User value evaluation method and device, equipment, medium and product

Abstract

The invention provides a user value evaluation method and device, equipment, a medium and a product, and the method comprises the steps: encrypting a local user identifier to obtain a first encrypted identifier, and carrying out the safety intersection with a second encrypted identifier of a data holder to obtain an intersection user; sending a query request to a data holder, and obtaining a user statistical quantity to construct a feature distribution data set; based on the feature distribution data set, screening to obtain an optimal feature set, and according to the optimal feature set, carrying out federated learning with a data holder to construct a user value evaluation model; and obtaining a to-be-evaluated user set, and inputting the to-be-evaluated user set into the user value evaluation model to obtain a user value evaluation result. Therefore, on the premise of protecting the data privacy of each participant, the modeling accuracy can be improved, the user value can be accurately and efficiently evaluated, and the effects of accurate marketing and further improvement of the user satisfaction can be achieved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, device, medium and product for evaluating user value. Background Technology

[0002] In today's digital age, user value assessment has become a crucial tool for businesses to achieve precise marketing and improve customer satisfaction. As business volume and diversification continue to increase, effectively identifying and analyzing user value to provide targeted services has become an inevitable trend in enterprise reform and development. However, existing user value assessment methods suffer from numerous technical problems. Currently, many user value assessment methods rely primarily on rule-based filtering and the collection and analysis of personal information. This approach is often limited to low-dimensional information, resulting in non-standardized data and an inability to accurately assess user value.

[0003] Furthermore, while some companies attempt to analyze historical user data and build user profiles using data warehouses and data analytics tools, this approach is prone to privacy breaches when handling sensitive information. Although user data is stored and transmitted with encryption, decryption is required during computation, still posing a risk of privacy leaks. Simultaneously, existing technologies fail to effectively capture highly relevant features, resulting in low accuracy in user value scoring and failing to meet the actual needs of enterprises. Different companies hold diverse attribute data in varying formats, and directly using these massive datasets often incurs high computational and communication costs. Moreover, traditional rule-based screening and data analysis methods often rely on superficial user characteristics, failing to capture complex user behavior patterns and potential needs, leading to unsatisfactory evaluation results.

[0004] In summary, existing technologies have technical problems in user value assessment, including high privacy risks, low assessment accuracy, and insufficient processing efficiency. Summary of the Invention

[0005] This application provides a method, apparatus, device, medium, and product for evaluating user value, in order to solve the technical problems of high privacy risks, low evaluation accuracy, and insufficient processing efficiency in the prior art for evaluating user value.

[0006] To solve the above-mentioned technical problems, this application is implemented as follows:

[0007] In a first aspect, embodiments of this application provide a method for evaluating user value, the method being applied to a data querying party, the method comprising:

[0008] The user identifier held locally is encrypted to obtain a first encrypted identifier. A secure intersection operation is performed based on the first encrypted identifier and the second encrypted identifier provided by the data holder to obtain the intersection users jointly owned by the data queryer and the data holder.

[0009] Send a query request to the data holder and receive the fragmentation information returned by the data holder based on the query request. The query request carries multiple features of the user to be queried and the conditions that each feature to be queried must satisfy.

[0010] Based on the query request and the sharding information, determine the statistical number of users in the intersection users when each of the multiple features to be queried satisfies the corresponding conditions;

[0011] Based on multiple statistical quantities for multiple features to be queried, a feature distribution dataset is constructed. Based on the feature distribution dataset, multicollinearity analysis and correlation evaluation are performed to select the optimal feature set.

[0012] The optimal feature set is sent to the data holder, and federated learning is performed with the data holder based on the optimal feature set to jointly construct a user value assessment model.

[0013] Obtain the set of users to be evaluated, input the set of users to be evaluated into the user value evaluation model, and obtain the user value evaluation result output by the user value evaluation model.

[0014] Optionally, a set of users to be evaluated is obtained, and the set of users to be evaluated is input into the user value evaluation model to obtain the user value evaluation result output by the user value evaluation model, including:

[0015] Obtain a set of users to be evaluated, construct a polynomial commitment based on the set of users to be evaluated, and send the polynomial commitment to the data holder, wherein the data holder is used to verify the batch validity of the number of users in the set of users to be evaluated being greater than a preset threshold based on the polynomial commitment;

[0016] Based on the set of users to be evaluated, legality verification data is generated and sent to the data holder, who uses the data holder to verify the legality of the set of users to be evaluated.

[0017] After the batch verification and legality verification are passed, the set of users to be evaluated is input into the user value evaluation model to obtain the user value evaluation results output by the user value evaluation model.

[0018] Optionally, the sharding information includes: a second secret shard and a second statistical result shard. A query request is sent to the data holder, and the sharding information returned by the data holder based on the query request is received. Based on the query request and the sharding information, the statistical number of users among the intersecting users whose respective queried features satisfy the corresponding conditions is determined includes:

[0019] The query request is split into a first secret shard set, a portion of the first secret shards in the first secret shard set is sent to the data holder, and a portion of the second secret shards corresponding to the feature is received from the data holder;

[0020] Homomorphic computation is performed on the remaining first secret fragments locally excluding the aforementioned portion of the first secret fragments, and on the received portion of the second secret fragments from the data holder, to obtain the difference fragment between the condition and the feature;

[0021] The difference fragments are converted into tag fragments using a secure comparison protocol, and the tag fragments of each user in the intersection users are homomorphically accumulated to obtain the first statistical result fragment.

[0022] The system receives the second statistical result fragment obtained from the data holder, merges the first statistical result fragment and the second statistical result fragment, and performs fragment reconstruction to restore the statistical number of the user in plaintext.

[0023] Optionally, based on the aforementioned feature distribution dataset, multicollinearity analysis and correlation assessment are performed to select the optimal feature set, including:

[0024] Each feature in the feature distribution dataset is taken as a target feature, and the following operations are performed on the target feature: the target feature is taken as the dependent variable, and the remaining features in the feature distribution dataset other than the target feature are taken as independent variables to construct a linear regression model; the coefficient of determination is determined based on the linear regression model, and the variance inflation factor value of the target feature is calculated based on the coefficient of determination;

[0025] The variance inflation factor values ​​corresponding to each feature are compared with the preset variance inflation factor thresholds.

[0026] Based on the comparison results, all features with variance inflation factor values ​​higher than the variance inflation factor threshold are deleted from the feature distribution dataset to generate a deredundant feature set.

[0027] Based on the preset user value tags of the data query party, calculate the correlation coefficient between each feature in the redundancy removal feature set and the user value assessment result;

[0028] The optimal feature set is formed by selecting features from the deredundancy feature set whose correlation coefficient is higher than a preset correlation coefficient threshold.

[0029] Optionally, obtaining a set of users to be evaluated, constructing a multinomial commitment based on the set of users to be evaluated, and sending the multinomial commitment to the data holder includes:

[0030] The set of users to be evaluated is used as the coefficients of a polynomial to construct the polynomial;

[0031] Based on elliptic curve cryptography, a polynomial commitment is generated for the polynomial, and the polynomial commitment is sent to the data holder.

[0032] Optionally, after obtaining the set of users to be evaluated, constructing a multinomial commitment based on the set of users to be evaluated, and sending the multinomial commitment to the data holder, the method further includes:

[0033] In response to a random challenge point sent by the data holder based on the polynomial commitment;

[0034] Calculate the value of the polynomial at the random challenge point, construct a quotient polynomial based on the value, and generate a commitment for the quotient polynomial;

[0035] The value and the commitment of the quotient polynomial are sent to the data holder, wherein the data holder is used to verify the batch nature of the number of users in the user set to be evaluated being greater than a preset threshold based on the value and the commitment of the quotient polynomial.

[0036] Optionally, based on the set of users to be evaluated, generating legitimacy verification data and sending the legitimacy verification data to the data holder includes:

[0037] Recursive hash calculations are performed on the set of users to be evaluated to construct a Merkle tree, and the root hash of the Merkle tree is calculated.

[0038] Based on the Merkle tree, a corresponding Merkle path is generated for each user in the set of users to be evaluated;

[0039] The root hash, all the Merkel paths, and the set of users to be evaluated are sent as the legitimacy verification data to the data holder, wherein the data holder is used to verify the legitimacy of the set of users to be evaluated based on the legitimacy verification data.

[0040] Optionally, in the presence of multiple data holders, receiving a portion of the second secret fragment corresponding to the feature from the data holder includes:

[0041] Receive Reed-Solomon encoded fragments from each data holder, where each encoded fragment is generated by the corresponding data holder by encoding all local second secret fragments into a polynomial;

[0042] Based on the aforementioned coded segments, the Berlekamp-Welch algorithm is used to construct the error localization polynomial;

[0043] Calculate the roots of the error-locating polynomial to pinpoint the evaluation point of the malicious data holder who provided the erroneous second secret fragment;

[0044] A set of repair equations is constructed based on the evaluation points and the redundancy information of the coded segments;

[0045] Solve the repair equations to recover the correct second secret fragment corresponding to the malicious data holder.

[0046] Optionally, after the batch verification and legality verification are passed, the method further includes inputting the set of users to be evaluated into the user value evaluation model and obtaining the user value evaluation results output by the user value evaluation model:

[0047] Based on the value assessment results, the users in the set of users to be assessed are divided into different value levels;

[0048] Implement differentiated marketing strategies for users at different value levels.

[0049] Secondly, embodiments of this application provide a method for evaluating user value, the method being applied to a data holder, the method comprising:

[0050] The user identifier held locally is encrypted to obtain a second encrypted identifier. A secure intersection operation is performed between the second encrypted identifier and the first encrypted identifier provided by the data query party to obtain the intersection users jointly owned by the data query party and the data holder.

[0051] The system receives a query request from the data querying party, generates sharding information based on the query request, and sends the sharding information to the data querying party. The query request carries multiple features to be queried by the user and the conditions that each feature to be queried must satisfy. The sharding information is used to determine the optimal feature set.

[0052] The system receives the optimal feature set sent by the data query party, and performs federated learning based on the optimal feature set and the data query party to jointly construct a user value assessment model; wherein, the data query party is used to input the set of users to be assessed into the user value assessment model to obtain the user value assessment result output by the user value assessment model.

[0053] Optionally, after jointly constructing the user value assessment model, the method further includes: receiving a polynomial commitment constructed based on the set of users to be assessed sent by the data query party, and verifying that the number of users in the set of users to be assessed is greater than a preset threshold batching based on the polynomial commitment;

[0054] Receive the legality verification data generated based on the set of users to be evaluated sent by the data query party, and verify the legality of the set of users to be evaluated based on the legality verification data;

[0055] After the batch verification and legality verification are passed, a verification pass instruction is sent to the data query party. The data query party is used to input the set of users to be evaluated into the user value evaluation model after receiving the verification pass instruction, and obtain the user value evaluation result output by the user value evaluation model.

[0056] Optionally, the sharding information includes: a second secret shard and a second statistical result shard. Receiving a query request sent by the data querying party, generating sharding information based on the query request, and sending the sharding information to the data querying party includes:

[0057] Receive a portion of the first secret fragment corresponding to the query request from the data querying party;

[0058] The user's characteristics held by itself are split into a second secret shard set, and a portion of the second secret shards in the second secret shard set are sent to the data query party;

[0059] Homomorphic computation is performed on the remaining second secret fragments locally excluding the aforementioned portion of the second secret fragments, and on the received portion of the first secret fragments from the data query party, to obtain the difference fragment between the condition and the feature;

[0060] The difference fragments are converted into tag fragments using a secure comparison protocol, and the tag fragments of each user in the intersection users are homomorphically accumulated to obtain the second statistical result fragments.

[0061] The second statistical result fragment is sent to the data query party, wherein the data query party is used to merge the second statistical result fragment with the first statistical result fragment obtained by itself, and perform fragment reconstruction to restore the statistical number of the user in plaintext.

[0062] Optionally, in the case of multiple data holders, after splitting the characteristics of the users held by each holder into a second set of secret shards, the method further includes:

[0063] Generate random numbers, and send each of the second secret fragments in the second secret fragments together with the generated random numbers to other data holders, while receiving the second secret fragments from other data holders and the random numbers generated by other data holders;

[0064] All locally held second secret fragments are encoded into a polynomial using Reed-Solomon encoding to generate encoded fragments with redundant information;

[0065] The encoded fragment is sent to the data query party; wherein, the data query party is used to locate the evaluation point of the malicious data holder who provided the incorrect second secret fragment based on the encoded fragment and the Berlekamp-Welch algorithm, construct a set of repair equations based on the evaluation point and the encoded fragment, and recover the correct second secret fragment by solving the set of repair equations.

[0066] Optionally, verifying the batch validity of the set of users to be evaluated, based on the polynomial commitment, if the number of users is greater than a preset threshold includes:

[0067] Based on the polynomial commitment, a random challenge point is sent to the data query party;

[0068] Receive the commitment from the data query party regarding the value determined based on the random challenge point and the quotient polynomial;

[0069] Using bilinear pairing techniques, it is verified whether the polynomial commitment, the random challenge point, the value, and the commitment of the quotient polynomial satisfy a predetermined equation relationship, so as to verify that the number of users in the user set to be evaluated is greater than a preset threshold batch size.

[0070] Optionally, the legitimacy data includes: the root hash of the Merkel tree generated by the data querying party based on the set of users to be evaluated, the Merkel paths of all users in the set of users to be evaluated, and the set of users to be evaluated; verifying the legitimacy of the set of users to be evaluated based on the legitimacy verification data includes:

[0071] For each user in the set of users to be evaluated, the legitimacy of the user is verified by reconstructing the hash path based on the Merkel path corresponding to the user and the root hash.

[0072] Thirdly, embodiments of this application provide a user value evaluation device, which is applied to a data querying party, and the device includes:

[0073] The first secure intersection module is used to encrypt the user identifier held locally to obtain a first encrypted identifier, and perform a secure intersection operation based on the first encrypted identifier and the second encrypted identifier provided by the data holder to obtain the intersection users jointly owned by the data query party and the data holder.

[0074] The sending module is used to send a query request to the data holder and receive the fragmentation information returned by the data holder according to the query request. The query request carries multiple features to be queried by the user and the conditions that each feature to be queried must satisfy.

[0075] The determination module is used to determine, based on the query request and the sharding information, the statistical number of users in the intersection users when each of the multiple features to be queried satisfies the corresponding conditions.

[0076] The building module is used to construct a feature distribution dataset based on multiple statistical counts for multiple features to be queried;

[0077] The filtering module is used to perform multicollinearity analysis and correlation evaluation based on the feature distribution dataset to filter out the optimal feature set;

[0078] The first model building module is used to send the optimal feature set to the data holder, and perform federated learning with the data holder based on the optimal feature set to jointly build a user value assessment model.

[0079] The value assessment module is used to input the set of users to be assessed into the user value assessment model and obtain the user value assessment results output by the user value assessment model.

[0080] Fourthly, embodiments of this application provide a user value assessment device, which is applied to a data holder, and the device includes:

[0081] The second secure intersection module is used to encrypt the user identifier held locally to obtain a second encrypted identifier, and perform a secure intersection operation based on the second encrypted identifier and the first encrypted identifier provided by the data query party to obtain the intersection users jointly owned by the data query party and the data holder.

[0082] A receiving module is used to receive a query request sent by the data querying party, generate sharding information according to the query request, and send the sharding information to the data querying party. The query request carries multiple features to be queried by the user and the conditions that each feature to be queried must satisfy. The sharding information is used to determine the optimal feature set.

[0083] The second model construction module is used to receive the optimal feature set sent by the data query party, and to perform federated learning based on the optimal feature set and the data query party to jointly construct a user value assessment model; wherein, the data query party is used to input the set of users to be assessed into the user value assessment model to obtain the user value assessment result output by the user value assessment model.

[0084] Fifthly, embodiments of this application provide a network device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of a user value assessment method as described in the first or second aspect.

[0085] In a sixth aspect, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a user value assessment method as described in the first or second aspect.

[0086] In a seventh aspect, embodiments of this application provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps of a user value assessment method as described in the first or second aspect.

[0087] In this embodiment, taking the data query party as an example, the data query party encrypts the user identifiers held locally and performs secure intersection with the encrypted identifiers provided by the data holder. Only the users shared by both parties are determined within the encrypted space, avoiding the exposure of the original user identifiers during the intersection process and reducing the risk of user identity information leakage from the source. On this basis, the data query party sends a query request to the data holder carrying multi-dimensional query features and their respective conditions. Both parties count the number of users who meet the conditions for each feature without disclosing plaintext data and construct a feature distribution dataset, so that the distribution of users in the key feature dimensions of the intersection can be fully grasped before model training. Subsequently, the data query party performs multicollinearity analysis and correlation evaluation based on the feature distribution dataset, removes highly collinear redundant features and retains the optimal feature set that is highly correlated with user value, thereby mitigating the impact of multicollinearity on the stability and generalization ability of parameter estimation. Then, the optimal feature set is used to jointly model with the data holder under the federated learning framework, fully integrating multi-source feature information without leaving the domain of the original data of each party, and constructing a user value assessment model that is more in line with the true value of users.

[0088] Therefore, while protecting the data privacy of all participants, the accuracy and stability of modeling can be improved, enabling accurate and efficient assessment of user value, and achieving the technical effect of precision marketing and thus improving user satisfaction. Attached Figure Description

[0089] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0090] Figure 1 A flowchart illustrating a user value assessment method provided in this application embodiment;

[0091] Figure 2 A flowchart illustrating a user value assessment method provided in this application embodiment;

[0092] Figure 3 This is a schematic diagram illustrating the interaction between a data queryer and a data holder, provided as an embodiment of this application.

[0093] Figure 4 A schematic diagram of a polynomial verification process provided for an embodiment of this application;

[0094] Figure 5 A structural block diagram of a user value assessment device provided in an embodiment of this application;

[0095] Figure 6 A structural block diagram of a user value assessment device provided in an embodiment of this application;

[0096] Figure 7 This is a structural block diagram of a network device provided in an embodiment of this application. Detailed Implementation

[0097] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0098] Figure 1 This application illustrates a method for evaluating user value according to an embodiment of the present application. The method is applied to a data query party and includes:

[0099] Step S101: Encrypt the user identifier held locally to obtain the first encrypted identifier. Perform a secure intersection operation based on the first encrypted identifier and the second encrypted identifier provided by the data holder to obtain the intersection users jointly owned by the data queryer and the data holder.

[0100] Step S102: Send a query request to the data holder and receive the fragmentation information returned by the data holder based on the query request;

[0101] The query request carries multiple features of the user to be queried and the conditions that each feature must satisfy.

[0102] Step S103: Based on the query request and sharding information, determine the statistical number of users in the intersection users when each of the multiple features to be queried satisfies the corresponding conditions.

[0103] Step S104: Construct a feature distribution dataset based on multiple statistical quantities for multiple features to be queried;

[0104] Step S105: Based on the feature distribution dataset, perform multicollinearity analysis and correlation assessment to select the optimal feature set;

[0105] Step S106: Send the optimal feature set to the data holder, and based on the optimal feature set, perform federated learning with the data holder to jointly construct a user value assessment model.

[0106] Step S107: Obtain the set of users to be evaluated, input the set of users to be evaluated into the user value evaluation model, and obtain the user value evaluation results output by the user value evaluation model.

[0107] It should be noted that, Figure 1 The method described revolves around the overall business process of the data querying party. First, a first encrypted identifier is obtained by encrypting the user identifier held locally. Then, a secure set intersection (PSI) operation is performed between the first encrypted identifier and a second encrypted identifier provided by the data holder. This determines the intersection users shared by both the data querying party and the data holder only within the encrypted space, thus limiting the user scope for subsequent calculations without exposing the original user identifiers. Subsequently, the data querying party sends a query request to the data holder, carrying multiple user features to be queried and the conditions that each feature must satisfy. Both parties, focusing on the intersection users, obtain the statistical number of users who satisfy each of the multiple queried features according to the query request, without revealing the original plaintext features. Based on this, a feature distribution dataset for the multiple queried features is constructed, allowing the data querying party to understand the distribution of the intersection users across each feature dimension.

[0108] Building upon this foundation, the data querying party performs multicollinearity analysis and correlation assessment on the feature distribution dataset, eliminating highly correlated redundant features and selecting the optimal feature set more closely related to the user value assessment results. This improves the stability and interpretability of subsequent modeling. Next, based on the optimal feature set, the data querying party and the data holder engage in federated learning (FL) to jointly construct a user value assessment model, enabling multi-party data to participate in training without leaving their respective domains. In the model usage phase after constructing the user value assessment model, the set of users to be evaluated can be input into the model to obtain the user value assessment results.

[0109] It should also be noted that the user value assessment model can be deployed by both parties or a third party, and can be configured according to actual needs.

[0110] Through the above overall process, while protecting the data privacy of all participants and the legitimacy of the user set, the modeling accuracy of the user value assessment model can be improved, accurately and efficiently assessing user value and providing a reliable foundation for subsequent precision marketing.

[0111] In one possible implementation, a set of users to be evaluated is obtained, and this set is input into a user value evaluation model. The user value evaluation results output by the user value evaluation model include:

[0112] Obtain the set of users to be evaluated, construct a multinomial commitment based on the set of users to be evaluated, and send the multinomial commitment to the data holder; wherein, the data holder is used to verify the batch validity of the number of users in the set of users to be evaluated being greater than a preset threshold based on the multinomial commitment;

[0113] Based on the set of users to be evaluated, legality verification data is generated and sent to the data holder, who uses it to verify the legality of the set of users to be evaluated.

[0114] After passing batch verification and legality verification, the set of users to be evaluated is input into the user value evaluation model to obtain the user value evaluation results output by the user value evaluation model.

[0115] It should be noted that during the user value assessment model usage phase, the data querying party obtains the set of users to be assessed, constructs a multinomial commitment based on the set, and sends it to the data holder. The data holder verifies the batch size of the user set to be assessed based on the multinomial commitment to ensure that the number of users exceeds a preset threshold, thus avoiding sensitive queries targeting individual users. Simultaneously, the data querying party generates legality verification data based on the user set to be assessed and sends it to the data holder, who verifies the legality of the user set to be assessed, thereby ensuring that the user set participating in the prediction is genuine and valid. Only after both batch size verification and legality verification pass can the data querying party input the user set to be assessed into the user value assessment model to obtain the user value assessment result output by the user value assessment model.

[0116] Therefore, by implementing batch verification and legality verification on the set of users to be evaluated, malicious queries at the individual user level can be effectively limited during the model usage phase, reducing the risk of model reverse inference and member inference attacks.

[0117] In one possible implementation, the sharding information includes: a second secret shard and a second statistical result shard. A query request is sent to the data holder, and the sharding information returned by the data holder based on the query request is received. Based on the query request and the sharding information, the statistical number of users among the intersecting users who satisfy the corresponding conditions for each of the multiple queried features is determined, including:

[0118] The query request is split into a first secret fragment set. A portion of the first secret fragments in the first secret fragment set is sent to the data holder, and a portion of the second secret fragments corresponding to the features are received from the data holder. Homomorphic computation is performed on the remaining first secret fragments (excluding the portion of the first secret fragments) locally, as well as the portion of the second secret fragments received from the data holder, to obtain the difference fragments between the conditions and the features. The difference fragments are converted into tag fragments using a secure comparison protocol, and the tag fragments of each user in the intersection are homomorphically accumulated to obtain the first statistical result fragments. The second statistical result fragments obtained from the data holder are received, the first statistical result fragments and the second statistical result fragments are merged, and fragment reconstruction is performed to restore the statistical number of users in plaintext.

[0119] It should be noted that this possible implementation provides a specific privacy-preserving computing (PPC) process based on secret sharding and homomorphic computation for the implementation details of "obtaining the statistical number of users when each of the multiple features to be queried satisfies the corresponding conditions according to the query request".

[0120] Specifically, after generating a query request, the data querying party splits the query request into a first set of secret fragments, sends a portion of the first secret fragments to the data holder, and receives a portion of the second secret fragments corresponding to the queried feature from the data holder, thus forming a shared secret fragment structure regarding conditions and features between the two parties. Based on this, the data querying party performs homomorphic computation on the remaining first secret fragments (excluding a portion of the first secret fragments) locally, as well as the received portion of the second secret fragments from the data holder, to obtain the difference fragments between the conditions and features, ensuring that the condition judgment logic is completed within the ciphertext domain. Next, the data querying party converts the difference fragments into flag fragments using a secure comparison protocol, and homomorphically accumulates the flag fragments of each user in the intersection users to obtain the first statistical result fragment representing the statistical contribution of that side; simultaneously, the data holder obtains the second statistical result fragment based on a symmetric process. Ultimately, the data querying party receives the second statistical result fragment from the data holder, merges the first and second statistical result fragments, and performs fragment reconstruction to restore the statistical number of users in plaintext, thereby completing the accurate statistics of the number of users who meet the conditions without disclosing the original feature data and condition plaintext of all parties.

[0121] Therefore, by combining the first secret sharding, the second secret sharding, homomorphic computation, homomorphic accumulation, and shard reconstruction, the privacy and correctness of the user statistical quantity calculation process are guaranteed, providing a safe and reliable data foundation for the subsequent construction of the feature distribution dataset.

[0122] In one possible implementation, multicollinearity analysis and correlation assessment are performed on the feature distribution dataset to select the optimal feature set. This includes: treating each feature in the feature distribution dataset as a target feature, and performing the following operations on the target features: constructing a linear regression model with the target feature as the dependent variable and the remaining features in the feature distribution dataset as independent variables; determining the coefficient of determination based on the linear regression model, and calculating the VIF (Variance Inflation Factor) value of the target feature based on the coefficient of determination; comparing the VIF values ​​of each feature with a preset VIF threshold; deleting all features with VIF values ​​higher than the VIF threshold from the feature distribution dataset based on the comparison results, generating a deredundant feature set; calculating the correlation coefficient between each feature in the deredundant feature set and the user value assessment result based on the preset user value tags of the data query party; and selecting features with correlation coefficients higher than a preset correlation coefficient threshold from the deredundant feature set to form the optimal feature set.

[0123] It should be noted that this possible implementation method specifically limits the implementation method of "conducting multicollinearity analysis and correlation evaluation based on feature distribution datasets to select the optimal feature set", and provides a feature selection scheme based on VIF value and correlation coefficient.

[0124] Specifically, the data queryer treats each feature in the feature distribution dataset as a target feature and performs the following operations on each target feature: A linear regression model is constructed using the target feature as the dependent variable and all other features in the feature distribution dataset as independent variables. The coefficient of determination is determined based on the linear regression model, and the VIF value of the target feature is calculated based on the coefficient of determination to quantify the degree of multicollinearity between the feature and other features. Subsequently, the VIF values ​​corresponding to each feature are compared with a preset VIF threshold. Based on the comparison results, all features with VIF values ​​higher than the VIF threshold are deleted from the feature distribution dataset, generating a de-redundant feature set, thereby removing highly collinear redundant features. Further, based on local preset user value tags, the data queryer calculates the correlation coefficient between each feature in the de-redundant feature set and the user value assessment result, and selects features with correlation coefficients higher than a preset correlation coefficient threshold from the de-redundant feature set to form the optimal feature set, achieving fine-grained selection of features that are neither overly redundant nor lack strong explanatory power for user value.

[0125] Therefore, without increasing the risk of privacy leakage, the dual standard of VIF value and correlation coefficient can be used to effectively reduce multicollinearity among features and improve the stability, interpretability and predictive accuracy of the user value assessment model.

[0126] In one possible implementation, obtaining the set of users to be evaluated, constructing a polynomial commitment based on the set of users to be evaluated, and sending the polynomial commitment to the data holder includes: using the set of users to be evaluated as coefficients of a polynomial to construct the polynomial; generating a polynomial commitment based on elliptic curve cryptography, and sending the polynomial commitment to the data holder.

[0127] It should be noted that this possible implementation provides a detailed description of the implementation details of "obtaining the set of users to be evaluated, constructing a multinomial commitment based on the set of users to be evaluated, and sending the multinomial commitment to the data holder".

[0128] Specifically, after obtaining the set of users to be evaluated, the data querying party uses this set as the coefficients of a polynomial to construct the polynomial. This allows the user information in the set to be encoded and uniformly mapped to the polynomial coefficient space. Subsequently, the data querying party generates a polynomial commitment based on elliptic curve cryptography. Through a cryptographic commitment mechanism, the coefficients of the polynomial are bound together into a commitment value that is difficult to forge. This polynomial commitment is then sent to the data holder, enabling the data holder to perform batch verification based solely on the polynomial commitment during subsequent challenges and verifications, without needing to know the specific plaintext content of the set of users to be evaluated.

[0129] Therefore, by using multinomial commitments to encode the set of users to be evaluated into a single commitment value, the privacy of the set of users to be evaluated is protected, while providing a reliable basis for subsequent batch verification based on random challenge points, thus laying the foundation for preventing malicious queries against a single user.

[0130] In one possible implementation, after obtaining the set of users to be evaluated, constructing a polynomial commitment based on the set of users to be evaluated, and sending the polynomial commitment to the data holder, the method further includes: responding to a random challenge point sent by the data holder based on the polynomial commitment; calculating the value of the polynomial at the random challenge point, constructing a quotient polynomial based on the value, and generating a commitment of the quotient polynomial; sending the value and the commitment of the quotient polynomial to the data holder, wherein the data holder uses the value and the commitment of the quotient polynomial to verify the batching of the number of users in the set of users to be evaluated to be greater than a preset threshold.

[0131] It should be noted that this possible implementation further limits the interaction process for batch verification based on polynomial commitments. After obtaining the set of users to be evaluated, constructing a polynomial commitment, and sending it to the data holder, the data querying party responds to the random challenge point sent by the data holder based on the polynomial commitment, calculates the value of the polynomial at the random challenge point, constructs a quotient polynomial based on the value, and then generates a commitment for the quotient polynomial. Subsequently, the data querying party sends the value and the commitment for the quotient polynomial to the data holder. The data holder, based on this, combines the original polynomial commitment, the random challenge point, and the quotient polynomial commitment, and verifies whether the degree of the polynomial corresponding to the set of users to be evaluated satisfies the constraint through a predetermined equation relationship, thereby verifying the batch validity of the set of users to be evaluated, which has a user count greater than a preset threshold.

[0132] Therefore, by using this verification method based on random challenge points and quotient polynomial commitments, the data holder can confirm that the query operation is indeed aimed at a sufficiently large set of users rather than a single user without disclosing the specific members of the user set to be evaluated. This effectively reduces the risk of model reverse inference and member inference attacks, and enhances the privacy protection capabilities of the user value evaluation model in the inference stage.

[0133] In one possible implementation, generating legitimacy verification data based on the set of users to be evaluated and sending the legitimacy verification data to the data holder includes: performing recursive hash calculations on the set of users to be evaluated to construct a Merkle tree and calculating the root hash of the Merkle tree; generating a corresponding Merkle path for each user in the set of users to be evaluated based on the Merkle tree; and sending the root hash, all Merkle paths, and the set of users to be evaluated as legitimacy verification data to the data holder, wherein the data holder uses the legitimacy verification data to verify the legitimacy of the set of users to be evaluated.

[0134] It should be noted that this possible implementation method provides a detailed explanation of the implementation method of "generating legality verification data based on the set of users to be evaluated and sending the legality verification data to the data holder", and introduces a Merkle tree structure for user set legality verification.

[0135] The data querying party first performs recursive hash calculations on the set of users to be evaluated to construct a Merkle tree and calculates the root hash of the Merkle tree, ensuring that each user's hash value is mapped layer by layer to a unique root hash. Next, based on the Merkle tree, the data querying party generates a corresponding Merkle path for each user in the set to be evaluated, recording the hash link from the user's leaf node to the root hash. Subsequently, the data querying party sends the root hash, all Merkle paths, and the set of users to be evaluated as validity verification data to the data holder. The data holder can then reconstruct the hash path for each user using the user's corresponding Merkle path and the root hash, verifying whether it matches the root hash provided by the data querying party, thereby determining whether the user truly exists in the set of users to be evaluated.

[0136] By introducing a Merkle tree structure, without exposing the complete user list in plaintext, the data holder can effectively detect whether the set of users to be evaluated has been tampered with or forged, prevent forged or unauthorized users from being mixed into the prediction set, ensure the authenticity and integrity of the prediction object, and ensure that the user set participating in the model inference is legitimate and valid.

[0137] In one possible implementation, when multiple data holders exist, receiving a partial second secret fragment corresponding to a feature from each data holder includes: receiving a Reed-Solomon encoded fragment from each data holder, wherein each encoded fragment is generated by the corresponding data holder by encoding all local second secret fragments into a polynomial; constructing an error-locating polynomial based on the encoded fragment using the Berlekamp-Welch algorithm; calculating the roots of the error-locating polynomial to locate the evaluation point of the malicious data holder providing the erroneous second secret fragment; constructing a system of repair equations based on the evaluation point and the redundancy information of the encoded fragment; and solving the system of repair equations to recover the correct second secret fragment corresponding to the malicious data holder.

[0138] It should be noted that this possible implementation is designed for multi-party scenarios with multiple data holders. It further extends the implementation steps of "receiving the second secret fragment corresponding to the feature from the data holder" by introducing Reed-Solomon encoding and the Berlekamp-Welch algorithm to achieve malicious data holder detection and error correction.

[0139] Specifically, in the presence of multiple data holders, the data querying party receives Reed-Solomon encoded fragments from each data holder. Each encoded fragment is generated by the corresponding data holder by encoding all local second-secret fragments into a polynomial, thereby introducing redundant information at the fragment level. Based on these encoded fragments, the data querying party constructs an error-location polynomial using the Berlekamp-Welch algorithm. By solving the roots of the error-location polynomial, the evaluation point of the malicious data holder providing the erroneous second-secret fragment is located. Subsequently, based on these evaluation points and the redundant information carried in the encoded fragments, the data querying party constructs a system of repair equations and recovers the correct second-secret fragment of the corresponding malicious data holder by solving the system of repair equations.

[0140] Through the above-mentioned combination of redundant coding, error location, and error repair processes, the system can automatically detect and correct malicious or erroneous fragmentation in privacy computing scenarios involving multiple data holders, thereby ensuring the correctness and robustness of the overall statistical computing results.

[0141] In one possible implementation, after batch verification and legality verification are passed, the set of users to be evaluated is input into the user value evaluation model. After obtaining the user value evaluation results output by the user value evaluation model, the method further includes: dividing the users in the set of users to be evaluated into different value levels according to the value evaluation results; and implementing differentiated marketing strategies for users of different value levels.

[0142] It should be noted that this possible implementation further describes how user value assessment results are applied and processed on the business side. Specifically, after batch verification and legality verification are passed, the data querying party inputs the set of users to be assessed into the user value assessment model, obtaining the user value assessment results output by the model. These results can be used in a structured manner: users in the set to be assessed are divided into different value levels based on the assessment results, for example, into different value categories. Subsequently, differentiated marketing strategies are implemented for users of different value levels, allowing high-value, medium-value, and low-value users to be matched with services and outreach methods appropriate to their respective value levels.

[0143] Therefore, through this post-processing step, the user value assessment results can be applied to differentiated marketing scenarios, enabling the user value assessment results obtained under the premise of privacy protection to directly support more refined user operation decisions, and indirectly improve user experience and resource allocation efficiency.

[0144] Figure 2 This application illustrates a method for evaluating user value according to an embodiment of the present application, such as... Figure 2 As shown, the method is applied to the data holder and includes:

[0145] Step S201: Encrypt the user identifier held locally to obtain a second encrypted identifier. Perform a secure intersection operation based on the second encrypted identifier and the first encrypted identifier provided by the data query party to obtain the intersection users jointly owned by the data query party and the data holder.

[0146] Step S202: Receive the query request sent by the data query party, generate fragmentation information according to the query request, and send the fragmentation information to the data query party;

[0147] The query request contains multiple features to be queried by the user and the conditions that each feature must meet; the sharding information is used to determine the optimal feature set.

[0148] Step S203: Receive the optimal feature set sent by the data query party, and perform federated learning based on the optimal feature set and the data query party to jointly construct a user value assessment model; wherein, the data query party is used to input the set of users to be assessed into the user value assessment model to obtain the user value assessment result output by the user value assessment model.

[0149] It should be noted that, Figure 2 The method shown describes a approach from the perspective of the data holder. Figure 1 The methods shown correspond to user value assessment methods, reflecting the division of responsibilities between the two parties in the same privacy computing protocol.

[0150] First, the data holder encrypts the user identifiers held locally to obtain a second encrypted identifier. Then, it performs a secure intersection operation with the first encrypted identifier provided by the data query party, based on the second encrypted identifier, to obtain the intersection users jointly owned by the data query party and the data holder. This ensures that subsequent statistical analysis and modeling are only performed on users shared by both parties. Next, the data holder receives a query request from the data query party. The query request carries multiple user features to be queried and the conditions that each feature must meet. Based on the query request, without disclosing its original local data, the data holder assists the data query party in obtaining the statistical number of users for each of the multiple queried features that meets the corresponding conditions. The data query party uses this to determine the optimal feature set. After the optimal feature set is determined, the data holder and the data query party perform federated learning based on the optimal feature set to jointly construct a user value assessment model. In the model usage phase after the user value assessment model is constructed, the data query party can input the set of users to be evaluated into the user value assessment model to obtain the user value assessment results.

[0151] Through the above process, a complete chain of responsibility is clearly defined on the data holder's side, including encrypted intersection, collaborative statistics, federated learning, and verification feedback. This ensures that the data holder can participate in building and controlling the secure use of the user value assessment model without exposing local data.

[0152] In one possible implementation, after jointly constructing the user value assessment model, the method also includes:

[0153] Receive a polynomial commitment from the data query party, constructed based on the set of users to be evaluated, and verify the batching of the set of users to be evaluated by the polynomial commitment to ensure that the number of users is greater than a preset threshold.

[0154] Receive the legality verification data generated based on the set of users to be evaluated sent by the data query party, and verify the legality of the set of users to be evaluated based on the legality verification data;

[0155] After the batch verification and legality verification are passed, a verification pass instruction is sent to the data query party. The data query party, upon receiving the verification pass instruction, inputs the set of users to be evaluated into the user value assessment model and obtains the user value assessment results output by the user value assessment model.

[0156] It should be noted that the data holder receives a multinomial commitment constructed based on the set of users to be evaluated from the data query party, and verifies the batch size of the user set to be evaluated by verifying that the number of users is greater than a preset threshold. The data holder also receives legality verification data generated based on the user set to be evaluated from the data query party, and verifies the legality of the user set to be evaluated by verifying the legality of the user set to be evaluated by verifying the legality of the user set to be evaluated. When both batch size verification and legality verification are passed, the data holder sends a verification pass instruction to the data query party. After receiving the verification pass instruction, the data query party inputs the user set to be evaluated into the user value evaluation model and obtains the user value evaluation result output by the user value evaluation model.

[0157] Therefore, by implementing batch verification and legality verification on the user set to be evaluated, sensitive queries on individual users can be effectively avoided during the model usage phase, and the user set participating in the prediction can be ensured to be real, complete and verifiable. This improves the credibility and usability of the evaluation results while protecting privacy and meeting compliance constraints.

[0158] In one possible implementation, the sharding information includes: a second secret shard and a second statistical result shard; receiving a query request sent by the data querying party; generating sharding information based on the query request; and sending the sharding information to the data querying party, including:

[0159] The system receives a portion of the first secret fragment corresponding to the query request from the data query party; it splits the user features it holds into a set of second secret fragments, and sends a portion of the second secret fragments in the second secret fragment set to the data query party; it performs homomorphic computation on the remaining second secret fragments locally (excluding the portion of the second secret fragments) and the portion of the first secret fragments received from the data query party to obtain the difference fragments between the conditions and features; it converts the difference fragments into tag fragments through a secure comparison protocol, and performs homomorphic accumulation on the tag fragments of each user in the intersection users to obtain the second statistical result fragments; it sends the second statistical result fragments to the data query party, whereby the data query party merges the second statistical result fragments with its own obtained first statistical result fragments and performs fragment reconstruction to restore the statistical number of users in plaintext.

[0160] It should be noted that this possible implementation method provides detailed limitations on the specific implementation process of the data holder side "assisting the data query party in obtaining the statistical number of users when each of the multiple features to be queried meets the corresponding conditions based on the query request".

[0161] The data holder first receives a partial first secret fragment corresponding to the query request from the data queryer. Then, it splits the features of the users it holds into a set of second secret fragments and sends a portion of these second secret fragments to the data queryer, establishing a secret fragment correspondence between the two parties regarding conditions and features. Next, the data holder performs homomorphic computation on the remaining second secret fragments (excluding the partial second secret fragments) and the received partial first secret fragments from the data queryer to obtain the difference fragments between the conditions and features, enabling condition judgment within the ciphertext domain. Then, the data holder converts the difference fragments into flagged fragments using a secure comparison protocol and homomorphically accumulates the flagged fragments of each user in the intersection to obtain the second statistical result fragment. Finally, the data holder sends the second statistical result fragment to the data queryer, who merges it with its own obtained first statistical result fragment and reconstructs the fragments to restore the statistical count of users in the plaintext.

[0162] This collaborative process allows for a more detailed understanding of the data holder's role in homomorphic computation, tag generation, and sharded output of statistical results. This enables the multi-party collaborative statistical process to accurately obtain the number of users who meet the criteria while protecting the data privacy of both parties.

[0163] In one possible implementation, when multiple data holders exist, after splitting the characteristics of the users held by the data holder into a set of second secret fragments, the method further includes: generating random numbers; sending each second secret fragment and the generated random numbers together to other data holders, while simultaneously receiving second secret fragments from other data holders and random numbers generated by other data holders; encoding all locally held second secret fragments into a polynomial using Reed-Solomon to generate encoded fragments with redundant information; sending the encoded fragments to the data query party; wherein the data query party uses the encoded fragments to locate the evaluation point of the malicious data holder providing the incorrect second secret fragments using the Berlekamp-Welch algorithm, constructs a repair equation system based on the evaluation point and the encoded fragments, and recovers the correct second secret fragments by solving the repair equation system.

[0164] It should be noted that this possible implementation further extends the handling of second secret shards in scenarios with multiple data holders, introducing random numbers and Reed-Solomon encoding mechanisms to support multi-party robustness.

[0165] Specifically, in the presence of multiple data holders, each data holder, after splitting the characteristics of the users they hold into a set of second-secret fragments, generates random numbers and sends each of the second-secret fragments along with the generated random numbers to other data holders. Simultaneously, they receive second-secret fragments from other data holders and random numbers generated by them, forming an interaction of fragments and random numbers among multiple data holders. Subsequently, each data holder encodes all locally held second-secret fragments into a polynomial using Reed-Solomon encoding, generating encoded fragments with redundant information, and sends these encoded fragments to the data query party. Based on the encoded fragments, the data query party uses the Berlekamp-Welch algorithm to locate the evaluation point of a malicious data holder providing incorrect second-secret fragments. Based on the evaluation point and the encoded fragments, a system of repair equations is constructed, and the correct second-secret fragments are recovered by solving the system of repair equations.

[0166] Therefore, in the privacy computation process involving multiple data holders, the introduction of random numbers, Reed-Solomon encoding, and the Berlekamp-Welch algorithm enables the recovery of the correct fragment through error location and repair mechanisms, even if a malicious data holder tamperes with the second secret fragment, thereby improving the overall statistical process's fault tolerance and security against malicious behavior.

[0167] In one possible implementation, verifying the batch nature of the user set to be evaluated based on the polynomial commitment includes: sending a random challenge point to the data queryer based on the polynomial commitment; receiving the commitment of the value determined by the random challenge point and the quotient polynomial returned by the data queryer; and using bilinear pairing technology to verify whether the commitment of the polynomial commitment, the random challenge point, the value, and the quotient polynomial satisfies a predetermined equation relationship, so as to verify that the number of users in the user set to be evaluated is greater than a preset threshold batch nature.

[0168] It should be noted that this possible implementation specifically defines the interactive process by which the data holder verifies the batch nature of the user set to be evaluated based on the polynomial commitment. It is a refinement of the batch verification steps in the aforementioned possible implementations. After receiving the polynomial commitment from the data queryer, the data holder sends a random challenge point to the data queryer based on the polynomial commitment, requiring the data queryer to prove the specified random challenge point. The data holder receives the commitment from the data queryer, based on the value determined by the random challenge point and the quotient polynomial, and uses bilinear pairing technology to substitute the polynomial commitment, the random challenge point, the value, and the quotient polynomial commitment into a pre-defined equation for verification. If the equation is satisfied, it can be determined that the polynomial corresponding to the user set to be evaluated satisfies the condition that the number of users exceeds a preset threshold for batch nature.

[0169] This bilinear pairing verification process ensures the cryptographic reliability of the batch verification process, allowing data holders to confirm that the data queryer is indeed making a prediction request for a sufficiently large user set without obtaining specific member information of the user set to be evaluated, thus reducing the risk of attacks targeting individual users from the source.

[0170] In one possible implementation, the legitimacy data includes: the root hash of the Merkle tree generated by the data queryer based on the set of users to be evaluated, the Merkle paths of all users in the set of users to be evaluated, and the set of users to be evaluated; verifying the legitimacy of the set of users to be evaluated based on the legitimacy verification data includes: for each user in the set of users to be evaluated, the legitimacy of the user is verified by reconstructing the hash path based on the Merkle path and root hash corresponding to the user.

[0171] It should be noted that this possible implementation specifically defines the process by which the data holder verifies the legitimacy of the user set to be evaluated based on the legitimacy verification data, and is a refinement of the legitimacy verification steps in the above possible implementations.

[0172] The legitimacy verification data includes the root hash of the Merkle tree generated by the data querying party based on the set of users to be evaluated, the Merkle paths of all users in the set, and the set itself. Upon receiving the legitimacy verification data, the data holder, for each user in the set, reconstructs the hash path based on the user's corresponding Merkle path and root hash, verifying whether the reconstructed root hash matches the root hash provided by the data querying party. If they match, it indicates that the user is indeed included in the set of users to be evaluated initially declared by the data querying party and has not been forged or tampered with. By performing this verification process on all users to be evaluated, the data holder can comprehensively determine whether the set of users to be evaluated matches the commitment made by the data querying party.

[0173] Therefore, by combining root hashing and Merkel path verification mechanisms, data holders can reliably verify the integrity and authenticity of the user set to be evaluated without exposing the entire plaintext list of users, thereby further ensuring the security and credibility of the user value assessment process.

[0174] From the perspective of data interaction between data queryers and data holders (see reference) Figure 3 This application provides a detailed description of the user value assessment method shown in the embodiments of this application.

[0175] Specifically, this application provides a user value assessment method based on privacy-preserving computation. Under certain threshold conditions, it uses Secret Sharing (SS) to calculate user statistical results, selects optimal features, constructs a model based on these features, and outputs a user value score. Furthermore, it employs technologies such as Zero-Knowledge Proof (ZKP) to ensure data integrity throughout the process. The steps include the following:

[0176] S1. Each participating party collects a wealth of user-related information using the data collection module. For example, when collecting basic user information, this may include name, age, gender, region, and contact information. Behavioral data includes browsing history, click behavior, dwell time, and search keywords on websites or applications. Consumption data may include purchase records, spending amounts, purchase frequency, and preferred product or service categories. Since data attributes can be categorized as discrete or continuous, the collected data requires separate data cleaning and preprocessing operations to remove duplicate, erroneous, or incomplete data. Simultaneously, relevant data collection regulations and privacy policies must be followed to ensure the legality and compliance of data collection.

[0177] S2, Data Query Party and data holders Perform separate operations on user mobile phone numbers After encryption, both parties perform a secure intersection based on the encrypted phone numbers to obtain the intersection users. , .

[0178] S3, Data Query Method query request It is split into secret shared shards. Among them, Inquiry party reserve Fragmentation, .Will Send in chunks , Data holder User characteristics Similarly, secret fragments are generated. Holder reserve Fragmentation, .Will Send in chunks , .

[0179] S4, Data Query Method With data holders Calculate the difference together The calculation formula is as follows: ;

[0180] in, The total number of feature segments, The modulus, usually a prime number, is used to normalize the calculation result. `mod` is the modulo operation, where the difference is obtained by taking the modulo.

[0181] S5. Based on the query request To clarify the threshold values ​​of each feature The relationship with 0, for example: when monthly active users > 1000, that is... If so, add the user to the statistics. Convert the difference to a binary sign bit, using the highest bit to represent positive or negative numbers: 0 for positive and 1 for negative. If Mark if conditions are met. ,otherwise Until the data query party With data holders Accumulate all users Fragmentation, , Both sides add up all users The values ​​of the flag bits are denoted as follows: , .

[0182]

[0183]

[0184] S6, Data Query Method With data holders After summing, we get , And calculate the final result. ,in, To satisfy the threshold condition for the total number of features in each class, The calculation formula is:

[0185] .

[0186] S7, Data Query Method Obtain the total number of users at each feature threshold, and then filter them based on their own features and other relevant information. and Modeling is performed using excellent features from both sides. To overcome the high correlation between features, which leads to unstable estimations of some features in the subsequent model, multicollinearity analysis is used, and the variance inflation factor is selected to measure the multicollinearity between each feature and other features. For each feature... Using it as the dependent variable and all other characteristics as independent variables, a linear regression analysis is performed, and the calculation formula is as follows:

[0187] ;

[0188] in, These are regression coefficients, representing the magnitude and direction of the influence of each feature on the target variable; The intercept term represents the expected value of the target variable when all features are 0. Corresponding to features The weight, i.e. the influence coefficient. For the error term, Other characteristics.

[0189] S8. For features with high VIF values, the data query method... Features that are more independent of the target variable and have lower VIF values ​​can be removed from the feature set. The calculation formula is as follows:

[0190] ;

[0191] in, It is a feature When obtained from regression with other features value.

[0192] S9. Considering both VIF and correlation, for each feature, combine the VIF value and the correlation coefficient with the target variable to select the most important features. (Data query method) The data query method filters out a relatively independent set of features that are strongly correlated with the model. With data holders The model is trained based on the obtained optimal features. The model construction process is not described in detail.

[0193] S10, If the data query party A new group of encrypted users It is necessary to predict its value score. To prevent privacy leaks or security risks caused by single-sample prediction through model reverse engineering attacks or member inference attacks, the data holder... It needs to be verified; please refer to the following: Figure 4 , to collect users As coefficients of a polynomial, the formula for calculating the polynomial is:

[0194] .

[0195] S11, Data Query Method Parameters are generated based on the BLS12-381 curve (a specific elliptic curve, named after the researcher and its parameters). and generator .in, These are secret points generated during system initialization. (Data query method) Calculate the KZG commitment value (a polynomial commitment scheme named after its researcher). ,promise All coefficients are bound together, and the commitment value is... Send to the data holder .

[0196] .

[0197] S12, Data Query Method Constructing quotient polynomial , used for the existence of non-zero higher-order terms. The calculation formula is:

[0198] ;

[0199] like Existence indicates that the denominator Divisible ,but It is at least a first-order polynomial, that is Meanwhile, the data query party Generate quotient polynomial commitments and the commitment value Send to the data holder Commitment value The calculation formula is:

[0200] .

[0201] S13, Data Holder Select random challenge point The location is determined by the data query party. Generate the proof point here. Data holder The user set is determined by verifying whether the equation holds true. The size is greater than 1.

[0202] ;

[0203] in, In order to be in The actual result calculated from the point; For the quotient polynomial in The predicted target value is used to verify whether it meets expectations; This is a hash value.

[0204] S14, At the same time, in order to ensure user aggregation The legitimacy of the user and the data query party set Each user data point is treated as a leaf node, and its hash value is calculated. Leaf nodes are paired, merged, and then hashed again to generate the parent node hash. This process is repeated until a unique root hash (Merkle Root) is generated. Then, the data query... User set Each user's Merkle tree path and root hash are sent to the data holder. . For each user Perform the following operations: Calculate the hash value of the user data and use the data query method. The provided Merkle path is merged with sibling hashes hierarchically and the hash is recalculated until the root node. If the final root hash does not match the data query... If the provided root hash is consistent, then the user It is legal and valid.

[0205] .

[0206] S15. After the polynomial verification is completed, the data query party... Get user set Based on the predicted values, user value is assessed, categorized as high, medium, and low-value users, and corresponding services are sent. If the data query party... With multiple data holders Modeling, User Sets After input, corresponding value scores are obtained. The weights of each participant can be set using the feature importance and number of important features in the multicollinearity analysis described above, and a weighted average method is used to obtain the corresponding user value score. Different marketing methods are then applied to high, medium, and low-value users, such as product SMS delivery, advertising pushes, and offline terminal development.

[0207] Furthermore, steps S1-S6 can be extended to scenarios with more participants. For example, multiple data holders may interact with the data queryer, allowing the queryer to obtain the calculation and statistical results from each data holder separately for individual business processing, or the feature calculation results from each data holder may be aggregated for further business processing. However, in business scenarios involving multiple participants, it's impossible to guarantee the security and trustworthiness of each participating node. If malicious nodes or missing nodes exist, it's difficult to guarantee the accuracy of the results and the security of the data features. Therefore, to detect and correct malicious nodes and ensure the acquisition of correct results, Reed-Solomon encoding is applied to the error detection, correction, and recovery of privacy data feature fragments from each participant.

[0208] D1: Data Holder Each of them will have its own privacy data characteristics Split into multiple secret fragments Select random number , each fragment Send to other data holders, among whom, Recorded as , This is a random number used for encryption and recovery verification of fragments. , The total number of secret shards is determined, and all shards will be sent to other data holders to ensure that each data holder receives only a portion of the data.

[0209] D2: Data Holder Its secret fragments The Reed-Solomon shards are encoded into a polynomial, and these encoded shards are sent to the querying party. The secret shards are... Representation in a polynomial:

[0210] ;

[0211] in, The coefficients of the polynomial, For evaluation points (identifiers or locations of each node). It is fragmented data. The degree encoded in Reed-Solomon.

[0212] D3: Constructing an error localization polynomial based on the improved Berlekamp-Welch algorithm The roots of the polynomial are the evaluation points of malicious nodes. The Berlekamp-Welch algorithm is mainly used to correct errors in Reed-Solomon coding. The improved Berlekamp-Welch algorithm expands the application scope of traditional algorithms, allowing the processing of data in the real number domain and solving for polynomial coefficients through matrix operations. It is a collection of malicious data holders. This is the assessment point for malicious nodes;

[0213] .

[0214] D4: Redundancy information and error localization polynomial generated by Reed-Solomon encoding Construct a linear equation to repair the tampered shard. The linear equation is: In which each row of matrix D Corresponding data holder Data And each column represents a polynomial. At different assessment points The value of represents the data shard position of each participant (usually the participant's ID or a specific evaluation point), and is the degree of the Reed-Solomon encoding. These are the polynomial coefficients obtained by solving a linear system. The received result contains errors;

[0215] .

[0216] D5: The queryer obtains the data coefficients. The original data was then recovered. Ultimately, the data query party... Finally, the final result is obtained by receiving the combined and corrected calculation results from all participants.

[0217] In summary, the method shown in the embodiments of this application has the following advantages compared with the prior art:

[0218] 1. The use of multinomial commitment technology to verify the legitimacy of the user set, and the verification in the encrypted state, thereby eliminating malicious and precise queries on the privacy-sensitive data of individual users, effectively preventing malicious users from tampering with or forging data.

[0219] 2. In this embodiment, the original data does not leave the domain, and the distribution of fragmented computational features ensures data security. Simultaneously, it can be applied to multi-tenant clients, where clients can obtain feature information under different business backgrounds through fragmented interaction. Even in the presence of malicious nodes, redundant information is generated through Reed-Solomon (RS) encoding, and an error localization polynomial is constructed using an improved Berlekamp-Welch algorithm to complete error correction and calculate the correct result, thus achieving the detection and correction of malicious nodes. Currently, most secret fragmentation technologies are simply applied between two parties, and there are no relevant cases regarding fragmented transmission, security, and efficiency between multiple parties. The method shown in this application proposes a related process for multi-party feature fragmented transmission. Even in the case of malicious nodes among multiple participating parties, an error localization polynomial can be constructed, enabling the detection of malicious nodes and error correction. This demonstrates the innovativeness of this technology.

[0220] 3. By combining multicollinearity analysis and calculating VIF values, highly correlated features are accurately identified, and feature selection is optimized. This is more suitable for scenarios where the initial model training results are not good, but precise user service marketing is still needed.

[0221] 4. Securely calculate and obtain intersections without exposing personal privacy data. In practice, compliance departments often require data providers to provide user authorization codes, authorization status, and expiration dates. This technology perfectly achieves these functions. Data is trained locally, avoiding centralized data storage and transmission issues; only model parameters are transmitted, without exposing the data itself.

[0222] Figure 5 This application illustrates a user value assessment apparatus according to an embodiment of the present application. The apparatus is applied to a data query party, and the apparatus 50 includes:

[0223] The first secure intersection module 501 is used to encrypt the user identifier held locally to obtain a first encrypted identifier, and perform a secure intersection operation based on the first encrypted identifier and the second encrypted identifier provided by the data holder to obtain the intersection users jointly owned by the data queryer and the data holder.

[0224] The sending module 502 is used to send a query request to the data holder and receive the fragmentation information returned by the data holder according to the query request. The query request carries multiple features to be queried by the user and the conditions that each feature to be queried must satisfy.

[0225] The determination module 503 is used to determine the statistical number of users in the intersection of users when each of the multiple features to be queried satisfies the corresponding conditions, based on the query request and the sharding information.

[0226] Module 504 is used to construct a feature distribution dataset based on multiple statistical counts for multiple features to be queried;

[0227] The filtering module 505 is used to perform multicollinearity analysis and correlation evaluation based on the feature distribution dataset to filter out the optimal feature set;

[0228] The first model building module 506 is used to send the optimal feature set to the data holder, and based on the optimal feature set, perform federated learning with the data holder to jointly build a user value assessment model.

[0229] The value assessment module 507 is used to input the set of users to be assessed into the user value assessment model and obtain the user value assessment results output by the user value assessment model.

[0230] Figure 6 An apparatus for evaluating user value according to an embodiment of this application is shown. The apparatus is applied to a data holder. The apparatus 60 includes:

[0231] The second secure intersection module 601 is used to encrypt the user identifier held locally to obtain a second encrypted identifier, and to perform a secure intersection operation based on the second encrypted identifier and the first encrypted identifier provided by the data query party to obtain the intersection users jointly owned by the data query party and the data holder.

[0232] The receiving module 602 is used to receive a query request sent by the data querying party, generate fragmentation information according to the query request, and send the fragmentation information to the data querying party. The query request carries multiple features to be queried by the user and the conditions that each feature to be queried must satisfy. The fragmentation information is used to determine the optimal feature set.

[0233] The second model construction module 603 is used to receive the optimal feature set sent by the data query party, and to perform federated learning with the data query party based on the optimal feature set to jointly construct a user value assessment model; wherein, the data query party is used to input the set of users to be assessed into the user value assessment model to obtain the user value assessment result output by the user value assessment model.

[0234] In summary, in this embodiment, taking the data querying party as an example, by encrypting the user identifier held locally on the data querying party's side and performing secure intersection with the encrypted identifier provided by the data holder, the intersection users jointly owned by both parties are determined only within the encrypted space, avoiding the exposure of the original user identifier during the intersection process and reducing the risk of user identity information leakage from the source. On this basis, the data querying party sends a query request to the data holder carrying multi-dimensional query features and their respective conditions. Both parties count the number of users that meet the conditions for each feature and construct a feature distribution dataset without disclosing plaintext data, so that the distribution of intersection users in key feature dimensions can be fully grasped before model training. Subsequently, the data querying party performs multicollinearity analysis and correlation evaluation based on the feature distribution dataset, removes highly collinear redundant features and retains the optimal feature set that is highly correlated with user value, thereby mitigating the impact of multicollinearity on parameter estimation stability and generalization ability. Then, the optimal feature set is used to jointly model with the data holder under the federated learning framework, fully integrating multi-source feature information without leaving the domain of the original data of each party, and constructing a user value assessment model that is more in line with the true value of users.

[0235] Therefore, while protecting the data privacy of all participants, the accuracy and stability of modeling can be improved, enabling accurate and efficient assessment of user value, and achieving the technical effect of precision marketing and thus improving user satisfaction.

[0236] This application provides a network device 70, such as... Figure 7 As shown, the network device 70 includes a processor 701, a memory 702, and a program stored in the memory 702 and executable on the processor 701. When the program is executed by the processor 701, it implements the steps of a user value assessment method as shown in the above embodiment.

[0237] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of a user value assessment method as shown in the above embodiments, achieving the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0238] This application also provides a computer program product, including computer instructions. When executed by a processor, the computer instructions implement the steps of the user value evaluation method shown in the above embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0239] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0240] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0241] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for evaluating user value, characterized in that, The method is applied to the data query party, and the method includes: The user identifier held locally is encrypted to obtain a first encrypted identifier. A secure intersection operation is performed based on the first encrypted identifier and the second encrypted identifier provided by the data holder to obtain the intersection users jointly owned by the data queryer and the data holder. Send a query request to the data holder and receive the fragmentation information returned by the data holder based on the query request. The query request carries multiple features of the user to be queried and the conditions that each feature to be queried must satisfy. Based on the query request and the sharding information, determine the statistical number of users in the intersection users when each of the multiple features to be queried satisfies the corresponding conditions; A feature distribution dataset is constructed based on multiple statistical counts for multiple features to be queried; Based on the aforementioned feature distribution dataset, multicollinearity analysis and correlation assessment are performed to select the optimal feature set. The optimal feature set is sent to the data holder, and federated learning is performed with the data holder based on the optimal feature set to jointly construct a user value assessment model. Obtain the set of users to be evaluated, input the set of users to be evaluated into the user value evaluation model, and obtain the user value evaluation result output by the user value evaluation model.

2. The method according to claim 1, characterized in that, Obtain the set of users to be evaluated, input the set of users to be evaluated into the user value evaluation model, and obtain the user value evaluation results output by the user value evaluation model, including: Obtain a set of users to be evaluated, construct a polynomial commitment based on the set of users to be evaluated, and send the polynomial commitment to the data holder, wherein the data holder is used to verify the batching of the number of users in the set of users to be evaluated to a preset threshold based on the polynomial commitment; Based on the set of users to be evaluated, legality verification data is generated and sent to the data holder, who uses the data holder to verify the legality of the set of users to be evaluated. After the batch verification and legality verification are passed, the set of users to be evaluated is input into the user value evaluation model to obtain the user value evaluation results output by the user value evaluation model.

3. The method according to claim 1, characterized in that, The sharding information includes: a second secret shard and a second statistical result shard. A query request is sent to the data holder, and the sharding information returned by the data holder based on the query request is received. Based on the query request and the sharding information, the statistical number of users among the intersecting users whose respective queried features satisfy the corresponding conditions is determined, including: The query request is split into a first secret shard set, a portion of the first secret shards in the first secret shard set is sent to the data holder, and a portion of the second secret shards corresponding to the feature is received from the data holder; Homomorphic computation is performed on the remaining first secret fragments locally excluding the aforementioned portion of the first secret fragments, and on the received portion of the second secret fragments from the data holder, to obtain the difference fragment between the condition and the feature; The difference fragments are converted into tag fragments using a secure comparison protocol, and the tag fragments of each user in the intersection users are homomorphically accumulated to obtain the first statistical result fragment. The system receives the second statistical result fragment obtained from the data holder, merges the first statistical result fragment and the second statistical result fragment, and performs fragment reconstruction to restore the statistical number of the user in plaintext.

4. The method according to claim 1, characterized in that, Based on the aforementioned feature distribution dataset, multicollinearity analysis and correlation assessment were performed to select the optimal feature set, which includes: Each feature in the feature distribution dataset is taken as a target feature, and the following operations are performed on the target feature: the target feature is taken as the dependent variable, and the remaining features in the feature distribution dataset other than the target feature are taken as independent variables to construct a linear regression model; the coefficient of determination is determined based on the linear regression model, and the variance inflation factor value of the target feature is calculated based on the coefficient of determination; The variance inflation factor values ​​corresponding to each feature are compared with the preset variance inflation factor thresholds. Based on the comparison results, all features with variance inflation factor values ​​higher than the variance inflation factor threshold are deleted from the feature distribution dataset to generate a deredundant feature set. Based on the preset user value tags of the data query party, calculate the correlation coefficient between each feature in the redundancy removal feature set and the user value assessment result; The optimal feature set is formed by selecting features from the deredundancy feature set whose correlation coefficient is higher than a preset correlation coefficient threshold.

5. The method according to claim 1, characterized in that, Obtaining a set of users to be evaluated, constructing a polynomial commitment based on the set of users to be evaluated, and sending the polynomial commitment to the data holder includes: The set of users to be evaluated is used as the coefficients of a polynomial to construct the polynomial; Based on elliptic curve cryptography, a polynomial commitment is generated for the polynomial, and the polynomial commitment is sent to the data holder.

6. The method according to claim 5, characterized in that, After obtaining the set of users to be evaluated, constructing a multinomial commitment based on the set of users to be evaluated, and sending the multinomial commitment to the data holder, the method further includes: In response to a random challenge point sent by the data holder based on the polynomial commitment; Calculate the value of the polynomial at the random challenge point, construct a quotient polynomial based on the value, and generate a commitment for the quotient polynomial; The value and the commitment of the quotient polynomial are sent to the data holder, wherein the data holder is used to verify the batch nature of the number of users in the user set to be evaluated being greater than a preset threshold based on the value and the commitment of the quotient polynomial.

7. The method according to claim 1, characterized in that, Based on the set of users to be evaluated, generating legitimacy verification data and sending the legitimacy verification data to the data holder includes: Recursive hash calculations are performed on the set of users to be evaluated to construct a Merkle tree, and the root hash of the Merkle tree is calculated. Based on the Merkle tree, a corresponding Merkle path is generated for each user in the set of users to be evaluated; The root hash, all the Merkel paths, and the set of users to be evaluated are sent as the legitimacy verification data to the data holder, wherein the data holder is used to verify the legitimacy of the set of users to be evaluated based on the legitimacy verification data.

8. The method according to claim 3, characterized in that, In the presence of multiple data holders, receiving a portion of the second secret fragment corresponding to the feature from the data holders includes: Receive Reed-Solomon encoded fragments from each data holder, where each encoded fragment is generated by the corresponding data holder by encoding all local second secret fragments into a polynomial; Based on the aforementioned coded segments, the Berlekamp-Welch algorithm is used to construct the error localization polynomial; Calculate the roots of the error-locating polynomial to pinpoint the evaluation point of the malicious data holder who provided the erroneous second secret fragment; A set of repair equations is constructed based on the evaluation points and the redundancy information of the coded segments; Solve the repair equations to recover the correct second secret fragment corresponding to the malicious data holder.

9. The method according to any one of claims 1-8, characterized in that, After passing batch verification and legality verification, the method further includes inputting the set of users to be evaluated into the user value evaluation model and obtaining the user value evaluation results output by the user value evaluation model: Based on the value assessment results, the users in the set of users to be assessed are divided into different value levels; Implement differentiated marketing strategies for users at different value levels.

10. A method for evaluating user value, characterized in that, The method is applied to the data holder, and the method includes: The user identifier held locally is encrypted to obtain a second encrypted identifier. A secure intersection operation is performed between the second encrypted identifier and the first encrypted identifier provided by the data query party to obtain the intersection users jointly owned by the data query party and the data holder. The system receives a query request from the data querying party, generates sharding information based on the query request, and sends the sharding information to the data querying party. The query request carries multiple features to be queried by the user and the conditions that each feature to be queried must satisfy. The sharding information is used to determine the optimal feature set. The system receives the optimal feature set sent by the data query party, and performs federated learning based on the optimal feature set and the data query party to jointly construct a user value assessment model; wherein, the data query party is used to input the set of users to be assessed into the user value assessment model to obtain the user value assessment result output by the user value assessment model.

11. The method according to claim 10, characterized in that, After jointly constructing the user value assessment model, the method further includes: Receive a polynomial commitment constructed based on the set of users to be evaluated sent by the data query party, and verify the batching of the set of users to be evaluated by the polynomial commitment to ensure that the number of users is greater than a preset threshold. Receive the legality verification data generated based on the set of users to be evaluated sent by the data query party, and verify the legality of the set of users to be evaluated based on the legality verification data; After the batch verification and legality verification are passed, a verification pass instruction is sent to the data query party. The data query party is used to input the set of users to be evaluated into the user value evaluation model after receiving the verification pass instruction, and obtain the user value evaluation result output by the user value evaluation model.

12. The method according to claim 10, characterized in that, The sharding information includes: a second secret shard and a second statistical result shard. The system receives a query request from the data querying party, generates sharding information based on the query request, and sends the sharding information to the data querying party, including: Receive a portion of the first secret fragment corresponding to the query request from the data querying party; The user's characteristics held by itself are split into a second secret shard set, and a portion of the second secret shards in the second secret shard set are sent to the data query party; Homomorphic computation is performed on the remaining second secret fragments locally excluding the aforementioned portion of the second secret fragments, and on the received portion of the first secret fragments from the data query party, to obtain the difference fragment between the condition and the feature; The difference fragments are converted into tag fragments using a secure comparison protocol, and the tag fragments of each user in the intersection users are homomorphically accumulated to obtain the second statistical result fragments. The second statistical result fragment is sent to the data query party, wherein the data query party is used to merge the second statistical result fragment with the first statistical result fragment obtained by itself, and perform fragment reconstruction to restore the statistical number of the user in plaintext.

13. The method according to claim 12, characterized in that, In the presence of multiple data holders, after splitting the characteristics of the users held by the holder into a second secret shard set, the method further includes: Generate random numbers, and send each of the second secret fragments in the second secret fragments together with the generated random numbers to other data holders, while receiving the second secret fragments from other data holders and the random numbers generated by other data holders; All locally held second secret fragments are encoded into a polynomial using Reed-Solomon encoding to generate encoded fragments with redundant information; The encoded fragment is sent to the data query party; wherein, the data query party is used to locate the evaluation point of the malicious data holder who provided the incorrect second secret fragment based on the encoded fragment and the Berlekamp-Welch algorithm, construct a set of repair equations based on the evaluation point and the encoded fragment, and recover the correct second secret fragment by solving the set of repair equations.

14. The method according to claim 10, characterized in that, Verifying the batch size of the user set to be evaluated, based on the polynomial commitment, as greater than a preset threshold includes: Based on the polynomial commitment, a random challenge point is sent to the data query party; Receive the commitment from the data query party regarding the value determined based on the random challenge point and the quotient polynomial; Using bilinear pairing techniques, it is verified whether the polynomial commitment, the random challenge point, the value, and the commitment of the quotient polynomial satisfy a predetermined equation relationship, so as to verify that the number of users in the user set to be evaluated is greater than a preset threshold batch size.

15. The method according to claim 10, characterized in that, The legitimacy data includes: the root hash of the Merkel tree generated by the data querying party based on the set of users to be evaluated, the Merkel paths of all users in the set of users to be evaluated, and the set of users to be evaluated; verifying the legitimacy of the set of users to be evaluated based on the legitimacy verification data includes: For each user in the set of users to be evaluated, the legitimacy of the user is verified by reconstructing the hash path based on the Merkel path corresponding to the user and the root hash.

16. A user value assessment device, characterized in that, The device is used by a data queryer, and the device includes: The first secure intersection module is used to encrypt the user identifier held locally to obtain a first encrypted identifier, and perform a secure intersection operation based on the first encrypted identifier and the second encrypted identifier provided by the data holder to obtain the intersection users jointly owned by the data query party and the data holder. The sending module is used to send a query request to the data holder and receive the fragmentation information returned by the data holder according to the query request. The query request carries multiple features to be queried by the user and the conditions that each feature to be queried must satisfy. The determination module is used to determine, based on the query request and the sharding information, the statistical number of users in the intersection users when each of the multiple features to be queried satisfies the corresponding conditions. The building module is used to construct a feature distribution dataset based on multiple statistical counts for multiple features to be queried; The filtering module is used to perform multicollinearity analysis and correlation evaluation based on the feature distribution dataset to filter out the optimal feature set; The first model building module is used to send the optimal feature set to the data holder, and perform federated learning with the data holder based on the optimal feature set to jointly build a user value assessment model. The value assessment module is used to input the set of users to be assessed into the user value assessment model and obtain the user value assessment results output by the user value assessment model.

17. A user value assessment device, characterized in that, The device is used by a data holder, and the device includes: The second secure intersection module is used to encrypt the user identifier held locally to obtain a second encrypted identifier, and perform a secure intersection operation based on the second encrypted identifier and the first encrypted identifier provided by the data query party to obtain the intersection users jointly owned by the data query party and the data holder. A receiving module is used to receive a query request sent by the data querying party, generate sharding information according to the query request, and send the sharding information to the data querying party. The query request carries multiple features to be queried by the user and the conditions that each feature to be queried must satisfy. The sharding information is used to determine the optimal feature set. The second model construction module is used to receive the optimal feature set sent by the data query party, and to perform federated learning based on the optimal feature set and the data query party to jointly construct a user value assessment model; wherein, the data query party is used to input the set of users to be assessed into the user value assessment model to obtain the user value assessment result output by the user value assessment model.

18. A network device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of a user value assessment method as claimed in any one of claims 1 to 9, or, when executed by the processor, the program implements the steps of a user value assessment method as claimed in any one of claims 10 to 15.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a user value assessment method as described in any one of claims 1 to 9, or, when executed by the processor, implements the steps of a user value assessment method as described in any one of claims 10 to 15.

20. A computer program product, characterized in that, The method includes computer instructions that, when executed by the processor, implement the steps of a user value assessment method as described in any one of claims 1 to 9, or, when executed by the processor, implement the steps of a user value assessment method as described in any one of claims 10 to 15.

Images