Task coverage balanced spatial crowdsourcing recommendation method

Abstract

The application discloses a task coverage balanced space crowdsourcing recommendation method, in the user preference modeling stage, through constructing a multi-channel decoupling attention computing mechanism, the semantic features, sequence position features, spatial features and time features are modeled respectively, and the user preference representation is obtained through adaptive fusion, so as to improve the accuracy of user intention modeling. In the task recommendation stage, by introducing the initial allocation strategy based on the task allocability and the iterative optimization mechanism based on the multi-objective constraint, the collaborative optimization of task coverage and recommendation diversity is realized under the premise of controlling the computational complexity. Compared with the prior art, the application can improve the overall task coverage capability of the system while ensuring the recommendation accuracy, and enhance the diversity of the recommendation results, thereby improving the user experience and the platform resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to a method for spatial crowdsourcing recommendation with balanced task coverage, and more particularly to a method for such recommendation. Background Technology

[0002] With the rapid development of mobile internet and location technology, spatial crowdsourcing, as an emerging distributed computing model, has been widely applied in areas such as urban inspection, crowdsourced data collection, instant delivery, and ride-hailing services. In spatial crowdsourcing systems, platforms typically need to recommend executable spatial tasks to users based on their historical behavior, interests, and current status, thereby improving task completion efficiency and user experience. Existing technologies often employ sequence recommendation models to model users' historical behavior, learning the temporal relationships in the user's access to points of interest to predict the next task or location the user might be interested in, and generating recommendations accordingly. However, because spatial crowdsourcing scenarios involve multiple heterogeneous information such as spatial, temporal, and semantic data, traditional sequence recommendation methods still have significant limitations in this context.

[0003] Specifically, existing technologies have the following shortcomings. First, heterogeneous feature coupling leads to significant modeling errors. Existing methods typically fuse semantic features, spatial location information, and temporal information at the model input stage, for example, by adding or concatenating vectors into a self-attention model. During subsequent attention calculations, cross-calculation occurs between different types of features, introducing interference terms from non-homologous features. Since semantic information, spatial coordinates, and temporal features differ significantly in data distribution and representation space, this direct coupling method easily introduces additional noise, reducing the model's ability to characterize the user's true intent and thus affecting recommendation accuracy.

[0004] Secondly, recommendation strategies focused on individual optimization lead to insufficient global coverage. Traditional recommendation methods typically employ a Top-K strategy, recommending tasks based on each user's highest preference score. This approach primarily focuses on the interest matching of individual users, neglecting the overall task allocation across the system. In practice, this strategy easily results in popular tasks being repeatedly recommended, while some long-tail tasks remain unassigned for extended periods, leading to uneven utilization of task resources and reducing the platform's overall task coverage and execution efficiency.

[0005] Furthermore, the homogenization of recommendation results negatively impacts the long-term user experience. Existing methods typically prioritize maximizing prediction accuracy, which can lead to recommendations concentrating on a few high-frequency task categories. This results in a lack of categorical diversity among individual users. Over long-term use, this phenomenon may reduce user engagement, impacting user retention and the system's sustainable operational capabilities.

[0006] To address the problems existing in the prior art, the present invention aims to provide a spatial crowdsourcing task recommendation method. By decoupling and modeling multi-dimensional features in user behavior and combining them with a multi-objective optimization scheduling strategy, the method can improve task coverage and enhance the diversity of recommendation results while ensuring recommendation accuracy. Summary of the Invention

[0007] This invention addresses the technical problems in existing spatial crowdsourcing recommendation technologies, such as large modeling errors caused by heterogeneous feature coupling, insufficient global coverage due to individual-optimization-oriented recommendation strategies, and the impact of homogeneous recommendation results on long-term user experience. It provides a spatial crowdsourcing recommendation method and system with balanced task coverage. By constructing a multi-channel decoupled spatiotemporal attention calculation mechanism, a task coverage-balanced recommendation mechanism based on task assignability, and a dynamic optimization mechanism for recommendation results, it significantly improves task coverage and enhances the diversity of recommendation results while ensuring recommendation accuracy.

[0008] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0009] This invention includes the following steps:

[0010] S1: Multidimensional spatiotemporal trajectory data modeling: The system accesses the user's historical trajectory sequence in real time, extracts semantic features, location features, spatial coordinate features and time features for each time step, and initializes them as independent tensors respectively;

[0011] S2: Decoupled spatiotemporal attention modeling: The system constructs four independent attention computation branches in parallel. For semantic tensors, position tensors, and spatial coordinate tensors, respectively, dedicated query and key-value tensors are generated through linear projection, and attention matrices are calculated using scaled dot products. For time tensors, strictly bounded analytic function attention kernels are constructed to calculate time intervals and generate temporal attention matrices.

[0012] S3: Adaptive Attention Fusion and User Preference Generation: The attention results of each channel are input into the fusion module. The importance of each channel is dynamically adjusted through learnable weights to obtain a unified fusion attention matrix. Then, the user's predicted preference tensor is output through a feedforward network and a residual readout layer.

[0013] S4: Candidate Task Set Construction: Based on the user's current location, reachable distance, and task time constraints, select tasks that meet the execution conditions and construct the user's candidate task set;

[0014] S5: Initial recommendation based on task allocatability: Count the number of allocatable users corresponding to each task as the task allocatability, sort the tasks in ascending order of task allocatability, select tasks in turn and assign them to users with the highest preference score for the task and whose recommendation limit has not been reached, and generate the initial task recommendation results.

[0015] S6: Recommendation result optimization based on multi-objective constraints: Define task coverage index and recommendation diversity index, traverse the replacement combination of tasks and candidate replacement tasks in the user recommendation results, calculate the impact of replacement operation on diversity and coverage, and perform replacement when the improvement in diversity is greater than or equal to the improvement in coverage. Repeat the replacement operation until the recommendation result is stable.

[0016] S7: Output Recommendation Results: Output a list of recommended tasks for each user.

[0017] The beneficial effects of this invention are:

[0018] This invention employs a multi-channel decoupled spatiotemporal attention computation mechanism, independently modeling semantic features, location features, spatial coordinate features, and temporal features, thus avoiding cross-interference between heterogeneous features and improving the accuracy of user preference modeling. Through an initial recommendation strategy based on task allocability, it prioritizes assigning long-tail tasks with fewer allocable users to suitable users, effectively solving the problem of repeated recommendations of popular tasks while long-tail tasks remain unassigned in traditional recommendation methods, significantly improving the overall task coverage of the platform. Furthermore, through a recommendation result optimization mechanism based on multi-objective constraints, it enhances the diversity of recommendation results while ensuring task coverage, preventing recommendations from concentrating on a few high-frequency task categories and improving the long-term user experience. Through the synergistic effect of these technologies, this invention achieves a comprehensive improvement in recommendation accuracy, task coverage, and recommendation diversity in spatial crowdsourcing recommendation scenarios. Attached Figure Description

[0019] Figure 1 A schematic diagram of the overall process of the spatial crowdsourcing task recommendation method provided by the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of the decoupled spatiotemporal attention model provided by the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0022] like Figure 1 , 2As shown: This invention provides a spatial crowdsourcing task recommendation method. Based on the system architecture and the temporal logical sequence of the data flow, the execution steps of this invention are as follows:

[0023] Step 1: Multidimensional Spatiotemporal Trajectory Data Modeling

[0024] The system accesses the user's historical trajectory sequences in real time. For each time step, it extracts independent features across four dimensions and initializes them as follows: dimensional tensor:

[0025] semantic tensor : Interest point features corresponding to historical tasks.

[0026] position tensor : Position encoding of the corresponding sequence.

[0027] Spatial coordinate tensor : Convert to latitude and longitude or projected coordinate space.

[0028] time tensor : The timestamp of the corresponding task execution.

[0029] Step 2: Decoupling Spatiotemporal Attention Modeling

[0030] To ensure numerical stability and eliminate inner product noise in heterogeneous tensors, the system constructs four independent attention computation branches in parallel:

[0031] Semantic / Location / Spatial Attention Matrix Calculation:

[0032] Specifically for Tensors, through linear projection, generate custom query and key-value tensors. For example, for semantic tensors, perform... and ,in and These are the projection matrices for the query and the key, respectively, and the attention matrix is ​​calculated using a scaled dot product. Similarly, precise output. and .

[0033] Steady-state periodic time attention calculation:

[0034] Since timestamps do not possess a rigorous dot product distance property, the system constructs a... Strictly bounded analytic function attention kernel, computation time interval Through the formula Generate Temporal Attention Matrix ,in The attenuation coefficient is set manually. This design ensures smooth attenuation and gradient stability during backpropagation.

[0035] Step 3: Adaptive Attention Fusion and User Preference Generation

[0036] The attention results of each channel are input into the fusion module, and the importance of each channel is dynamically adjusted through learnable weights to obtain a unified fused attention matrix. Specifically, this invention does not use a rigid weighted summation, but instead... The input is fed into a feedforward network, for these four Dynamic generation of channel weights using a matrix Output the fused alignment attention matrix The predicted preference tensor of the user is then output through a feedforward network and a residual readout layer. .

[0037] Step 4: Constructing the Candidate Task Set

[0038] Based on the user's current location, reachability distance, and task time constraints, tasks that meet the execution conditions are filtered to construct a candidate task set for the user.

[0039] Step 5: Initial Recommendation Based on Task Assignability

[0040] The candidate task set is processed as follows:

[0041] The number of users who can be assigned to each task is counted, and this number is used as the task's assignability.

[0042] Sort by task assignability from smallest to largest;

[0043] Select tasks in sequence and assign them to users who have the highest preference scores for the task and whose recommendation limit has not been reached, thus generating initial task recommendation results.

[0044] Step Six: Optimization of Recommendation Results Based on Multi-Objective Constraints

[0045] Based on the initial recommendation results, the recommendation results are optimized:

[0046] Indicator definition:

[0047] Task coverage: This represents the proportion of recommended tasks out of all tasks.

[0048] Recommendation diversity: refers to the degree of difference in categories among individual users' recommendation tasks;

[0049] Define a multi-objective joint benefit function ,in This is the balance coefficient.

[0050] Replacement decision mechanism: For tasks in the user recommendation results, candidate replacement tasks are selected, and the impact of the replacement operation on diversity and coverage is calculated. Replacement is executed when the following condition is met: the improvement in diversity is greater than or equal to the improvement in coverage. Specifically, when traversing replacement combinations, the system dynamically calculates the coverage loss boundary caused by the replacement task. and theoretical diversity benefits Only when In other words, task replacement is only triggered when the diversity gains are mathematically determined to exceed the coverage loss.

[0051] Iterative optimization: Repeat the replacement operation until the recommendation result is stable or the termination condition is met.

[0052] Step 7: Output Recommendation Results

[0053] Output a task recommendation list for each user, which can be used for task distribution or user selection for execution.

[0054] This invention proposes a spatial crowdsourcing recommendation method with balanced task coverage. By constructing a decoupled spatiotemporal attention modeling mechanism and a task coverage balancing optimization mechanism, it can effectively improve the recommendation accuracy, task coverage, and category diversity of recommendation results in spatial crowdsourcing scenarios, thereby improving the overall task allocation efficiency of the platform and the long-term user experience. The method of this invention can be applied to spatial crowdsourcing scenarios such as ride-hailing services, instant delivery, urban inspection, and crowdsourced data collection. In specific applications, the platform first acquires user historical trajectory data, point-of-interest location data, and task time information, and constructs a sequence of user historical behavior. Then, it independently models the semantic features, sequence location features, spatial location features, and temporal features, generating user preference representations through an adaptive fusion mechanism. Next, it combines the user's current location, task reachability, and task time constraints to generate a candidate task set, and generates initial recommendation results based on task allocability. Finally, it iteratively adjusts the recommendation results through a coverage balancing optimization mechanism, thereby improving the overall task coverage capability and enhancing the category diversity of the recommendation results.

[0055] To verify the effectiveness of this invention, this embodiment uses two real-world spatial trajectory datasets, NYC and TKY, for experimental verification. The NYC dataset contains 1084 users, 5136 points of interest (POIs), and 146855 behavioral records; the TKY dataset contains 2294 users, 7874 POIs, and 445277 behavioral records. During the experiment, the user's historical behavior sequences were divided into training, validation, and test sets according to chronological order, and corresponding spatial crowdsourcing tasks were generated based on the POIs. The experimental hardware environment used an Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz processor, 128GB of memory, and an NVIDIA GeForce GTX 1080 graphics processor. During the experiment, the reachable distance for workers was set to 3km, the effective task time was set to 1 hour, and the recommended number of tasks was set to 10.

[0056] The method of this invention was compared and validated with existing recommendation methods such as SASRec, Flashback, and STiSAN. Experimental results show that the method of this invention achieves a Hit@10 index of 0.5749 and an NDCG@10 index of 0.3648 on the NYC dataset; and a Hit@10 index of 0.5291 and an NDCG@10 index of 0.3141 on the TKY dataset, both of which are superior to the existing comparative methods.

[0057] Furthermore, based on the same user preference modeling results, this invention designs a task coverage balancing optimization mechanism. By combining task assignability and the category differences of recommendation results, the initial recommendation results are dynamically adjusted, thereby avoiding the problems of repeated recommendations of popular tasks and long-tail tasks remaining unassigned for a long time, which are caused by traditional Top-K recommendation methods that only focus on matching the preferences of a single user. Specifically, this invention prioritizes recommending tasks with fewer assignable users to improve overall task coverage; at the same time, by iteratively optimizing the task category distribution in the recommendation results, the category differences of the recommendation results are improved, thereby achieving a synergistic improvement in the overall task allocation efficiency and recommendation balance of the platform while maintaining a high user preference matching capability. In addition, this invention adopts a recommendation strategy based on task assignability and a coverage balancing optimization mechanism, which does not require a complex graph matching exhaustive solution process, and can still maintain low computational overhead in large-scale task recommendation scenarios, improving the overall execution efficiency of the system.

[0058] This invention constructs a decoupled spatiotemporal attention modeling mechanism, which independently calculates the attention of semantic features, sequence position features, spatial position features and temporal features in user historical behavior, and generates a unified user preference representation through an adaptive fusion mechanism, thereby reducing the interference caused by heterogeneous feature coupling and improving the accuracy of user preference modeling.

[0059] This invention addresses the problems of repeated recommendations of popular tasks and long-tail tasks remaining unassigned in traditional recommendation methods. It designs a task coverage balancing recommendation mechanism based on task assignability, which improves the overall task coverage of the platform by prioritizing the recommendation of tasks with fewer assignable users.

[0060] This invention further constructs a dynamic optimization mechanism for recommendation results. By iteratively adjusting the distribution of task categories in the recommendation results, it improves the category diversity of recommendation results while maintaining the ability to match user preferences, thereby improving the long-term user experience.

[0061] This invention adopts an optimization strategy based on incremental recommendation result updates. During the recommendation result adjustment process, there is no need to use a complex graph matching exhaustive solution process, thereby reducing the computational complexity of the system and improving the task recommendation efficiency in large-scale spatial crowdsourcing scenarios.

[0062] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A spatial crowdsourcing recommendation method with balanced task coverage, characterized in that, Includes the following steps: S1: Multidimensional spatiotemporal trajectory data modeling: The system accesses the user's historical trajectory sequence in real time, extracts semantic features, location features, spatial coordinate features and time features for each time step, and initializes them as independent tensors respectively; S2: Decoupled spatiotemporal attention modeling: The system constructs four independent attention computation branches in parallel, which generate dedicated query and key-value tensors for semantic tensors, position tensors, and spatial coordinate tensors through linear projection, and calculate the attention matrix using scaled dot products; For the time tensor, a strictly bounded analytic function attention kernel is constructed to calculate the time interval and generate the time attention matrix; S3: Adaptive Attention Fusion and User Preference Generation: The attention results of each channel are input into the fusion module. The importance of each channel is dynamically adjusted through learnable weights to obtain a unified fusion attention matrix. Then, the user's predicted preference tensor is output through a feedforward network and a residual readout layer. S4: Candidate Task Set Construction: Based on the user's current location, reachable distance, and task time constraints, select tasks that meet the execution conditions and construct the user's candidate task set; S5: Initial recommendation based on task allocatability: Count the number of allocatable users corresponding to each task as the task allocatability, sort the tasks in ascending order of task allocatability, select tasks in turn and assign them to users with the highest preference score for the task and whose recommendation limit has not been reached, and generate the initial task recommendation results. S6: Recommendation result optimization based on multi-objective constraints: Define task coverage index and recommendation diversity index, traverse the replacement combination of tasks and candidate replacement tasks in the user recommendation results, calculate the impact of replacement operation on diversity and coverage, and perform replacement when the improvement in diversity is greater than or equal to the improvement in coverage. Repeat the replacement operation until the recommendation result is stable. S7: Output Recommendation Results: Output a list of recommended tasks for each user.

2. The spatial crowdsourcing recommendation method with balanced task coverage according to claim 1, characterized in that, In step S2, the attention matrix is ​​calculated. ,in and , and These are the projection matrices of the query and the key, respectively; the semantic tensor. : Interest point features corresponding to historical tasks; position tensor : Position encoding of the corresponding sequence; spatial coordinate tensor : Convert to latitude and longitude or projected coordinate space.

3. The spatial crowdsourcing recommendation method with balanced task coverage according to claim 1, characterized in that, The time interval is calculated in step S2. ; through formula Generate Temporal Attention Matrix ,in The attenuation coefficient is set by the user.

4. The spatial crowdsourcing recommendation method with balanced task coverage according to claim 1, characterized in that, Step S3 specifically involves: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] The input is fed into a feedforward network, for these four Dynamic generation of channel weights using a matrix Output the fused alignment attention matrix The predicted preference tensor of the user is then output through a feedforward network and a residual readout layer. .

5. The spatial crowdsourcing recommendation method with balanced task coverage according to claim 1, characterized in that, In step S6, when traversing the replacement combinations, the system dynamically calculates the coverage loss boundary caused by the replacement task. and theoretical diversity benefits Only when In other words, task replacement is only triggered when the diversity gains are mathematically determined to exceed the coverage loss.

Images