Space crowdsourcing task allocation method based on trajectory prediction

Abstract

The invention relates to a space crowdsourcing task allocation method based on trajectory prediction, and the method comprises the steps: constructing an MMLSTM model based on a meta path and a Mogrimer LSTM module in combination with end-to-end training, and enabling the MMLSTM model to be used for predicting a user trajectory; and according to the predicted user track, adopting a dynamic strategy of combining grouping insertion and single task insertion to distribute the published new task to the corresponding user on line. Compared with the prior art, on one hand, by introducing the space-time diagram and the multi-type meta path, the complex space-time interaction relationship between the users can be explicitly and structurally modeled, and by combining with the Mogrifier LSTM, the time sequence evolution can be dynamically captured, so that the prediction of the future trajectory of the user is more accurate; and on the other hand, a dynamic strategy of combining grouping insertion and single task insertion is adopted, geographically adjacent tasks are packaged and allocated, and an access sequence is optimized, so that the total detour overhead when a user executes a plurality of tasks can be remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of spatial crowdsourcing technology, and in particular to a spatial crowdsourcing task allocation method based on trajectory prediction. Background Technology

[0002] Spatial crowdsourcing is a new model that distributes location-based tasks to workers via the internet, primarily used in fields such as news, disaster response, and urban planning. The main process includes task posting and task execution. During task posting, requesters submit tasks to the platform, which then assigns them to workers based on the given pattern. During task execution, workers must travel to designated locations to complete the task, typically requiring real-time location reporting. In spatial crowdsourcing applications, task posters can dynamically initiate various location-based tasks, such as photo capture, video recording, traffic monitoring, or hotspot information reporting. The platform needs to be aware of user locations in real time and dynamically and efficiently assign tasks to suitable users based on the spatiotemporal constraints of the tasks. While traditional task allocation methods have achieved some success in static, offline scenarios, they generally assume that user and task location information is known at system initialization, making it difficult to reflect the dynamic characteristics of users and tasks continuously joining and leaving the system. In practical applications, the future location of users is highly uncertain, leading to significant limitations in accuracy and adaptability for allocation strategies based solely on the current state.

[0003] Task allocation, as the core mechanism of spatial crowdsourcing systems, directly impacts task completion rates and user participation. Traditional methods typically recommend nearby tasks based on a user's current location or their preset starting and destination points, aiming to minimize additional travel distances or detours. While these methods are computationally efficient and suitable for static task environments, they often overlook users' personalized travel preferences and struggle to cope with the complex and ever-changing movement patterns in reality.

[0004] With the development of mobile devices and location technology, platforms can continuously acquire users' travel trajectory data, making it possible to introduce richer spatiotemporal information into task allocation. Increasing research is focusing on task allocation strategies based on users' historical trajectories, namely trajectory-aware task assignment. These methods assume that users are more inclined to accept tasks near their usual paths to reduce unnecessary detours, thereby increasing their willingness to accept tasks and execution efficiency. They allow users to deviate from their original movement trajectory to perform tasks and then return to their original trajectory.

[0005] Existing trajectory sensing task allocation methods have the following main shortcomings: 1. Traditional methods (such as the PNN algorithm) assume that the user's trajectory is fixed or based only on the current position, forcing the user to return to the original path from the deviation point, which increases unnecessary movement overhead and cannot adapt to the dynamic changes of users and tasks in real-world scenarios.

[0006] 2. Methods based on historical trajectories (such as the TCA algorithm) are essentially "static trajectory matching" and cannot predict the user's future location. This results in the inability to effectively schedule tasks in advance before the user arrives at the task area, leading to low task allocation rate and efficiency.

[0007] 3. Most existing deep learning-based trajectory prediction models (such as ST-RNN, Social-LSTM, etc.) implicitly capture user interactions and lack modeling of explicit, multi-level spatiotemporal relationships between users, resulting in insufficient prediction accuracy and thus affecting task allocation performance. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a spatial crowdsourcing task allocation method based on trajectory prediction, which can accurately predict the user's future trajectory and achieve efficient and low-overhead online task allocation.

[0009] The objective of this invention can be achieved through the following technical solution: a spatial crowdsourcing task allocation method based on trajectory prediction, comprising the following steps: S1. Based on the metapath and Mogrifier LSTM module, and combined with end-to-end training, an MMLSTM model is constructed to predict user trajectories. S2. Based on the predicted user trajectories, a dynamic strategy combining group insertion and single-task insertion is adopted to assign the newly released tasks to the corresponding users online.

[0010] Further, step S1 includes the following process: S11. Model user movement as a spatiotemporal graph and design meta-paths to display the spatiotemporal interaction relationships between modeled users; S12. Combining meta-paths, user-defined location embeddings, and the Mogrifier LSTM module, construct an MMLSTM framework that includes path MLSTM and node MLSTM modules. S13. Perform end-to-end training on the MMLSTM framework to obtain the MMLSTM model.

[0011] Furthermore, step S11 specifically involves modeling the user's movement as a spatiotemporal graph. , where the set of nodes This indicates the user's spatial location at each time step. ; Space edge Indicates users at the same time and Spatial distance between them Time edge Connect the distance traveled by the same user between adjacent time steps. ; The metapath is designed to include the following types: Space-Spatial Metapath Through mutual neighbors Create user With users exist Indirect spatial relationships at different times: Space-Time Metapath , indicating user Current position relative to user Relative positional relationship at the previous moment: Time-space metapath Associated users movement trajectory and user Position at the previous moment: Time-Time Metapath Characterizing users The movement trend of the last two steps: .

[0012] Furthermore, in step S12, the path MLSTM module specifically performs temporal modeling using the Mogrifier LSTM module for each type of meta-path, generating high-order features representing the relationships between users. The working process of the path MLSTM module includes: Path aggregation, for each type of meta-path (such as space-space, time-space, etc.), will aggregate the user... With all other users Meta-path features between Perform a summation operation to obtain the aggregated representation. ; Relation embedding, through nonlinear transformation functions The aggregated features are mapped to the embedding space, represented as ; Temporal modeling, using the Mogrifier LSTM module, embeds features The hidden state of the previous moment Fusion, outputting the relation representation at the current moment. .

[0013] Furthermore, in step S12, the node MLSTM module specifically fuses the user's own location embedding with all meta-path relationship features, and then decodes and outputs the user's future location coordinates through the Mogrifier LSTM module. The working process of the node MLSTM module includes: Feature fusion embeds the user's current location into a representation. Six types of meta-path features output by path MLSTM (in The components are concatenated to form a joint representation. ; Temporal modeling, features Compared to the previous hidden state The input is fed into the MogrifierLSTM module to capture the evolution of user behavior and output the updated hidden state. ; Position decoding maps the hidden state to an offset vector using a decoder. This is then added to the current coordinates to obtain the predicted position for the next moment.

[0014] Furthermore, during training in step S13, the model is optimized by minimizing the mean square error (MSE) loss function between the predicted trajectory and the true trajectory. The MMLSTM model first synchronously predicts the next position of all users at each time step, and then proceeds to the next time step.

[0015] Further, step S2 includes the following process: S21. Update all predicted user trajectories to the current time, and obtain the current location and subsequent available paths for each user; S22. For each user, determine whether the new task can be merged into a task group with any task that the user has already accepted and is geographically close. If the determination is yes, execute the grouping insertion strategy and obtain the user's detour cost increment. Otherwise, execute the single-task insertion strategy and obtain the user's detour cost increment; S23. From all users who meet the detour budget constraint, task time constraint, and capacity constraint, select the user with the smallest detour cost increment to execute the new task, and update the user's trajectory and remaining resources.

[0016] Furthermore, in step S22, if a new task... With users Task group accepted The spatial distance between any task point and a preset threshold is less than the threshold. Then determine the new task Able to merge users task group The new task group was formed by merging and reorganizing. .

[0017] Furthermore, the process of executing the grouping insertion strategy in step S22 includes: For task groups ,turn up Zhongyu The most recent task is inserted after it to obtain the initial task sequence. ; based on -OPT optimization The sequence generates both ascending and descending access orders. Finding user tracks Up to the new mission group The nearest trajectory point Based on the traversal window Determine the traversal trajectory Insertion segment within the window distance range Calculate the detour cost increment: in, and These are the first and last task points of the access sequence, respectively. For the task group Path distance after determining the access order; Selecting detour cost increment The minimal insertion method enables joint scheduling of multiple tasks, thereby spreading the cost of detours.

[0018] Furthermore, the process of executing the single-task insertion strategy in step S22 includes: Traverse all adjacent path segments Calculate the corresponding detour cost increment: And choose to make The smallest insertion point.

[0019] Compared with the prior art, the present invention has the following advantages: This invention first combines metapath and Mogrifier LSTM modules to construct an MMLSTM model for predicting user trajectories. The MMLSTM model achieves an effective balance between modeling individual dynamics and neighbor interaction dependencies through structured spatiotemporal modeling and a multi-round Mogrifier adjustment mechanism, significantly improving the accuracy and robustness of trajectory prediction. Then, a dynamic strategy combining group insertion and single-task insertion is adopted to allocate newly issued tasks to the corresponding users online, enabling efficient allocation in complex dynamic environments and ensuring that users' detour costs are minimized.

[0020] This invention introduces spatiotemporal graphs and multi-type meta-paths to explicitly and structurally model the complex spatiotemporal interaction relationships between users. By combining the Mogrifier LSTM module to construct path MLSTM modules and node MLSTM modules respectively, it is possible to dynamically capture temporal evolution, making the prediction of users' future trajectories more accurate and laying a solid foundation for subsequent high-quality task allocation.

[0021] When allocating online tasks based on predicted trajectories, this invention integrates multiple optimization mechanisms such as local window search, task group aggregation, and candidate user filtering. By packaging and allocating geographically adjacent tasks and optimizing the access order, it significantly reduces the total detour cost for users to perform multiple tasks, achieving the effect of "completing tasks along the way". Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram illustrating the application process of an example. Figure 3 A schematic diagram of the Mogrifier LSTM module framework; Figure 4 A schematic diagram of the path MLSTM module framework; Figure 5 A schematic diagram of the node MLSTM module framework; Figure 6 This is a schematic diagram of the MMLSTM model framework; Figure 7 This is a schematic diagram of the algorithm flow for online task allocation in this invention; Figure 8 A schematic diagram of the algorithm flow for the grouping insertion strategy in online task allocation. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0024] Example like Figure 1As shown, a spatial crowdsourcing task allocation method based on trajectory prediction includes the following steps: S1. Based on the metapath and Mogrifier LSTM module, and combined with end-to-end training, an MMLSTM model is constructed to predict user trajectories. S2. Based on the predicted user trajectories, a dynamic strategy combining group insertion and single-task insertion is adopted to assign the newly released tasks to the corresponding users online.

[0025] This embodiment applies the above-described solution, such as Figure 2 As shown, a two-stage processing framework is constructed: first, a trajectory prediction model MMLSTM is used to accurately predict the user's future movement trajectory, and then the online task allocation method GILO (Group Insertion with Local Optimization) is executed based on the trajectory prediction results.

[0026] Specifically, in the first stage, an MMLSTM model is constructed based on meta-paths and Mogrifier LSTM modules (such as...). Figure 3 As shown, trajectory prediction is achieved.

[0027] 1) Spatiotemporal graph construction: Modeling user movement as a spatiotemporal graph , where the set of nodes This indicates the user's spatial location at each time step. ; Space edge Indicates that at the same time, the user and Spatial distance between them Time edge Connect the distance moved by the same user between adjacent time steps, where, .

[0028] 2) Meta-path design: To explicitly model complex spatiotemporal interactions between users, four meta-path types are defined: Space-Spatial Metapath Through common neighbors Create user With users exist Indirect spatial relationship of time

[0029] Space-Time Metapath :user Current position relative to user Relative positional relationship at the previous moment

[0030] Time-space metapath Associated users movement trajectory and user The position at the previous moment

[0031] Time-Time Metapath Characterizing users The trend of the last two steps

[0032] These metapath structures model the dynamic trajectories between users and within the user itself at multiple scales, both spatially and temporally.

[0033] 3) MMLSTM model prediction: Path MLSTM module (e.g.) Figure 4 As shown): For each type of meta-path, the Mogrifier LSTM module is used for temporal modeling to generate high-order features representing the relationships between users. The entire process includes the following three key steps: Path aggregation: For each type of meta-path (e.g., space-space, time-space, etc.), the user... With all other users Meta-path features between Perform a summation operation to obtain the aggregated representation. ; Relational embedding: through nonlinear transformation functions The aggregated features are mapped to the embedding space, represented as ; Temporal modeling: Using the Mogrifier LSTM module, embedded features are... The hidden state of the previous moment Fusion, outputting the relation representation at the current moment. .

[0034] Through the above process, the path MLSTM module can effectively capture spatial or temporal dependencies between users and integrate dynamic features of multiple meta-path types within a unified framework. The formula is defined as follows:

[0035] Here, the Mogrifier LSTM module is applied to capture user movement factors by selectively emphasizing the most salient information in the user's behavioral trajectory. Specifically, the input vector... This is provided to the Mogrifier LSTM module to preserve order dependencies and extract contextual features. Formally, it computes... The current output Depends on the current input and the previous hidden state .

[0036] Node MLSTM modules (such as) Figure 5 As shown): The user's own location is embedded and fused with all meta-path relationship features, and then decoded and outputted by the Mogrifier LSTM module to output the user's future location coordinates. The entire process consists of the following three main steps: Feature fusion: embedding the user's current location Six types of meta-path features output by path MLSTM (in The components are concatenated to form a joint representation. ; Temporal modeling: combining features Compared to the previous hidden state The input is fed into the MogrifierLSTM module to capture the evolution of user behavior and output the updated hidden state. ; Position decoding: The hidden state is mapped to an offset vector using a decoder. This is then added to the current coordinates to obtain the predicted position for the next moment.

[0037] This module integrates node dynamics and neighbor information to achieve fine-grained modeling of user location, significantly improving trajectory prediction accuracy. The specific calculation process is shown below:

[0038] 4) MMLSTM model training To achieve efficient and accurate user trajectory prediction, the entire MMLSTM framework (such as...) Figure 6 (As shown) Training is performed in an end-to-end manner. The embedding layer uses linear transformation combined with... The activation function performs feature compression, and Dropout is applied during the training phase to alleviate overfitting; while the decoding layer adopts a similar structure but does not use Dropout to ensure prediction stability.

[0039] The model is optimized by minimizing the mean squared error (MSE) loss function between the predicted and actual trajectories. Notably, since the meta-path representation depends on the location states of all users at the current time step, during the prediction phase, the model must simultaneously predict the next location of all users at each time step before proceeding to the next time step. This strategy ensures the consistency of path dependencies across the temporal dimension, improving the robustness of the prediction.

[0040] Ultimately, the MMLSTM model outputs to the user. In the future Predicted trajectory within the time period This provides a high-confidence location prior for the subsequent task allocation module.

[0041] In summary, the MMLSTM model, through structured spatiotemporal modeling and combined with a multi-round Mogrifier adjustment mechanism, achieves an effective balance between modeling individual dynamics and neighbor interaction dependencies, significantly improving the accuracy and robustness of trajectory prediction and providing a foundation for subsequent task allocation.

[0042] Phase Two: Online Task Allocation Based on Predicted Trajectories – GILO. When a new task is released, based on the user trajectory predicted in the previous phase, the following online task allocation process is executed (e.g., ...). Figure 7 (as shown) User status update: Update all users trajectory Updated to the current time Get its current location and subsequent available paths.

[0043] Insertion strategy selection: For each user, determine whether the new task can be merged into a task group with any geographically adjacent task already accepted by the user (the geographical distance between tasks is less than a preset threshold). ).

[0044] Grouping insertion strategy (e.g.) Figure 8 (As shown): If they can be merged, then: If new task With users Task group accepted The spatial distance between any task point and the threshold is less than the threshold. Then Merge into this group to create a new task group. ; For task groups ,turn up Zhongyu The most recent task is inserted after it to obtain the initial task sequence. ; based on -OPT optimization The sequence generates both ascending and descending access orders. Finding user tracks Up to the new mission group The nearest trajectory point Based on the traversal window Determine the traversal trajectory Insertion segment within the window distance range Calculate the detour cost increment: in, and These are the first and last task points of the access sequence, respectively. For the task group Path distance after determining the access order; Selecting detour cost increment The minimal insertion method enables joint scheduling of multiple tasks, thereby spreading the cost of detours.

[0045] Single-task insertion strategy: If merging is not possible, when the task... If a task cannot be combined with any existing task, the algorithm reverts to a single-task insertion strategy, which involves traversing all adjacent path segments. Calculate the corresponding detour cost increment: And choose to make The minimum insertion point. The computational complexity of this strategy is O(n log n). It is suitable for online application scenarios with high requirements for response latency.

[0046] User selection and update: From all users who meet the detour budget constraint, task time constraint, and capacity constraint, select the user with the smallest detour cost increment to perform the task, and update the user's trajectory and remaining resources.

[0047] To verify that the computational complexity of the GILO method described above is controllable, this embodiment sets the maximum trajectory length to be... The number of users is The number of tasks is The complexity analysis of the GILO method is as follows: Single-task insertion: Each task requires at most one iteration. Each user's trajectory contains [number] user trajectories, and each trajectory includes [number] user trajectories. There are insertion segments, so the complexity is O(n). ; Grouping insertion: For tasks Each corresponding candidate task group First, determine the access order of the task group, then call the 2-OPT algorithm, with a complexity of O(n log n). Then, it is necessary to determine the insertion segment of the trajectory and calculate the detour cost increment based on the traversal window, with a complexity of O(n). In actual algorithm execution, If the window size is less than the user's maximum capacity, traverse the window. If it is a constant, then it applies to a single task. The complexity is Each task requires at most one iteration. With a user base, the complexity is O(n). ; Worst-case scenario: The overall worst-case time complexity is... .

[0048] In summary, the GILO method proposed in this scheme integrates multiple optimization mechanisms such as local window search, task group aggregation, and candidate user screening, achieving efficient allocation in complex dynamic environments with controllable computational complexity. It can meet the real-time response requirements of spatial crowdsourcing environments and has high practical application value.

[0049] To verify the performance of this solution for user trajectory prediction, experiments were conducted on three public datasets (ETH dataset, UCY dataset, and Geolife dataset). The proposed MMLSTM model was compared with ST-RNN, STRNN (Structural-RNN), and MESRNN models. Training data was divided into training and validation sets, with 80%-20% allocated to each set. All models were trained for 10 epochs, and convergence was verified by the convergence trend of the training loss. The optimizer used was ADAM, with an initial learning rate of 0.001. Meta-path features were calculated based on a simple difference operator as the distance function. Mean squared error (MSE) was used as the loss function during training to improve the accuracy of location prediction. The evaluation metrics were Average Displacement Error (ADE) and Final Displacement Error (FDE). ADE measures the overall accuracy of trajectory prediction, while FDE focuses more on the accuracy of the model's judgment of the final target location.

[0050] As shown in Table 1, the MMLSTM model significantly outperforms other comparative methods in both ADE and FDE metrics across all five datasets, demonstrating superior trajectory prediction performance. This model achieves the lowest ADE and FDE across all prediction tasks, indicating its strong ability to capture the temporal dynamics and spatial dependencies in user trajectories, while also exhibiting good generalization and prediction stability. In contrast, ST-RNN shows significantly higher error values, reflecting its shortcomings in handling complex movement patterns and multi-factor interactions. Although improved models such as SRNN and MESRNN enhance prediction accuracy to some extent by introducing structure-aware mechanisms and multi-scale feature enhancements, their overall performance still falls short of MMLSTM. In summary, MMLSTM maintains high prediction accuracy while also considering the model's structural expressive power, making it particularly suitable for spatiotemporal crowdsourcing task allocation scenarios with high trajectory accuracy requirements, and possessing strong practical application potential.

[0051] Table 1. Comparative experimental results of trajectory prediction models on ADE / FDE indices. To verify the performance of this solution for task allocation, experiments were conducted on the publicly available Geolife dataset in this embodiment. The proposed GILO method was compared and analyzed with PPN, E-Greedy, and G-Dijkstra methods. This embodiment evaluates the effectiveness and efficiency of various task allocation methods from two dimensions: the number of tasks allocated and the total detour cost for users. The experiment, based on the Geolife dataset, selected data within a spatial region between latitude 39.9–40.0 and longitude 116.3–116.5 to construct user trajectories, with a trajectory length of 40. Task locations were randomly generated within this spatial region. The maximum task capacity for users was set to 10, the average user speed was 10 km / h, and the task grouping threshold was 0.5 km.

[0052] Table 2 Performance of Task Allocation Method: Number of Tasks Allocated / Detour Cost Index Impact Table 3 Performance of Task Allocation Methods: Number of Tasks Allocated / Number of Users as a Metric for Detour Cost Impact Table 2 shows the performance of different algorithms as the number of tasks varies from 1000 to 5000. The table reveals that G-Dijkstra and GILO assigned the most tasks across all task sizes, indicating that their task grouping and insertion strategies effectively improved task acceptance rates. In contrast, E-Greedy and PNN showed weaker task allocation capabilities, especially PNN, whose performance declined more significantly with increasing task numbers, demonstrating its limited allocation ability in task-intensive scenarios. Regarding total detour cost, G-Dijkstra and GILO still performed superiorly, significantly lower than E-Greedy and PNN, indicating that their allocation strategies were more efficient and their task insertion was more reasonable.

[0053] Table 3 shows the performance of different algorithms with user numbers ranging from 100 to 500. Experimental results show that, under different user numbers, GILO and G-Dijkstra methods outperform PNN and E-Greedy in both task allocation quantity and total detour cost. Specifically, GILO and G-Dijkstra consistently allocate the most tasks, indicating that their task grouping and path optimization mechanisms make fuller use of limited user resources. Simultaneously, they consistently maintain low detour costs, outperforming other methods and demonstrating higher path planning efficiency and more rational task allocation. This fully demonstrates that the proposed GILO method, while maximizing the number of tasks allocated, ensures low total detour costs for users, thus optimizing task allocation efficiency.

Claims

1. A trajectory prediction based spatial crowdsourcing task allocation method, characterized in that, The method comprises the following steps: S1, constructing an MMLSTM model based on a meta-path and a Mogrifier LSTM module in combination with end-to-end training, for predicting user trajectories; S2, according to the predicted user trajectories, using a dynamic strategy combining group insertion and single-task insertion to online assign a new published task to a corresponding user. 2.The trajectory prediction based spatial crowdsourcing task allocation method of claim 1, wherein, The step S1 comprises the following process: S11, modeling the movement of a user as a space-time graph, and designing a meta-path to show the space-time interaction relationship between the modeled users; S12, combining the meta-path, the position embedding of the user itself and the Mogrifier LSTM module to construct an MMLSTM framework comprising a path MLSTM module and a node MLSTM module; S13, performing end-to-end training on the MMLSTM framework to obtain an MMLSTM model. 3.The trajectory prediction based spatial crowdsourcing task allocation method of claim 2, wherein, Specifically, step S11 involves modeling the user's movement as a spatiotemporal graph. , where the set of nodes This indicates the user's spatial location at each time step. ; Space edge Indicates users at the same time and Spatial distance between them Time edge Connect the distance traveled by the same user between adjacent time steps. ; The process of designing a meta-path comprises the following types: Space-Space meta-path By common neighbors Establishing users With users Indirect spatial relationship at Time Space-time meta-path , representing the user current location relative to the user relative location relationship at the previous time point: Time-space cell path , associated user 's movement trajectory and user 's position at the previous time instant: Time-time meta-path , characterizing users Recent two-step movement trends: 。 4.The trajectory prediction based spatial crowdsourcing task allocation method of claim 3, wherein, The path MLSTM module in the step S12 is specifically to use the Mogrifier LSTM module to perform time series modeling for each type of meta-path, generate high-order features representing the relationship between users, and the working process of the path MLSTM module comprises: Path aggregation, for each type of meta-path (e.g. space-space, time-space, etc.), summing the meta-path features between the user and all other users to obtain an aggregated representation ;​ relationship embedding, by a non-linear transformation function mapping the aggregated features to an embedding space, denoted as ; Temporal modeling, with the help of the Mogrifier LSTM module, embeds features from the previous time step fusion, output the relationship representation at the current time .

5. The trajectory prediction based spatial crowdsourcing task allocation method according to claim 4, characterized in that, The node MLSTM module in the step S12 is specifically to fuse the position embedding of the user itself with all meta-path relationship features, and then decode and output the future position coordinates of the user through the Mogrifier LSTM module again, and the working process of the node MLSTM module comprises: Feature fusion, embedding representation of the current location of the user six class meta-path features of the path MLSTM output (Where ) are spliced to form a joint representation ; Temporal modeling, features from the previous time step are input to a Mogrifier LSTM module that captures the evolution of user behavior and outputs an updated hidden state ; Position decoding, mapping of the hidden state by a decoder to an offset vector and added to the current coordinates to obtain the position prediction for the next time instant. 6.The trajectory prediction based spatial crowdsourcing task allocation method of claim 2, wherein, In the step S13, the training is optimized by minimizing the mean square error loss function between the predicted trajectory and the real trajectory, and the MMLSTM model synchronously predicts the next position of all users at each time step, and then enters the next time step.

7. The trajectory prediction based spatial crowdsourcing task allocation method according to claim 5, characterized in that, The step S2 comprises the following process: S21, updating all predicted user trajectories to the current time, obtaining the current position and subsequent available paths of each user; S22, for each user, determining whether a new task can be merged into a task group with any task accepted by the user and geographically adjacent to the new task, if the determination is yes, executing a group insertion strategy and obtaining the detour cost increment of the user; otherwise, executing a single-task insertion strategy and obtaining the detour cost increment of the user; S23, selecting a user with the minimum detour cost increment from all users satisfying the detour budget constraint, the task time constraint and the capacity constraint to execute the new task, and updating the trajectory and the remaining resources of the user. 8.The method of claim 7, wherein, If the spatial distance between the new task and any task point in the task group accepted by the user is less than a preset threshold in the step S22 , it is determined that the new task can be incorporated into the task group of the user , and a new task group is obtained by combining the new task and the task group accepted by the user . ​​​​​ 9.The trajectory prediction based spatial crowdsourcing task allocation method of claim 8, wherein, The process of executing the group insertion strategy in the step S22 comprises: For the task group , find the nearest task in , insert it after, get the initial task sequence ; Based on -OPT optimization sequences, generating both forward and reverse order access sequences; Finding user tracks Up to the new mission group The nearest trajectory point Based on the traversal window Determine the traversal trajectory Insertion segment within the window distance range Calculate the detour cost increment: wherein, and are the first and last task points of the access sequence, respectively, is the task group the path distance after determining the access order; Increment of detour cost The minimum insertion mode realizes the joint scheduling of multiple tasks, thereby amortizing the detour cost. 10.The method of claim 7, wherein, The process of executing the single-task insertion strategy in the step S22 comprises: traversing all adjacent path segments , calculating a corresponding detour cost increment: and selecting the point of insertion that minimizes the minimum insertion point.

Images