Trusted user recruitment method based on multi-arm gambling machine and tripartite game

Abstract

The invention discloses a credible user recruitment method based on a multi-arm gambling machine and a tripartite game, and aims to maximize the utility of participants while realizing high-quality user recruitment. According to the method, a worker selection problem is modeled into a combined multi-arm bandit problem, a trust evaluation mechanism based on active truth value discovery is introduced, normal users and malicious users are effectively distinguished, collusion attacks are resisted at the same time, and by comprehensively considering the reliability and the stability of the users, the reliability of the users is improved. An improved confidence interval upper bound algorithm is adopted to select high-quality users to execute tasks, and the balance of exploration and utilization is achieved. Besides, based on the trust evaluation and user selection results, a Starkolberg game composed of the requester, the platform and the user is constructed, an optimized decision group is obtained by adopting reverse induction solution, and a win-win situation is realized. The credible user recruitment method is provided for the mobile crowd sensing system, the overall user recruitment quality is effectively improved, and meanwhile effectiveness maximization is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile crowd intelligence perception, and in particular relates to a trusted user recruitment method based on a multi-armed gambling machine and a three-party game. Background Art

[0002] As an emerging crowdsourcing technology, mobile crowd sensing makes full use of the extensive communication capabilities of smart mobile devices, and by recruiting a large number of mobile users to participate, it realizes the perception, collection and processing of large-scale data, and then provides corresponding services. With the popularization of smart mobile devices such as smartphones, smart bracelets, and connected vehicles, the use of the diverse sensors built into these devices makes it more efficient and convenient to perceive the surrounding environment. Mobile crowd sensing technology makes full use of the mobility of users and the diversity of sensors embedded in their devices, and handles various large-scale perception tasks by gathering the wisdom of the crowd. Compared with traditional wireless sensor networks, mobile crowd sensing has significant advantages such as wider coverage, lower deployment cost, and greater flexibility. At present, mobile crowd sensing technology has been widely used in many fields, greatly facilitating our daily lives. Typical applications include environmental pollution detection, noise monitoring, urban road traffic monitoring, and health care. These applications provide users with valuable information and services by collecting and analyzing a large amount of perception data uploaded by users, significantly improving the efficiency of public services and social management.

[0003] The realization of reliable mobile crowdsensing applications relies on sufficient participants and high-quality sensing data. Due to the different skills and efforts of users, the heterogeneity of sensing devices and the uncertainty of participant behavior, the credibility of different users and the quality of uploaded data vary significantly, and even produce contradictory results, which seriously affects the reliability of mobile crowdsensing services. Existing mobile crowdsensing user recruitment methods usually assume that most users are reliable or the quality distribution of users is known. However, in actual situations, the platform usually lacks prior information about candidate users, and the platform can only perform trust assessment based on the quality of the data submitted by the user after the user uploads the data. Due to limited budget, the platform cannot conduct multiple rounds of testing on each user to verify its quality. Therefore, the platform faces a key challenge: how to strike a balance between data quality and recruitment cost and select high-quality users to complete the sensing task, that is, the platform needs to weigh whether to continue to select known high-quality users or explore users of unknown quality in order to find better participants. Summary of the invention

[0004] The purpose of the present invention is to provide a trusted user recruitment method based on a multi-armed bandit and a three-party game, which improves the overall user recruitment quality in mobile crowd intelligence perception and achieves high-quality data collection by selecting reliable users to perform tasks.

[0005] To achieve the above purpose, the technical solution provided by the present invention is as follows:

[0006] A trusted user recruitment method based on a multi-armed bandit and a three-party game, the method comprising the following steps:

[0007] (1) Step 1: The requester sends a task request to the crowd-sensing platform. The platform publishes the corresponding task. The sensing task is executed cycle by cycle and lasts until Γ. Then the platform recruits L users based on the improved confidence interval upper bound algorithm. The specific process of user recruitment is as follows: First, the platform selects users according to the quality index Sort in descending order, and then adopt a greedy selection strategy to select in each cycle t The top L users in the ranking perform tasks, among which user W i Quality indicators The calculation formula is as follows:

[0008]

[0009]

[0010] Where, ΔT i (t) and Represents user W i The estimated trust value and estimated stability are used to indicate the reliability and stability of the user, respectively. Represents the upper bound of the confidence interval, which is used to reflect the selection of user W i uncertainty, among which represents the total number of user recruitments, τ i (t) indicates recruiting the user W in period t i The number of times user W i The estimated trust value ΔT i (t) and estimated stability The specific calculation method is as follows:

[0011]

[0012]

[0013]

[0014] Among them, the indicator variable Represents user W i Whether the jth time slot generates sensing data in period t, Head λ i (t) represents user W i The upload frequency in period t, τ i(t) represents the total number of times the user uploads in period t. Represents user W i The variance of the data submitted in period t is calculated as: Represents user W i The variance accumulated to period t, Represents user W i The stability of represents the user normalized variance;

[0015] (2) Step 2: The selected L users arrive at the designated location to complete the perception according to the task requirements, and each user uses their own optimal frequency λ′ i (t) Upload perception data, λ′ i (t)∈[0,Ψ], Ψ is the maximum upload frequency in each period t, and user W i The optimal decision λ′ i The calculation formula of (t) is as follows:

[0016]

[0017] in, Indicates that the platform pays the user W i The reward, i and η i For user W i The adjustment parameters of the cost model for executing tasks depend on the performance of the sensing device carried by the user. In order to conveniently represent and calculate all users, it is expressed in matrix form as follows: but

[0018] (3) Step 3: The platform aggregates the received data and obtains the estimated true value by aggregating the data submitted by L users. At the same time, calculate each user W i Trust value in period t The specific method is as follows: First, for the task task, use Represents user W i The sensor data submitted in cycle t, User W i Trust value The calculation formula is as follows:

[0019]

[0020] Among them, μ and κ are trust adjustment parameters, 0<μ<1, and κ depends on the allowed perception error value. Represents user W i The distance between the submitted perception data and the platform's estimated true value, in Represents the estimated true value, and then obtains the weighted cumulative sum of each user data based on the trust value The formula is as follows:

[0021]

[0022] The above two steps of trust value update and estimated truth value calculation are iterated alternately until the difference of the estimated truth value meets the convergence condition, that is, it is less than the set threshold σ, and the final user trust value and platform data aggregation results are returned;

[0023] (4) Step 4: The platform pays additional costs C td (t) For the perception subtask of Ψε(t), the truth discovery technology is used to obtain the baseline data of the subtask. ε(t) represents the proportion of active truth discovery selected by the platform. This part of the perception data is used as the true value of the corresponding subtask. The trust value of the user is actively verified and corrected through the trust calculation formula. The truth discovery cost calculation formula is as follows:

[0024]

[0025] Among them, γ and δ are the truth discovery cost adjustment parameters, γ>0, δ≥0, C p (t) represents the remuneration paid by the platform to the user unit for uploading data, I i (t) represents user W i Whether it is selected to execute the task in cycle t, I i (t) = {0, 1}, I i (t) is 1, which means the platform selects the user to perform the task; otherwise, it means the user is not selected. It represents the reward paid by the platform to all recruited users. ε(t) changes dynamically with the average quality of recruited users. The calculation formula is as follows:

[0026]

[0027] (5) Step 5: After the platform receives the data submitted by the user, the optimal reward for the unit user is C′ p (t), the calculation formula is: Where m is an L*1 dimensional vector, n is an L*L dimensional matrix, = [1] L*1 , n = [1] L*L , I is the L*L dimensional identity matrix, α and β represent the adjustment parameters of the cost model of platform user data, α>0, β≥0,

[0028] (6) Step 6: The platform feeds back the aggregated data to the requester. After receiving the aggregated data, the requester pays the platform the optimal unit data reward C′ r (t), the calculation formula is as follows: make in

[0029] Beneficial Effects

[0030] The present invention proposes a method for recruiting users based on crowd intelligence perception of multi-armed bandits. When the quality of users is unknown in advance, the user quality distribution is determined by exploring different options, and the information obtained by exploration is used to select high-quality users to perform tasks. This method models the worker selection problem as a combined multi-armed bandit problem, introduces a trust evaluation mechanism based on active truth value discovery, effectively distinguishes normal users from malicious users, and resists collusion attacks. The platform comprehensively considers the reliability and stability of users, and achieves a balance between exploration and utilization based on an improved confidence interval upper bound algorithm, thereby maximizing the quality of overall recruited users.

[0031] The present invention constructs the interaction between the requester, the platform and the user into a three-party Stackelberg game. Each participant makes the best decision under the current conditions, so that the data requester can obtain higher quality data services at a lower remuneration cost, the platform pays the user less salary, and earns more difference from the data remuneration paid by the requester, and the user maximizes the difference between the salary paid by the platform and its own cost. According to the decision-making of the interactive behavior of the participants in the game process, the equilibrium achieved by the decision is analyzed to maximize the utility of all parties and finally reach the game equilibrium. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of the method for recruiting users based on the crowd intelligence perception of the multi-armed bandit in the present invention;

[0033] Figure 2 This is a comparison chart of the total quality of users recruited by the platform in four cases of adopting the method of the present invention, random strategy, greedy-∈ strategy and first-∈ strategy in the embodiment of the present invention;

[0034] Figure 3 A comparison chart of the total quality of users recruited by the platform under different numbers of candidate users in four cases of using the method of the present invention, random strategy, greedy-∈ strategy and first-∈ strategy in an embodiment of the present invention;

[0035] Figure 4This is a comparison chart of the total quality of users recruited by the platform under different numbers of recruited users in four cases where the method of the present invention, random strategy, greedy-∈ strategy and first-∈ strategy are used in the embodiments of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments. Figure 1 The flowchart of the method for recruiting users based on the multi-armed bandit group intelligence perception of the present invention is as follows: Figure 1 As shown, the method for recruiting crowd intelligence perception users includes:

[0037] The requester sends a task request to the crowd-sensing platform, and the platform publishes the corresponding task. The sensing task is executed cycle by cycle and lasts until Γ. Then the platform recruits L users based on the improved confidence interval upper bound algorithm. The specific process of user recruitment is as follows: First, the platform selects users according to the quality index Sort in descending order, and then adopt a greedy selection strategy to select in each cycle t The top L users in the ranking perform tasks, among which user W i Quality indicators The calculation formula is as follows: Where, ΔT i (t) and Represents user W i The estimated trust value and estimated stability are used to indicate the reliability and stability of the user, respectively. Represents the upper bound of the confidence interval, which is used to reflect the selection of user W i uncertainty, among which represents the total number of user recruitments, τ i (t) indicates recruiting the user W in period t i The number of times user W i The estimated trust value ΔT i (t) and estimated stability The specific calculation method is as follows: The indicator variable Represents user W i Whether the jth time slot generates sensing data in period t, Head λ i (t) represents user W i The upload frequency in period t, τ i (t) represents the total number of times the user uploads in period t. Represents user W i The variance of the data submitted in period t is calculated as: Represents user W i The variance accumulated to period t, Represents user W i The stability of represents the user normalized variance.

[0038] The selected L users arrive at the designated location to complete the perception according to the task requirements, and then the interaction between the requester, the platform and the user is modeled as a three-party Stackelberg game. First, the profit function of each party is expressed as follows: The requester's profit is the gain brought by the data service in each cycle minus the reward it pays to the platform, expressed as P r (t) = R r (t)- The platform's revenue is the reward paid by the requester minus the total cost, expressed as The user's benefit is the reward obtained from the platform minus the cost of performing the sensing task, expressed as The requester and the platform act as leaders, and the user acts as a follower. The leader makes decisions based on demand and actual conditions, and the followers play non-cooperative games under the leader's strategy and make decisions to maximize their own utility. Then the leader further adjusts its own decisions based on the follower's decisions to maximize its own utility in the current state. The requester, as the leader in the first stage, first determines the optimal unit data payment cost C′ r (t). As the leader of the second stage, the platform r (t), determine the optimal unit recruitment reward C′ paid to users p (t). As a follower, the user, given C p (t), determine its own optimal upload frequency λ′ i (t). In the equilibrium state, the optimal decision array [C′ r (t), C′ p (t),λ′ i (t)], at this point, the benefits of all parties involved in the crowd sensing task are optimal.

[0039] Each user has his or her own optimal frequency λ′ i (t) Upload perception data, λ′ i (t)∈[0,Ψ], Ψ is the maximum upload frequency in each period t, and user W i The optimal decision λ′ i The calculation formula of (t) is as follows: in, Indicates that the platform pays the user W i The reward, iand η i For user W i The adjustment parameters of the cost model for executing tasks depend on the performance of the sensing device carried by the user. but

[0040] The platform aggregates the received data and aggregates the data submitted by L users to obtain the estimated true value At the same time, calculate each user W i Trust value T in period t i (t) , The specific method is as follows: First, for the task task, use Represents user W i The sensor data submitted in cycle t, User W i Trust value The calculation formula is as follows: Among them, μ and κ are trust adjustment parameters, 0<μ<1, and κ depends on the allowed perception error value. Represents user W i The distance between the submitted perception data and the platform's estimated true value, in Represents the estimated true value, and then obtains the weighted cumulative sum of each user data based on the trust value The formula is as follows: The above two steps of trust value update and estimated truth value calculation are executed alternately and iteratively until the change difference of the estimated truth value meets the convergence condition, that is, it is less than the set threshold σ, and the final user trust value and platform data aggregation results are returned.

[0041] The platform pays additional costs C td (t) For the perception subtask of Ψε(t), the truth discovery technology is used to obtain the baseline data of the subtask. ε(t) represents the proportion of active truth discovery selected by the platform. This part of the perception data is used as the true value of the corresponding subtask. The trust value of the user is actively verified and corrected through the trust calculation formula. The truth discovery cost calculation formula is as follows: Among them, γ and δ are the truth discovery cost adjustment parameters, γ>0, δ≥0, C p (t) represents the remuneration paid by the platform to the user unit for uploading data, I i (t) represents user W i Whether it is selected to execute the task in cycle t, I i (t) = {0, 1}, I i(t) is 1, which means the platform selects the user to perform the task; otherwise, it means the user is not selected. represents the reward paid by the platform to all recruited users. ε(t) changes dynamically with the average quality of recruited users. The calculation formula is as follows: if ε(t)>1, ε(t)=1, otherwise, ε(t)=ε(t-1)(1+(T′ i (t-1)-T′ i (t))). After the platform receives the data submitted by the user, the optimal reward for the unit user is C′ p (t), the calculation formula is: Where m is an L*1 dimensional vector, n is an L*L dimensional matrix, = [1] L*1 , n = [1] L*L , I is the L*L dimensional identity matrix, α and β represent the adjustment parameters of the cost model of platform user data, α>0, β≥0,

[0042] The platform feeds back the aggregated data to the requester. After receiving the aggregated data, the requester pays the platform the optimal unit data reward C′ r (t), the calculation formula is: make in

[0043] In order to verify the feasibility and effectiveness of the trusted user recruitment method based on multi-armed bandit and three-party game proposed in the present invention, an experimental simulation is carried out to verify the method, and the method of the present invention, random strategy, greedy-∈ strategy and first-∈ strategy are compared, where ∈ takes values ​​of [0.2, 0.4, 0.6] respectively. Figures 2 to 4 The experimental results lead to the following conclusions:

[0044] 1. Figure 2 The experimental results of the total quality comparison of platform recruited users under different strategies are given when the number of candidate users is 50 and the number of recruited users is 14. As can be seen from the figure, the total quality of users recruited by the method of the present invention is always higher than that of other comparison schemes, and the total quality of recruitment under the random strategy is the lowest and the performance is the worst. Comparing the two strategies of First-∈ and Greedy-∈, the overall recruitment quality under the initial First-∈ strategy is lower than that of the Greedy-∈ strategy. At the period t=∈Γ, the recruitment quality growth curves under the three Greedy-∈ strategies turn, and finally obtain a total quality higher than that of the First-∈ strategy. This is because the First-∈ strategy only explores but does not utilize in the later stage, while the Greedy-∈ strategy still explores with ∈ probability, exploring and utilizing at the same time, so the growth rate of the recruitment quality of the former becomes faster.

[0045] 2. Figure 3 The experimental results of the total quality comparison of platform recruited users under different strategies are given when the number of candidate users changes from 50 to 100. It can be seen from the figure that the total quality of users recruited by the method of the present invention is always higher than that of other comparison schemes. Compared with different numbers of candidate users, the difference in recruitment quality performance is small, because the number of recruited users remains unchanged. However, due to the increase in the number of candidate users, the number of low-quality users increases, and the probability of selecting low-quality users increases, so the total quality performance may decrease.

[0046] 3. Figure 4 The experimental results of the total quality comparison of platform recruited users under different strategies are given when the number of recruited users changes from 10 to 30. It can be seen from the figure that the total quality of user recruitment under different strategies increases with the increase of the number of recruitments. Under different recruitment numbers, the total quality of users recruited by the method of the present invention is always higher than that of other comparison schemes. Compared with the three strategies of First-∈ and Greedy-∈ of the method of the present invention, the growth rate of the total quality of recruited users slows down as the number of recruited users increases. This is because the number of recruited users gradually approaches the number of candidate users, while the number of high-quality users among users is certain, so the growth rate slows down.

Claims

1. A trusted user recruitment method based on multi-armed bandit and three-party game, It is characterized by the following steps: (1) Step 1: The requester sends a task request to the crowd-sensing platform. The platform publishes the corresponding task. The sensing task is executed cycle by cycle and lasts until Γ. Then the platform recruits L users based on the improved confidence interval upper bound algorithm. The specific process of user recruitment is as follows: First, the platform selects users according to the quality index Sort in descending order, and then adopt a greedy selection strategy to select in each cycle t i∈[1,N] ranked top L users perform tasks, among which user W i Quality indicators The calculation formula is as follows: Where, ΔT i (t) and Represents user W i The estimated trust value and estimated stability are used to indicate the reliability and stability of the user, respectively. Represents the upper bound of the confidence interval, which is used to reflect the selection of user W i uncertainty, among which represents the total number of user recruitments, τ i (t) indicates recruiting the user W in period t i The number of times user W i The estimated trust value ΔT i (t) and estimated stability The specific calculation method is as follows: Among them, the indicator variable Represents user W i Whether the jth time slot generates sensing data in period t, and λ i (t) represents user W i The upload frequency in period t, τ i (t) represents the total number of times the user uploads in period t. Represents user W i The variance of the data submitted in period t is calculated as: Represents user W i The variance accumulated to period t, Represents user W i The stability of represents the user normalized variance; (2) Step 2: The selected L users arrive at the designated location to complete the perception according to the task requirements, and each user uses their own optimal frequency λ′ i (t) Upload perception data, λ′ i (t)∈[0,Ψ], Ψ is the maximum upload frequency in each period t, and user W i The optimal decision λ ′ i The calculation formula of (t) is as follows: in, Indicates that the platform pays the user W i The reward, i and η i For user W i The adjustment parameters of the cost model for executing tasks depend on the performance of the sensing device carried by the user. In order to conveniently represent and calculate all users, it is expressed in matrix form as follows: but (3) Step 3: The platform aggregates the received data and obtains the estimated true value by aggregating the data submitted by L users. At the same time, calculate each user W i Trust value in period t The specific method is as follows: First, for the task task, use Represents user W i The sensor data submitted in cycle t, User W i Trust value The calculation formula is as follows: Among them, μ and κ are trust adjustment parameters, 0<μ<1, and κ depends on the allowed perception error value. Represents user W i The distance between the submitted perception data and the platform's estimated true value, in Represents the estimated true value, and then obtains the weighted cumulative sum of each user data based on the trust value The formula is as follows: The above two steps of trust value update and estimated truth value calculation are iterated alternately until the difference of the estimated truth value meets the convergence condition, that is, it is less than the set threshold σ, and the final user trust value and platform data aggregation results are returned; (4) Step 4: The platform pays additional costs C td (t) For the perception subtask of Ψε(t), the truth discovery technology is used to obtain the baseline data of the subtask. ε(t) represents the proportion of active truth discovery selected by the platform. This part of the perception data is used as the true value of the corresponding subtask. The trust value of the user is actively verified and corrected through the trust calculation formula. The truth discovery cost calculation formula is as follows: Among them, γ and δ are the truth discovery cost adjustment parameters, γ>0, δ≥0, C p (t) represents the remuneration paid by the platform to the user unit for uploading data, I i (t) represents user W i Whether it is selected to execute the task in cycle t, I i (t)={0,1},I i (t) is 1, which means the platform selects the user to perform the task; otherwise, it means the user is not selected. It represents the reward paid by the platform to all recruited users. ε(t) changes dynamically with the average quality of recruited users. The calculation formula is as follows: (5) Step 5: After the platform receives the data submitted by the user, the optimal reward for the unit user is C′ p (t), the calculation formula is: Where m is an L*1 dimensional vector, n is an L*L dimensional matrix, = [1] L*1 ,n=[1] L*L , is the L*L dimensional identity matrix, α and β represent the adjustment parameters of the cost model of platform user data, α>0, β≥0, (6) Step 6: The platform feeds back the aggregated data to the requester. After receiving the aggregated data, the requester pays the platform the optimal unit data reward C′ r (t), the calculation formula is as follows: make in

Images