. In the fine-tuning phase, the model adjusts the fusion method of spatiotemporal features according to the specific requirements of the task, thereby optimizing the task performance. This task adaptation approach enables GTR to achieve excellent results in multiple different trajectory analysis tasks, such as trajectory similarity search, trajectory classification, and travel time estimation.

[0064] The Spatiotemporal Expert Hybrid (STP) mechanism is one of the key innovations in the GTR framework. By introducing a spatiotemporal expert hybrid module, it achieves dynamic fusion of spatiotemporal features and task adaptation. This mechanism not only overcomes the limitations of traditional trajectory learning methods in spatiotemporal feature fusion but also enables GTR to perform exceptionally well across multiple downstream tasks. Through pre-training and fine-tuning, GTR can generate general trajectory representations and further optimize their accuracy based on task adaptation. Through self-supervised learning and task adaptation, the STP mechanism allows GTR to exhibit powerful flexibility and accuracy when handling dynamic and complex trajectory data.

[0065] Step S14: Based on the dynamic evolution of trajectory data, the model encoder is divided into a frozen layer (preserving historical patterns) and a hot update layer (learning new features), and the parameter update ratio is dynamically adjusted.

[0066] The dynamic nature of trajectory data requires models to be updated as the data changes. Traditional trajectory representation learning methods often cannot effectively cope with real-time data changes, leading to unstable performance in dynamic environments. The online freeze-hot update strategy proposed by the GTR framework can update the model using new data while maintaining the stability of the original model. Specifically, GTR divides the model into a freeze part and a hot update part. During each update, only the hot update part is trained, thereby achieving rapid adaptation to new data. This method not only reduces training costs but also ensures that the model can continuously optimize in dynamic environments.

[0067] In GTR, the online freeze-hot update strategy is a significant innovation, designed to address the inherent dynamism of trajectory data, particularly in real-time data streams, and how to maintain model update capability without impacting existing model performance. Trajectory data is typically continuously generated, especially in complex urban traffic systems where GPS trajectories change rapidly. Traditional model training methods often struggle to handle this rapidly changing data, making online updates crucial.

[0068] The OFU strategy, by flexibly freezing and updating different layers of the model, can efficiently utilize newly arrived data for online model optimization without causing computational inefficiency or model instability due to frequent updates. The OFU design not only maintains the importance of historical data but also adapts to new data patterns in real time, providing GTR with continuous learning capabilities in dynamic environments. To better understand the design and implementation of the OFU strategy, we will explore its core mechanisms in detail, including the principles of online updates, the specific implementation of freeze-hot updates, the efficiency of model updates, and how to maintain model robustness in constantly changing data environments.

[0069] The dynamic nature of trajectory data is a significant challenge in the field of trajectory representation learning, particularly in areas like urban traffic where trajectory data exhibits strong temporal variations. As traffic conditions, weather, and other factors change, the trajectory generation patterns also shift. This makes traditional trajectory learning models perform poorly when handling such dynamic data, as they typically rely on static training datasets and struggle to adapt to newly generated data once training is complete.

[0070] GTR's OFU strategy addresses this problem by providing a dynamic online learning method. This strategy allows GTR to dynamically update the model's representation as data arrives, without retraining the entire model. This approach adapts to new trajectory data in real time while preserving the influence of historical data, avoiding the high computational overhead of retraining, and ensuring the model's stability and robustness.

[0071] One of the key challenges of the OFU strategy is how to efficiently update the model without affecting the existing task performance. Specifically, how to selectively freeze and update some layers of the model so that the update process can make the most of the effective information of the new data and gradually enhance the model's adaptability to new data without destroying the original model's capabilities.

[0072] The core idea of ​​the OFU strategy is freeze-hot update, which involves "freezing" certain parts of the model to maintain its learning ability on historical data, while "hot updating" other parts to adapt to new data streams in a timely manner. This method effectively reduces the amount of computation in the online update process and enables efficient learning in dynamic environments.

[0073] In the implementation of GTR, the OFU strategy divides the different layers of the Transformer model into two parts: a frozen part and a hot-update part. The frozen part consists of layers that have learned enough knowledge and are highly adaptable to new data; these layers remain unchanged during training to preserve the influence of historical data. The hot-update part consists of layers that need to adapt to new data promptly; these layers are updated based on the latest trajectory data, allowing the model to continuously learn new trajectory patterns.

[0074] The design of the freeze-hot update mechanism must ensure the following: ① Computational efficiency: By freezing some layers, the entire model is avoided from being retrained, thus reducing computational complexity.

[0075] ② Preservation of historical knowledge: Freezing the parts can preserve historical data information, avoiding the loss of existing knowledge due to frequent updates.

[0076] ③ Adaptation to new data: The hot update part can make timely adjustments to new data, thereby ensuring the model's ability to learn new trajectory patterns.

[0077] The execution process of the OFU strategy includes several key steps: data inflow, layer selection and freezing, and updating and optimization. These steps ensure that the model can flexibly adapt to the constantly changing data environment and maintain efficient computation and stability.

[0078] First, when new trajectory data comes in, GTR cleans and transforms the data through a preprocessing module before inputting it into the model. During this process, GTR selectively freezes or updates certain model layers based on the current task requirements and data characteristics.

[0079] Next, GTR dynamically adjusts different layers in the Transformer model according to the freeze-hot update mechanism. For example, for layers with relatively stable spatiotemporal features (such as low-level graph convolutional network layers), these layers can be frozen to preserve their representational ability in historical data; while for high-level layers with large spatiotemporal changes (such as self-attention layers), hot updates are performed to adapt to the spatiotemporal relationships of new data in a timely manner.

[0080] During the update process, GTR optimizes the model's performance through incremental training. This training method differs from traditional batch training. It allows the model to be locally adjusted each time new data arrives, without having to retrain the entire model. This incremental optimization method can effectively avoid wasting computing resources and ensure that the model always stays up-to-date with real-time data streams.

[0081] In addition, GTR employs an early stopping strategy to avoid overfitting. When the model is updated, if it is found that the addition of new data has little effect on the model's performance, GTR will automatically stop the update process, thereby avoiding unnecessary computational overhead.

[0082] The flexibility of the freeze-hot update strategy lies in its ability to dynamically adjust the frozen and updated parts based on the model's training progress and data characteristics. To achieve this, GTR introduces a dynamic adjustment mechanism that automatically selects the updated layers based on changes in the data flow and the model's training state. For example, in the initial stages, the model may have poor adaptability to new data, so the hot update portion will be relatively larger to quickly adapt to new trajectory patterns. In the later stages, as the model's representational power gradually increases, the proportion of the frozen portion will gradually increase, ensuring that the model does not over-adjust existing knowledge and avoid overfitting.

[0083] Furthermore, GTR optimizes the freeze-hot update strategy by introducing task relevance assessment. In different downstream tasks, the model's requirements for features at different levels vary. Therefore, GTR flexibly adjusts which layers should be frozen and which should be hot-updated based on the specific requirements of the task. In this way, GTR can maintain efficient learning capabilities in multi-task learning while avoiding mutual interference between tasks.

[0084] One of the biggest advantages of the OFU strategy is its efficiency, especially in dynamic data environments. Compared to traditional batch training methods, OFU can significantly reduce computational overhead while ensuring model updates and optimization. In GTR, the model training and update process does not rely on retraining on all historical data, but rather improves the model's accuracy incrementally through incremental updates.

[0085] To ensure the stability of model updates, GTR is optimized in the following ways: First, by freezing some layers, overfitting caused by frequent updates is avoided; second, GTR employs regularization techniques, such as L2 regularization and Dropout, to prevent overfitting during hot updates; finally, GTR introduces validation set evaluation during the update process to ensure that the model's performance improves after each update, thereby maintaining the stability of the update process.

[0086] The OFU strategy provides an efficient and flexible online learning method for model training through a dynamic mechanism of freezing and hot updating. In GTR, this strategy helps the model adapt to new data patterns in real time when faced with constantly changing trajectory data, while maintaining the stability of historical data. The freeze-hot update mechanism not only improves the computational efficiency of the model and reduces training overhead, but also ensures the balance between old and new data through adaptive adjustment, enabling GTR to exhibit strong adaptability and stability in dynamic environments. Through this strategy, GTR can maintain high performance and robustness in complex trajectory analysis tasks, meeting the needs of multi-task and multi-data stream operations.

[0087] Step S15: Whenever new trajectory data triggers a model update, starting from the spatiotemporal fusion representation, the incremental trajectory representation vector is generated and a multi-task analysis report is generated by adaptively matching the downstream task requirements through gating weights.

[0088] In GTR, model training and fine-tuning are crucial processes to ensure that the model can effectively learn trajectory data features and achieve excellent performance across multiple downstream tasks. During training and fine-tuning, GTR not only improves the quality of trajectory representation through its powerful spatiotemporal feature fusion and multi-view encoding capabilities, but also further enhances the model's performance in practical applications through self-supervised learning and task-specific fine-tuning strategies.

[0089] The training process of GTR is divided into two main stages: the pre-training stage and the fine-tuning stage. The goal of the pre-training stage is to learn a general trajectory representation, enabling the model to capture key features in trajectory data. The goal of the fine-tuning stage is to adjust the model according to specific downstream tasks (such as trajectory classification, travel time estimation, trajectory generation, etc.) so that it can perform better on these tasks.

[0090] First, GTR extracts spatiotemporal features from trajectory data from two main perspectives—the road network perspective and the free space perspective—using a multi-view encoder. Building on this, GTR employs a spatiotemporal feature fusion pre-training mechanism, using a self-supervised learning method to pre-train the model and enhance its understanding of these features. Specifically, GTR uses a masked language model to train the model to capture the contextual information of the trajectory and employs triplet loss to enhance the discriminativeness of the trajectory representation.

[0091] During the pre-training phase, GTR employs a self-supervised learning approach, aiming to learn trajectory representations without relying on labeled data. Self-supervised learning tasks typically include masked language models and triplet loss. Masked language models are a common self-supervised learning method that randomly masks a portion of the input data (such as certain points in the trajectory), training the model to predict this masked information to capture the contextual dependencies of the data. In this way, GTR is able to learn rich spatiotemporal features from trajectory data.

[0092] In GTR, the task of the masking language model is to mask various spatiotemporal points of a trajectory. The model needs to predict the masked points. The masking task enables the model to capture the overall structure of the trajectory and learn the dependencies between trajectory points. To enhance the model's representational power, GTR further employs triplet loss. The goal of triplet loss is to make the representations of similar trajectories closer in the embedding space, while keeping different trajectories at greater distances. Through triplet loss, GTR can improve the discriminativeness of trajectory representations and provide better performance in downstream tasks.

[0093] With the combined effect of masked language model and triple loss, GTR is able to learn effective trajectory representations. This pre-training stage lays a solid foundation for subsequent fine-tuning, enabling the model to demonstrate good performance in different trajectory analysis tasks.

[0094] After pre-training, GTR enters the fine-tuning phase. In this phase, the model is further optimized for specific downstream tasks (such as trajectory classification, travel time estimation, and trajectory generation). The goal of fine-tuning is to enable the model to achieve optimal performance on these tasks, rather than simply capturing the spatiotemporal features of trajectories. To achieve this, GTR performs specialized optimizations for different tasks by adjusting hyperparameters such as some layers and the learning rate in the network.

[0095] In trajectory classification tasks, GTR uses trajectory representations learned during the pre-training phase as input and classifies trajectories through fully connected layers. The fine-tuning phase aims to adjust network parameters so that the model can accurately predict trajectory categories based on the trajectory representations. To improve classification accuracy, GTR employs a cross-entropy loss function and uses optimization algorithms to train the model. During fine-tuning, GTR achieves better performance on trajectory classification tasks by adjusting the loss function and optimization methods.

[0096] For the travel time estimation task, GTR uses a regression model to predict the travel time of a trajectory. In this task, GTR models the spatiotemporal characteristics of the trajectory, learns its temporal properties, and predicts the travel time from the origin to the destination based on these properties. For the regression task, GTR uses the mean squared error loss function to calculate the difference between the predicted and actual values, thereby adjusting the model parameters to optimize the prediction performance.

[0097] In the trajectory generation task, GTR fine-tuning employs generative models such as generative adversarial networks for training. The trajectory generation task requires the model to generate new trajectories based on input trajectory information. GTR further optimizes the generative network by fine-tuning the pre-trained model, enabling it to generate trajectory sequences similar to real trajectories. During this process, GTR uses generative adversarial loss to train the generator and discriminator, improving the quality and diversity of generated trajectories.

[0098] GTR's fine-tuning strategy is tailored to the specific task, employing different training strategies for different tasks. This task-adaptive approach allows GTR to achieve good performance across multiple tasks. For example, some tasks may require a greater focus on spatial features, while others may need a greater focus on temporal features. To address these differences, GTR dynamically adjusts the feature fusion method to suit the needs of different tasks.

[0099] Specifically, GTR introduces a task-adaptive mechanism, enabling it to dynamically adjust the network structure based on the characteristics of different tasks. For example, when performing trajectory classification, GTR may increase its focus on spatial features, improving classification performance by strengthening the weights of spatial features; while when performing travel time estimation tasks, GTR will pay more attention to learning temporal features, improving prediction accuracy by adjusting the ratio of spatiotemporal feature fusion.

[0100] Furthermore, GTR combines incremental learning and transfer learning methods to further improve the model's generalization ability. In incremental learning, GTR can continuously incorporate new data for training, keeping the model up-to-date. In transfer learning, GTR borrows pre-trained model parameters and performs rapid fine-tuning on new tasks, thereby improving the model's performance on those new tasks.

[0101] To prevent overfitting, GTR incorporates various regularization methods during fine-tuning, with L2 regularization and Dropout being commonly used. L2 regularization adds a squared penalty term to the loss function to help the model maintain parameter balance and avoid overfitting caused by excessively large parameters. Dropout, on the other hand, randomly discards some connections in the neural network during training to prevent the network from becoming overly reliant on certain neurons, thereby enhancing the model's generalization ability.

[0102] Furthermore, GTR employs learning rate scheduling and early stopping strategies to further optimize the fine-tuning process. In learning rate scheduling, GTR dynamically adjusts the learning rate based on changes in loss during training to avoid excessively rapid convergence and thus maintain training stability. The early stopping strategy stops training prematurely when there is no significant improvement in performance on the validation set, thus preventing overtraining of the model.

[0103] During fine-tuning, GTR ensures optimal model performance on each task by independently evaluating and optimizing each task. When evaluating task performance, GTR uses a range of common metrics to measure model performance. For example, in trajectory classification, accuracy and F1 score are used to assess classification performance; in travel time estimation, mean squared error and mean absolute error are used to assess prediction accuracy.

[0104] Based on the task evaluation results, GTR will adjust the model's training strategy, including adjusting the loss function, optimization algorithm, and learning rate during the fine-tuning phase, in order to further improve task performance. Through this task-driven fine-tuning strategy, GTR can demonstrate strong adaptability and efficiency in different trajectory analysis tasks.

[0105] GTR's model training and fine-tuning methods, through self-supervised learning and task adaptation mechanisms, enable the model to perform exceptionally well across various trajectory analysis tasks. During the pre-training phase, GTR learns a general trajectory representation through a multi-view encoder and spatiotemporal feature fusion mechanism. In the fine-tuning phase, GTR flexibly adjusts the model's structure and parameters according to the specific needs of the task to achieve optimal performance. By combining regularization methods, incremental learning, and transfer learning, GTR ensures the model's stability and robustness in dynamic and complex data environments.

[0106] To verify the effectiveness of the GTR framework of this invention, we conducted extensive experiments on two real-world trajectory datasets: the Porto, Portugal dataset and the Beijing, China dataset. Experimental results show that GTR outperforms existing state-of-the-art methods in multiple trajectory analysis tasks, including trajectory similarity calculation, trajectory classification, and travel time estimation. In the trajectory similarity calculation task, GTR achieved a low mean rank and a high hit rate on both the Porto and Beijing datasets. In the travel time estimation task, GTR significantly reduced errors compared to other methods, demonstrating its powerful ability to capture spatiotemporal features. Furthermore, GTR's online update strategy significantly improves the model's real-time adaptability, enabling continuous optimization in dynamic data environments.

[0107] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A pre-training and online update method for multi-view dynamic trajectory representation learning, characterized in that, Includes the following steps: (1) Obtain GPS streaming data of the moving object. Each data point contains longitude, latitude and timestamp to form the original trajectory sequence; (2) The road network is modeled as a directed graph, and the POI semantic features of the free space grid are combined to generate a multi-view joint trajectory embedding representation by fusing information from the road network perspective and the free space perspective. (3) Based on the spatiotemporal hybrid expert model, road topology, regional semantics and multi-granular temporal features are dynamically fused, and a general trajectory embedding is generated through span mask pre-training and triplet comparison learning; (4) Based on the dynamic evolution of trajectory data, the model encoder is divided into a frozen layer and a hot update layer, and the parameter update ratio is dynamically adjusted; (5) Whenever new trajectory data triggers a model update, starting from the spatiotemporal fusion representation, the incremental trajectory representation vector is generated and a multi-task analysis report is generated by adaptively matching the downstream task requirements through gating weights.

2. The pre-training and online update method for multi-view dynamic trajectory representation learning according to claim 1, characterized in that, The specific implementation method of step (2) is as follows: S21: Embedding learning of road segments is performed based on graph neural networks. The graph convolutional network of the Transformer model is used to capture the spatiotemporal dependencies in the road network. The static features of the road segments are encoded based on the graph attention network to generate the embedding vector of each road segment. S22: The entire trajectory space is divided into multiple grid cells. Each point in the trajectory is mapped to the corresponding grid according to its geographical coordinates. The spatiotemporal dependencies of the trajectory are represented by location embedding and time embedding. S23: The four different features of road embedding, grid embedding, location embedding and time embedding are fused together and combined using a weighted sum to obtain the final trajectory embedding representation.

3. The pre-training and online update method for multi-view dynamic trajectory representation learning according to claim 2, characterized in that: In step S21, the road network graph is input into a graph convolutional network. Each road segment in the road network graph is a node, and the connection relationship between road segments is an edge. Through the hierarchical propagation of the graph convolutional network, the spatiotemporal features of the road network graph can be learned. The information of different road segments is weighted through an attention mechanism, and the attention coefficient is calculated using a graph attention network to determine the strength of the relationship between different road segments, thereby generating an embedding vector for each road segment.

4. The pre-training and online update method for multi-view dynamic trajectory representation learning according to claim 1, characterized in that, The specific implementation method of step (3) is as follows: S31: The multi-head attention mechanism of the Transformer model is used to extract the global association between the road segment topology and regional semantics in the trajectory embedding representation, so as to obtain features in the spatial and temporal dimensions; S32: Model the dynamic temporal dependencies between trajectory points using gated cyclic units, and calculate the weights of spatial and temporal features; S33: Generate gating weights based on task IDs, dynamically adjust the fusion ratio of spatial and temporal features, and fuse spatial and temporal features through a weighted fusion strategy; S34: In the pre-training stage of spatiotemporal feature fusion, the contextual information of the trajectory is captured by the masking language model. Some trajectory information is randomly masked. The task of this model is to predict the masked part, thereby learning the spatiotemporal dependency of the trajectory. Then, triple loss is used to enhance the discriminativeness of the trajectory representation, so that the representations of similar trajectories are closer in the feature space, while the representations of different trajectories are further apart.

5. The pre-training and online update method for multi-view dynamic trajectory representation learning according to claim 1, characterized in that, The specific implementation method of step (4) is as follows: S41: The Transformer model in the multi-head attention mechanism is divided into frozen layers and hot update layers. Frozen layers are some layers in the model that have learned enough knowledge and are more adaptable to new data. These layers remain unchanged during training. Hot update layers are those layers in the model that need to adapt to new data in a timely manner. These layers are updated according to the latest trajectory data, so that the model can continuously learn new trajectory patterns. S42: New trajectory data is filtered based on a sliding time window, and incremental training is performed only on the hot update layer; S43: By using comparative loss and regularization techniques to constrain the update process, we ensure the consistency between new features and historical patterns.

6. The pre-training and online update method for multi-view dynamic trajectory representation learning according to claim 1, characterized in that, The specific implementation method of step (5) is as follows: S51: For newly added trajectory data, activate the corresponding spatiotemporal hybrid expert branch based on the task ID to generate task-specific trajectory embeddings. S52: Lightweight fine-tuning modules are designed for different downstream tasks. For trajectory classification tasks, this module takes the trajectory representation learned in the pre-training stage as input and classifies the trajectory through a fully connected layer. For travel time estimation tasks, this module uses a regression model to predict the travel time of the trajectory. For trajectory generation tasks, this module uses a generative adversarial network to generate new trajectories. S53: Visualize the impact of key road segments on task results through attention weights and generate explanatory reports.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor is used to execute the computer program to implement the pre-training and online update method for multi-view dynamic trajectory representation learning as described in any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the pre-training and online update method for learning dynamic trajectory representations for multiple views as described in any one of claims 1 to 6.

Images