WO2022126916A1 - Product recommendation method, system, computer device and storage medium - Google Patents

Abstract

The present application relates to artificial intelligence, and discloses a product recommendation method, a system, a computer device and a storage medium. The method comprises: a server side receives parameter data; a server obtains a product factor matrix and parameters after initialization, and sends the product factor matrix parameters and the parameter data to multiple clients; the clients update a user factor matrix accordingly, calculate weighted differences and normalized constants by using a q-FedAvg algorithm, and return the weighted differences and normalized constants to the server; the server updates the product factor matrix parameters by using the product factor matrix parameters, the weighted differences and the normalized constants; the server determines a target product factor matrix according to the updated product factor matrix parameters, and sends the target product factor matrix to a target client; and the target client uses the target product factor matrix and the updated user factor matrix to calculate predicted score values and recommend the first N products having a high ranking to a user. The present application improves the fitting degree between recommended products and users while protecting user data from leakage.

Description

A product recommendation method, system, computer equipment and storage medium

This application claims the priority of the Chinese patent application filed on December 18, 2020 with the application number 202011504951.4 and the invention title is "A product recommendation method, system, computer equipment and storage medium", the entire content of which is approved by Reference is incorporated in this application.

technical field

The present application relates to the technical field of artificial intelligence, and in particular, to a product recommendation method, system, computer equipment and storage medium.

Background technique

With the continuous development of society, people pay more and more attention to the protection of privacy, and various countries have also promulgated various provisions to protect privacy data, such as the EU's GDPR (General Data Protection Regulation), and Facing this situation, the inventors realized that it is a problem how to recommend users to complete products without leaking user data. In the prior art, the collaborative filtering algorithm based on matrix decomposition has the problem of high computing and storage costs, and also has the problem that user data may be leaked by the server, so how to complete the product recommendation for users without leaking user privacy , and reducing the computational cost of the recommendation model has become an urgent problem to be solved.

SUMMARY OF THE INVENTION

The present application provides a product recommendation method, system, computer equipment and storage medium to solve the problems of user privacy leakage and high model calculation cost in the prior art when recommending products to users.

In order to solve the above problems, the present application provides a product recommendation method, including:

The server side receives user input parameter data;

The server side obtains the product factor matrix V and the product factor matrix parameter V ^t through initialization processing, and sends the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

The multiple clients update the user factor matrix u _i according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , Returning the weighted difference value Δ _i ^t and the normalized constant h _i ^t to the server side;

Based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , to the product factor matrix parameter V ^t Iteratively update;

The server side determines the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sends the target product factor matrix to the target client;

The target client uses the target product factor matrix and the updated user factor matrix to calculate the predicted score value of the product by the user, sort the predicted score value, and recommend the top N products with the highest ranking to the product. the user.

In order to solve the above problems, the present application also provides a product recommendation system, the system includes:

server and client;

The server includes:

The receiving module is used to receive the parameter data input by the user;

an initialization module, configured to obtain a product factor matrix V and a product factor matrix parameter V ^t through initialization, and send the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

The update module, based on the remaining part of the parameter data, uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t to perform the product factor matrix parameter V ^t . iterative update;

A determination module, for determining a target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sending the target product factor matrix to the target client;

The client includes:

The calculation module is used to update the user factor matrix u _i according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , The weighted difference Δ _i ^t and the normalized constant h _i ^t are returned to the server;

The output module, using the target product factor matrix and the updated user factor matrix, calculates the predicted score value of the product by the user, sorts the predicted score value, and recommends the top N products with the highest ranking to the user.

To solve the above problem, an embodiment of the present application further provides a computer device, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, and the processor executes The computer-readable instructions implement the following steps:

The server side receives the parameter data input by the user;

The server side obtains the product factor matrix V and the product factor matrix parameter V ^t through initialization processing, and sends the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

The multiple clients update the user factor matrix u _i according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , Returning the weighted difference value Δ _i ^t and the normalized constant h _i ^t to the server side;

Based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , to the product factor matrix parameter V ^t Iteratively update;

The server side determines the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sends the target product factor matrix to the target client;

The target client uses the target product factor matrix and the updated user factor matrix to calculate the predicted score value of the product by the user, sort the predicted score value, and recommend the top N products with the highest ranking to the product. the user.

In order to solve the above problem, an embodiment of the present application further provides a computer-readable storage medium, where computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, the processing The device performs the following steps:

The server side receives the parameter data input by the user;

The server side obtains the product factor matrix V and the product factor matrix parameter V ^t through initialization processing, and sends the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

The multiple clients update the user factor matrix u _i according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , Returning the weighted difference value Δ _i ^t and the normalized constant h _i ^t to the server side;

Based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , to the product factor matrix parameter V ^t Iteratively update;

The server side determines the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sends the target product factor matrix to the target client;

The target client uses the target product factor matrix and the updated user factor matrix to calculate the predicted score value of the product by the user, sort the predicted score value, and recommend the top N products with the highest ranking to the product. the user.

Compared with the prior art, the product recommendation method, system, computer equipment and storage medium provided according to the embodiments of the present application have at least the following beneficial effects:

Through the cooperation between the server and the client, the server receives the parameter data input by the user; the server is initialized and sends the initialized parameter data to the client, and the client calculates according to the parameter data and combined with its own data, and obtains The weighted difference Δ _i ^t and the normalized constant h _i ^t are fed back to the server; the client uses local data for calculation, avoiding the problem of data leakage when sending data to the server, and protecting the user's data; The server side updates the matrix parameters based on the weighted difference Δ _i ^t and the normalization constant h _i ^t , determines the target product factor matrix based on the updated matrix parameters and sends it to the client, and the client sends the target product factor matrix based on the target product factor matrix. User recommended products. While protecting the user data, the user data is used to complete the accurate recommendation for the user, and the server side only uses the weighted difference Δ _i ^t sent by the client and the normalized constant h _i ^t to calculate, which reduces the model calculation cost.

Description of drawings

In order to illustrate the solutions in the present application more clearly, the following will briefly introduce the accompanying drawings used in the description of the embodiments of the present application. For those of ordinary skill, other drawings can also be obtained from these drawings without any creative effort.

1 is a schematic flowchart of a product recommendation method provided by an embodiment of the present application;

FIG. 2 is a schematic diagram of a module of a product recommendation device provided by an embodiment of the present application;

FIG. 3 is a schematic structural diagram of a computer device according to an embodiment of the present application.

Detailed ways

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field of this application; the terms used herein in the specification of the application are for the purpose of describing specific embodiments only It is not intended to limit the application; the terms "comprising" and "having" and any variations thereof in the description and claims of this application and the above description of the drawings are intended to cover non-exclusive inclusion. The terms "first", "second" and the like in the description and claims of the present application or the above drawings are used to distinguish different objects, rather than to describe a specific order.

The present application provides a product recommendation method. Referring to FIG. 1 , it is a schematic flowchart of a product recommendation method according to an embodiment of the present application.

In this embodiment, the product recommendation method includes:

S1. The server side receives the parameter data input by the user;

The specific parameter data includes the total number of clients n, the number of epochs E, that is, the number of stochastic gradient descent is E times, the number of loops is T, that is, the number of iterations is T times, the q-FedAvg algorithm parameter q, the Lipschitz constant L (K for functions on a subset of the set of real numbers

If there is a constant such that |f(a)-f(b)|≤K|ab|

Then f is said to meet the Lipschitz condition, and the constant K with the smallest f is called the Lipschitz constant of f), step size α, regularization parameter λ and gain parameter γ.

Part of the parameter data is used on the server side, that is, the total number of clients n, the number of cycles T and other parameters, and a part is sent to the client through the server side, that is, the number of epochs E, the q-FedAvg algorithm parameter q, the Lipschitz constant L, the step The length α, regularization parameter λ and gain parameter γ will be sent to the client.

The server side receives parameter data input by the user.

S2, the server side obtains the product factor matrix V and the product factor matrix parameter V ^t through initialization processing, and sends the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

The server side obtains the product factor matrix parameter V ^t and the product factor matrix V by randomly initializing the product factor matrix and the product factor matrix parameters ^, and the E, q, L, α, λ and γ parameters are sent to multiple clients.

The random initialization can be initialized by means of Gaussian distribution, Xavier, etc., the product factor matrix parameter V ^t and the product factor matrix V are randomly generated, and the model is more likely to find the global optimal solution through random initialization, and if the product factor matrix is fully assigned. If it is 0, symmetry will be generated, which will result in two neurons in the hidden layer generated by the input value. Due to the symmetry of the matrix, regardless of forward propagation or back propagation, these two neurons that should represent different characteristics are If the operations are done the same, there will be problems.

Further, the sending of the product factor matrix parameter V ^t and part of the parameter data to multiple clients includes:

Take n clients as the set S _t ;

The product factor matrix parameter V ^t and the parameter data are sent to n clients in the set S _t .

Specifically, the product factor matrix parameter V ^t and the E, q, L, α, λ, γ parameters in the parameter data are sent to n clients in the set S _t through the number of clients n, specified in the parameter data.

At the same time, the client also needs to be initialized, and the user factor matrix and the user factor matrix parameters are also randomly initialized to obtain the user factor matrix parameter u _i ^t and the user factor matrix u _i , and the random initialization method is the same as that of the server side.

In this case, by using horizontal federated learning, n clients are used as the samples for this update iteration. Using n clients to improve the uncertainty of the data, it is convenient for subsequent training to obtain the universality of the data.

Before the sending the product factor matrix parameter V ^t to multiple clients, the method further includes:

Homomorphically encrypt the product factor matrix parameter V ^t .

Data security is guaranteed after homomorphic encryption. And the advantage of homomorphic encryption is that there is no need to decrypt the data, and the data can be used for operations.

S3. Multiple clients update the user factor matrix u _i according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , Returning the weighted difference value Δ _i ^t and the normalized constant h _i ^t to the server side;

The partial parameter data includes the epoch number E, the q-FedAvg algorithm parameter q, the Lipschitz constant L, the step size α, the regularization parameter λ and the gain parameter γ. When the server receives all parameter data, the total number of clients n, the number of cycles T and other parameters are assigned to the server for use; epoch times E, q-FedAvg algorithm parameter q, Lipschitz constant L, step size α, regularization The parameter λ and the gain parameter γ are used by the client.

The client needs to undergo homomorphic encryption before sending the weighted difference Δ _i ^t and the normalized constant h _i ^t to the server to ensure data security.

Further, updating the user factor matrix _ui according to the product factor matrix parameter V ^t and the parameter data by the multiple clients includes:

Based on the user factor matrix _ui and the product factor matrix parameter V ^t , the loss function F _i is calculated according to formula 1, and the formula 1 is:

Based on the loss function F _i , the gradient of the user factor matrix _ui and the gain parameter γ, the updated user factor matrix parameter u _i ^t+1 is calculated according to the formula 2, and the formula 2 is:

Update the user factor matrix _ui according to the iterative user factor matrix parameter _ui ^t+1 ;

Among them, t represents the number of iterations, γ represents the gain parameter,

Represents the gradient value of _ui ^t after the last iteration, j=1,2,..., m,m is the number of products, r _ij represents the user i's rating value for product j, λ represents the regularization parameter, v _j represents the product In the jth row of the factor matrix V, Fi is the loss function, _ui ^t+1 is the user factor matrix parameter after iteration, and _ui ^t is the user factor matrix parameter.

Specifically, the loss function F _i is calculated by using the user factor matrix _ui , the product factor matrix parameter V ^t and the regularization parameter λ on the client; and then based on the loss function F _i , the user factor matrix _ui The gradient and gain parameter γ of , calculate the updated user factor matrix parameter u _i ^t+1 ; r _ij is the existing one of the client.

use

The formula specifically calculates F _i (V ^t , _ui ^t ), and this formula is a specific expansion of formula one.

The client maintains its own data to avoid leakage caused by data transmission.

Still further, the use of the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t includes:

Obtain the updated user factor matrix parameters

and the preset step size α, perform E times of stochastic gradient descent on the loss function F _i to obtain the model parameters

based on the model parameters

Calculate the parameter difference ΔV _i ^t of the adjacent product factor matrix;

Calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t according to the following q-FedAvg algorithm according to the difference between the adjacent product factor matrix parameters;

h _i ^t =qF _i ^q-1 (V ^t , _ui ^t+1 )||ΔV _i ^t || ² +LF _i ^q (V ^t , _ui ^t+1 );

Among them, ΔV _i ^t represents the difference between adjacent product factor matrix parameters, L represents the Lipschitz constant, q represents the q-FedAvg algorithm parameter, k=1,2,...,E.

Specifically, through V ₀ ^t+1 =V ^t ;

When doing the first stochastic gradient descent, i.e.

Since V ₀ ^t+1 =V ^t , so

V ₁ ^t+1 can be obtained, so as to keep iterating E times, when the E time,

to get the model parameters

Based on model parameters

Then the difference ΔV _i ^t of the adjacent product factor matrix parameters can be obtained;

The client calculates the weighted difference Δ _i ^t and the normalized constant h _i ^t based on its own data characteristics and sends them to the client. The data features are displayed for subsequent product recommendations.

S4. Based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , to the product factor matrix parameter V ^t is iteratively updated;

The server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalized constant h _i ^t to iteratively update the product factor matrix parameter V ^t , including:

The server uses the

The updated product factor matrix parameter V ^t+1 , i represents the number of clients, and S _t represents the client set.

The product factor matrix parameter V ^t+1 is updated by collecting the weighted difference Δ _i ^t and normalized constant h _i ^t data fed back by all clients in the client set S _t . Improves the universality of parameters.

S5, the server side determines the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sends the target product factor matrix to the target client;

Further, determining the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , including:

Determine whether the difference between the updated product factor matrix parameter V ^t+1 and the pre-updated product factor matrix parameter V ^t is less than a preset value;

If the difference is greater than the preset value, re-send the updated product factor matrix parameter V ^t+1 to the client to continue iterative update until the difference is less than the preset value;

If the difference is less than or equal to the preset value, a target product factor matrix is generated by using the updated product factor matrix parameter V ^t+1 .

Specifically, in an embodiment of the present application, it is determined by comparing the difference between the product factor matrix parameters obtained by two adjacent iterations with a preset value to determine whether iterative update is needed, and if it is greater than the preset value , then the updated product factor matrix parameter V ^t+1 is re-sent to the client to continue the iterative update, that is, V ^t+1 is replaced with the original V ^t , so as to obtain V ^t+2 , and iterate continuously until adjacent iterations get The difference of the parameters of the product factor matrix is less than the preset value; when the difference is less than the preset value, the updated product factor matrix parameters will be used to generate the target product factor matrix; the target client needs to perform The client corresponding to the user recommended by the product, and among the clients participating in the iteration.

By comparing the difference with the preset value, the product factor matrix parameters that meet the conditions are finally obtained, which improves the quality of product recommendation. Recommend products that are closer to their actual needs.

In another embodiment of the present application, the product factor matrix parameters obtained in the last iteration are outputted only after completing the preset number of iterations, and the target product factor matrix is generated by using the product factor matrix parameters.

S6. The target client uses the target product factor matrix and the updated user factor matrix to calculate the predicted score value of the product by the user, sort the predicted score value, and sort the top N products with the highest ranking recommended to the user.

The updated user factor matrix is the last user factor matrix after iterative update.

The clients are in one-to-one correspondence with the users to which they belong.

It should be emphasized that, in order to further ensure the privacy and security of the data, the target product factor matrix data can also be stored in a node of a blockchain.

The blockchain referred to in this application is a new application mode of computer technologies such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm. Blockchain, essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information to verify its Validity of information (anti-counterfeiting) and generation of the next block. The blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.

Through the cooperation between the server side and the client side, the server side receives the parameter data input by the user; the server side obtains the product factor matrix V and the product factor matrix parameter V ^t through initialization processing, and combines the product factor matrix parameter V ^t with the said product factor matrix parameter V t The parameter data is sent to multiple clients; the multiple clients update the user factor matrix u _i according to the product factor matrix parameter V ^t and the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalized value. Normalize the constant h _i ^t , and return the weighted difference Δ _i ^t and the normalized constant h _i ^t to the server side; the client uses the data to perform calculations by itself, avoiding the occurrence of data when the data is sent to the server side The leakage problem protects the user's data; the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , to the product factor matrix parameter V ^t is iteratively updated; the server side determines a target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sends the target product factor matrix to the target client; the target client uses the The target product factor matrix and the updated user factor matrix are used to calculate the predicted score value of the product by the user, sort the predicted score value, and recommend the top N products with the highest ranking to the user. By receiving the product factor matrix parameters sent by the server, the user data is used locally to calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t on the client side, while protecting the user data, the user data is used to complete accurate user recommendations. And the server side only uses the weighted difference Δ _i ^t sent by the client and the normalized constant h _i ^t to calculate, which reduces the model calculation cost.

As shown in FIG. 2 , it is a functional module diagram of the product recommendation system of the present application.

The product recommendation system 100 described in this application can be installed in an electronic device, and specifically can be installed in a server and a client. According to the implemented functions, the product recommendation system 100 includes a server end 110 and a client end 120. The server end 110 includes a receiving module 111, an initialization module 112, an update module 113, and a determination module 114; the client end 120 includes a computing module 121 and output module 122. The module in the present invention can also be called a unit, which refers to an instruction segment of a series of computer-readable instructions that can be executed by the electronic device processor and can perform fixed functions, and is stored in the memory of the electronic device.

In this embodiment, the functions of each module/unit are as follows:

The server side 110 includes:

a receiving module 111, configured to receive parameter data input by a user;

Specifically, the receiving module 111 receives the total number of clients n and the number of epochs E, that is, the number of stochastic gradient descent is E times and the number of loops is T, that is, the number of iterations is T times, the q-FedAvg algorithm parameter q, the Lipschitz constant L , step size α, regularization parameter λ and gain parameter γ parameter.

an initialization module 112, configured to obtain a product factor matrix V and a product factor matrix parameter V ^t through initialization processing, and send the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

The initialization module 112 obtains the product factor matrix parameter V ^t and the product factor matrix V through random initialization.

The server side 110 further includes a homomorphic encryption module;

The homomorphic encryption module is used to perform homomorphic encryption on the product factor matrix parameter V ^t .

The homomorphic encryption module ensures the security of the data.

Update module 113, based on the remaining part of the parameter data, using the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , for the product factor matrix parameter V ^t Iteratively update;

Further, the update module 113 utilizes

The updated product factor matrix parameter V ^t+1 is obtained, i represents the number of clients, and St represents the set of clients.

A determination module 114, configured to determine a target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and send the target product factor matrix to the target client;

Further, the determining module 114 includes a difference judgment sub-module, an iterative sub-module and a generation sub-module;

The difference judgment submodule is used to judge whether the difference between the updated product factor matrix parameter V ^t+1 and the before updated product factor matrix parameter V ^t is less than a preset value;

The iterative submodule is used to re-send the updated product factor matrix parameter V ^t+1 to the client to continue iterative update if the difference is greater than the preset value, until the difference is less than the preset value until;

The generating sub-module is configured to generate a target product factor matrix by using the updated product factor matrix parameter V ^t+1 if the difference is less than or equal to the preset value.

The difference judging sub-module compares and judges the difference between the product factor matrix parameters obtained by two adjacent iterations and the preset value to determine whether iterative update is needed. If it is greater than the preset value, the iterative sub-module will update The final product factor matrix parameter V ^t+1 is re-sent to the client to continue the iterative update, that is, V ^t+1 is replaced with the original V ^t , so as to obtain V ^t+2 , and iterate continuously until the product factor matrix obtained by adjacent iterations The difference value of the parameter is less than the preset value; when the difference value is less than the preset value, the generating sub-module will use the updated product factor matrix parameter to generate the target product factor matrix;

Through the cooperation of the difference judgment sub-module, iterative sub-module and generation sub-module, the difference value is compared with the preset value, and finally the product factor matrix parameters that meet the conditions are obtained, which improves the product recommendation quality, that is, using the product factor The target product factor matrix generated by the matrix parameters combined with the user factor matrix will recommend products that are closer to their actual needs to users.

The client 120 includes:

The calculation module 121 is used to update the user factor matrix u _i according to the product factor matrix parameter V ^t , and use the q-FedAvg algorithm to calculate the weighted difference value Δ _i ^t and the normalization constant h _i ^t , and the weighted difference value Δ _i ^t and normalized constant h _i ^t are returned to the server;

The calculation module 121 includes a loss function calculation submodule, a user factor matrix parameter update submodule and a user factor matrix update submodule;

The loss function calculation sub-module calculates the loss function F _i (V ^t , u _i ^t ) according to the formula 1 based on the user factor matrix _ui and the product factor matrix parameter V ^t , and the formula 1 is:

The user factor matrix parameter update submodule calculates the updated user factor matrix parameter u _i ^t+ according to formula 2 based on the loss function F _i (V ^t , u _i ^t ), the gradient of the user factor matrix _ui and the gain parameter γ ¹ , the second formula is:

The user factor matrix updating submodule updates the user factor matrix u _i according to the updated user factor matrix parameter u _i ^t+1 ;

Among them, t represents the number of iterations, γ represents the gain parameter,

Represents the gradient value of _ui ^t after the last iteration, j=1,2,..., m,m is the number of products, r _ij represents the user i's rating value for product j, λ represents the regularization parameter, v _j represents the product The jth row of the factor matrix V, F _i is the loss function, _ui ^t+1 is the user factor matrix parameter after iteration, and _ui ^t is the user factor matrix parameter.

Through the cooperation of the loss function calculation sub-module, the user factor matrix parameter update sub-module and the user factor matrix update sub-module, the client can maintain its own data and avoid leakage caused by data transmission.

The calculation module 121 further includes a model parameter calculation sub-module, a difference calculation sub-module and a calculation sub-module based on the q-FedAvg algorithm;

The model parameter calculation submodule obtains the updated user factor matrix parameters by

and the preset step size α, perform E times of stochastic gradient descent on the loss function F _i to obtain the model parameters

The difference calculation submodule is based on the model parameters

Calculate the parameter difference ΔV _i ^t of adjacent product factor matrices:

The calculation sub-module based on the q-FedAvg algorithm calculates the weighted difference Δ _i ^t and the normalized constant h _i ^t according to the difference of the adjacent product factor matrix parameters according to the following q-FedAvg algorithm:

h _i ^t =qF _i ^q-1 (V ^t , _ui ^t+1 )||ΔV _i ^t || ² +LF _i ^q (V ^t , _ui ^t+1 );

Among them, ΔV _i ^t represents the difference between adjacent product factor matrix parameters, L represents the Lipschitz constant, q represents the q-FedAvg algorithm parameter, k=1,2,...,E.

Through the cooperation of the parameter calculation sub-module, the difference calculation sub-module and the calculation sub-module based on the q-FedAvg algorithm, the client can calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t based on its own data characteristics And send it to the client, by using its own data to perform calculations on the local end, while ensuring data security, it displays its own data characteristics for subsequent product recommendations.

The output module 122, using the target product factor matrix and the updated user factor matrix, calculates the predicted score value of the product by the user, sorts the predicted score value, and recommends the top N products with the highest ranking to all the products. described user.

It should be emphasized that, in order to further ensure the privacy and security of the data, the target product factor matrix data can also be stored in a node of a blockchain.

By using the above device, the product recommendation system 100 includes a server end 110 and a client end 120, the server end 110 includes a receiving module 111, an initialization module 112, an update module 113, and a determination module 114; the client end 120 includes The calculation module 121 and the output module 122 are used together to calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t by receiving the product factor matrix parameters sent by the server and using user data locally at the client. At the same time of user data, user data is used to complete accurate recommendation for users, and the server side only uses the weighted difference Δ _i ^t sent by the client and the normalized constant h _i ^t to calculate, which reduces the model calculation cost.

To solve the above technical problems, the embodiments of the present application also provide computer equipment. For details, please refer to FIG. 3 , which is a block diagram of a basic structure of a computer device according to this embodiment.

The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that communicate with each other through a system bus. It should be pointed out that only the computer device 4 with components 41-43 is shown in the figure, but it should be understood that it is not required to implement all of the shown components, and that more or less components may be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculation and/or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, special-purpose Integrated circuit (Application Specific Integrated Circuit, ASIC), programmable gate array (Field-Programmable Gate Array, FPGA), digital processor (Digital Signal Processor, DSP), embedded equipment, etc.

The computer equipment may be a desktop computer, a notebook computer, a palmtop computer, a cloud server and other computing equipment. The computer device can perform human-computer interaction with the user through a keyboard, a mouse, a remote control, a touch pad or a voice control device.

The memory 41 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static Random Access Memory (SRAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Programmable Read Only Memory (PROM), Magnetic Memory, Magnetic Disk, Optical Disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4 , such as a hard disk or a memory of the computer device 4 . In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, flash memory card (Flash Card), etc. Of course, the memory 41 may also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4 , such as computer-readable instructions for a product recommendation method. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

The processor 42 may be a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. This processor 42 is typically used to control the overall operation of the computer device 4 . In this embodiment, the processor 42 is configured to execute computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for executing the product recommendation method.

The network interface 43 may include a wireless network interface or a wired network interface, and the network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

In this embodiment, when the processor executes the computer-readable instructions stored in the memory, the steps of the product recommendation method in the above-mentioned embodiment are implemented. Through the cooperation between the server and the client, the server receives the parameter data input by the user; the server is initialized The product factor matrix V and the product factor matrix parameter V ^t are obtained by processing, and the product factor matrix parameter V ^t and the parameter data are sent to multiple clients; the multiple clients are based on the product factor matrix parameter V ^t and the parameter data to update the user factor matrix u _i , and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t , and the weighted difference Δ _i ^t and the normalized constant h _i ^{t is} returned to the server; the client uses the data to calculate by itself, which avoids the problem of data leakage when the data is sent to the server, and protects the user's data; the server utilizes the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalized constant h _i ^t , iteratively update the product factor matrix parameter V ^t ; the server side determines according to the updated product factor matrix parameter V ^t+1 target product factor matrix, and send the target product factor matrix to the target client; the target client uses the target product factor matrix and the updated user factor matrix to calculate the user's predicted score for the product value, and sort the predicted score value, and recommend the top N products with the highest ranking to the user. By receiving the product factor matrix parameters sent by the server, the client locally uses the user data to calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t . While protecting the user data, the user data is used to complete accurate user recommendations. , and the server side only uses the weighted difference Δ _i ^t sent by the client and the normalized constant h _i ^t to calculate, which reduces the model calculation cost.

The present application also provides another embodiment, that is, to provide a computer-readable storage medium, where the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to Make the at least one processor perform the steps of the above-mentioned product recommendation method, and through the cooperation between the server and the client, the server receives the parameter data input by the user; the server obtains the product factor matrix V and the product factor matrix through initialization processing parameter V ^t , and send the product factor matrix parameter V ^t and the parameter data to multiple clients; the multiple clients update the user factor matrix u according to the product factor matrix parameter V ^t and the parameter data _i , and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t , and return the weighted difference Δ _i ^t and the normalized constant h _i ^t to the server; the client By using the data for calculation by itself, the problem of data leakage when the data is sent to the server side is avoided, and the user's data is protected; the server side uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , iteratively update the product factor matrix parameter V ^t ; the server side determines the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and the target product factor matrix is The product factor matrix is sent to the target client; the target client uses the target product factor matrix and the updated user factor matrix to calculate the predicted score value of the product by the user, and sorts the predicted score value to The top N products with the highest ranking are recommended to the user. By receiving the product factor matrix parameters sent by the server, the client locally uses the user data to calculate the weighted difference Δ _i ^t and the normalized constant h _i ^t . While protecting the user data, the user data is used to complete accurate user recommendations. , and the server side only uses the weighted difference Δ _i ^t sent by the client and the normalized constant h _i ^t to calculate, which reduces the model calculation cost. The computer-readable storage medium may be non-volatile or volatile.

From the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general hardware platform, and of course hardware can also be used, but in many cases the former is better implementation. Based on this understanding, the technical solutions of the present application can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products are stored in a storage medium (such as ROM/RAM, magnetic disk, CD-ROM), including several instructions to make a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in the various embodiments of this application.

Obviously, the above-described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The accompanying drawings show the preferred embodiments of the present application, but do not limit the scope of the patent of the present application. This application may be embodied in many different forms, rather these embodiments are provided so that a thorough and complete understanding of the disclosure of this application is provided. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or perform equivalent replacements for some of the technical features. . Any equivalent structures made by using the contents of the description and drawings of the present application, which are directly or indirectly used in other related technical fields, are all within the scope of the patent protection of the present application.

Claims (20)

1. A product recommendation method, the method includes:

   The server side receives the parameter data input by the user;

   The server side obtains the product factor matrix V and the product factor matrix parameter V ^t through initialization processing, and sends the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

   The multiple clients update the user factor matrix u _i according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , Returning the weighted difference value Δ _i ^t and the normalized constant h _i ^t to the server side;

   Based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , to the product factor matrix parameter V ^t Iteratively update;

   The server side determines the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sends the target product factor matrix to the target client;

   The target client uses the target product factor matrix and the updated user factor matrix to calculate the predicted score value of the product by the user, sort the predicted score value, and recommend the top N products with the highest ranking to the product. the user.
2. The product recommendation method according to claim 1, wherein, determining the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , comprising:

   Determine whether the difference between the updated product factor matrix parameter V ^t+1 and the pre-updated product factor matrix parameter V ^t is less than a preset value;

   If the difference is greater than the preset value, re-send the updated product factor matrix parameter V ^t+1 to the client to continue iterative update until the difference is less than the preset value;

   If the difference is less than or equal to the preset value, a target product factor matrix is generated by using the updated product factor matrix parameter V ^t+1 .
3. The product recommendation method according to claim 1, wherein before the sending the product factor matrix parameter V ^t to a plurality of clients, the method further comprises:

   Homomorphically encrypt the product factor matrix parameter V ^t .
4. The product recommendation method according to claim 1, wherein, based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constants h _i ^t , iteratively update the product factor matrix parameter V ^t , including:

   The server uses

   

   The updated product factor matrix parameter V ^t+1 is obtained, i represents the number of clients, and S _t represents the set of clients.
5. The product recommendation method according to claim 1, wherein the plurality of clients update the user factor matrix _ui according to the product factor matrix parameter V ^t and part of the parameter data, including:

   Based on the user factor matrix _ui and the product factor matrix parameter V ^t , the loss function F _i is calculated according to formula 1, and the formula 1 is:

   

   Based on the loss function F _i , the gradient of the user factor matrix _ui and the gain parameter γ, the updated user factor matrix parameter u _i ^t+1 is calculated according to the formula 2, and the formula 2 is:

   

   Update the user factor matrix _ui according to the updated user factor matrix parameter _ui ^t+1 ;

   Among them, t represents the number of iterations, γ represents the gain parameter,

   

   Represents the gradient value of _ui ^t after the last iteration, j=1,2,..., m,m is the number of products, r _ij represents the user i's rating value for product j, λ represents the regularization parameter, v _j represents the product In the jth row of the factor matrix V, F _i is the loss function, _ui ^t+1 is the user factor matrix parameter after iteration, and _ui ^t is the user factor matrix parameter.
6. The product recommendation method according to claim 5, wherein the calculating the weighted difference Δ _i ^t and the normalized constant h _i ^t by using the q-FedAvg algorithm comprises:

   Obtain the updated user factor matrix parameters

   

   and the preset step size α in the parameter data, perform E times of stochastic gradient descent on the loss function F _i to obtain the model parameters

   

   based on the model parameters

   

   Calculate the parameter difference ΔV _i ^t of adjacent product factor matrices:

   

   According to the adjacent product factor matrix parameter difference ΔV _i ^t , the weighted difference value Δ _i ^t and the normalization constant h _i ^t are calculated according to the following q-FedAvg algorithm:

   Δ _i ^t =Fi ^q (V ^t , u _i ^t+1 _{)ΔV i t} ^;

   h _i ^t =qF _i ^q-1 (V ^t , _ui ^t+1 )||ΔV _i ^t || ² +LF _i ^q (V ^t , _ui ^t+1 );

   Among them, ΔV _i ^t represents the difference between adjacent product factor matrix parameters, L represents the Lipschitz constant, q represents the q-FedAvg algorithm parameter, k=1,2,...,E.
7. A product recommendation system, the product recommendation system includes a server and a client;

   The server includes:

   The receiving module is used to receive the parameter data input by the user;

   an initialization module, configured to obtain a product factor matrix V and a product factor matrix parameter V ^t through initialization, and send the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

   The update module, based on the remaining part of the parameter data, uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t to perform the product factor matrix parameter V ^t . iterative update;

   A determination module, for determining a target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sending the target product factor matrix to the target client;

   The client includes:

   The calculation module is used to update the user factor matrix _ui according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , The weighted difference Δ _i ^t and the normalized constant h _i ^t are returned to the server;

   The output module, using the target product factor matrix and the updated user factor matrix, calculates the predicted score value of the product by the user, sorts the predicted score value, and recommends the top N products with the highest ranking to the user.
8. The product recommendation system according to claim 7, wherein the update module comprises a difference judgment submodule, an iterative submodule and a generation submodule;

   The difference judgment submodule is used to judge whether the difference between the updated product factor matrix parameter V ^t+1 and the before updated product factor matrix parameter V ^t is less than a preset value;

   The iterative submodule is configured to re-send the updated product factor matrix parameter V ^t+1 to the client to continue iterative update if the difference is greater than a preset value, until the difference is less than the until the preset value;

   The generating sub-module is configured to generate a target product factor matrix by using the updated product factor matrix parameter V ^t+1 if the difference is less than or equal to the preset value.
9. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, the processor implementing the following steps when executing the computer-readable instructions:

   The server side receives the parameter data input by the user;

   The server side obtains the product factor matrix V and the product factor matrix parameter V ^t through initialization processing, and sends the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

   The multiple clients update the user factor matrix u _i according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , Returning the weighted difference value Δ _i ^t and the normalized constant h _i ^t to the server side;

   Based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , to the product factor matrix parameter V ^t Iteratively update;

   The server side determines the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sends the target product factor matrix to the target client;

   The target client uses the target product factor matrix and the updated user factor matrix to calculate the predicted score value of the product by the user, sort the predicted score value, and recommend the top N products with the highest ranking to the product. the user.
10. The computer device according to claim 9, wherein the determining the target product factor matrix according to the updated product factor matrix parameter V ^t+1 comprises:

    Determine whether the difference between the updated product factor matrix parameter V ^t+1 and the pre-updated product factor matrix parameter V ^t is less than a preset value;

    If the difference is greater than the preset value, re-send the updated product factor matrix parameter V ^t+1 to the client to continue iterative update until the difference is less than the preset value;

    If the difference is less than or equal to the preset value, a target product factor matrix is generated by using the updated product factor matrix parameter V ^t+1 .
11. The computer device according to claim 9, wherein before the sending the product factor matrix parameter V ^t to a plurality of clients, further comprising:

    Homomorphically encrypt the product factor matrix parameter V ^t .
12. The computer device according to claim 9, wherein, based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , iteratively update the product factor matrix parameter V ^t , including:

    The server uses

    

    The updated product factor matrix parameter V ^t+1 is obtained, i represents the number of clients, and S _t represents the set of clients.
13. The computer device according to claim 9, wherein the plurality of clients update the user factor matrix _ui according to the product factor matrix parameter V ^t and part of the parameter data, comprising:

    Based on the user factor matrix _ui and the product factor matrix parameter V ^t , the loss function F _i is calculated according to formula 1, and the formula 1 is:

    

    Based on the loss function F _i , the gradient of the user factor matrix _ui and the gain parameter γ, the updated user factor matrix parameter u _i ^t+1 is calculated according to the formula 2, and the formula 2 is:

    

    Update the user factor matrix _ui according to the updated user factor matrix parameter _ui ^t+1 ;

    Among them, t represents the number of iterations, γ represents the gain parameter,

    

    Represents the gradient value of _ui ^t after the last iteration, j=1,2,..., m,m is the number of products, r _ij represents the user i's rating value for product j, λ represents the regularization parameter, v _j represents the product In the jth row of the factor matrix V, F _i is the loss function, _ui ^t+1 is the user factor matrix parameter after iteration, and _ui ^t is the user factor matrix parameter.
14. The computer device according to claim 13, wherein said calculating the weighted difference value Δ _i ^t and the normalization constant h _i ^t using the q-FedAvg algorithm comprises:

    Obtain the updated user factor matrix parameters

    

    and the preset step size α in the parameter data, perform E times of stochastic gradient descent on the loss function F _i to obtain the model parameters

    

    based on the model parameters

    

    Calculate the parameter difference ΔV _i ^t of adjacent product factor matrices:

    

    According to the adjacent product factor matrix parameter difference ΔV _i ^t , the weighted difference value Δ _i ^t and the normalization constant h _i ^t are calculated according to the following q-FedAvg algorithm:

    Δ _i ^t =Fi ^q (V ^t , u _i ^t+1 _{)ΔV i t} ^;

    h _i ^t =qF _i ^q-1 (V ^t , _ui ^t+1 )||ΔV _i ^t || ² +LF _i ^q (V ^t , _ui ^t+1 );

    Among them, ΔV _i ^t represents the difference between adjacent product factor matrix parameters, L represents the Lipschitz constant, q represents the q-FedAvg algorithm parameter, k=1,2,...,E.
15. A computer-readable storage medium, where computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, the processor is caused to perform the following steps:

    The server side receives the parameter data input by the user;

    The server side obtains the product factor matrix V and the product factor matrix parameter V ^t through initialization processing, and sends the product factor matrix parameter V ^t and part of the parameter data to multiple clients;

    The multiple clients update the user factor matrix u _i according to the product factor matrix parameter V ^t and part of the parameter data, and use the q-FedAvg algorithm to calculate the weighted difference Δ _i ^t and the normalization constant h _i ^t , Returning the weighted difference value Δ _i ^t and the normalized constant h _i ^t to the server side;

    Based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalization constant h _i ^t , to the product factor matrix parameter V ^t Iteratively update;

    The server side determines the target product factor matrix according to the updated product factor matrix parameter V ^t+1 , and sends the target product factor matrix to the target client;

    The target client uses the target product factor matrix and the updated user factor matrix to calculate the predicted score value of the product by the user, sort the predicted score value, and recommend the top N products with the highest ranking to the product. the user.
16. The computer-readable storage medium according to claim 15, wherein the determining the target product factor matrix according to the updated product factor matrix parameter V ^t+1 comprises:

    Determine whether the difference between the updated product factor matrix parameter V ^t+1 and the pre-updated product factor matrix parameter V ^t is less than a preset value;

    If the difference is greater than the preset value, re-send the updated product factor matrix parameter V ^t+1 to the client to continue iterative update until the difference is less than the preset value;

    If the difference is less than or equal to the preset value, a target product factor matrix is generated by using the updated product factor matrix parameter V ^t+1 .
17. The computer-readable storage medium of claim 15, wherein before the sending the product factor matrix parameter V ^t to a plurality of clients, further comprising:

    Homomorphically encrypt the product factor matrix parameter V ^t .
18. The computer-readable storage medium according to claim 15, wherein, based on the remaining part of the parameter data, the server uses the product factor matrix parameter V ^t and all received weighted differences Δ _i ^t and normalized The constant h _i ^t is used to iteratively update the product factor matrix parameter V ^t , including:

    The server uses

    

    The updated product factor matrix parameter V ^t+1 is obtained, i represents the number of clients, and S _t represents the set of clients.
19. The computer-readable storage medium of claim 15, wherein the plurality of clients update the user factor matrix _ui according to the product factor matrix parameter V ^t and part of the parameter data, comprising:

    Based on the user factor matrix _ui and the product factor matrix parameter V ^t , the loss function F _i is calculated according to formula 1, and the formula 1 is:

    

    Based on the loss function F _i , the gradient of the user factor matrix _ui and the gain parameter γ, the updated user factor matrix parameter u _i ^t+1 is calculated according to the formula 2, and the formula 2 is:

    

    Update the user factor matrix _ui according to the updated user factor matrix parameter _ui ^t+1 ;

    Among them, t represents the number of iterations, γ represents the gain parameter,

    

    Represents the gradient value of _ui ^t after the last iteration, j=1,2,..., m,m is the number of products, r _ij represents the user i's rating value for product j, λ represents the regularization parameter, v _j represents the product In the jth row of the factor matrix V, F _i is the loss function, _ui ^t+1 is the user factor matrix parameter after iteration, and _ui ^t is the user factor matrix parameter.
20. The computer-readable storage medium according to claim 19, wherein the product recommendation method according to claim 5, wherein the weighted difference value Δ _i ^t and the normalization constant h _i ^t are calculated using the q-FedAvg algorithm include:

    Obtain the updated user factor matrix parameters

    

    and the preset step size α in the parameter data, perform E times of stochastic gradient descent on the loss function F _i to obtain the model parameters

    

    based on the model parameters

    

    Calculate the parameter difference ΔV _i ^t of adjacent product factor matrices:

    

    According to the adjacent product factor matrix parameter difference ΔV _i ^t , calculate the weighted difference according to the following q-FedAvg algorithm

    Δ _i ^t and normalizing constant h _i ^t :

    Δ _i ^t =Fi ^q (V ^t , u _i ^t+1 _{)ΔV i t} ^;

    h _i ^t =qF _i ^q-1 (V ^t , _ui ^t+1 )||ΔV _i ^t || ² +LF _i ^q (V ^t , _ui ^t+1 );

    Among them, ΔV _i ^t represents the difference between adjacent product factor matrix parameters, L represents the Lipschitz constant, q represents the q-FedAvg algorithm parameter, k=1,2,...,E.

Images