Incentive method and device based on rule alignment under federal data control

Abstract

The present application discloses a rule-aligned incentive method and device under federated data control, which first specifies the federated data control rules, obtains the user's behavior data from the first federated node and the second federated node at the same time, and uses the data processing technology based on deep learning to perform time series analysis on the user behavior data to capture the user behavior characteristics of the first federated node and the second federated node respectively, and then, by performing semantic association analysis on the user behavior characteristics of the first federated node and the user behavior characteristics of the second federated node with the federated data control rules respectively, the incentive results of the federated nodes are intelligently determined. In this way, by quantifying the consistency of user behavior with the federated data control rules to adjust the weight distribution of each node in the federated learning process, fair incentives for federated learning participants are achieved, which can effectively promote the efficient use of data resources in the federated learning environment and improve the overall performance and privacy of federated learning.

Description

Technical Field

[0001] The present application relates to the field of intelligent incentives, and more specifically, to a rule-aligned incentive method and device under federal data management. Background Art

[0002] With the advent of the big data era, data has become a new factor of production, and its value in various industries has become increasingly prominent. However, in practical applications, due to factors such as data privacy protection and regulatory restrictions, data is often distributed among different organizations or institutions, forming so-called "data islands". This not only hinders the effective integration and utilization of data resources, but also limits the depth and breadth of cross-institutional cooperation.

[0003] To solve this problem, federated learning has emerged as an emerging distributed learning framework. Federated learning allows multiple participants (i.e., federated nodes, usually clients with local data sets) to jointly train a shared learning model without directly exchanging raw data, which not only ensures data security and privacy, but also realizes the value mining of multi-source data.

[0004] However, there are also a series of challenges in the process of federated learning, such as how to implement an effective incentive mechanism while maintaining data privacy, so that each federated node can actively participate and contribute high-quality data. Specifically, in federated learning, the contribution of nodes directly affects the effect and efficiency of model training. Due to differences in factors such as data quality, computing power, and network conditions of each node, how to fairly and reasonably evaluate the contribution of nodes and design an effective incentive mechanism based on this to stimulate the participation enthusiasm and willingness of nodes to cooperate is an urgent problem to be solved. Traditional incentive mechanisms often rely on direct data exchange or sharing, which is no longer applicable in the scenario of federated learning that protects data privacy.

[0005] Therefore, a rule-aligned incentive method and device under federal data control is expected. Summary of the invention

[0006] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides an incentive method and device based on rule alignment under federal data control, which first specifies the federal data control rules, obtains the user's behavior data from the first federal node and the second federal node at the same time, and uses the data processing technology based on deep learning to perform time series analysis on the user behavior data to capture the user behavior characteristics of the first federal node and the second federal node respectively, and then, by performing semantic association analysis on the user behavior characteristics of the first federal node and the user behavior characteristics of the second federal node with the federal data control rules respectively, the incentive results of the federal node are intelligently determined. In this way, by quantifying the consistency of user behavior with federal data control rules to adjust the weight distribution of each node in the federated learning process, fair incentives for federated learning participants are achieved, which can effectively promote the efficient use of data resources in the federated learning environment and improve the overall performance and privacy of federated learning.

[0007] According to one aspect of the present application, a rule-aligned incentive method under federated data management is provided, which includes:

[0008] Specify the federated data control rules, and simultaneously obtain the time queue of the user behavior data of the first federated node and the time queue of the user behavior data of the second federated node;

[0009] Embedding the time queue of user behavior data of the first federated node and the time queue of user behavior data of the second federated node using a user behavior embedding matrix to obtain a time queue of embedded coding vectors of user behavior data of the first federated node and a time queue of embedded coding vectors of user behavior data of the second federated node;

[0010] Performing feature time series propagation aggregation on the time queue of the first federated node user behavior data embedded coding vector and the time queue of the second federated node user behavior data embedded coding vector respectively to obtain the first federated node user behavior time series propagation aggregation representation vector and the second federated node user behavior time series propagation aggregation representation vector;

[0011] Based on the semantic association between the first federal node user behavior temporal propagation aggregate representation vector and the second federal node user behavior temporal propagation aggregate representation vector relative to the federal data management rules, the incentive results of the first federal node and the second federal node are determined.

[0012] According to another aspect of the present application, a rule-based alignment incentive device under federated data management is provided, which includes:

[0013] A data collection module, used to specify federated data management rules and simultaneously obtain a time queue of user behavior data of a first federated node and a time queue of user behavior data of a second federated node;

[0014] An embedding coding module, configured to embed the time queue of the user behavior data of the first federated node and the time queue of the user behavior data of the second federated node using a user behavior embedding matrix to obtain a time queue of the first federated node user behavior data embedding coding vector and a time queue of the second federated node user behavior data embedding coding vector;

[0015] A feature time series propagation aggregation module, used to perform feature time series propagation aggregation on the time queue of the first federated node user behavior data embedded in the coding vector and the time queue of the second federated node user behavior data embedded in the coding vector respectively to obtain the first federated node user behavior time series propagation aggregation representation vector and the second federated node user behavior time series propagation aggregation representation vector;

[0016] An incentive result determination module is used to determine the incentive results of the first federal node and the second federal node based on the semantic association between the user behavior time series propagation aggregation representation vector of the first federal node and the user behavior time series propagation aggregation representation vector of the second federal node relative to the federal data management rules.

[0017] Compared with the prior art, the present application provides a rule-aligned incentive method and device under federated data control, which first specifies the federated data control rules, obtains the user's behavior data from the first federated node and the second federated node at the same time, and uses the data processing technology based on deep learning to perform time series analysis on the user behavior data to capture the user behavior characteristics of the first federated node and the second federated node respectively, and then, by performing semantic association analysis on the user behavior characteristics of the first federated node and the user behavior characteristics of the second federated node with the federated data control rules respectively, the incentive results of the federated nodes are intelligently determined. In this way, by quantifying the consistency of user behavior with the federated data control rules to adjust the weight distribution of each node in the federated learning process, fair incentives for federated learning participants are achieved, which can effectively promote the efficient use of data resources in the federated learning environment and improve the overall performance and privacy of federated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 It is a flowchart of a rule-aligned incentive method under federated data management according to an embodiment of the present application;

[0020] Figure 2 A schematic diagram of data flow of a rule-aligned incentive method under federated data control according to an embodiment of the present application;

[0021] Figure 3 It is a flowchart of sub-step S3 of the incentive method based on rule alignment under the federated data management according to an embodiment of the present application;

[0022] Figure 4 It is a block diagram of a rule-aligned incentive device under federal data control according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0024] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0025] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0026] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0027] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0028] In the technical solution of the present application, a rule-alignment-based incentive method under federal data management is proposed. Figure 1 This is a flowchart of a rule-aligned incentive method under federated data management according to an embodiment of the present application. Figure 2 FIG. 1 is a data flow diagram of a rule-aligned incentive method under federated data control according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to the example of the present application, the rule-aligned incentive method under the federal data control includes the following steps: S1, specifying the federal data control rules, and simultaneously obtaining the time queue of the user behavior data of the first federal node and the time queue of the user behavior data of the second federal node; S2, using the user behavior embedding matrix to embed the time queue of the user behavior data of the first federal node and the time queue of the user behavior data of the second federal node to obtain the time queue of the embedded coding vector of the user behavior data of the first federal node and the time queue of the embedded coding vector of the user behavior data of the second federal node; S3, respectively performing feature time series propagation aggregation on the time queue of the embedded coding vector of the user behavior data of the first federal node and the time queue of the embedded coding vector of the user behavior data of the second federal node to obtain the first federal node user behavior time series propagation aggregation representation vector and the second federal node user behavior time series propagation aggregation representation vector; S4, based on the semantic association between the first federal node user behavior time series propagation aggregation representation vector and the second federal node user behavior time series propagation aggregation representation vector relative to the federal data control rules, determining the incentive results of the first federal node and the second federal node.

[0029] In particular, the S1 specifies the federal data control rules, and simultaneously obtains the time queue of the user behavior data of the first federal node and the time queue of the user behavior data of the second federal node. It should be understood that in the context of federated learning, in order to promote cooperation among different participants in accordance with certain rules or standards, it is first necessary to clarify the federal data control rules to guide and constrain the behavior of each node (i.e., data holder) participating in federated learning in the process of data collection, processing, transmission and use. In other words, the federal data control rules define the code of conduct of each federal node in the model training process, including but not limited to data quality standards, data usage specifications, and model update frequency. Therefore, it can be used as a contribution evaluation benchmark for the user behavior of each federal node, so as to facilitate corresponding feedback incentives based on the consistency of user behavior with the federal data control rules, and dynamically adjust the weights of each federal node in the model training process to improve the efficiency and quality of federated learning. It is worth mentioning that in this application, all actions to obtain information or data are carried out under the premise of complying with the corresponding data protection laws and policies of the country of residence and obtaining authorization from the owner of the corresponding device.

[0030] In particular, S2 uses the user behavior embedding matrix to embed the time queue of the user behavior data of the first federal node and the time queue of the user behavior data of the second federal node to obtain the time queue of the embedded coding vector of the user behavior data of the first federal node and the time queue of the embedded coding vector of the user behavior data of the second federal node. Considering that the user behavior data may contain a large amount of discrete, non-numerical behavior data, directly processing these high-dimensional data will increase the computational complexity and storage costs. Therefore, the present application adopts embedded coding technology to process each user behavior data separately, so as to map the high-dimensional raw data to a unified semantic feature space, capture the core semantic features of each user behavior, and convert it into a low-dimensional dense vector representation, effectively reducing the dimension of the data, thereby obtaining the time queue of the embedded coding vector of the user behavior data of the first federal node and the time queue of the embedded coding vector of the user behavior data of the second federal node, providing a more efficient data processing basis for subsequent user behavior evaluation.

[0031] In particular, the S3 performs feature time series propagation aggregation on the time queue of the first federated node user behavior data embedded in the coding vector and the time queue of the second federated node user behavior data embedded in the coding vector to obtain the first federated node user behavior time series propagation aggregation representation vector and the second federated node user behavior time series propagation aggregation representation vector. Since the user behavior data has obvious time series characteristics, that is, some behaviors of the user have a certain order and dependency in time. Therefore, in order to fully understand the changes in user behavior, the present application further performs time series analysis processing on the time queue of the first federated node user behavior data embedded in the coding vector and the time queue of the second federated node user behavior data embedded in the coding vector to reveal the dynamic characteristics of user behavior. It is worth mentioning that considering that the degree of correlation dependence between user behavior data at different time points is usually affected by the time interval, that is, the correlation between the behavior data at a closer time point and the current behavior data is usually stronger than the correlation between the behavior data at a distant time point. Based on this, the present application proposes a user behavior analysis method based on a time series decay mechanism, which introduces a time decay factor so that in the process of user behavior time series analysis, more attention is paid to recent user behavior characteristics, thereby more accurately capturing the time series change pattern of user behavior and improving the timeliness and accuracy of data analysis. In a specific example of this application, Figure 3As shown, the S3 includes: S31, dividing the time queue of the first federal node user behavior data embedded coding vector into time nodes and extracting the principal components to obtain the time queue of the current first federal node user behavior principal component representation vector and the historical first federal node user behavior principal component representation vector; S32, based on the time span of each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector relative to the current first federal node user behavior principal component representation vector, performing feature transfer aggregation based on temporal gradient attenuation on the time queue of the first federal node user behavior data embedded coding vector to obtain the first federal node user behavior temporal propagation aggregation representation vector.

[0032] Specifically, the S31 performs time node division and principal component extraction on the time queue of the first federal node user behavior data embedded coding vector to obtain the time queue of the current first federal node user behavior principal component representation vector and the historical first federal node user behavior principal component representation vector. That is, the time queue of the first federal node user behavior data embedded coding vector is divided into time nodes and principal component extraction is performed to obtain the time queue of the current first federal node user behavior principal component representation vector and the historical first federal node user behavior principal component representation vector; that is, the last time node is regarded as the current time node, and the other time nodes are regarded as historical time nodes, and the PCA analysis is performed on the first federal node user behavior data embedded coding vector at each time node to remove noise interference in the data, highlight key information, and extract the most representative principal component.

[0033] In an embodiment of the present application, the specific process of performing time node division and principal component extraction on the time queue of the first federal node user behavior data embedded coding vector includes: extracting the current first federal node user behavior data embedded coding vector from the time queue of the first federal node user behavior data embedded coding vector, and performing eigenvalue-based principal component extraction on the current first federal node user behavior data embedded coding vector to obtain the current first federal node user behavior principal component representation vector; defining other first federal node user behavior data embedded coding vectors in the time queue of the first federal node user behavior data embedded coding vector as historical first federal node user behavior data embedded coding vectors, and performing eigenvalue-based principal component extraction on each historical first federal node user behavior data embedded coding vector to obtain the time queue of the historical first federal node user behavior principal component representation vector.

[0034] Among them, the eigenvalue-based principal component extraction is performed on the current first federal node user behavior data embedded coding vector to obtain the current first federal node user behavior principal component representation vector, including: calculating the covariance matrix of the current first federal node user behavior data embedded coding vector to obtain the current first federal node user behavior covariance matrix; performing eigenvalue decomposition on the current first federal node user behavior covariance matrix to obtain a set of current first federal node user behavior principal component eigenvalues ​​and a corresponding set of current first federal node user behavior principal component eigenvectors; selecting the current first federal node user behavior principal component eigenvalue corresponding to the maximum eigenvalue in the set of current first federal node user behavior principal component eigenvalues ​​as the current first federal node user behavior principal component representation vector.

[0035] Specifically, the S32, based on the time span of each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector relative to the current first federal node user behavior principal component representation vector, performs feature transfer aggregation based on time gradient attenuation on the time queue of the first federal node user behavior data embedded encoding vector to obtain the first federal node user behavior time series propagation aggregation representation vector. Specifically, first, calculate the time series propagation absolute attenuation factor of each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector relative to the current first federal node user behavior principal component representation vector to obtain the time queue of the absolute attenuation factor of the time series propagation of the historical first federal node user behavior feature; that is, in order to capture the trend of the first federal node user behavior feature of each historical time node evolving over time and the information loss, further calculate the corresponding absolute attenuation factor of the time series propagation based on the principal component change rate of the behavior feature of each historical time node relative to the current time node, so as to enhance the model's ability to understand information transmission over long time intervals. Next, based on the time span between each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector and the current first federal node user behavior principal component representation vector, the time queue of the historical first federal node user behavior feature time series propagation absolute attenuation factor is modulated by time span to obtain a time queue of time span modulated propagation attenuation factor; the above time series propagation absolute attenuation factor is adjusted by introducing the time span factor to simulate the gradual forgetting process of data in the time dimension, so that the model can pay more attention to the most recent user behavior features and reduce excessive dependence on early data. Furthermore, based on the time queue of the time span modulated propagation attenuation factor, the time queue of the first federal node user behavior data embedded coding vector is subjected to feature time series transmission aggregation to obtain the first federal node user behavior time series propagation aggregation representation vector. Here, based on the gating mechanism, the information of the absolute attenuation factor of the time series propagation after time span modulation is screened, the most influential propagation path is identified and the corresponding weight is assigned accordingly to improve the distinguishability of the features, and the principal component features of the user behavior of the first federal node at each time node are weighted and aggregated based on the generated weights. In this way, while capturing the long-term change trend of the user behavior of the first federal node, the short-term dynamic pattern of the user behavior is effectively mined, and the time series propagation aggregation representation vector of the user behavior of the first federal node is generated.Through the above-mentioned data processing method, the time queue of the first federal node user behavior data embedded coding vector and the time queue of the second federal node user behavior data embedded coding vector are processed respectively, so as to obtain the first federal node user behavior timing propagation aggregation representation vector and the second federal node user behavior timing propagation aggregation representation vector containing the global timing correlation change information of user behavior.

[0036] Among them, the specific process of calculating the time series propagation absolute attenuation factor of each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector relative to the current first federal node user behavior principal component representation vector includes: calculating the dot division vector between the historical first federal node user behavior principal component representation vector and the current first federal node user behavior principal component representation vector, and calculating the base-two logarithmic value of the absolute value of each eigenvalue in the dot division vector to obtain the time series propagation attenuation absolute standard representation vector; calculating the dot product vector between the time series propagation attenuation absolute standard representation vector and the historical first federal node user behavior principal component representation vector, and calculating the exponential function value with base e and the sum of the eigenvalues ​​of the dot product vector as the exponent to obtain the time series propagation absolute attenuation factor of the historical first federal node user behavior feature.

[0037] More specifically, the specific process of performing time span modulation on the time queue of the absolute attenuation factor of the temporal propagation of the historical first federal node user behavior feature based on the time span between each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector and the current first federal node user behavior principal component representation vector includes: calculating the timestamp difference between the historical first federal node user behavior principal component representation vector and the current first federal node user behavior principal component representation vector, rounding down the timestamp difference to obtain the time span factor; dividing the temporal propagation absolute attenuation factor of the historical first federal node user behavior principal component representation vector by the natural exponential function value of the time span factor to obtain the time span modulated propagation attenuation factor.

[0038] More specifically, based on the time queue of the time span modulated propagation attenuation factor, the specific process of performing feature temporal transmission aggregation on the time queue of the first federal node user behavior data embedded coding vector includes: inputting the time queue of the time span modulated propagation attenuation factor into the information transmission screening module based on the gating function to obtain the time queue of the historical first federal node user behavior feature temporal gradient attenuation weights; based on the time queue of the historical first federal node user behavior feature temporal gradient attenuation weights, calculating the weighted sum of the time queue of the historical first federal node user behavior principal component representation vector to obtain the historical first federal node user behavior principal component significant transmission aggregation representation vector; calculating the positional sum of the historical first federal node user behavior principal component significant transmission aggregation representation vector and the current first federal node user behavior principal component representation vector to obtain the first federal node user behavior temporal propagation aggregation representation vector. Among them, the time queue of the time span modulated propagation attenuation factor is input into the information transmission screening module based on the gating function to obtain the time queue of the temporal gradient attenuation weight of the user behavior feature of the historical first federal node, including: based on a preset threshold, masking the time queue of the time span modulated propagation attenuation factor to obtain the time queue of the masked time span modulated propagation attenuation factor; inputting the time queue of the masked time span modulated propagation attenuation factor into the Sigmoid function for normalization processing to obtain the time queue of the temporal gradient attenuation weight of the user behavior feature of the historical first federal node.

[0039] In summary, in the above embodiment, similarly, taking the time queue of the first federal node user behavior data embedded in the coding vector as an example, the time queue of the first federal node user behavior data embedded in the coding vector is subjected to feature time series propagation aggregation, including: processing the time queue of the first federal node user behavior data embedded in the coding vector using the following information time series dynamic propagation formula to obtain the first federal node user behavior time series propagation aggregation representation vector, wherein the information time series dynamic propagation formula is:

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047] in, represents the time queue of the first federated node user behavior data embedded in the encoding vector, is the number of vectors in the time queue of the first federated node user behavior data embedded encoding vector, represents the current first federation node user behavior data embedding coding vector, , and They respectively represent the first, second and third in the time queue of the historical first federal node user behavior data embedded in the encoding vector The historical first federation node user behavior data embedding coding vector, represents the principal component analysis function, Indicates The principal component representation vector of the user behavior of the first historical federated node, Represents the principal component representation vector of the user behavior of the current first federal node, represents the transpose of a vector, represents the length of the embedding encoding vector of the user behavior data of the first historical federated node, Indicates the The covariance matrix of the historical first federated node user behavior data embedded in the encoding vector, that is, the historical first federated node user behavior covariance matrix, Indicates that through The matrix composed of the set of principal component eigenvectors of the historical first federal node user behavior obtained by eigenvalue decomposition of the historical first federal node user behavior covariance matrix, Indicates that by The diagonal matrix composed of the set of the principal component eigenvalues ​​of the historical first federal node user behavior obtained by eigenvalue decomposition of the historical first federal node user behavior covariance matrix, is the diag function, which is used to construct a diagonal matrix. , , and Respectively represent The first, second, and third and The principal component eigenvalues ​​of the user behavior of the first historical federation node, is the number of eigenvalues ​​in the set of principal component eigenvalues ​​of the user behavior of the first historical federated node, , , and Respectively represent The first, second, and third and The principal component feature vector of the user behavior of the first historical federation node, Indicates the index corresponding to the maximum value. represents the index of the maximum eigenvalue in the set of principal component eigenvalues ​​of the user behavior of the first historical federated node, Indicates The first historical federated node user behavior principal component representation vector eigenvalues, The first principal component representation vector of the user behavior of the current first federal node eigenvalues, represents the absolute value, represents the logarithmic function with base 2, represents the length of the principal component representation vector of the user behavior of the current first federal node, represents the exponential function with base e, Indicates The absolute attenuation factor of the time series propagation of the user behavior characteristics of the first historical federation node, and Respectively represent the current first federation node user behavior data embedding coding vector and the The timestamp of the first historical federated node user behavior data embedded in the encoding vector, Indicates rounding down. Indicates The time span modulates the propagation attenuation factor, Indicates mask processing, represents the sigmoid function, Indicates the preset threshold value, Indicates The time-series gradient attenuation weight of the user behavior characteristics of the first historical federation node, A representation vector representing the time series propagation aggregation of the user behavior of the first federated node.

[0048] In particular, the S4 determines the incentive results of the first federated node and the second federated node based on the semantic association between the first federated node user behavior time series propagation aggregation representation vector and the second federated node user behavior time series propagation aggregation representation vector relative to the federated data management rule. That is, in the technical solution of the present application, in order to realize the contribution evaluation of the user behavior of each federated node in federated learning, the first federated node user behavior time series propagation aggregation representation vector and the second federated node user behavior time series propagation aggregation representation vector are further compared and analyzed with the behavior criteria defined by the federated data management rule.

[0049] Specifically, first, the federal data control rules are semantically encoded to obtain a semantic encoding vector of the federal data control rules. That is, in order to convert the federal data control rules into a unified feature vector form, so as to facilitate comparative analysis with the user behavior characteristics of each federal node, the present application further uses a pre-trained language model to semantically encode the federal data control rules to extract the contextual semantic information in the federal data control rules, and convert the rule text into a vector representation in the semantic feature space, thereby obtaining a semantic encoding vector of the federal data control rules. In a specific example of the present application, a pre-trained language model based on the Transformer architecture is used to implement this semantic encoding process.

[0050] Next, the first federal node user behavior temporal propagation aggregation representation vector and the federal data control rule semantic encoding vector are input into the behavior-rule association network to obtain the first federal node user behavior-rule association feature matrix; and, the second federal node user behavior temporal propagation aggregation representation vector and the federal data control rule semantic encoding vector are input into the behavior-rule association network to obtain the second federal node user behavior-rule association feature matrix. Specifically, in the technical solution of the present application, the behavior-rule association network captures the complex interactive relationship between user behavior features and control rule semantic features by calculating the Kronecker product between the user behavior temporal propagation aggregation representation vector of each federal node and the transposed vector of the federal data control rule semantic encoding vector, learns the intrinsic connection between the two, and generates the corresponding first federal node user behavior-rule association feature matrix and the second federal node user behavior-rule association feature matrix, thereby achieving a quantitative evaluation of the degree of association between user behavior and control rules.

[0051] Finally, the first federated node user behavior-rule association feature matrix and the second federated node user behavior-rule association feature matrix are respectively input into the classifier-based incentive module to obtain incentive results, and the incentive results are used to indicate whether to increase the weight in model training. In particular, the classifier can effectively mine the association pattern information between user behavior and control rules contained therein by learning the feature pattern of the input user behavior-rule association feature matrix, and classify user behavior accordingly to guide the weight adjustment of federated nodes in the federated learning model, thereby giving higher weights to federated nodes that meet the control rules during the model training process, so as to optimize the performance of the federated learning model.

[0052] That is, when the first federal node user behavior temporal propagation aggregation representation vector and the federal data management rule semantic coding vector respectively represent the temporal dynamic propagation aggregation characteristics of the user behavior data of the first federal node and the semantic coding characteristics of the federal data management rules, when they are input into the behavior-rule association network, the difference in semantic coding characteristics between the first federal node user behavior temporal propagation aggregation representation vector and the federal data management rule semantic coding vector will also lead to sparse distribution of behavior-rule associations, resulting in the lack of decision-making of association feature instances in the obtained first federal node user behavior-rule association feature matrix, thereby affecting the accuracy of the incentive results obtained through the classifier-based incentive module.

[0053] Preferably, inputting the first federated node user behavior-rule association feature matrix into a classifier-based incentive module to obtain the incentive result comprises:

[0054] Expanding the first federated node user behavior-rule association feature matrix into a first federated node user behavior-rule association feature vector;

[0055] Calculate the distance between each pair of eigenvalues ​​of the first federated node user behavior-rule association feature vector, such as the L2 distance, and take the square root of the distance to obtain the federated node user behavior-rule association feature distance representation matrix, that is,

[0056]

[0057] in, represents the user behavior-rule association feature vector of the first federated node, and The first federated node user behavior-rule association feature vector and Eigenvalues, The first federated node user behavior-rule association feature vector and The distance between eigenvalues, Represents the federated node user behavior-rule association feature distance representation matrix The eigenvalue of the position;

[0058] The federated node user behavior-rule association feature auto-association matrix of the first federated node user behavior-rule association feature vector as a row vector is obtained, that is, ,in, represents the user behavior-rule association feature vector of the first federated node, represents the transpose of a vector, represents matrix multiplication, Represents the self-correlation matrix of the user behavior-rule association features of the federated nodes;

[0059] The first federated node user behavior-rule association feature vector is matrix-multiplied with the federated node user behavior-rule association feature distance representation matrix to obtain a first-level mapping vector of the federated node user behavior-rule association feature, that is, ,in, represents the user behavior-rule association feature vector of the first federated node, represents the distance representation matrix of the user behavior-rule association feature of the federated node, represents matrix multiplication, Represents the first-level mapping vector of the user behavior-rule association feature of the federated node;

[0060] The first-level mapping vector of the federated node user behavior-rule association feature is matrix-multiplied by the matrix product of the distance representation matrix of the federated node user behavior-rule association feature and the self-association matrix of the federated node user behavior-rule association feature to obtain a multi-level mapping vector of the federated node user behavior-rule association feature ,in, Represents the first-level mapping vector of the user behavior-rule association feature of the federated node, represents the distance representation matrix of the user behavior-rule association feature of the federated node, represents matrix multiplication, represents the self-correlation matrix of the user behavior-rule association characteristics of the federated nodes, Represents the multi-level mapping vector of the user behavior-rule association features of the federated node;

[0061] The federated node user behavior-rule association feature multi-level mapping vector is interpolated with the federated node user behavior-rule association feature associated eigenvector composed of the eigenvalues ​​of the federated node user behavior-rule association feature auto-association matrix to obtain an optimized first federated node user behavior-rule association feature vector, wherein interpolation or zero padding is performed when the eigenvalue is insufficient;

[0062] The optimized first federated node user behavior-rule association feature vector is input into a classifier-based incentive module to obtain an incentive result.

[0063] Accordingly, by using the linear target mapping representation of the similarity distance representation matrix of the first federal node user behavior-rule association feature vector obtained by expanding the first federal node user behavior-rule association feature matrix, the self-correlation of the first federal node user behavior-rule association feature vector is instantiated with a secondary target mapping representation based on a multi-level distribution hierarchy, and the association mismatch negative impact factor is compensated by the association fusion kernel bias to improve the degree of instance determination of the eigenvalue of the first federal node user behavior-rule association feature vector under similarity constraints, that is, the significance of the eigenvalue as an instance for classification regression determination, and to improve the accuracy of the excitation result obtained by the first federal node user behavior-rule association feature vector through the classifier-based excitation module.

[0064] In summary, the incentive method based on rule alignment under the federated data control of the embodiment of the present application is explained, which first specifies the federated data control rules, obtains the user's behavior data from the first federated node and the second federated node at the same time, and uses the data processing technology based on deep learning to perform time series analysis on the user behavior data to capture the user behavior characteristics of the first federated node and the second federated node respectively, and then, by performing semantic association analysis on the user behavior characteristics of the first federated node and the user behavior characteristics of the second federated node with the federated data control rules respectively, the incentive results of the federated nodes are intelligently determined. In this way, by quantifying the consistency of user behavior with the federated data control rules to adjust the weight distribution of each node in the federated learning process, fair incentives for federated learning participants are achieved, which can effectively promote the efficient use of data resources in the federated learning environment and improve the overall performance and privacy of federated learning.

[0065] Furthermore, a rule-aligned incentive device under federal data management is also provided.

[0066] Figure 4 FIG. 1 is a block diagram of a rule-based alignment incentive device under federated data control according to an embodiment of the present application. Figure 4As shown, according to the example of the present application, the rule-aligned incentive device 300 under the federated data control includes: a data acquisition module 310, which is used to specify the federated data control rules and simultaneously obtain the time queue of the user behavior data of the first federated node and the time queue of the user behavior data of the second federated node; an embedding coding module 320, which is used to embed the time queue of the user behavior data of the first federated node and the time queue of the user behavior data of the second federated node using the user behavior embedding matrix to obtain the time queue of the first federated node user behavior data embedded coding vector and the time queue of the second federated node user behavior data embedded coding vector; a feature timing propagation aggregation module 330, which is used to perform feature timing propagation aggregation on the time queue of the first federated node user behavior data embedded coding vector and the time queue of the second federated node user behavior data embedded coding vector respectively to obtain the first federated node user behavior timing propagation aggregation representation vector and the second federated node user behavior timing propagation aggregation representation vector; an incentive result determination module 340, which is used to determine the incentive results of the first federated node and the second federated node based on the semantic association of the first federated node user behavior timing propagation aggregation representation vector and the second federated node user behavior timing propagation aggregation representation vector relative to the federated data control rules.

[0067] As described above, the incentive device 300 based on rule alignment under the federal data control according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with an incentive algorithm based on rule alignment under the federal data control. In one possible implementation, the incentive device 300 based on rule alignment under the federal data control according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the incentive device 300 based on rule alignment under the federal data control can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the incentive device 300 based on rule alignment under the federal data control can also be one of the many hardware modules of the wireless terminal.

[0068] Alternatively, in another example, the rule-based alignment incentive device 300 under federal data control and the wireless terminal may also be separate devices, and the rule-based alignment incentive device 300 under federal data control may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0069] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A rule-aligned incentive method under federated data management, characterized in that: include: Specify the federated data control rules, and simultaneously obtain the time queue of the user behavior data of the first federated node and the time queue of the user behavior data of the second federated node; Embedding the time queue of user behavior data of the first federated node and the time queue of user behavior data of the second federated node using a user behavior embedding matrix to obtain a time queue of embedded coding vectors of user behavior data of the first federated node and a time queue of embedded coding vectors of user behavior data of the second federated node; Performing feature time series propagation aggregation on the time queue of the first federated node user behavior data embedded coding vector and the time queue of the second federated node user behavior data embedded coding vector respectively to obtain the first federated node user behavior time series propagation aggregation representation vector and the second federated node user behavior time series propagation aggregation representation vector; Determine incentive results for the first federated node and the second federated node based on the semantic association between the first federated node user behavior temporal propagation aggregate representation vector and the second federated node user behavior temporal propagation aggregate representation vector relative to the federated data management rule; Among them, the time queue of the first federal node user behavior data embedded in the coding vector is subjected to feature temporal propagation aggregation, including: dividing the time queue of the first federal node user behavior data embedded in the coding vector by time nodes and extracting principal components to obtain a time queue of the current first federal node user behavior principal component representation vector and a time queue of the historical first federal node user behavior principal component representation vector; based on the time span of each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector relative to the current first federal node user behavior principal component representation vector, the time queue of the first federal node user behavior data embedded in the coding vector is subjected to feature transfer aggregation based on temporal gradient attenuation to obtain the first federal node user behavior temporal propagation aggregation representation vector.

2. The rule-aligned incentive method under federated data management according to claim 1, characterized in that: The time queue of the first federated node user behavior data embedded coding vector is divided into time nodes and the principal component is extracted to obtain the current first federated node user behavior principal component representation vector and the time queue of the historical first federated node user behavior principal component representation vector, including: Extracting the current first federal node user behavior data embedding coding vector from the time queue of the first federal node user behavior data embedding coding vector, and performing eigenvalue-based principal component extraction on the current first federal node user behavior data embedding coding vector to obtain the current first federal node user behavior principal component representation vector; The other first federal node user behavior data embedding coding vectors in the time queue of the first federal node user behavior data embedding coding vector are defined as historical first federal node user behavior data embedding coding vectors, and principal component extraction based on eigenvalues ​​is performed on each historical first federal node user behavior data embedding coding vector to obtain the time queue of the historical first federal node user behavior principal component representation vector.

3. The rule-aligned incentive method under federated data management according to claim 2, characterized in that: Performing principal component extraction based on eigenvalues ​​on the current first federal node user behavior data embedded coding vector to obtain the current first federal node user behavior principal component representation vector, including: Calculating the covariance matrix of the current first federal node user behavior data embedded coding vector to obtain the current first federal node user behavior covariance matrix; Performing eigenvalue decomposition on the current first federal node user behavior covariance matrix to obtain a set of current first federal node user behavior principal component eigenvalues ​​and a corresponding set of current first federal node user behavior principal component eigenvectors; The current first federal node user behavior principal component feature vector corresponding to the maximum eigenvalue in the set of the current first federal node user behavior principal component feature values ​​is selected as the current first federal node user behavior principal component representation vector.

4. The incentive method based on rule alignment under federated data management according to claim 3 is characterized in that: Based on the time span of each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector relative to the current first federal node user behavior principal component representation vector, the time queue of the first federal node user behavior data embedding encoding vector is subjected to feature transfer aggregation based on temporal gradient attenuation to obtain the first federal node user behavior temporal propagation aggregation representation vector, including: Calculate the time series propagation absolute attenuation factor of each historical first federal node user behavior principal component representation vector in the time series queue of the historical first federal node user behavior principal component representation vector relative to the current first federal node user behavior principal component representation vector to obtain a time series queue of the historical first federal node user behavior feature time series propagation absolute attenuation factor; Based on the time span between each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector and the current first federal node user behavior principal component representation vector, the time queue of the absolute attenuation factor of the temporal propagation of the historical first federal node user behavior feature is modulated by time span to obtain a time queue of the time span modulated propagation attenuation factor; Based on the time queue of the time span modulated propagation attenuation factor, feature time series transfer aggregation is performed on the time queue of the first federal node user behavior data embedded coding vector to obtain the first federal node user behavior time series propagation aggregation representation vector.

5. The rule-aligned incentive method under federated data management according to claim 4, characterized in that: The method comprises: performing time span modulation on the time queue of the absolute attenuation factor of the temporal propagation of the historical first federal node user behavior feature based on the time span between each historical first federal node user behavior principal component representation vector in the time queue of the historical first federal node user behavior principal component representation vector and the current first federal node user behavior principal component representation vector to obtain a time queue of the time span modulated propagation attenuation factor, including: Calculating a timestamp difference between the historical first federated node user behavior principal component representation vector and the current first federated node user behavior principal component representation vector, and rounding down the timestamp difference to obtain a time span factor; The time-series propagation absolute attenuation factor of the historical first federated node user behavior principal component representation vector is divided by the natural exponential function value of the time span factor to obtain the time span modulated propagation attenuation factor.

6. The rule-aligned incentive method under federated data management according to claim 5, characterized in that: Based on the time queue of the time span modulated propagation attenuation factor, feature time series transfer aggregation is performed on the time queue of the first federated node user behavior data embedded coding vector to obtain the first federated node user behavior time series propagation aggregation representation vector, including: Input the time queue of the time span modulated propagation attenuation factor into the information transmission screening module based on the gating function to obtain the time queue of the temporal gradient attenuation weight of the user behavior feature of the historical first federation node; Based on the time queue of the temporal gradient attenuation weight of the historical first federal node user behavior feature, calculating the weighted sum of the time queue of the historical first federal node user behavior principal component representation vector to obtain the historical first federal node user behavior principal component significant transfer aggregation representation vector; The position-wise sum of the historical first federated node user behavior principal component significant transfer aggregation representation vector and the current first federated node user behavior principal component representation vector is calculated to obtain the first federated node user behavior time series propagation aggregation representation vector.

7. The incentive method based on rule alignment under federal data management according to claim 6, characterized in that: Determining incentive results of the first federated node and the second federated node based on the semantic association of the first federated node user behavior temporal propagation aggregate representation vector and the second federated node user behavior temporal propagation aggregate representation vector relative to the federated data management rule includes: Performing semantic encoding on the federal data control rule to obtain a federal data control rule semantic encoding vector; Inputting the first federated node user behavior time series propagation aggregation representation vector and the federated data management rule semantic encoding vector into a behavior-rule association network to obtain a first federated node user behavior-rule association feature matrix; Inputting the second federated node user behavior time series propagation aggregation representation vector and the federated data management rule semantic encoding vector into the behavior-rule association network to obtain a second federated node user behavior-rule association feature matrix; The first federated node user behavior-rule association feature matrix and the second federated node user behavior-rule association feature matrix are respectively input into a classifier-based incentive module to obtain an incentive result, and the incentive result is used to indicate whether to increase the weight in model training.

8. The rule-aligned incentive method under federated data management according to claim 7, characterized in that: Inputting the first federated node user behavior time series propagation aggregation representation vector and the federated data management rule semantic encoding vector into the behavior-rule association network to obtain the first federated node user behavior-rule association feature matrix, including: The Kronecker product between the temporal propagation aggregation representation vector of the user behavior of the first federated node and the transposed vector of the semantic encoding vector of the federated data management rule is calculated to obtain the user behavior-rule association feature matrix of the first federated node.

9. A rule-based alignment incentive device under federated data control, used to execute the rule-based alignment incentive method under federated data control as claimed in any one of claims 1 to 8, characterized in that: include: A data collection module, used to specify federated data management rules and simultaneously obtain a time queue of user behavior data of a first federated node and a time queue of user behavior data of a second federated node; An embedding coding module, configured to embed the time queue of the user behavior data of the first federated node and the time queue of the user behavior data of the second federated node using a user behavior embedding matrix to obtain a time queue of the first federated node user behavior data embedding coding vector and a time queue of the second federated node user behavior data embedding coding vector; A feature time series propagation aggregation module, used to perform feature time series propagation aggregation on the time queue of the first federated node user behavior data embedded in the coding vector and the time queue of the second federated node user behavior data embedded in the coding vector respectively to obtain the first federated node user behavior time series propagation aggregation representation vector and the second federated node user behavior time series propagation aggregation representation vector; An incentive result determination module is used to determine the incentive results of the first federal node and the second federal node based on the semantic association between the user behavior time series propagation aggregation representation vector of the first federal node and the user behavior time series propagation aggregation representation vector of the second federal node relative to the federal data management rules.

Images