A multi-person cooperation task allocation method for crowd sensing

Abstract

The present application relates to a kind of multi-person cooperation task allocation method for crowd sensing, belong to mobile crowd sensing field.It is related to the task allocation method based on user social network and graph neural network, the potential cooperation between crowd sensing users and the potential relationship between user and task are mined, and the quality of task executed by crowd sensing user cooperation is improved.The method of the present application includes two stages: first, to obtain cooperation preference and ability, the embedding of group and task is learned respectively;Second, search the optimal group for each task, maximize overall utility.Through the above method, the potential information of interaction between task and user is effectively mined, which provides support for subsequent task allocation and improves the execution quality of allocated task.

Description

Technical Field

[0001] This invention belongs to the field of mobile crowd sensing and relates to a multi-person collaborative task allocation method for crowd sensing, specifically a task allocation method based on user social networks and graph neural networks. Background Technology

[0002] With the widespread adoption of smart mobile devices (such as smartphones, tablets, smart bracelets, and smartwatches), mobile crowd sensing has received increasing attention in recent years. Many mobile crowd sensing platforms have been proposed and applied in scientific research, industrial production, and daily life. These platforms recruit mobile workers to perform various tasks, such as environmental sensing, traffic information collection, and indoor positioning. However, workers often need to spend a considerable amount of time and energy selecting tasks they are capable of performing and that they prefer. Therefore, task allocation is a key research problem in mobile crowd sensing, which automatically assigns suitable tasks to workers through task allocation algorithms. In recent years, there has been much research on task allocation in mobile crowd sensing. Based on the dependencies between tasks, task allocation work can be divided into single-task allocation and multi-task allocation. We focus on multi-task allocation, where different tasks are interconnected because they share a limited resource pool (e.g., a shared work pool or total budget). Existing multi-task allocation work can be categorized into three types based on the number of workers required for each task and whether the workers collaborate: independent multi-task allocation, non-collaborative multi-task allocation, and multi-person collaborative multi-task allocation. 1) In independent multi-task allocation, each task only needs to hire one worker. These jobs solve the independent multi-task allocation problem with different objectives and constraints. 2) In non-cooperative multi-task allocation, each task needs to hire multiple workers, who complete the task independently in a non-cooperative manner. To ensure successful task execution, a task is redundantly assigned to multiple swarm intelligence users, and the platform ultimately integrates the task execution results from multiple users. 3) In multi-person collaborative multi-task allocation, due to the complexity of the tasks, each task needs to hire multiple workers to collaboratively execute the task. However, these three jobs employ a centralized approach (first selecting a leader, then selecting the remaining workers accordingly) to form a swarm, limiting the search space of possible swarms. Summary of the Invention

[0003] Technical problems to be solved

[0004] To avoid the shortcomings of existing technologies, this invention provides a method for allocating collaborative tasks among multiple users based on crowd intelligence perception.

[0005] Technical solution

[0006] A method for allocating collaborative tasks among multiple users based on crowd intelligence perception, characterized by the following steps:

[0007] Step 1: User team statement

[0008] A user team refers to a group of multiple collectively intelligent sensing users who collaborate to perform a specific task. The user team aims to learn the latent factors of users and group them into a user group (G). i The learned latent factors are represented as h i ∈R d ;

[0009] Step 2: Task Representation

[0010] The purpose of task representation is to learn task z i The potential factors in are denoted as z. i ∈R d The user task bipartite diagram includes the interaction between the team and the task, as well as the team's task execution quality; therefore, information on the interaction between the team and the task, as well as information on task execution quality, should be captured together to learn the underlying factors of the task.

[0011] Step 3: Task Execution Quality Prediction

[0012] The potential representation of the team and the task h j and z i Connect them together, and then feed them into a three-layer neural network to make the following predictions:

[0013]

[0014] Step 4: Combinatorial optimization solution

[0015] Using the predicted task execution quality as a constant, the task allocation problem of maximizing the total task execution quality can be solved by finding the optimal team combination. First, the task execution quality of each task is sorted according to the candidate teams, and then the team with the highest task execution quality score is selected as the candidate team for each task.

[0016] A further technical solution of the present invention: The user team representation in step 1 includes the following three parts:

[0017] Part 1: Aggregating information from team members themselves. Since each user in the team has different confidence levels, professional knowledge, and social relationships, the contribution of team members to task execution is different. Therefore, we assign different weights to different users to accurately represent their contribution to the team.

[0018] Part Two: Team and Task Interaction Information. Since the user task bipartite graph includes not only team and task interactions but also the team's task execution quality, to extract information reflecting team interests and capabilities, we consider aggregating information with the target team G. i Interactive task information and team task execution quality information;

[0019] Part Three: Information Interaction Between Teams; Aggregating Team G in the Task-User Bipartite Diagram j Information on second-order neighbor teams, and using C 2-hop (j) is represented as G j A set of second-order neighbors; members of a neighbor group are typically located in the same neighborhood as G. j Interactive tasks are located near each other; intuitively, neighboring team members and the target team will have similar task choices; therefore, learning C... 2-hop Information about the team in (j).

[0020] A further technical solution of the present invention: The task representation in step 2 is as follows:

[0021] For each task t i It is necessary to start from the task t i Information is aggregated in the set B(i) of interactive teams, using f il To describe the target task t i and team G in set B(i) l The interaction between them, and by bringing the team G l The representation of g l And the team carried out the task t i The vector representation of the mass e′ p By cascading and inputting into a three-layer neural network, we obtain:

[0022]

[0023] In order to learn z i It also proposes the team and target task t in B(i). i Interactions between them; team aggregation functions are described as Aggre groups It is similar to the aggregation task function Aggre. tasks , z i Represented as:

[0024]

[0025] An attention mechanism is introduced to assign different weights u to different teams in set B(i). il The input to the attention network is the team interaction embedding f il and target task t iThe embedding is represented as q i ,

[0026]

[0027]

[0028]

[0029] A computer system is characterized by comprising: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method described above.

[0030] A computer-readable storage medium is characterized by storing computer-executable instructions, which, when executed, are used to implement the above-described method.

[0031] Beneficial effects

[0032] This invention provides a method for allocating collaborative tasks among users based on crowd-aware perception. By mining the potential cooperative relationships among users and the potential relationships between users and tasks through users' social networks and a bipartite graph of users and tasks, suitable users are assigned to newly emerging tasks, thereby improving the quality of collaborative task execution by users based on crowd-aware perception.

[0033] Extensive experiments were conducted on two real-world datasets. Results show that the method implemented in this invention achieves an average task performance improvement of 155.26% compared to four baseline methods. Attached Figure Description

[0034] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0035] Figure 1 This is an overall framework diagram of task allocation in an example of the present invention;

[0036] Figure 2 The study demonstrates the effectiveness of a crowd-aware collaborative task allocation method and a baseline method in improving task execution quality. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0038] In multi-person collaborative task allocation, some tasks require hiring a group of workers to achieve 1) multi-person collaboration, where workers in the team each complete a different part of a task, collaborating to complete the task; and 2) personalized team size, where a fixed number of workers are recruited based on the task's requirements. Based on this, the goal of this invention is to design a multi-person collaborative task allocation method that assigns each task to a well-coordinated group with a specified number of workers. We face two challenges: first, how to ensure distributed cooperation, meaning any two workers in the team can collaborate effectively; and second, how to achieve personalized and differentiated team sizes, where the team size is fixed for each task, but different tasks recruit different team sizes.

[0039] To achieve the above objectives, the present invention employs the following technical solution:

[0040] This invention proposes a novel multi-person collaborative task assignment (CTA) framework. This framework considers distributed collaboration and personalized, differentiated team sizes for assignment. First, we model the interaction information of historical team combinations and tasks as a heterogeneous graph, which includes not only the interactions between teams and tasks but also the task execution quality of teams. We model the social relationships of users within a team as a user social network graph. Through a graph neural network (GNN), we can learn the group's collaborative preferences and task execution capabilities from the heterogeneous graph and the social graph. Second, to maximize the overall task execution performance, we use a greedy algorithm to solve the combinatorial optimization problem.

[0041] The logic framework diagram of this invention is as follows: Figure 1 As shown. It includes the following steps:

[0042] Step 1: The user team stated. We used w t To represent a group of intelligent sensing users, use G j This refers to a user team, which is a group of multiple collective sensing users collaborating to perform a specific task. The user team concept aims to learn the latent factors of users, which we will explore from the user portfolio G. j The latent factors learned in the middle are represented as h j ∈R d The acquisition of information from the user team includes the following three parts:

[0043] Part 1: Team G (G) j Information about the internal members themselves is represented by g. j This represents the aggregated team information.

[0044]

[0045] Where, μ t Team G j User w in t The vector representation of o jt User w t For team G j The vector representation of g j The contribution of each user in the team to the task execution varies because each user possesses different confidence levels, expertise, and social connections. Therefore, we assign different weights to different users to accurately represent their contributions to the team. We use a two-layer network to parameterize this contribution. jt We call this an attention network.

[0046]

[0047] Where w1 and w2 represent the weights of each layer of the neural network, b1 and b2 represent the biases of each layer of the neural network, and σ represents the activation function. In the experiment, we used the sigmoid function as the activation function.

[0048] The final attention weights are obtained by normalizing the attention scores using the Softmax function, which can be interpreted as user w t For team G j The contribution expressed

[0049]

[0050] Part Two: Team and Task Interaction Information. Since the user-task bipartite graph includes not only team and task interactions but also the team's task execution quality, to extract information reflecting team interests and capabilities, we consider aggregating information related to the target team G. j Interactive task information and team task execution quality information. Mathematically, aggregation functions are defined as...

[0051]

[0052] Where C(j) represents team G j The collection of tasks that have been completed, x ja Indicates and team G j Interactive task t a The vector representation of Aggre tasks The expression represents the aggregation function, and w3 and b3 represent the weights and biases of the neural network, respectively.

[0053] We use p ja Team G j Execute task t a The execution quality is represented by the vector e.p ∈R d Task t a The vector representation of q a and task execution quality e p The interaction between them is represented by x. ja express,

[0054]

[0055] in, MLP represents an element-wise concatenation operation between two vectors, and MLP represents a three-layer neural network.

[0056] Since a team performs differently across various tasks, we should consider the varying contributions of different interactions to the team's representation. Therefore, we introduce an attention mechanism into the aggregation function to assign different weights to the interactions between the team and different tasks.

[0057]

[0058] Where, α ja Team G j The interaction weights between the tasks and the task set C(j), where w4 and b4 represent the weights and biases of the neural network, respectively.

[0059] In mathematics, task attention neural networks are represented by the following formula:

[0060]

[0061]

[0062] Part Three: Information Exchange Between Teams. Our goal is to aggregate team C in the task-user bipartite diagram. j Information on second-order neighbor teams, and using C 2-hop (j) is represented as G j The second-order neighbor set. Members of the neighbor team are typically located in the same neighborhood as G. j Interacting tasks are near each other. Intuitively, neighboring team members and the target team will have similarities in their task choices. Therefore, we learn C... 2-hop Information about the team in (j). Mathematically, we use... To represent the target team G j Aggregation of neighbor information:

[0063]

[0064] Agger neighbors Let g represent an aggregate function. bk It is C 2-hopTeam G in (j) k The vector representation of .

[0065] Because of the differences between teams, we still use attention mechanisms to learn the neighbor information of the target team. Therefore, The mathematical expression is

[0066]

[0067] Where, β bk This represents the weights for different user combinations. We calculate β using an attention mechanism. bk ,

[0068]

[0069]

[0070] By aggregating the target team G j Member information g j Interaction information with tasks and information about the target team's neighbors. The team ultimately stated that through cascading g j , and Then input it into a fully connected layer:

[0071]

[0072] Where σ is the aggregation formula;

[0073] Step Two: Task Representation. The purpose of task representation is to learn task t. i The potential factors in it, we denote them as z i ∈R d The user task bipartite diagram encompasses the interaction between the team and the task, as well as the team's task execution quality. Therefore, information on team and task interactions, along with information on task execution quality, should be captured collectively to learn the underlying factors of the task.

[0074] For each task t i We need to start from task t i Information is aggregated in the set B(i) of interacting teams. We use f il To describe the target task t i and team G in set B(i) l The interaction between them, and by bringing the team G l The representation of g l And the team carried out the task t i The vector representation of the mass e′ p By cascading and inputting into a three-layer neural network, we obtain:

[0075]

[0076] In order to learn z i We also proposed the team and target task t in B(i). i The interaction between them. Team aggregation functions are described as Aggre. groups It is similar to the aggregation task function Aggre. tasks , z i Represented as:

[0077]

[0078] We introduce an attention mechanism to assign different weights u to different teams in set B(i). il The input to the attention network is the team interaction embedding f il and target task t i The embedding, we denote as q i ,

[0079]

[0080]

[0081]

[0082] Step 3: Task Execution Quality Prediction. We apply the proposed model to predict the task execution quality of teams. This is achieved by utilizing the latent representations of the team and the task (i.e., h...). j and z i We can first connect them, and then feed them into a three-layer neural network to make the following predictions:

[0083]

[0084] After obtaining the predicted task execution quality p′ ij Next, we need to optimize the model by minimizing the loss function. The loss function is defined as p′ ij and the true value p ij Mean squared error (MSE) between:

[0085]

[0086] Where P is the set of all execution quality values, p ij For Team G j To task t i The true value of the execution quality.

[0087] Step 4: Combinatorial Optimization Solution. Using the predicted task execution quality as a constant, we can solve the task allocation problem by finding the optimal team combination to maximize the overall task execution quality. We employ a greedy approach to assign a suitable team to each task: first, we rank the task execution quality of each task according to the candidate teams; then, we select the team with the highest task execution quality score as the candidate team for each task.

[0088] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Images