A distributed SLB service recommendation method and system based on an incentive mechanism

Abstract

The application discloses a distributed SLB service recommendation method and system based on an incentive mechanism. The method comprises the following steps: acquiring client data and constructing a distributed SLB model, solving by a ridge regression estimation algorithm, and determining the upload data volume of the client; determining an incentive cost and constructing a cooperation proportion relationship formula according to the upload data volume of the client, and determining a client participant set; determining a marginal cost, the server issues an incentive to the client in the client participant set according to the marginal cost, acquires the upload data of the client, and obtains a distributed SLB service recommendation result. The embodiment of the application can encourage the participation of the client through the incentive mechanism, meet the privacy demand of the client before obtaining the incentive, and resist the trick of a malicious user to cheat higher income by falsely reporting the cost. The application can be widely applied to the technical field of distributed personalized recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distributed personalized recommendation, and particularly relates to a distributed SLB service recommendation method and system based on an incentive mechanism. BACKGROUND

[0002] In recent years, various media applications on the Internet provide increasingly rich media content for users to browse, and one of the main tasks of these applications is to recommend content that meets the user's preferences to attract and retain users. For example, popular video platforms such as TikTok provide timely and accurate content recommendations by inferring users' personal information, browsing history and other data. Similarly, online learning platforms such as Coursera recommend appropriate courses and learning content by collecting users' course selection, learning progress, performance, etc. using machine learning algorithms.

[0003] With the continuous improvement of the computing power of terminal devices, the paradigm of using multiple clients to collaborate in learning has been widely applied, which is called distributed learning technology. However, distributed learning requires the transmission of local data or models by multiple clients, and some malicious attackers can directly obtain or indirectly infer the private information of client users, such as gender, age, preferences, etc. through these data or models, so the client needs to consider the privacy leakage cost brought by participating in distributed learning. At the same time, the client participating in distributed learning will also generate local learning cost and communication cost, and in a real scenario, the terminal device needs to consider its own benefits, and only in the case of obtaining sufficient incentives can the participation motivation of the client be rationalized, so the distributed learning algorithm needs to consider the design of the incentive mechanism.

[0004] At present, related methods extend the SLB algorithm to distributed recommendation systems, and these distributed methods require clients to upload data or models related to actions and rewards, and then the server aggregates and synchronizes the models. However, most of these methods assume that the client is willing to participate, and ignore the various costs faced by the terminal device participating in distributed learning, and lack the design of an incentive mechanism to rationalize the participation motivation of the client. A few studies introduce an incentive mechanism in the distributed SLB algorithm, but these mechanisms also require the client to upload relevant data about the action in advance to develop appropriate incentives, and this part of the data still has privacy leakage considerations.

[0005] In summary, the technical problems existing in the related art need to be improved. SUMMARY

[0006] The main purpose of the embodiments of the present application is to provide a distributed SLB service recommendation method and system based on an incentive mechanism, which can motivate the participation of the client through the incentive mechanism, meet the privacy needs of the client before obtaining the incentive, and resist malicious users from cheating higher benefits by falsely reporting costs.

[0007] To achieve the above object, an aspect of the embodiment of the present application proposes a distributed SLB service recommendation method based on incentive mechanism, which comprises:

[0008] acquiring client data and constructing a distributed SLB model, solving by ridge regression estimation algorithm, determining the upload data volume of the client;

[0009] determining the incentive cost and constructing the cooperation proportion relationship according to the upload data volume of the client, determining the client participant set;

[0010] determining the marginal cost, the server issues incentives to the clients in the client participant set according to the marginal cost, acquires the upload data of the client, and obtains the distributed SLB service recommendation result.

[0011] In some embodiments, the acquiring client data and constructing a distributed SLB model, solving by ridge regression estimation algorithm, determining the upload data volume of the client, comprises:

[0012] acquiring client data, and determining the historical action information of the client data and the historical reward information of the client data;

[0013] constructing a covariance matrix according to the historical action information of the client data and constructing a reward vector according to the historical action information and historical reward information of the client data;

[0014] constructing a distributed SLB model according to the covariance matrix and the reward vector, and estimating by ridge regression estimation algorithm to obtain a distributed SLB model estimation result;

[0015] based on the optimistic strategy, selecting the action with the highest reward confidence upper bound of the distributed SLB model estimation result, and acquiring the corresponding reward as the upload data of the client, and recording the upload data volume of the client.

[0016] In some embodiments, the client data comprises first type data, second type data and third type data, wherein:

[0017] the first type data represents the local data newly added by the client after the last data service communication, if the newly added local data has an incentive, it is aggregated as the third type data, if the newly added local data does not have an incentive, it is aggregated as the second type data;

[0018] the second type data represents the local data that has not been uploaded to participate in data service communication;

[0019] the third type data represents the synchronization data for participating in data service communication.

[0020] In some embodiments, the determination of the incentive cost and the construction of the collaborative proportion relationship according to the upload data amount of the client determine a set of client participants, comprising:

[0021] According to the upload data amount of the client, a collaborative proportion relationship is constructed;

[0022] The incentive cost of the client is determined and sorted in ascending order to obtain a sorted client incentive cost;

[0023] A threshold ratio is set, and the sorted client incentive cost is sequentially divided into the set of client participants that meet the corresponding client of the collaborative proportion relationship.

[0024] In some embodiments, the expression of the communication trigger condition is as follows:

[0025]

[0026] In the above formula, Δt k represents the time elapsed from the last communication to the current round of communication, represents a discrimination threshold, represents a covariance matrix containing upload data, represents the covariance matrix of the first type of data.

[0027] In some embodiments, the expression of the collaborative proportion relationship is as follows:

[0028]

[0029] In the above formula, α represents the collaborative proportion parameter, P k,t represents the cumulative number of samples participating in collaboration, Q k,t represents the total number of samples in the current system, and α h represents a threshold ratio.

[0030] In some embodiments, the determination of the marginal cost, the server issues an incentive to the client in the set of client participants according to the marginal cost, obtains the upload data of the client, and obtains a distributed SLB service recommendation result, comprising:

[0031] According to the custom reporting cost of the client, the marginal cost is determined;

[0032] According to the data real cost and the custom reporting cost, the client is divided and processed to obtain malicious clients and honest clients;

[0033] According to the covariance matrix of the upload data amount of the client, a communication trigger condition is set;

[0034] distributing an incentive to the malicious client and the honest client according to the marginal cost based on the communication trigger condition;

[0035] if the malicious client and the honest client are selected to communicate, uploading the covariance matrix corresponding to the newly added local data and the reward vector to a server for data communication aggregation to obtain a first service recommendation result;

[0036] if the malicious client and the honest client are not selected to communicate, performing local data aggregation on the covariance matrix corresponding to the newly added local data and the reward vector to obtain a second service recommendation result;

[0037] combining the first service recommendation result and the second service recommendation result to obtain the distributed SLB service recommendation result.

[0038] In some embodiments, the marginal cost is calculated by keeping the custom reporting cost of the remaining clients unchanged, assuming that the custom reporting cost of the selected client changes, and sequentially changing the custom reporting cost of each unselected client, repeatedly determining the client participant set, so that the minimum custom reporting cost of the selected client that will not be selected by the server is the marginal cost.

[0039] In some embodiments, the distributing an incentive to the malicious client and the honest client according to the marginal cost based on the communication trigger condition comprises:

[0040] for the honest client, if the custom reporting cost is less than the marginal cost, the honest client is selected to communicate for data service and obtains the incentive, and if the custom reporting cost is greater than the marginal cost, the honest client is not selected to communicate for data service and cannot obtain the incentive;

[0041] for the malicious client, if the custom reporting cost is less than the data truth cost which is less than the marginal cost, or the data truth cost is less than the custom reporting cost which is less than the marginal cost, or the custom reporting cost is less than the marginal cost which is less than or equal to the data truth cost, the malicious client is selected to communicate for data service and obtains the incentive, and if the data truth cost is less than the marginal cost which is less than or equal to the custom reporting cost, or the marginal cost is less than the custom reporting cost which is less than or equal to the data truth cost, or the marginal cost is less than the data truth cost which is less than or equal to the custom reporting cost, the malicious client is not selected to communicate for data service and cannot obtain the incentive.

[0042] To achieve the above object, another aspect of the embodiment of the present application provides a distributed SLB service recommendation system based on an incentive mechanism, which comprises:

[0043] A first module is configured to acquire client data and build a distributed SLB model, solve by ridge regression estimation algorithm, determine the upload data volume of the client;

[0044] A second module is configured to determine an incentive cost and build a collaboration proportion relationship according to the upload data volume of the client, and determine a client participant set;

[0045] A third module is configured to determine a marginal cost, the server issues an incentive to the client in the client participant set according to the marginal cost, acquires the upload data of the client, and obtains a distributed SLB service recommendation result.

[0046] The embodiment of the present application has at least the following beneficial effects: the present application provides a distributed SLB service recommendation method and system based on an incentive mechanism, which acquires client data and builds a distributed SLB model, solves by ridge regression estimation algorithm, determines the upload data and upload data volume of the client, further determines an incentive cost and builds a collaboration proportion relationship according to the upload data volume of the client, determines a client participant set, considers the model heterogeneity of distributed learning, designs an incentive mechanism to encourage the participation of the client, and at the same time meets the privacy demand of the client before obtaining the incentive, further determines a marginal cost, the server issues an incentive to the client according to the marginal cost, acquires the upload data of the client, and is used to find the optimal recommendation strategy, at the same time guarantees the privacy protection demand of the client to the action and reward data, until the server provides a satisfactory incentive, the incentive mechanism is designed to meet the individual rationality and incentive compatibility, and can resist malicious users from cheating higher income by falsely reporting the cost. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a flowchart of a distributed SLB service recommendation method based on an incentive mechanism provided by the embodiment of the present application;

[0048] Figure 2 is a structural schematic diagram of a distributed SLB service recommendation system based on an incentive mechanism provided by the embodiment of the present application;

[0049] Figure 3 is a schematic diagram of a distributed SLB service recommendation model built by the embodiment of the present application;

[0050] Figure 4 is a schematic diagram of a distributed personalized recommendation algorithm provided by the embodiment of the present application. DETAILED DESCRIPTION

[0051] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application 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 only used to explain the present application and not intended to limit the present application. When the following description refers to the accompanying drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with embodiments of the present application. They are only examples of systems and methods consistent with some aspects of the embodiments of the present application as detailed in the appended claims.

[0052] It can be understood that the terms "first", "second" and the like used in the present application can be used herein to describe various concepts, but unless specifically stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another concept. For example, without departing from the scope of the embodiments of the present application, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information. Depending on the context, the word "if" as used herein can be interpreted as "when" or "when" or "in response to determining".

[0053] The terms "at least one", "multiple", "each", "any" and the like used in the present application include one, two or more than two, multiple includes two or more than two, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0055] Before the embodiments of the present application are described in detail, first, some nouns and terms involved in the embodiments of the present application are described, and the nouns and terms involved in the embodiments of the present application are applicable to the following explanations.

[0056] 1) Model heterogeneity: The difference between the local models of the clients, referred to as model heterogeneity, will make the total model obtained by collaboration deviate from the local model, resulting in the decision generated by the recommendation algorithm deviating from the optimal decision, affecting the performance of the recommendation system, and is also an important problem that needs to be considered by the distributed algorithm.

[0057] 2) Stochastic Linear Bandit (SLB): One of the most popular approaches for personalized recommendation is to formulate it as a stochastic linear bandit problem, where the candidate media content is formulated as a set of candidate actions, the user feedback obtained from recommending a certain media content is formulated as a reward, and a linear model parameter is learned from the history actions and rewards to predict the optimal decision for the next round.

[0058] In the related art, some deficiencies exist, such as the existing model ignores various costs faced by terminal devices participating in distributed learning, lacks the design of an incentive mechanism to rationalize the participation motivation of the client, and has the problem of privacy leakage.

[0059] Therefore, in the embodiments of the present application, a distributed SLB service recommendation method based on an incentive mechanism is provided, which introduces an incentive mechanism in distributed learning, simultaneously describes the preference differences of users of a recommendation system in different regions through a multi-task structure, and meets the privacy demand of some institutions that do not want data to be transmitted externally. A distributed personalized recommendation algorithm is proposed based on the model to find the optimal recommendation strategy, while ensuring the privacy protection demand of the client for action and reward data until the server provides satisfactory incentives. In addition, the algorithm designs an incentive mechanism to meet individual rationality and incentive compatibility, which can resist malicious users from cheating higher benefits by falsely reporting costs.

[0060] First of all, it needs to be pointed out that, as shown in Figure 3 The embodiments of the present application propose a multi-task distributed learning model for a personalized recommendation system. From the horizontal direction, each client group corresponds to a specific region, such as a city, town or institution. These regions differ in the number of users, participation level and preference patterns, and each region is represented as a different task to capture the model heterogeneity between different regions. Vertically, the framework combines a distributed learning architecture and an integrated incentive mechanism to promote client participation, where the system consists of a server and K client groups corresponding to K tasks, where task k∈{1,2,…,K} corresponds to the client set The cardinality of the set of all users is denoted as The cardinality of the set of all users is denoted as The cardinality of the set of all users is denoted as The global unknown model parameter shared by the users in the set is denoted as θ k , which is a key parameter in the linear reward structure of the SLB model.

[0061] Further, the client model of the embodiments of the present application is described as follows:

[0062] In real-world scenarios, the recommendation behavior on the clients does not always happen, i.e., the local learning process of the clients does not always happen. We assume that the recommendation behavior of the clients is intermittent, rather than happening at every round in a time horizon. At each round, only the active clients will interact with the media files, perform local learning, while the inactive clients only maintain the previous data in that round without further operations. To describe this intermittent interaction of the clients, we construct the following client model.

[0063] Let the round limit be denoted as T, the client group be denoted as For client i k , let its state at each time step t∈{1, 2,..., T} be denoted as If client i k is active at round t, then Otherwise Let the round of the last communication be denoted as t last . Let the number of active times of client i k since the last communication be denoted as , which is defined as Intuitively, can be regarded as the number of newly added samples of client i k since the last communication, reflecting the value that client i k can bring to the next communication.

[0064] Let be denoted as the set of active clients in the client group at round t. At round t, for the active client , many media files are candidate objects to be interacted with, which are constructed into an action set Then client i k selects the action to be performed in this round according to the historical interaction actions and results. Subsequently, client i k observes a reward with a linear structure is an unknown model parameter to be learned by task k, and is a Gaussian variable with mean 0 and variance σ 2 . To simplify the expression and analysis, we assume that: and ||θ k ||2≤1.

[0065] Further, the server-side model of the embodiments of the present application is described as follows:

[0066] Most of the work on distributed SLB learning assumes that clients participate in collaboration without requiring compensation, and few studies explore the rationality of client participation in collaboration, resulting in a lack of incentive mechanisms for distributed SLB learning to motivate clients. In the server-side model, we consider the cost of client participation in collaboration, and build the following model to describe the incentive mechanism and the communication process.

[0067] Distributed systems use collaboration between clients to facilitate learning, which brings local learning costs, communication costs and privacy leakage costs to clients, hindering the participation of clients. In our distributed learning model, we consider the collaboration cost of clients in each communication round. In the local learning with a total number of rounds T, the set of time indices of the rounds in which communication occurs is denoted as In round , if the distributed system is in the communication state, the client i k will pay a unit collaboration cost for each sample observed in the environment. Therefore, the total cost is The client i k needs to inform the server and

[0068] to establish pricing. In each communication round , the server will select a part of the clients to participate in collaboration, and assign appropriate incentives to each client i k in . If the incentive given by the server is sufficient to pay the collaboration cost, the client i k participates in collaboration and uploads the local data is rational. In formal terms, when a certain incentive mechanism is used, if each participant has a non-negative profit, that is, any client participating in collaboration satisfies , then it is said that the incentive mechanism satisfies individual rationality (IR).

[0069] In the design of the incentive mechanism, IR is the most basic property that needs to be satisfied. In addition, incentive compatibility (IC) is also a property worth attention, which aims to resist dishonest behavior of malicious users in the distributed system. In formal terms, if a certain incentive mechanism is used, for all clients, the income obtained when reporting the cost truthfully is optimal, and reporting false costs will not make them obtain more income, or even reduce income, then it is said that the incentive mechanism has the property of IC.

[0070] Finally, the performance measurement indicators of the embodiments of the application are described:

[0071] In the SLB model, the performance of the algorithm is usually measured by the regret value, and in the distributed learning SLB model, the target is to minimize the cumulative regret value of all clients participating in the distributed learning task, defined as:

[0072]

[0073] In the above formula, is the client i k The optimal action at round t can be selected.

[0074] Referring to Figure 1 and Figure 4 , Figure 1 The flowchart of a distributed SLB service recommendation method based on an incentive mechanism provided by an embodiment of the application, referring to Figure 1 The method comprises the following steps:

[0075] S100, obtaining client data and constructing a distributed SLB model, solving by a ridge regression estimation algorithm, and determining the upload data amount of the client;

[0076] It should be noted that in some embodiments, step S100 can comprise: S110, obtaining client data and determining the historical action information of the client data and the historical reward information of the client data; S120, constructing a covariance matrix according to the historical action information of the client data and constructing a reward vector according to the historical action information and the historical reward information of the client data; S130, constructing a distributed SLB model according to the covariance matrix and the reward vector, and estimating by a ridge regression estimation algorithm to obtain a distributed SLB model estimation result; S140, based on an optimistic strategy, selecting the action with the highest reward confidence upper bound of the distributed SLB model estimation result, and obtaining the corresponding reward as the upload data of the client, and recording the upload data amount of the client.

[0077] In some specific embodiments, at round t∈[T], the active client can perform local learning, and under the SLB setting, the local learning is equivalent to executing the classic non-distributed SLB algorithm, i.e., the LinUCB algorithm, in the local environment. At the same time, the inactive client only maintains the data collected previously in this round and does not perform any other operation.

[0078] The purpose of local learning is to collect historical action and reward data, and use these data to learn the model θ k , so as to select the optimal decision in this round. Specifically, let t' represent the round in which the last communication occurs, and the historical data saved by the local end includes three categories:

[0079] The first type of data: round t ′The newly added local data, which has not been uploaded for collaboration, will be uploaded for collaboration (aggregated into the third type of data) if the client receives the incentive issued by the server, otherwise it will not be uploaded and only recorded in the local data to promote local learning (aggregated into the second type of data);

[0080] The second type of data: round t ′ and the local data recorded before that has not been uploaded for collaboration;

[0081] The third type of data: round t ′ and the synchronized data obtained through collaboration before that, which includes both the data uploaded by the client for collaboration and the data uploaded by other clients in the same learning task for collaboration, is the source of the learning effect gain of the client.

[0082] Use and S k,t to represent the covariance matrices constructed by the action information in the three types of data, and the specific construction form of these matrices can be given by the following expressions:

[0083]

[0084]

[0085]

[0086] Similarly, use and s k，t to represent the reward vectors constructed by the action and reward information in the three parts of data. The construction of these vectors can be given by the following expressions:

[0087]

[0088]

[0089]

[0090] In the repeated local learning phase, the client i k learns the intrinsic model θ k The specific process is as follows. First, the client i k will obtain a set of candidate actions The client wants to predict the optimal action in this set to obtain the highest reward. In order to estimate the reward that can be obtained by selecting each action, the client first needs to estimate the model parameters θ k by ridge regression. The model parameters estimated by the client i k are denoted as Then, the ridge regression is used to estimate θk The specific expression is:

[0091]

[0092] Wherein And Meanwhile, the above-mentioned three types of data are included. Based on the estimated model The algorithm adopts an optimistic strategy to select the optimal action of prediction, that is, the action with the highest upper confidence bound (UCB) of estimated reward, that is:

[0093]

[0094] Wherein is a proper exploration parameter related to the confidence interval, and corresponding to the 1-δ confidence interval is set to:

[0095]

[0096] Select the action After that, the client i k observes the actual obtained reward and updates the local data and and is the key data containing new information, which can participate in cooperation in the upcoming communication round. After each communication round, and are merged into the other two types of data, and are reset to 0 d×d and 0 d respectively, so as to collect and cooperate subsequent data.

[0097] S200, determine the incentive cost and construct the cooperation proportion relationship according to the upload data amount of the client, and determine the client participant set;

[0098] It should be noted that in some embodiments, step S200 can include: S210, constructing the cooperation proportion relationship according to the upload data amount of the client; S220, determining the incentive cost of the client and sorting in ascending order to obtain the sorted client incentive cost; S230, setting a threshold ratio, and dividing the sorted client incentive cost into the client participant set according to the client that meets the cooperation proportion relationship.

[0099] In some embodiments, the client selects an optimization that aims to achieve the collaboration effect with limited cost. Obviously, without considering the incentive expenditure, the higher the degree of cooperation, the more learning samples it brings, and the more beneficial it is to the improvement of algorithm performance. However, when considering the cost of collaboration, blindly increasing the degree of cooperation may bring excessive incentive expenditure. Therefore, we need to balance the degree of cooperation and the cost of incentives.

[0100] In order to measure the trade-off between the degree of cooperation and the cost of incentives, we define a parameter a to represent the degree of cooperation, called the collaboration ratio, defined as follows:

[0101] a = P k,t / Q k,t

[0102] wherein P represents the cumulative number of samples participating in collaboration, Q represents the total number of samples in the current system (whether or not participating in collaboration).

[0103] In order to achieve the expected level of cooperation and thus achieve the expected performance improvement, the collaboration ratio needs to reach a threshold ratio a h . a h is a pre-input parameter, and the server sorts the costs of the clients in ascending order, and in this order, the clients are sequentially included in the participant set until the following condition is met, the expression of which is:

[0104]

[0105] In the above formula, a represents the collaboration ratio parameter, P k,t represents the cumulative number of samples participating in collaboration, Q k,t represents the total number of samples in the current system, and a h represents the threshold ratio.

[0106] S300, determine the marginal cost, the server issues an incentive to the client in the client participant set according to the marginal cost, obtains the upload data of the client, and obtains the distributed SLB service recommendation result;

[0107] It should be noted that in some embodiments, step S300 can include:

[0108] S310, determine the marginal cost according to the custom reporting cost of the client;

[0109] S320, divide and process the client according to the data real cost and the custom reporting cost, to obtain the malicious client and the honest client;

[0110] In some embodiments, the present application adopts a metric that exposes less private information to achieve the effect of protecting privacy. Specifically, we require the client to upload the number of new sample data before obtaining the incentive without uploading the client's historical actions and other highly privacy-related data. The client uploads and to reflect the value and cost of its participation in collaboration, and strives for appropriate incentives from the server. In order to distinguish from the actual cost we define the reported cost as

[0111] S330, according to the covariance matrix of the uploaded data volume of the client, set the communication trigger condition;

[0112] In some embodiments, in order to reduce communication overhead, communication only occurs in specific rounds, determined by the communication trigger condition, rather than every round. The communication trigger condition is used to evaluate whether the additional local data contains enough new information to be worth a new round of collaboration, and is calculated according to the following formula, specifically:

[0113]

[0114] where αt k represents the time elapsed from the last communication to the current round, is the decision threshold. The value of will affect the performance of the algorithm, in order to ensure that the algorithm achieves sublinear regret value, the value of

[0115]

[0116] If any client satisfies the communication trigger condition, the server and the client will communicate and collaborate in this round.

[0117] In some embodiments, the server calculates the appropriate incentive that satisfies IR and IC for the selected client by determining the marginal cost of each client. Use as the reported cost of all other clients except client i k The marginal cost is defined as: for any client its marginal cost satisfies The calculation method of the marginal cost is: keep the reported cost of other clients unchanged, traverse all cases of , repeat the steps of client selection, so that client i k will not be the smallest selected by the server. Using the marginal cost, the reasonable incentive value is set as:

[0118]

[0119] In the above formula, represents a reasonable incentive value, represents marginal cost, represents the client i k The number of samples added since the last round of communication.

[0120] Where, the honest client reports the real cost And the malicious client reports the false cost

[0121] S340, based on the communication trigger condition, according to the marginal cost to the malicious client and the honest client to issue incentives;

[0122] Specifically, for the honest client, if the self-defined reported cost is less than the marginal cost, the honest client is selected for data service communication and obtains the incentive, if the self-defined reported cost is greater than the marginal cost, the honest client is not selected for data service communication and cannot obtain the incentive;

[0123] For the malicious client, if the self-defined reported cost is less than the data real cost is less than the marginal cost or the data real cost is less than the self-defined reported cost is less than the marginal cost or the self-defined reported cost is less than the marginal cost is less than or equal to the data real cost, the malicious client is selected for data service communication and obtains the incentive, but the benefit brought by the incentive will not be higher than the benefit obtained when the malicious client behaves as an honest client, if the data real cost is less than the marginal cost is less than or equal to the self-defined reported cost or the marginal cost is less than the self-defined reported cost is less than or equal to the data real cost or the marginal cost is less than the data real cost is less than or equal to the self-defined reported cost, the malicious client is not selected for data service communication and cannot obtain the incentive, and the benefit brought by the incentive will not be higher than the benefit obtained when the malicious client behaves as an honest client.

[0124] In some embodiments, for i k The case of the honest client:

[0125] If Since the reported cost is less than the marginal cost, the client i k Obtains the incentive

[0126] If Because the reported cost is not less than the marginal cost, the client i k Is not selected, and the incentive is 0.

[0127] For i k The case of the malicious client:

[0128] If the reported cost is still less than the marginal cost, then the client i is still selected, and the incentive is still k unchanged.

[0129] If the reported cost is no longer less than the critical cost, then the client i k is no longer selected, and the incentive becomes 0, which is less than the incentive obtained in the honest case.

[0130] If the reported cost is less than the critical cost, then the client i k is selected by lying about the cost, and the incentive is At this point, the client i k will obtain a non-positive net gain.

[0131] If or the reported cost is still not less than the critical cost, then the client i k is still not selected, and the incentive is still 0.

[0132] Therefore, the marginal cost method prevents the client from obtaining more profit by cheating the server and resists the dishonest behavior of malicious clients.

[0133] S350, if the malicious client and the honest client are selected to communicate, upload the covariance matrix corresponding to the newly added local data and the reward vector to the server for data communication aggregation to obtain a first service recommendation result;

[0134] S360, if the malicious client and the honest client are not selected to communicate, locally aggregate the covariance matrix corresponding to the newly added local data and the reward vector to obtain a second service recommendation result;

[0135] S370, combine the first service recommendation result and the second service recommendation result to obtain a distributed SLB service recommendation result.

[0136] In some specific embodiments, the algorithm adopts a star-shaped communication network, and N clients directly interact with the center server. In a communication round, the client selected by the server will obtain sufficient incentive and participate in cooperation. If the client i k is selected, the client i k will upload the newly added local data and to the server, and the uploaded data will be aggregated as synchronous data. If the client i​​k If not selected, the newly added local data And Will only be aggregated to local data And

[0137] The data obtained by the server from the client And After aggregation to S k，t And s k，t , S k，t And s k,t Synchronize to the client to participate in the learning of the local model. Because the goal of the server is to promote the learning process of all clients, the client will receive updated synchronization data S k，t And S k，t .

[0138] In summary, the embodiments of the present application design an incentive mechanism to encourage the participation of clients, while meeting the privacy needs of clients before obtaining incentives. Considering the model heterogeneity of distributed learning, we propose a distributed learning model based on SLB. This model introduces an incentive mechanism in distributed learning, while depicting the preference differences of users of different regional recommendation systems through the structure of multiple tasks, and meets the privacy needs of some institutions that do not want data to be transmitted outside. Based on this model, a distributed personalized recommendation algorithm is proposed to find the optimal recommendation strategy, while ensuring the privacy protection needs of clients for action and reward data until the server provides satisfactory incentives. In addition, the algorithm designs an incentive mechanism to meet IR and IC, which can resist malicious users from cheating higher benefits by falsely reporting costs.

[0139] Please refer to Figure 2 The embodiments of the present application also provide a distributed SLB service recommendation system based on an incentive mechanism, which can implement the above-mentioned distributed SLB service recommendation method based on an incentive mechanism. The system comprises:

[0140] A first module 201 is configured to obtain client data and construct a distributed SLB model, solve by a ridge regression estimation algorithm, and determine the upload data amount of the client;

[0141] A second module 202 is configured to determine an incentive cost and construct a collaboration proportion relationship according to the upload data amount of the client, and determine a client participant set;

[0142] A third module 203 is configured to determine a marginal cost, the server issues an incentive to the client in the client participant set according to the marginal cost, obtains the upload data of the client, and obtains a distributed SLB service recommendation result.

[0143] It can be understood that the contents in the method embodiments are applicable to the system embodiments, the system embodiments specifically implement the functions same as those of the method embodiments, and achieve the beneficial effects same as those of the method embodiments.

[0144] The preferred embodiments of the application are described above with reference to the accompanying drawings, and are not intended to limit the scope of the application. Any modification, equivalent replacement and improvement made by those skilled in the art without departing from the scope and essence of the application shall be within the scope of the application.

Claims

1. A distributed SLB service recommendation method based on an incentive mechanism, characterized in that, The method includes the following steps: Acquire client data and construct a distributed stochastic linear slot machine model, solve it using the ridge regression estimation algorithm, and determine the amount of data uploaded by the client; Determine the incentive cost and construct a collaboration ratio formula based on the amount of data uploaded by the client to determine the set of client participants; Determine the marginal cost, and the server issues incentives to clients in the client participant set based on the marginal cost, obtains the uploaded data of the clients, and obtains the distributed random linear slot machine service recommendation results; The step of acquiring client data and constructing a distributed stochastic linear slot machine model, solving it using a ridge regression estimation algorithm, and determining the amount of data uploaded by the client includes: Acquire client data and determine the historical action information and historical reward information of the client data; A covariance matrix is ​​constructed based on the historical action information of the client data, and a reward vector is constructed based on the historical action information and historical reward information of the client data. A distributed stochastic linear slot machine model is constructed based on the covariance matrix and the reward vector, and then estimated using the ridge regression estimation algorithm to obtain the estimation result of the distributed stochastic linear slot machine model. Based on an optimistic strategy, the action with the highest reward confidence upper bound in the estimation results of the distributed stochastic linear slot machine model is selected, and the corresponding reward is obtained as the client's uploaded data. The amount of uploaded data from the client is recorded. The process of determining the incentive cost and constructing a collaboration ratio based on the amount of data uploaded by the client, thereby determining the set of client participants, includes: A collaborative ratio formula is constructed based on the amount of data uploaded by the client; Determine the incentive cost of the client and sort it in ascending order to obtain the sorted client incentive cost; Set a threshold ratio, and divide the clients whose incentive costs satisfy the collaboration ratio into the client participant set according to the sorted client incentive costs; The specific expression for the communication trigger condition is as follows: ； In the above formula, This indicates the time elapsed from the last communication to the current round of communication. Indicates the discrimination threshold. This represents the covariance matrix containing the uploaded data. This represents the covariance matrix of the first type of data.

2. The method according to claim 1, characterized in that, The client data includes a first type of data, a second type of data, and a third type of data, wherein: The first type of data represents the local data added by the client after the last data service communication. If the newly added local data has an incentive, it is aggregated into the third type of data. If the newly added local data does not have an incentive, it is aggregated into the second type of data. The second type of data represents local data that has not been uploaded for participation in data service communication; The third type of data refers to synchronization data used for participating in data service communication.

3. The method according to claim 1, characterized in that, The specific expression for the cooperation ratio is shown below: ； In the above formula, This represents the collaboration ratio parameter. This represents the cumulative number of samples participating in the collaboration. This represents the total number of samples in the current system. This indicates the threshold ratio.

4. The method according to claim 2, characterized in that, The determination of marginal cost involves the server distributing incentives to clients in the client participant set based on the marginal cost, acquiring the uploaded data from the clients, and obtaining the distributed random linear slot machine service recommendation results, including: The marginal cost is determined based on the client's custom reporting cost; Based on the actual cost of the data and the custom reporting cost, the clients are divided into malicious clients and honest clients. Based on the covariance matrix of the amount of data uploaded by the client, the communication triggering conditions are set; Incentives are distributed to the malicious client and the honest client based on the communication triggering conditions and the marginal cost. If the malicious client and the honest client are selected to communicate, the covariance matrix and the reward vector corresponding to the newly added local data are uploaded to the server for data communication aggregation to obtain the first service recommendation result; If the malicious client and the honest client are not selected to communicate, the covariance matrix corresponding to the newly added local data and the reward vector are aggregated locally to obtain the second service recommendation result. The distributed random linear slot machine service recommendation result is obtained by combining the first service recommendation result and the second service recommendation result.

5. The method according to claim 4, characterized in that, The marginal cost is calculated by keeping the custom reporting cost of the other clients constant, assuming that the custom reporting cost of the selected client changes, and then changing sequentially to the custom reporting cost of each unselected client. The client participant set is repeatedly determined so that the minimum custom reporting cost that prevents the selected client from being selected by the server is the marginal cost.

6. The method according to claim 4, characterized in that, The provision of incentives to the malicious client and the honest client based on the marginal cost includes: For the honest client, if the custom reporting cost is less than the marginal cost, the honest client is selected to perform data service communication and receives the incentive; if the custom reporting cost is greater than the marginal cost, the honest client is not selected to perform data service communication and cannot receive the incentive. For the malicious client, if the custom reporting cost is less than the actual data cost, which is less than the marginal cost, or the actual data cost is less than the custom reporting cost, which is less than the marginal cost, or the custom reporting cost is less than the marginal cost and less than or equal to the actual data cost, then the malicious client is selected to perform data service communication and receives the incentive. If the actual data cost is less than the marginal cost, which is less than or equal to the custom reporting cost, or the marginal cost is less than the custom reporting cost, which is less than or equal to the actual data cost, or the marginal cost is less than the actual data cost, which is less than or equal to the custom reporting cost, then the malicious client is not selected to perform data service communication and cannot receive the incentive.

7. A distributed SLB service recommendation system based on an incentive mechanism, characterized in that, The system includes: The first module is used to acquire client data and construct a distributed stochastic linear slot machine model, which is solved by the ridge regression estimation algorithm to determine the amount of data uploaded by the client. The second module is used to determine the incentive cost and construct a collaboration ratio formula based on the amount of data uploaded by the client, thereby determining the set of client participants; The third module is used to determine the marginal cost. The server issues incentives to the clients in the client participant set based on the marginal cost, obtains the uploaded data of the clients, and obtains the distributed random linear slot machine service recommendation results. The step of acquiring client data and constructing a distributed stochastic linear slot machine model, solving it using a ridge regression estimation algorithm, and determining the amount of data uploaded by the client includes: Acquire client data and determine the historical action information and historical reward information of the client data; A covariance matrix is ​​constructed based on the historical action information of the client data, and a reward vector is constructed based on the historical action information and historical reward information of the client data. A distributed stochastic linear slot machine model is constructed based on the covariance matrix and the reward vector, and then estimated using the ridge regression estimation algorithm to obtain the estimation result of the distributed stochastic linear slot machine model. Based on an optimistic strategy, the action with the highest reward confidence upper bound in the estimation results of the distributed stochastic linear slot machine model is selected, and the corresponding reward is obtained as the client's uploaded data. The amount of uploaded data from the client is recorded. The process of determining the incentive cost and constructing a collaboration ratio based on the amount of data uploaded by the client, thereby determining the set of client participants, includes: A collaborative ratio formula is constructed based on the amount of data uploaded by the client; Determine the incentive cost of the client and sort it in ascending order to obtain the sorted client incentive cost; Set a threshold ratio, and divide the clients whose incentive costs satisfy the collaboration ratio into the client participant set according to the sorted client incentive costs; The specific expression for the communication trigger condition is as follows: ； In the above formula, This indicates the time elapsed from the last communication to the current round of communication. Indicates the discrimination threshold. This represents the covariance matrix containing the uploaded data. This represents the covariance matrix of the first type of data.

Images