Personalized task allocation algorithm based on graph neural network

Abstract

The invention relates to the technical field of task allocation of mobile crowd sensing, in particular to a personalized task allocation algorithm based on a graph neural network, and the algorithm comprises the steps: extracting a worker capability element and a task demand element to construct a worker-task recommendation allocation graph, and proposing a task recommendation allocation model based on the graph neural network, estimating a recommendation allocation possibility between the to-be-allocated task and an employee; screening out a worker-task pre-allocation pair set from the worker-task recommendation allocation pair set according to a constraint condition of task execution time; a construction method of a worker-task allocation bipartite graph is designed, and a worker-task allocation pair set is obtained by adopting a bipartite graph optimal matching algorithm with the purpose of minimizing the task execution comprehensive cost. According to the method, nonlinear potential allocation rules between workers and tasks are mined, an end-to-end personalized task allocation strategy is obtained, and the reasonability of task allocation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of task allocation based on mobile crowd intelligence perception, and in particular to a personalized task allocation algorithm based on graph neural network. Background Art

[0002] Mobile Crowd Sensing (MCS) uses the intelligent mobile devices carried by a large number of ordinary users, such as smartphones and wearable devices, to complete large-scale and complex perception data collection. It is an emerging paradigm for collaborative knowledge extraction in the modern intelligent society and plays an important role in smart cities, industrial sensing, social networks and other fields.

[0003] Personalized task allocation is one of the business requirements in the process of mobile crowd intelligence perception, and is committed to obtaining a task allocation strategy that can simultaneously meet the economic benefits of the platform and the satisfaction of workers. However, due to the sparsity of task allocation-related data, it is often difficult to capture the matching relationship between workers and tasks, resulting in high task execution costs and low worker satisfaction, which leads to poor rationality of task allocation.

[0004] First, due to the scarcity of historical samples of executors of a single task, data characterizing the personal capabilities of workers is scarce; second, for privacy reasons, workers’ attribute information is less disclosed, and data characterizing their credibility is insufficient; in addition, due to the low frequency of tasks with certain attributes, data characterizing the characteristics of task requirements are also sparse.

[0005] In the face of sparse data, traditional heuristic algorithms have difficulty obtaining the global optimal solution, especially in complex or dynamic task allocation scenarios. For example, LGA assigns workers the tasks with the highest historical completion frequency, and cannot obtain the global optimal result. Matrix decomposition-based algorithms cannot learn effective features when data is sparse. For example, LF uses singular value decomposition of the worker's historical task record matrix, and uses the worker feature matrix vector and task feature matrix operation to obtain the task allocation prediction score. Since it cannot fully learn the personal ability characteristics of the workers, it results in a low task allocation completion rate. In summary, heuristic algorithms and matrix decomposition-based methods are often not applicable to task allocation when the allocation data is sparse, resulting in poor rationality of task allocation. Summary of the invention

[0006] In order to solve the technical problem of poor rationality of task allocation, the present invention proposes a personalized task allocation algorithm based on graph neural network.

[0007] The sparsity of task allocation related data can be compensated by extracting personalized elements of tasks and users from the historical task allocation, the social situation of workers, and the demand for tasks. The difficulty in solving this problem lies in how to organically integrate the above multi-dimensional elements, fully express them as personalized characteristics of workers and tasks, and mine the potential allocation rules between the two. Therefore, it is urgent to propose an end-to-end personalized task allocation algorithm suitable for sparse data.

[0008] As we all know, Graph Neural Network (GCN) is a deep learning model specifically used to process graph structured data. It mainly uses convolution operations on the graph to aggregate neighbor information to the node itself to achieve feature extraction and update, and can easily solve the end-to-end edge prediction problem in the case of sparse data. Based on this, the present invention provides a personalized task allocation algorithm based on graph neural network, which solves the optimal personalized task allocation strategy based on the GCN tool. Specifically, the algorithm includes:

[0009] Step 1: Extract worker capability factors and task requirement factors to construct a worker-task recommendation allocation graph, and propose a task recommendation allocation model based on graph neural network to estimate the possibility of recommendation allocation between tasks to be assigned and available workers. The model learns the personalized characteristics of workers and tasks from task attributes, worker social relationships, and task historical execution status, captures the nonlinear potential allocation rules between the two, and obtains a set of worker-task recommendation allocation pairs.

[0010] Step 2: Based on the result of step 1 and according to the constraint condition on the task execution time, a set of worker-task pre-assignment pairs is selected from the set of worker-task recommended assignment pairs;

[0011] Step three, based on the set of worker-task pre-allocation pairs, design a method for constructing a bipartite graph of worker-task allocation, and use the bipartite graph optimal matching algorithm to obtain the set of worker-task allocation pairs with the goal of minimizing the comprehensive cost of task execution.

[0012] Optionally, the estimated recommended allocation possibilities between the tasks to be allocated and the available workers include:

[0013] Graph neural networks are used to learn the characteristics of workers and tasks using graph convolution and attention mechanisms, and to estimate the possibility of recommended allocation between tasks to be assigned and available workers.

[0014] Optionally, the method of constructing a bipartite graph of worker-task allocation is designed based on the set of worker-task pre-allocation pairs, and the bipartite graph optimal matching algorithm is used to obtain the set of worker-task allocation pairs with the goal of minimizing the comprehensive cost of task execution, including:

[0015] The tasks to be assigned are regarded as the first type of nodes, and the workers available for employment are regarded as the second type of nodes;

[0016] The edge is constructed based on the allocation possibility between the tasks to be assigned and the available workers. If the possibility is established, the edge is generated, and the corresponding estimated comprehensive cost of task execution is used as the edge weight;

[0017] The optimal matching algorithm is used to obtain the set of worker-task assignment pairs.

[0018] The present invention has the following beneficial effects:

[0019] A personalized task allocation algorithm (PTA-GNNTR) based on graph neural network of the present invention models the personalized task allocation problem as a resource allocation problem and solves it using bipartite graph matching. PTA-GNNTR focuses on solving the problem of estimating the possibility of recommended allocation between workers and tasks, and converts it into a graph edge prediction problem. A task recommendation allocation model GNNTR based on graph neural network is designed to extract personalized task allocation features from factors such as task historical execution status, worker social relationships, and task requirements, and uses graph convolution, attention mechanism, multi-layer perceptron and other operations to mine nonlinear potential allocation rules between workers and tasks, and obtain an end-to-end personalized task allocation strategy. This allocation strategy can improve worker satisfaction while minimizing the comprehensive cost of task execution, thereby improving the rationality of task allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 A flowchart of a personalized task allocation algorithm based on a graph neural network of the present invention;

[0022] Figure 2 A more detailed flowchart of a personalized task allocation algorithm based on a graph neural network of the present invention;

[0023] Figure 3 A schematic diagram of MCS personalized task allocation of the present invention;

[0024] Figure 4 A schematic diagram of the decomposition steps of the method for solving the personalized task allocation problem of the present invention;

[0025] Figure 5A schematic diagram of a task recommendation allocation possibility estimation model based on a graph neural network of the present invention;

[0026] Figure 6 It is a task completion rate comparison diagram of the PTA-GNNTR algorithm of the present invention and the baseline algorithm;

[0027] Figure 7 It is a comparison chart of the comprehensive cost of task execution of the PTA-GNNTR algorithm of the present invention and the baseline algorithm. DETAILED DESCRIPTION

[0028] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0029] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0030] Task allocation is an important part of the mobile crowd-sensing process. It refers to the cloud platform achieving the optimal match between workers and tasks based on multi-dimensional task allocation factor information with the goal of minimizing the task execution cost. Compared with traditional task allocation requirements, personalized task allocation pays more attention to the worker's satisfaction with the task allocation strategy, which is an emerging business requirement in the field of perception task allocation. Therefore, personalized task allocation needs to balance the relationship between the two stakeholders, the platform and the workers. Among them, the platform mainly focuses on whether the task can be completed on time, whether the task execution cost can be effectively reduced, and whether the hired workers are trustworthy; while the workers pay more attention to whether the task meets their personal preferences, whether they can get rewards or reputation for executing the task, etc. In summary, in the personalized task allocation mechanism, it is necessary to design different task allocation mechanisms for the needs of the above two stakeholders, and realize the organic integration of the two mechanisms, so as to balance the interests of both.

[0031] Some names related to the present invention are described as follows:

[0032] Mobile Crowd Sensing (MCS): Utilizes the intelligent mobile devices carried by a large number of ordinary users to complete large-scale and complex perception data collection.

[0033] Personalized task allocation: The cloud platform adopts a reasonable task allocation strategy based on multi-dimensional task allocation factor information to minimize the overall cost of task execution while ensuring that workers are satisfied with the task allocation results. Compared with traditional task allocation requirements, personalized task allocation pays more attention to workers' satisfaction with task allocation strategies and is an emerging business requirement in the field of perceptual task allocation.

[0034] Graph neural network: It is a neural network that learns graph structured data. It learns the representation of nodes by transferring and aggregating information between nodes. Typical operation methods include graph convolution and attention mechanism, which can efficiently complete clustering, classification, prediction and other needs on the graph.

[0035] Bipartite graph matching: In a bipartite graph consisting of two types of vertex sets, a subset of edges is obtained. The edges in this subset are not adjacent to each other, that is, any two edges have no common vertices.

[0036] Optimal bipartite graph matching: A bipartite graph matching is a subset of edges whose total weight is either maximal or minimal.

[0037] Satisfaction: It is calculated based on the actual task allocation results and the task allocation results expected by the model. It is one of the indicators to measure the worker's personalization. It is mainly evaluated by the accuracy of the neural network model. The indicators include the following two: MAE and RMSE.

[0038] Task completion rate: The ratio between the number of tasks actually completed and the number of tasks to be assigned, which is one of the indicators of platform benefits.

[0039] Comprehensive cost of task execution: It is the sum of basic cost and incentive cost during task execution and is one of the indicators of platform benefits.

[0040] refer to Figure 1, showing the process of some embodiments of a personalized task allocation algorithm based on graph neural network (PTA-GNNTR, Personalized Task Allocation-Graph Neural Network Task Recommendation) according to the present invention. The personalized task allocation algorithm based on graph neural network models the personalized task allocation problem as a resource allocation problem and solves it using bipartite graph matching. The problem solving focuses on the estimation problem of the possibility of recommended allocation between workers and tasks, and converts it into a graph edge prediction problem. A task recommendation allocation model GNNTR based on graph neural network is designed. Personalized task allocation features are extracted from factors such as task historical execution status, worker social relationships, and task requirements. Graph convolution, attention mechanism, multi-layer perceptron and other operations are used to mine nonlinear potential allocation rules between workers and tasks, and an end-to-end personalized task allocation strategy is obtained, which specifically includes the following steps:

[0041] Step 1: Extract worker capability factors and task requirement factors to construct a worker-task recommendation allocation graph, and propose a task recommendation allocation model based on graph neural network to estimate the possibility of recommended allocation between tasks to be assigned and available workers. The model learns the personalized characteristics of workers and tasks from task attributes, worker social relationships, and historical task execution status, captures the nonlinear potential allocation rules between the two, and obtains a set of worker-task recommendation allocation pairs.

[0042] In some embodiments, a graph neural network can be used to learn the characteristics of workers and tasks using operations such as convolution and attention mechanisms to estimate the possibility of recommended allocation between tasks to be assigned and available workers.

[0043] Step 2: Based on the result of step 1 and in accordance with the constraints on task execution time, a set of worker-task pre-assignment pairs is selected from the set of worker-task recommended assignment pairs.

[0044] In some embodiments, a set of pre-assigned worker-task pairs may be selected from a set of recommended worker-task assignment pairs according to constraint conditions.

[0045] Step three, based on the set of worker-task pre-allocation pairs, design a method for constructing a bipartite graph of worker-task allocation, and use the bipartite graph optimal matching algorithm to obtain the set of worker-task allocation pairs with the goal of minimizing the comprehensive cost of task execution.

[0046] In some embodiments, a method for constructing a bipartite graph of worker-task assignments can be designed based on the results of the screening step, with the goal of minimizing the overall cost of task execution, and a bipartite graph optimal matching algorithm can be used to obtain a set of worker-task assignment pairs, that is, an optimal task assignment strategy.

[0047] It should be noted that in step 3, the tasks to be assigned can be regarded as the first type of nodes, and the available workers can be regarded as the second type of nodes. Then, the edges are constructed based on the allocation possibilities between the tasks to be assigned and the available workers. If the possibility is established, the edges are generated, and the corresponding estimated task execution comprehensive cost is used as the edge weight. The optimal matching algorithm is used to obtain the final allocation result.

[0048] As an example, this step may include the following steps:

[0049] In the first step, tasks to be assigned are regarded as the first type of nodes, and workers that can be hired are regarded as the second type of nodes.

[0050] In the second step, edges are constructed based on the allocation possibilities between the tasks to be assigned and the available workers. If the possibility is established, an edge is generated, and the corresponding estimated comprehensive cost of task execution is used as the edge weight.

[0051] In the third step, the optimal matching algorithm is used to obtain the set of worker-task assignment pairs.

[0052] refer to Figure 2 , shows a more detailed process of a personalized task allocation algorithm based on a graph neural network according to the present invention, which may include the following steps:

[0053] Step 201, problem description.

[0054] Figure 3 This is a schematic diagram of the personalized task allocation problem in the crowd sensing scenario. As shown in the figure, in the urban sensing service area, there are a large number of candidate workers waiting to be assigned to complete several sensing tasks. Assume that all sensing tasks are simple tasks and are independent and indivisible. For example, noise detection, temperature perception, light intensity perception, road condition information shooting, etc. All workers have the ability to independently complete a single sensing task. For example, workers can directly complete the task using the mobile sensing device they carry with them. Assume that in each task allocation decision period, the platform can only assign one sensing task to a single worker. Then, the personalized task allocation problem can be described as follows: The platform formulates the optimal task allocation strategy, that is, the optimal allocation matching scheme between workers and tasks, based on various information related to task allocation, including the worker's ability status, the specific attributes of the task, and the historical execution of the task in the past, by adopting an efficient task allocation method. This scheme minimizes the comprehensive cost of task execution while ensuring that the worker is satisfied with the task allocation result.

[0055] Step 202: problem modeling.

[0056] The first step is task model.

[0057] The main attribute elements of a task include basic execution cost, location, deadline, etc. Let v j =(id j ,w j ,l j ,t j ) represents task v j , where id j It is task v j The unique identifier of j It is task v j The basic execution cost of j It is task v j The location of the task, usually refers to the geographical location that the worker needs to reach when performing the task; j It is task v j The deadline for execution is the latest completion time of the task.

[0058] The second step is the worker model.

[0059] The main attribute elements of workers include historical task execution and social relationship. i =(H i ,F i ,l i ,s i ) represents worker u i , where H i Is a worker i The historical mission execution status, F i Is a worker i Social relations, i Is a worker i The position of i Is a worker i moving speed.

[0060] The historical task execution status. This status is recorded in several historical task execution records. Each record includes the unique identifier of the executed record, historical execution time, historical remuneration and other information. Represents worker u i Historical task execution records, where Is a worker i The number of times the task has been executed in history; Is a worker i The vth historical task execution record, where id iv Is a worker iThe code of the vth task executed historically, Is a worker i The time for carrying out the historical tasks in Article V, Is a worker i The actual remuneration obtained for performing the historical tasks in Article V.

[0061] Social relationship status. This status reflects the social relationships of the worker, including the following two types: one is virtual social relationship. This part records the worker's interaction activities with other workers on the social platform, reflecting the worker's interpersonal relationships in the virtual world. The other is physical social relationship. This part records the worker's encounters with other workers during the execution of historical tasks, reflecting the worker's interpersonal relationships in the physical world. Figure 3 As shown, worker u1 and worker u2 are virtual friends of each other; worker u1 and worker u3 are physical friends of each other.

[0062] make Represents worker u i social relationships, among which Is a worker i Virtual social relationships, Is a worker i Virtual Friend U k , Is a worker i Total number of virtual friends; Is a worker i physical social relationships, Is a worker i Physics Friend U k , Is a worker i The total number of physical friends.

[0063] The third step is cost model.

[0064] The comprehensive cost of task execution is the remuneration required to assign a specific task to a specific worker. It includes the following two parts: the first is the basic execution cost, which is an inherent attribute element of the task, only related to the task itself, and is a fixed value; the second is the incentive cost, which is the additional remuneration for workers to complete difficult tasks. Its value is related to the specific task allocation plan and is a variable value. For example, the incentive cost may be related to factors such as the distance or time required for the worker to move to perform the task.

[0065] make Indicates that the task v j Assigned to worker u i The overall cost of execution, including It is task v j The basic execution cost, It is by worker u i Complete the task j The incentive cost is , and α is the cost adjustment coefficient.

[0066] Step 203, problem model.

[0067] Let U={u i |1≤i≤n} represents the set of all workers, where n is the total number of workers. Assuming that all social friends of a worker belong to the worker set, we have Let V = {v j |1≤j≤m} represents the set of all perception tasks, where

[0068] m is the total number of all tasks. Then V = V old ∪V new , where V new represents the current set of tasks to be assigned, V old Represents the set of tasks that have completed assignments historically.

[0069] In summary, the personalized task assignment problem P can be expressed as:

[0070]

[0071] in, Indicates that the task Assigned to worker u i The overall cost of execution, Indicates whether the task Assigned to worker u i , if assigned, then On the contrary, C1 to C5 are the constraints of problem P. Among them, C1 is the constraint on the indivisibility of tasks, that is, each task must be assigned as a whole; C2 is the constraint on the number of task assignments, that is, each task can only be assigned once; C3 is the constraint on the number of tasks assigned to workers, that is, a single worker can only be assigned one task at a time; C4 is the constraint on the number of tasks, where m new =|V new | is the total number of tasks to be assigned, that is, all tasks to be assigned must be assigned; C5 is the constraint on the task execution time, where, Is a worker i Arrival Mission The time required for the location, Represents worker u i With the task The distance between them, f() is the distance calculation function, s i Is a worker i The moving speed, It's a task The deadline for execution is the worker must arrive at the task location before the deadline for execution.

[0072] Step 204, solution step.

[0073] As shown in formula (1), problem P is a typical constrained resource allocation problem. This problem is to obtain the best solution to allocate limited resources (i.e., available workers) to the tasks to be executed, with the goal of optimizing the overall task execution comprehensive cost under the constraints of C1-C5. As we all know, bipartite graph matching is a classic solution to this type of problem. Therefore, the overall idea of ​​solving problem P in this paper is to construct a bipartite graph for worker-task recommendation allocation and use the bipartite graph matching method to solve the optimal task allocation strategy. In this solution process, constructing a bipartite graph that can accurately depict the essence of the demand for personalized task allocation problem is the key to solving problem P.

[0074] The construction method of the bipartite graph can be briefly described as follows: First, the tasks to be assigned are taken as the first type of nodes, and the available workers are taken as the second type of nodes. Then, the edges are constructed based on the allocation possibilities between the tasks to be assigned and the available workers. If the possibility is established, the edge is generated, and the corresponding estimated task execution comprehensive cost is used as the edge weight. Based on this, designing an efficient edge generation method, that is, accurately estimating the recommended allocation possibilities between the tasks to be assigned and the available workers, is a key step in constructing the bipartite graph.

[0075] Figure 4 It is a schematic diagram of the steps of the method for solving problem P. As shown in the figure, the method includes the following two sub-steps and a screening step.

[0076] The method of step one is briefly described as follows: extract worker capability elements and task requirement elements to construct a worker-task recommendation allocation graph, and propose a task recommendation allocation model based on graph neural network to estimate the possibility of recommendation allocation between tasks to be assigned and available workers. The model learns the personalized characteristics of workers and tasks from task attributes, worker social relationships, and historical task execution, captures the nonlinear potential allocation rules between the two, and obtains a set of worker-task recommendation allocation pairs.

[0077] Screening step: Based on the result of step 1, according to the constraint condition C5 on the task execution time in formula (1), a set of worker-task pre-assignment pairs is screened out from the set of worker-task recommended assignment pairs.

[0078] The method of step two is briefly described as follows: Based on the results of the screening step, a method for constructing a bipartite graph of worker-task assignment is designed. With the goal of minimizing the comprehensive cost of task execution, a bipartite graph optimal matching algorithm is used to obtain a set of worker-task assignment pairs, that is, the optimal task assignment strategy.

[0079] In summary, the first step is the stage of recommending and allocating tasks based on the needs of workers, focusing on whether the ability and credibility of workers meet the execution requirements of tasks, and is committed to recommending tasks with high satisfaction for workers to allocate; the second step is the stage of recommending and allocating tasks based on the interests of the platform, focusing on how to reduce the comprehensive cost of task execution, and striving to maximize the benefits of the platform. Therefore, the above method is used to obtain a personalized task allocation strategy that takes into account both the needs of workers and the interests of the platform.

[0080] Step 205: Task recommendation allocation model based on graph neural network.

[0081] The first step is the task recommendation model process.

[0082] Figure 5 This is a flowchart of the task recommendation allocation model based on graph neural network. In the figure, the circular nodes represent workers, the square nodes represent tasks, and the circular nodes filled with horizontal and vertical lines represent the virtual social friends and physical social friends of the workers, respectively. The specific process of the model is as follows: First, the worker-historical task allocation graph, the worker virtual social relationship graph and the physical social relationship graph, as well as the task attribute information are extracted from the worker-task recommendation allocation graph, and the node feature embedding method is used to obtain the initialization features of the worker and task nodes; then, the convolution method is used to learn the personalized feature representation of workers and tasks, and to explore the potential rules of the relationship between the two; finally, the updated personalized features of the two are sent to the multi-layer perceptron to estimate the recommendation allocation relationship between the two.

[0083] The second step is to analyze the task allocation problem.

[0084] Let G A =(U∪V,E A ) represents the worker-task recommendation allocation graph, where U is the set of all workers; V is the set of all tasks; E A represents the combination of the allocation relationship between workers and tasks, and the edge weight is the comprehensive cost corresponding to the allocation relationship. Represents the same worker u i The relevant edges are Among them, u i ∈U,id iv =h iv [1] Here, h iv [k] means take The kth element in the triple. Then, Accordingly, if the worker u i and task v j If there is a historical distribution relationship between Allocate the comprehensive cost to Then, worker u i and tasks to be assigned The recommended allocation possibility between can be expressed as The estimated value of the allocation comprehensive cost is If the worker u i and tasks If there is a possibility of recommended allocation between them, then a worker-task recommended allocation pair is generated In summary, the task recommendation allocation possibility estimation problem is transformed into an edge prediction problem.

[0085] The third step is worker feature extraction.

[0086] The worker's personalized elements include personal ability elements and credibility elements. Therefore, the worker u i The initial comprehensive features can be expressed as

[0087]

[0088] Among them, f i I Represents worker u i Personal ability characteristics, f i N Represents worker u i credibility characteristics.

[0089] The first sub-step is personal ability characteristics.

[0090] The personal ability characteristics of workers can be extracted and learned from the workers' historical task execution records. The specific method is as follows: First, extract the workers' historical task allocation information from the worker-task recommendation allocation graph, and construct the worker-task historical task allocation graph. in, is the edge corresponding to the historical task allocation, which exists between the worker and the historically allocated task. Next, in this graph, with a single worker as the center, attention convolution is used to aggregate its historical task allocation information to obtain the worker's personal ability characteristics. The formula is as follows:

[0091]

[0092] Among them, q j Represents task v j The vector embedding of Represents worker ui Complete the task j The comprehensive cost The vector embedding of , g represents a multi-layer perceptron, Represents task v j For workers i The influence weight of personal ability representation, Is a worker i The set of all historical assigned tasks, p i Represents worker u i The vector embedding of represents a concatenation operation, W, W1, and W2 represent weights, and b, b1, and b2 represent bias terms.

[0093] The second sub-step is the credibility feature representation.

[0094] The credibility characteristics of workers can be extracted and learned from social relationships. Since social relationships include virtual and physical relationships, the credibility characteristics of workers are:

[0095]

[0096] Among them, f i S Represents worker u i Virtual social credibility characteristics, f i G Represents worker u i Physical social credibility characteristics.

[0097] The virtual social credibility features of workers can be extracted and learned from the virtual social relationship graph, while the physical social credibility features can be extracted and learned from the physical social relationship graph. Since the acquisition methods of the two are similar, only the acquisition method of the virtual social credibility features is introduced in detail here. The specific method is as follows: First, the virtual social information of the worker is extracted from the worker's social relationship and a virtual social relationship graph is constructed. Represents worker u i A virtual social relationship graph, where U is the set of all workers; are the edges between all workers and their virtual friends, i.e. where u i ,u j ∈U,u j ∈F i S Is a worker i Virtual Friend U j Then, in the virtual social relationship graph, with a single worker as the center, attention convolution is used to aggregate the information of all its virtual social friends to obtain the worker's virtual social credibility characteristics. The formula is as follows:

[0098]

[0099] Among them, p j Indicates virtual friend u j The vector embedding of It's a virtual friend j For workers i The influence weight of virtual social credibility representation, p i Represents worker u i The vector embedding representation of N S (i) is a worker u i A collection of virtual friends.

[0100] The fourth step is task feature extraction.

[0101] The attribute elements of a task include demand attribute elements and execution attribute elements. Among them, the basic attribute elements reflect the core elements such as the type, complexity and resource requirements of the task, while the execution attribute elements reveal key information such as its execution difficulty and completion time. j The initial features can be expressed as

[0102]

[0103] in, Represents task v j The execution attribute characteristics of Represents task v j demand attribute characteristics.

[0104] The required attribute features of tasks can be directly extracted from the worker-task recommendation assignment graph. By aggregating the basic element embeddings of tasks, we can obtain the task v j Demand attribute characteristics

[0105]

[0106] Where θ = 1 / |M j |, M j It is task v j The number of attributes, Represents task v j The vector embedding of the demand attributes, N F (j) is task v j The set of required attributes.

[0107] The execution attribute characteristics of a task can be extracted from its historical task execution. The specific extraction method is as follows: Based on the worker-historical task allocation graph G R, centered on a single task, uses attention convolution to aggregate the information of task execution workers to obtain the execution attribute characteristics of the task. j The execution attribute characteristics of can be expressed as:

[0108]

[0109] in, Represents worker u k Complete historical mission v j Total compensation The vector embedding of k Is the task execution worker u k The embedding representation of MLP represents multi-layer perceptron, N U (j) has performed task v j A collection of workers, Represents worker u k For task v j The influence weight of W, W1, and W2 represent weights, and b, b1, and b2 represent bias terms.

[0110] Step 5, GNNTR model.

[0111] Algorithm 1 is the pseudo code of the GNNTR algorithm, and its specific process is shown in Table 1. The input is the worker-task recommendation allocation graph G A The output is the estimated possible allocation result between workers and tasks, that is, the set of worker-task recommended allocation pairs C R Among them, the first line is to obtain the worker-task historical task allocation graph G R , Virtual social relationship graph G T and the physical social relationship graph G P ; Lines 2 to 5 are to obtain the initial features F of the worker i (0) ; Lines 6 to 9 are to obtain the initial features of the task Lines 10 to 13 are the operations of the convolutional graph neural network. After multiple rounds of learning, the worker u is obtained. i Comprehensive features:

[0112] F i =σ(W l ·σ(W l-1 ···σ(W1·F i (0) +b1)···+b l-1 )+b l ) (9)

[0113] And task v jComprehensive characteristics of

[0114]

[0115] Line 14 uses concatenation, convolution and other operations to obtain the probability matrix for recommending specific tasks for all workers.

[0116]

[0117] Among them, F U The worker feature matrix is ​​obtained by concatenating the comprehensive features of all workers, F V It is to concatenate all the comprehensive features of the tasks to obtain the task feature matrix. Represents the dot product operation between matrices, W1 T 、W2 T is the transpose of the weight matrix, b1 and b2 are bias terms. Lines 15 to 18 are the obtained worker-task recommendation pairs, where line 16 is based on the probability matrix, which is the worker u i Get recommended tasks Line 17 calculates worker u i Execute the task The comprehensive cost Line 18 forms a worker-task recommendation pair And add to set C R Based on this, the GNNTR algorithm obtains a set of worker-task recommendation pairs by learning the implicit allocation rules between workers and tasks.

[0118] Table 1

[0119]

[0120]

[0121] Step 206: Personalized task allocation algorithm.

[0122] The problem of obtaining the optimal task allocation strategy between tasks and workers can be transformed into an optimal matching problem on a bipartite graph. This section will describe the task allocation algorithm based on bipartite graph matching, including the construction method of the worker-task allocation bipartite graph and the solution method of the optimal task allocation strategy based on the bipartite graph.

[0123] The first step is to analyze the task allocation problem.

[0124] Based on the worker-task recommended assignment pair set, a worker-task pre-assignment pair set is screened out. The specific method is as follows: Using the constraint condition C5 in formula (1), from C RRemove the assignment pairs whose arrival time is greater than the deadline required by the task, and form a set of recommended worker-task assignment pairs.

[0125] Let G=(U new ,V new ,E) represents the bipartite graph of worker-task assignment, where U new = {u i |u i ∈C' R [1]} is the set of workers that can be hired in the current decision period, as the first type of node, where C' R [i] represents the set of the i-th triplet element in the set of worker-task recommendation assignment pairs; V new = {v j |v j ∈C' R [2]} is the set of tasks to be assigned in the current decision period, which is the second type of node; represents the edge set, Indicates that the task Assigned to worker u i implement, represents the total execution cost. Then the optimal task allocation strategy for problem P is the optimal matching M on graph G. opt . Satisfy the conditions: in, for u i ≠u k And v j ≠v l Finally, for each Generate triples Put it into the worker-task assignment pair set C, that is,

[0126] In summary, the personalized task assignment problem is transformed into a bipartite graph optimal matching problem.

[0127] The second step is PTA-GNNTR algorithm.

[0128] Algorithm 2 is the pseudo code of the PTA-GNNTR algorithm. The specific process is shown in Table 2. The input is the set of all workers and tasks, and the output is the optimal task allocation strategy. The first line is to build the worker-task recommendation allocation graph G according to the description. A The second line calls the GNNTR model to obtain the worker-task recommendation pair set C R ; Line 3 is from C R Filter out the worker-task pre-allocation set C' R ; Line 4 is based on C'R Construct a bipartite graph G for worker-task assignment; line 5 calls the bipartite graph matching algorithm to obtain the optimal match M for graph G. opt ; Line 6 is by M opt Get worker-task assignment pairs and put them into set C.

[0129] Table 2

[0130]

[0131]

[0132] Step 207, experimental process.

[0133] The first step is experimental verification.

[0134] Verify the feasibility and effectiveness of the PTA-GNNTR algorithm, including experiments on analyzing factors affecting task allocation and comparing algorithm performance.

[0135] The second step is to set up the experimental environment.

[0136] The first sub-step is the data set. Specifically, the statistical information of the data set after processing is shown in Table 3.

[0137] Table 3

[0138]

[0139] All experiments are conducted on the following two open real-world datasets: Gowalla and Brightkite. Gowalla is from a social networking site, while Brightkite is from a social service provider that collects user check-in information. This paper extracts data related to the New York area from the above two datasets and uses a ten-core setting to screen out sufficient employable workers.

[0140] The experimental data is obtained based on the user check-in records and social records in the above two data sets. The specific data includes user ID, check-in time, check-in location and other information in the user check-in record, and user social friends and other information in the social record. The specific data processing method is as follows: the user is regarded as a worker, the check-in location is regarded as a perception task, and the user's check-in behavior is regarded as a record of the execution task. The user's friends on the social platform are regarded as their virtual social friends, and the number of social friends of each user is in the range of [10,175]. The workers who keep a certain distance from the user during a certain encounter period are regarded as physical social friends, and the encounter period T≤10min; the distance is r∈(5,10)m, and the Vincenty algorithm is used [5] Calculate the basic execution cost w of task j j∈(0,5), its value is proportional to the number of times the task has been completed in history. The reason is that the more times a task has been completed in history, the more attractive the reward for the task is. The incentive cost of task j is Its value is proportional to the distance between the worker and the task.

[0141] Table 3 shows the statistical information of the dataset after processing. The statistical information includes the number of workers, the number of tasks, the historical task execution records of workers, virtual relationships and physical relationships, etc. As shown in the table, compared with the Brightkite dataset, the Gowalla dataset has fewer average check-ins and is therefore more sparse.

[0142] This paper uses the following three baseline algorithms as performance comparison algorithms for GNNTR: random assignment algorithm (RAA), local greedy algorithm (LGA) and matrix decomposition algorithm (LF). Among them, RAA is a random assignment algorithm from a worker that randomly selects a task from the worker's historical execution and assigns it to the worker; LGA is to assign the worker the task with the highest historical completion frequency; LF is to perform singular value decomposition on the worker's historical task record matrix, multiply the worker's feature matrix vector by the transpose of the task feature matrix to obtain the prediction score, and then assign the next task based on the highest score.

[0143] The third step is experimental parameters.

[0144] The important parameters of the GNNTR model are set as follows: The objective function of the loss value Where |O| is the number of tasks, c ij is the actual task allocation, c' ij is the predicted task allocation. The ratio of training set to test set is 7:3. The vector feature dimension d = 128, the learning rate is 0.01, and the normalization coefficient is 10 -5 The hidden layer uses the ReLU activation function, the number of layers is two, the optimizer is RMSprop, and the dropout strategy is used to alleviate the overfitting problem.

[0145] Step 4: Performance indicators.

[0146] The experiment uses three indicators to evaluate the task allocation strategy. Among them, satisfaction is mainly used to evaluate the performance of the GNNTR model, which is measured by the accuracy of the neural network model, and mainly reflects the impact of the task allocation results on the interests of workers; task completion rate and task execution comprehensive cost are used to evaluate the performance of the PTA-GNNTR algorithm, which mainly reflects the impact of task allocation results on the interests of the platform.

[0147] The first sub-step is satisfaction.

[0148] Mean absolute error (MAE):

[0149] Root mean square error RMSE:

[0150] Among them, c ij is the predicted allocation result, that is, it is recommended to assign task v j Assigned to worker u i , c i ' k ∈C' is the actual allocation result, that is, task v k Assigned to worker u i .

[0151] The above two indicators mainly measure the workers' satisfaction with the allocation results. The smaller the value, the higher the workers' satisfaction and the more the task allocation strategy meets the workers' personalized needs.

[0152] The second sub-step is the task completion rate.

[0153]

[0154] Among them, N c is the number of tasks assigned, N t is the total number of tasks in the perceived service area at the decision moment. The higher the task completion rate, the better the task allocation strategy.

[0155] The third sub-step is the comprehensive cost of task execution.

[0156]

[0157] Among them, W all is the comprehensive cost of task execution, and the symbols in the formula are explained in formula (1). The smaller the index is, the better the task allocation strategy is.

[0158] Step 5, Experiment 1.

[0159] This section verifies the performance of the GNNTR model, including experiments on worker satisfaction, task completion rate, and comprehensive cost of task execution.

[0160] Table 4 shows the satisfaction results of the GNNTR model and the baseline algorithm. As shown in the table, the satisfaction of the LF, RAA, LGA and GNNTR models increases in turn. This is because LF requires sufficient historical data to accurately obtain the potential relationship between workers and tasks, so it is the worst when the data is sparse; LGA lacks a global perspective and can only obtain local optimality in each step of selection, so it performs best in the baseline algorithm; RAA randomly selects tasks from the worker's historical execution tasks, and to a certain extent considers the worker's personal preferences, so the allocation result is better than the matrix decomposition algorithm, but it cannot obtain choices other than historical tasks, resulting in limitations in the allocation result. The GNNTR model uses a graph neural network method, which can effectively process sparse task allocation data and accurately capture the potential relationship between workers and tasks.

[0161] Table 4

[0162]

[0163]

[0164] Step 6, Experiment 2.

[0165] This section verifies the performance of the PTA-GNNTR algorithm, including experiments on task completion rate and comprehensive cost of task execution.

[0166] The first sub-step is task completion rate.

[0167] Figure 6 It is a schematic diagram of the task completion rate of the PTA-GNNTR algorithm and the baseline algorithm as the task deadline changes. As shown in the figure, the task completion rate of PTA-GNNTR increases as the task deadline increases, and it is always higher than other baseline methods. The task completion rates of other algorithms decrease in the following order: LGA, LF, RAA. On the Gowalla dataset, the task completion rate of LF is lower than that of RAA, while on the Brightkite dataset, the task completion rate of LF is higher than that of RAA. This is because the Gowalla dataset contains more sparse data than the Brightkite dataset, so the effect of matrix decomposition is worse.

[0168] The second sub-step is the comprehensive cost of task execution.

[0169] Figure 7This is a comparison of the comprehensive cost of task execution of the PTA-GNNTR algorithm and the baseline algorithm as the number of tasks changes. As shown in the figure, the comprehensive cost of tasks of PTA-GNNTR increases as the number of tasks increases, and it is always lower than other baseline algorithms. The comprehensive cost of tasks of other algorithms increases in the following order: LGA, LF, RAA. This is because the LGA algorithm pursues local optimality, the LF algorithm does not adequately explore the complex allocation relationship between workers and tasks, and the RAA algorithm randomly allocates tasks, all of which lead to unreasonable task allocation and generate additional incentive costs.

[0170] The key improvements of the present invention include:

[0171] The present invention proposes a personalized task allocation algorithm based on graph neural network (Personalized Task Allocation-Graph Neural Network Task Recommendation, PTA-GNNTR) to obtain the optimal task allocation strategy, while minimizing the comprehensive cost of task execution, improving the worker's satisfaction. The two key tasks are as follows: First, the task allocation problem is modeled as a resource allocation problem with constraints, a bipartite graph of worker-task allocation is constructed, and the optimal task allocation strategy is solved by the bipartite graph optimal matching method, and the platform interests are fully considered in the task allocation process. Second, the problem of estimating the possibility of recommended allocation between workers and tasks is converted into a graph edge prediction problem, and a task recommendation allocation model based on graph neural network (Graph Neural Network Task Recommendation, GNNTR) is designed to fully consider the personalized needs of users in the task allocation process. The working principle of GNNTR is to construct a worker-task recommendation allocation graph based on factors such as the historical allocation of tasks, the social situation of workers, and the demand for tasks, and adopt operations such as graph convolution, multi-layer perceptron, and graph attention mechanism to learn the characteristics of workers and tasks and explore the potential allocation relationship between the two.

[0172] The beneficial effects of the present invention include:

[0173] First, compared with baseline algorithms such as matrix decomposition algorithm, random algorithm, and greedy algorithm, the GNNTR model has the highest worker satisfaction.

[0174] Second, the task completion rate of the PTA-GNNTR algorithm is higher than that of other baseline methods.

[0175] Third, the comprehensive cost of task execution of the PTA-GNNTR algorithm is lower than that of other baseline algorithms.

[0176] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features can be replaced by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A personalized task allocation algorithm based on graph neural network, characterized in that: The following steps are involved: Step 1: Extract worker capability factors and task requirement factors to construct a worker-task recommendation allocation graph, and propose a task recommendation allocation model based on graph neural network to estimate the possibility of recommendation allocation between tasks to be assigned and available workers. The model learns the personalized characteristics of workers and tasks from task attributes, worker social relationships, and task historical execution status, captures the nonlinear potential allocation rules between the two, and obtains a set of worker-task recommendation allocation pairs. Step 2: Based on the result of step 1 and according to the constraint condition on the task execution time, a set of worker-task pre-assignment pairs is selected from the set of worker-task recommended assignment pairs; Step three, based on the set of worker-task pre-allocation pairs, design a method for constructing a bipartite graph of worker-task allocation, and use the bipartite graph optimal matching algorithm to obtain the set of worker-task allocation pairs with the goal of minimizing the comprehensive cost of task execution.

2. According to claim 1, a personalized task allocation algorithm based on graph neural network is characterized in that: The recommended allocation possibilities between the estimated tasks to be allocated and the available workers include: Graph neural networks are used to learn the characteristics of workers and tasks using convolution and attention mechanisms, and to estimate the possibility of recommended allocation between tasks to be assigned and available workers.

3. According to claim 1, a personalized task allocation algorithm based on graph neural network is characterized in that: The method for constructing a bipartite graph of worker-task allocation is designed based on a set of worker-task pre-allocation pairs, and a bipartite graph optimal matching algorithm is used to obtain a set of worker-task allocation pairs with the goal of minimizing the comprehensive cost of task execution, including: The tasks to be assigned are regarded as the first type of nodes, and the workers available for employment are regarded as the second type of nodes; The edge is constructed based on the allocation possibility between the tasks to be assigned and the available workers. If the possibility is established, the edge is generated, and the corresponding estimated comprehensive cost of task execution is used as the edge weight; The optimal matching algorithm is used to obtain the set of worker-task assignment pairs.

Images