WO2021121106A1 - Procédé, appareil et dispositif de recommandation personnalisée basées sur un apprentissage fédéré, et support - Google Patents

Procédé, appareil et dispositif de recommandation personnalisée basées sur un apprentissage fédéré, et support Download PDF

Info

Publication number
WO2021121106A1
WO2021121106A1 PCT/CN2020/135030 CN2020135030W WO2021121106A1 WO 2021121106 A1 WO2021121106 A1 WO 2021121106A1 CN 2020135030 W CN2020135030 W CN 2020135030W WO 2021121106 A1 WO2021121106 A1 WO 2021121106A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
personalized recommendation
model
federated learning
gradient
Prior art date
Application number
PCT/CN2020/135030
Other languages
English (en)
Chinese (zh)
Inventor
黄福华
刘畅
郑文琛
Original Assignee
深圳前海微众银行股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳前海微众银行股份有限公司 filed Critical 深圳前海微众银行股份有限公司
Publication of WO2021121106A1 publication Critical patent/WO2021121106A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9538—Presentation of query results

Definitions

  • This application relates to the field of artificial intelligence technology of financial technology (Fintech), and in particular to a method, device, device, and medium for personalized recommendation based on federated learning.
  • personalized recommendation technology With the continuous development of computer software and artificial intelligence, personalized recommendation technology has become more and more widely used.
  • personalized recommendation providers usually obtain user attribute data, user behavior data, user behavior context data and other data obtained by themselves. , Predict the user’s personalized behavior or item, for example, predict the user’s favorite mobile phone and predict the user’s web page click-through rate, etc.
  • the feature richness of user data in this method often has a great impact on the prediction result. Due to the tightening of data privacy protection legislation, user data cannot be shared in clear text between different data owners. Therefore, the feature richness of user data of a single personalized recommendation provider is often low, which in turn leads to personalized behavior or The prediction accuracy of items is also low, which in turn leads to a poor personalized recommendation effect. Therefore, there is a technical problem in related technologies that the personalized recommendation effect is poor.
  • the main purpose of this application is to provide a method, device, device and medium for personalized recommendation based on federated learning, aiming to solve the technical problem of poor personalized recommendation effect in related technologies.
  • the embodiments of the present application provide a personalized recommendation method based on federated learning.
  • the personalized recommendation method based on federated learning is applied to a personalized recommendation device based on federated learning.
  • the personalized recommendation method based on federated learning Recommended methods include:
  • the output results of the model are screened to obtain personalized recommendation results.
  • the target recall set includes a list of items to be recommended
  • the upload data includes user data, item data, and behavior data
  • the step of obtaining the data to be predicted corresponding to the uploaded data and the target recall set, and inputting the data to be predicted into a personalized recommendation model obtained based on the federated learning, and obtaining the model output result includes:
  • the data to be predicted is input into the personalized recommendation model to score and sort the recommended items in the list of recommended items to obtain model output results.
  • the step of inputting the to-be-predicted data into the personalized recommendation model to score and sort the to-be-recommended items in the to-be-recommended item list, and obtaining model output results includes:
  • the items to be recommended are sorted, and the model output result is obtained.
  • the personalized recommendation method based on federated learning is applied to the first device that performs the federated learning
  • the step of obtaining the to-be-predicted data corresponding to the uploaded data and the target recall set, and inputting the to-be-predicted data into the personalized recommendation model obtained based on the federated learning, before the step of obtaining the model output result includes:
  • the personalized recommendation model is obtained by interacting with the second device for federated learning.
  • the personalized recommendation model includes a logistic regression model
  • the step of obtaining the personalized recommendation model by interacting with the second device for federated learning based on the public training sample ID includes:
  • Receive the second weight value sent by the second device calculate the gradient auxiliary variable corresponding to the first weight value and the second weight value through a preset intermediate parameter formula, and send the gradient auxiliary variable to A second device, wherein the second device is used to calculate a second gradient corresponding to the gradient auxiliary variable;
  • the first gradient is calculated, and the first gradient is sent to a preset federation server, where the preset federation server is used to send data based on the first gradient and the second device
  • the second gradient calculates the total gradient of the federated model
  • the total gradient of the federation model fed back by the federation server is received, and the local training model of the first device is iteratively updated based on the total gradient of the federation model to obtain the logistic regression model.
  • the step of receiving uploaded data and extracting a target recall set corresponding to the uploaded data from a preset recall set storage database includes:
  • the personalized recommendation method based on federated learning is applied to the first device that performs the federated learning
  • the step of receiving uploaded data and extracting the target recall set corresponding to the uploaded data from a preset recall set storage database includes:
  • the second device associated with the first device performs federated learning to obtain a federated recall algorithm model
  • the present application also provides a personalized recommendation device based on federated learning.
  • the personalized recommendation device based on federated learning is applied to a personalized recommendation device based on federated learning.
  • the personalized recommendation device based on federated learning includes:
  • the extraction module is configured to receive the uploaded data, and extract the target recall set corresponding to the uploaded data from a preset recall set storage database;
  • a prediction module configured to obtain the data to be predicted corresponding to the uploaded data and the target recall set, and input the data to be predicted into a personalized recommendation model obtained based on the federated learning to obtain a model output result;
  • the screening module is used to screen the output results of the model to obtain personalized recommendation results.
  • the prediction module includes:
  • the screening unit is configured to screen the item data based on the list of items to be recommended to obtain the item data to be recommended, and set the user data, the item data to be recommended and the behavior data as all State the data to be predicted;
  • the prediction unit is used for inputting the data to be predicted into the personalized recommendation model, so as to score and sort the recommended items in the list of to-be-recommended items, and obtain model output results.
  • the prediction unit includes:
  • a scoring subunit for inputting the data to be predicted into the personalized recommendation model to score the item to be recommended based on the user data and the behavior data to obtain a scoring result
  • the sorting subunit is used for sorting the items to be recommended based on the scoring result to obtain a model output result.
  • the personalized recommendation device based on federated learning further includes:
  • a sample matching module configured to perform sample matching on the second device associated with the first device to obtain a public training sample ID
  • the first federated learning module is used to obtain the personalized recommendation model by interacting with the second device to perform federated learning based on the public training sample ID.
  • the federated learning module includes:
  • a first calculation unit configured to extract the first sample data corresponding to the public training sample ID based on the public training sample ID, and calculate the first weight corresponding to the first sample data
  • the second calculation unit is configured to receive the second weight value sent by the second device, and calculate the gradient auxiliary variable corresponding to the first weight value and the second weight value by using a preset intermediate parameter formula, And sending the gradient auxiliary variable to a second device, where the second device is used to calculate a second gradient corresponding to the gradient auxiliary variable;
  • the third calculation unit is configured to calculate a first gradient based on the gradient auxiliary variable, and send the first gradient to a preset federated server, where the preset federated server is used to calculate the first gradient based on the first gradient.
  • An iterative update unit configured to receive the total gradient of the federation model fed back by the federation server, and to iteratively update the local training model of the first device based on the total gradient of the federation model to obtain the logistic regression model .
  • the extraction module includes:
  • An extraction unit configured to receive the uploaded data and extract the sample ID in the uploaded data
  • the query unit is configured to query the corresponding target recall set in the preset recall set storage database based on the sample ID.
  • the personalized recommendation device based on federated learning includes:
  • a second federated learning module configured to perform federated learning for the second device associated with the first device to obtain a federated recall algorithm model
  • the storage module is used to obtain the sample upload data and input the sample upload data into the federated recall algorithm model to obtain a target recall set set, and store the target recall set set in the preset recall set storage In the database.
  • the personalized recommendation device based on federated learning includes: a memory, a processor, and the The program of the personalized recommendation method based on federated learning, when the program of the personalized recommendation method based on federated learning is executed by the processor, can realize the steps of the above-mentioned personalized recommendation method based on federated learning.
  • This application also provides a medium, the medium is a readable storage medium, the medium stores a program for implementing a personalized recommendation method based on federated learning, and the program for the personalized recommendation method based on federated learning is processed by a processor During execution, the steps of the above-mentioned personalized recommendation method based on federated learning are realized.
  • This application receives uploaded data, extracts the target recall set corresponding to the uploaded data from a preset recall set storage database, and then obtains the to-be-predicted data corresponding to the uploaded data and the target recall set, and combines the
  • the data to be predicted is input based on the personalized recommendation model obtained by the federated learning, and the model output result is obtained, and then the model output result is filtered to obtain the personalized recommendation result. That is, this application first receives the upload data, and then extracts the target recall set corresponding to the uploaded data from the preset recall set storage database, and then performs the common correspondence between the upload data and the target recall set.
  • the personalized recommendation model is acquired based on federated learning, that is, through the federated learning, the training of the personalized recommendation model can be combined with multi-party data, thereby increasing the feature richness of the training samples of the personalized recommendation model, and The data privacy of each data provider and data user will not be disclosed, thereby improving the robustness and broadness of the personalized recommendation model, thereby improving the prediction accuracy of the personalized recommendation model, and avoiding the result of the personalized recommendation model
  • the prediction accuracy rate of which leads to poor personalized recommendation effect, and by inputting the data to be predicted into the personalized recommendation model obtained based on the federated learning, the joint multi-party data for prediction is achieved when only local data is used for prediction.
  • the prediction effect of the model reduces the amount of calculation for personalized recommendation, thereby improving the response speed during personalized recommendation, and enhancing the effect of personalized recommendation. Therefore, the technical problem of poor personalized recommendation effect is solved.
  • the embodiment of the present application provides a personalized recommendation method based on federated learning.
  • the personalized recommendation method based on federated learning is applied to a personalized recommendation device based on federated learning.
  • the personalized recommendation method based on federated learning includes:
  • Step S10 receiving uploaded data, and extracting a target recall set corresponding to the uploaded data from a preset recall set storage database;
  • the upload data includes sample ID, user data, item data, and behavior data.
  • the upload data is uploaded by the client, and the user data includes user natural attribute data and user data. Interest attribute data, etc.
  • the item data includes data such as the item name and item attributes of the item to be personalized and recommended
  • the behavior data includes user behavior data on the item and context data when the behavior occurs, such as user behavior on the item
  • the data includes browsing, clicking, etc.
  • the context data when the behavior occurs includes geographic location, network type, etc.
  • the sample ID includes the user's name, user ID number, and user phone number.
  • the database is queried for the target recall set corresponding to the user data, where the target recall set includes one or more initial items to be recommended, where the initial items to be recommended refer to items that may be of interest to the user.
  • the step of receiving uploaded data and extracting a target recall set corresponding to the uploaded data from a preset recall set storage database includes:
  • Step S11 receiving uploaded data, and extracting the sample ID in the uploaded data
  • the sample ID includes identity tags such as user IDs and character strings
  • the uploaded data includes one or more data samples
  • the data samples are represented by a preset data sample format.
  • id, label, user_a_feature_i, item_feature_i, action_i, other_a_i is a preset data sample format, where id is the sample ID and label is the sample label, where the sample label identifies the type of user, for example, The user is a good customer or a bad customer, user_a_feature_i is user data, item_feature_i is item data, action_i is behavior data, and other_a_i is other data.
  • Step S12 based on the sample ID, query a corresponding target recall set in the preset recall set storage database.
  • the sample ID includes identity tags such as user ID, character string, phone number, etc. That is, the sample ID can be represented by a character string, or can be directly represented by a phone number .
  • the corresponding target recall set is queried in the preset recall set storage database.
  • the sample ID is used as a keyword to search in the preset recall set storage database to obtain all The target recall set.
  • the personalized recommendation method based on federated learning is applied to the first device that performs the federated learning
  • the step of receiving uploaded data and extracting the target recall set corresponding to the uploaded data from a preset recall set storage database includes:
  • Step A10 the second device associated with the first device performs federated learning to obtain a federated recall algorithm model
  • the federated learning includes vertical federated learning and horizontal federated learning, and the first device and the second device may be communicatively connected.
  • the second device associated with the first device performs federated learning to obtain a federated recall algorithm model, specifically, performs sample alignment with the second device to align the sample ID of the first device with the second device Align the sample ID of the sample to obtain the public training sample ID, train the first training model of the first device based on the uploaded data corresponding to the public training sample ID, obtain the first training result, and calculate the first training result
  • the error from the theoretical training result is further based on the error and the network weight of the first training model to obtain a partial derivative of the preset objective function to obtain the first gradient, wherein the preset objective function is related to the error And the network weight function, and then send the first gradient to a preset federation server, so that the preset federation server sends the first gradient and the second device to the first gradient and the second device based on preset federation rules.
  • the two gradients are federated to obtain the federated gradient, and then the federated gradient fed back by the preset federated server is received, and the first training model is iteratively updated based on the federated gradient to obtain the federated recall algorithm model.
  • the preset federal rules include weighted average, and the conditions for stopping the iterative update include reaching the maximum number of iterations, and the model converges to a preset error threshold.
  • Step A20 Obtain sample upload data, and input the sample upload data into the federal recall algorithm model to obtain a target recall set set, and store the target recall set set in the preset recall set storage database.
  • the target recall set includes one or more target recall sets, wherein each sample ID corresponds to a target recall set.
  • sample upload data and input the sample upload data into the federal recall algorithm model to obtain a target recall set set, and store the target recall set set in the preset recall set storage database, specifically, obtain Sample upload data, where the sample upload data includes upload data of each user stored in the first device, and each user corresponds to a sample ID, where the upload data includes user data, item data, and behavior data Etc., and then input the sample upload data into the federal recall algorithm model to predict the items that may be of interest to the user of the first device, obtain one or more items to be recommended, and assign them to the items to be recommended corresponding to the same sample ID.
  • the recommended item is divided into a target recall set, the target recall set is obtained, and the target recall set is stored in the preset recall set storage database, where the target recall set is in the preset recall set
  • the collection storage database is stored in the form of item lists, item collections, etc., and the query keywords corresponding to the target recall collection include sample IDs.
  • Step S20 Obtain the data to be predicted corresponding to the uploaded data and the target recall set, and input the data to be predicted into a personalized recommendation model obtained based on the federated learning to obtain a model output result;
  • the participants include a first device and a second device, and the personalized learning obtained based on the federated learning
  • the recommendation model includes a logistic regression model that has been trained based on the federated learning, the output of the model includes items to be recommended and their ratings and rankings, and the data to be predicted includes user data, behavior data, and corresponding target recall sets. Item data, etc.
  • the data to be predicted corresponding to the target recall set is input based on the personalized recommendation model obtained by the federated learning to score and rank the recommended items corresponding to the target recall set, and obtain the corresponding items to be recommended
  • the sorted list of items and the scores corresponding to the items to be recommended, that is, the output results of the model are obtained.
  • the target recall set includes a list of items to be recommended
  • the upload data includes user data, item data, and behavior data
  • the step of obtaining the data to be predicted corresponding to the uploaded data and the target recall set, and inputting the data to be predicted into a personalized recommendation model obtained based on the federated learning, and obtaining the model output result includes:
  • Step S21 Screen the item data based on the list of items to be recommended to obtain item data to be recommended, and set the user data, the item data to be recommended and the behavior data as the data to be predicted ;
  • the items in the list of items to be recommended are associated with the item data to be recommended.
  • Step S22 Input the to-be-predicted data into the personalized recommendation model to score and sort the to-be-recommended items in the to-be-recommended item list to obtain model output results.
  • the to-be-predicted data is input into the personalized recommendation model to score and sort the to-be-recommended items in the to-be-recommended item list to obtain the model output result
  • the The prediction data is input into the personalized recommendation model to score the items in the list of items to be recommended based on the user data, the behavior data, and the item data to be recommended to obtain a scoring result, which is then based on the score
  • the items in the list of items to be recommended are sorted, and the output result of the model is obtained.
  • step S22 the step of inputting the to-be-predicted data into the personalized recommendation model to score and sort the to-be-recommended items in the to-be-recommended item list, and obtaining model output results includes:
  • Step S221 Input the to-be-predicted data into the personalized recommendation model to score the to-be-recommended item based on the user data and the behavior data to obtain a scoring result;
  • the to-be-predicted data is input into the personalized recommendation model to score the to-be-recommended item based on the user data and the behavior data to obtain a scoring result
  • the The data to be predicted is input into the personalized recommendation model to score the items to be recommended based on the user data and the behavior data, to obtain a first scoring result and a second scoring result, based on the first scoring result
  • the scoring result is calculated by a preset calculation rule, where the preset calculation rule includes weighted average, summation, quadrature, etc., for example, assuming that the calculation rule is a summation, If the first scoring result is 1 point and the second scoring result is 2 points, then the scoring result is 3 points.
  • Step S222 Sort the items to be recommended based on the scoring result to obtain a model output result.
  • the items to be recommended are sorted to obtain a model output result, specifically, based on the scoring result, the scoring items to be recommended are performed in a preset sorting manner Sorting, wherein the preset sorting method includes sorting from small to large, sorting from large to small, etc., to obtain model output results.
  • Step S30 Screen the output results of the model to obtain personalized recommendation results.
  • the model output results are filtered to obtain personalized recommendation results.
  • the model output results are filtered based on preset business logic, and the items to be recommended in the model output results are extracted as The personalized recommendation result, for example, extracting a preset number of items with the highest score in the model output result as the personalized recommendation result.
  • upload data is received, and the target recall set corresponding to the uploaded data is extracted from a preset recall set storage database, and then the to-be-predicted data corresponding to the uploaded data and the target recall set are obtained, and all The data to be predicted is input based on the personalized recommendation model obtained by the federated learning, and the model output result is obtained, and then the model output result is filtered to obtain the personalized recommendation result. That is, this embodiment first receives the upload data, and then extracts the target recall set corresponding to the uploaded data from the preset recall set storage database, and then performs the common correspondence between the upload data and the target recall set.
  • the personalized recommendation model is obtained based on federated learning, that is, through the federated learning, the training of the personalized recommendation model can be combined with multi-party data, thereby increasing the feature richness of the training samples of the personalized recommendation model, The data privacy of each data provider and data user will not be leaked, thereby improving the robustness and broadness of the personalized recommendation model, thereby improving the prediction accuracy of the personalized recommendation model, and avoiding the result of personalized recommendation
  • the prediction accuracy of the model leads to poor personalized recommendation effects, and by inputting the data to be predicted into the personalized recommendation model obtained based on the federated learning, joint multi-party data processing is achieved when only local data is used for prediction.
  • the prediction effect of the predicted model reduces the amount of calculation when making personalized recommendation, thereby improving the response speed during personalized recommendation and enhancing the effect of personalized recommendation. Therefore, the technical problem of poor personalized recommendation effect is solved.
  • the personalized recommendation method based on federated learning is applied to the federated learning
  • the first device
  • the step of obtaining the to-be-predicted data corresponding to the uploaded data and the target recall set, and inputting the to-be-predicted data into the personalized recommendation model obtained based on the federated learning, before the step of obtaining the model output result includes:
  • Step B10 Perform sample matching with the second device associated with the first device to obtain a public training sample ID
  • the public training sample ID includes identity tags such as user IDs and identification strings.
  • the second device associated with the first device performs sample matching to obtain a common training sample ID, specifically, the first training sample ID in the first device is compared with the second training sample ID in the second device Perform intersection processing to obtain the public training sample ID.
  • Step B20 Based on the public training sample ID, the personalized recommendation model is obtained by interacting with the second device for federated learning.
  • the personalized recommendation model is obtained by interacting with the second device for federated learning.
  • the public The first sample data corresponding to the training sample ID is obtained, and the first gradient is obtained based on the first sample data, and the second device interacts with the second device to assist the second device in obtaining the second gradient, and to The first gradient is sent to the preset federation server, and further, the federated model total gradient fed back by the preset federated server is received, wherein the federated model total gradient is calculated by the preset federated server through preset federation rules.
  • the second gradient sent by the second device where the preset federation rules include weighted average, summation, etc.
  • the first training model in the first device is further determined based on the total gradient of the federation model. Iterative update is performed to obtain the personalized recommendation model.
  • the personalized recommendation model includes a logistic regression model
  • the step of obtaining the personalized recommendation model by interacting with the second device for federated learning based on the public training sample ID includes:
  • Step B21 based on the public training sample ID, extract the first sample data corresponding to the public training sample ID, and calculate the first weight corresponding to the first sample data;
  • the first sample data is upload data of the user client
  • the upload data includes one or more data samples
  • the first weight value is the same as the first The product of the first sample feature corresponding to this data and the first weight in the first training model, where the first weight is the network weight of the first training model, for example, suppose the first weight Is WB, and the first sample feature is XB, then the first weight value is WBXB.
  • Step B22 Receive the second weight value sent by the second device, calculate the gradient auxiliary variable corresponding to the first weight value and the second weight value through a preset intermediate parameter formula, and assist the gradient The variable is sent to a second device, where the second device is used to calculate a second gradient corresponding to the gradient auxiliary variable;
  • the second weight value sent by the second device is received, and the gradient auxiliary variable corresponding to the first weight value and the second weight value is calculated through a preset intermediate parameter formula, and the all The gradient auxiliary variable is sent to a second device, where the second device is used to calculate a second gradient corresponding to the gradient auxiliary variable, specifically, receiving a second weight encrypted and sent by the second device, and encrypting The latter second weight is decrypted to obtain the second weight, where the second weight is the second sample feature corresponding to the second sample data and the second in the second training model in the second device.
  • the product of weights wherein the second weight is the network weight of the second training model, and further, the first weight and the second weight are combined to obtain the first device and the second The total weight corresponding to the two devices, and then substitute the total weight variable into the preset intermediate parameter formula to obtain the gradient auxiliary variable, and send the gradient auxiliary variable to the second device, and the The gradient auxiliary variable is associated with the sample label, wherein the preset intermediate parameter formula is as follows:
  • wTx is the total weight
  • y is the sample label
  • Step B23 Calculate a first gradient based on the gradient auxiliary variable, and send the first gradient to a preset federated server, where the preset federated server is configured to be based on the first gradient and the second
  • the second gradient sent by the device calculates the total gradient of the federated model
  • the first gradient is calculated based on the gradient auxiliary variable, and the first gradient is sent to a preset federated server, where the preset federated server is used to calculate the first gradient based on the first gradient and the
  • the second gradient sent by the second device calculates the total gradient of the federated model, specifically, the product of the gradient auxiliary variable and the first sample feature is calculated to obtain the first gradient, where the first gradient can be obtained by the following formula Calculated,
  • gB is the first sampling gradient
  • wTx is the total weight variable
  • y is the sample label
  • xB is the first sample feature of the first device.
  • the first gradient is sent to the preset federation server to pass the pre- It is assumed that the federation server calculates the total gradient of the federation model corresponding to the first gradient and the second gradient sent by the second device based on a preset federation rule, where the preset federation rule includes weighted average, summation, etc.,
  • the second gradient is the product of the gradient auxiliary variable and the second sample feature.
  • Step B24 Receive the total gradient of the federation model fed back by the federation server, and iteratively update the local training model of the first device based on the total gradient of the federation model to obtain the logistic regression model.
  • the local training model is the first training model.
  • the federation server Receive the total gradient of the federation model fed back by the federation server, and iteratively update the local training model of the first device based on the total gradient of the federation model to obtain the logistic regression model, specifically, receive the federation
  • the total gradient of the federated model fed back by the server, and based on the total gradient of the federated model, the local training model of the first device is trained and updated, and it is judged whether the updated local training device meets the preset training completion condition. If the local training device meets the preset training completion condition, the updated local training device is used as the logistic regression model.
  • the next federated learning The local training device performs training updates until the local training device reaches a preset training completion condition, where the training completion condition includes the model reaching the maximum number of iterations, the model error is less than the preset error threshold and converging.
  • a public training sample ID is obtained by performing sample matching with a second device associated with the first device, and then based on the public training sample ID, by interacting with the second device to perform federated learning, the Personalized recommendation model. That is, this embodiment first performs sample matching of the second device associated with the first device to obtain a public training sample ID, and then based on the public training sample ID, through interaction with the second device, In order to perform federated learning, the personalized recommendation model is obtained. That is, this embodiment provides a method for obtaining a personalized recommendation model through federated learning. That is, by interacting with the second device, the training sample data of the second device can be combined to perform federated learning.
  • FIG. 3 is a schematic diagram of the device structure of the hardware operating environment involved in the solution of the embodiment of the present application.
  • the personalized recommendation device based on federated learning may include a processor 1001, such as a CPU, a memory 1005, and a communication bus 1002.
  • the communication bus 1002 is used to implement connection and communication between the processor 1001 and the memory 1005.
  • the memory 1005 can be a high-speed RAM memory or a stable memory (non-volatile memory), such as disk storage.
  • the memory 1005 may also be a storage device independent of the aforementioned processor 1001.
  • the personalized recommendation device based on federated learning may also include a rectangular user interface, a network interface, a camera, and RF (Radio Frequency (radio frequency) circuits, sensors, audio circuits, WiFi modules, etc.
  • the rectangular user interface may include a display screen (Display) and an input sub-module such as a keyboard (Keyboard), and the optional rectangular user interface may also include a standard wired interface and a wireless interface.
  • the optional network interface can include standard wired interface and wireless interface (such as WI-FI interface).
  • the structure of the personalized recommendation device based on federated learning shown in FIG. 3 does not constitute a limitation on the personalized recommendation device based on federated learning, and may include more or fewer components than shown in the figure. Or some parts are combined, or different parts are arranged.
  • the memory 1005 as a computer storage medium may include an operating system, a network communication module, and a personalized recommendation program based on federated learning.
  • the operating system is a program that manages and controls the hardware and software resources of the personalized recommendation device based on federated learning, and supports the running of the personalized recommendation program based on federated learning and other software and/or programs.
  • the network communication module is used to realize the communication between the components in the memory 1005 and the communication with other hardware and software in the personalized recommendation system based on federated learning.
  • the processor 1001 is configured to execute the personalized recommendation program based on federated learning stored in the memory 1005 to realize the personalized recommendation based on federated learning described in any of the above items. Recommended method steps.
  • the specific implementation of the personalized recommendation device based on federated learning in this application is basically the same as the above-mentioned embodiments of the personalized recommendation method based on federated learning, and will not be repeated here.
  • An embodiment of the present application also provides a personalized recommendation device based on federated learning.
  • the personalized recommendation device based on federated learning includes:
  • the extraction module is configured to receive the uploaded data, and extract the target recall set corresponding to the uploaded data from a preset recall set storage database;
  • a prediction module configured to obtain the data to be predicted corresponding to the uploaded data and the target recall set, and input the data to be predicted into a personalized recommendation model obtained based on the federated learning to obtain a model output result;
  • the screening module is used to screen the output results of the model to obtain personalized recommendation results.
  • the prediction module includes:
  • the screening unit is configured to screen the item data based on the list of items to be recommended to obtain the item data to be recommended, and set the user data, the item data to be recommended and the behavior data as all State the data to be predicted;
  • the prediction unit is used for inputting the data to be predicted into the personalized recommendation model, so as to score and sort the recommended items in the list of to-be-recommended items, and obtain model output results.
  • the prediction unit includes:
  • a scoring subunit for inputting the data to be predicted into the personalized recommendation model to score the item to be recommended based on the user data and the behavior data to obtain a scoring result
  • the sorting subunit is used for sorting the items to be recommended based on the scoring result to obtain a model output result.
  • the personalized recommendation device based on federated learning further includes:
  • a sample matching module configured to perform sample matching on the second device associated with the first device to obtain a public training sample ID
  • the first federated learning module is used to obtain the personalized recommendation model by interacting with the second device to perform federated learning based on the public training sample ID.
  • the federated learning module includes:
  • a first calculation unit configured to extract the first sample data corresponding to the public training sample ID based on the public training sample ID, and calculate the first weight corresponding to the first sample data
  • the second calculation unit is configured to receive the second weight value sent by the second device, and calculate the gradient auxiliary variable corresponding to the first weight value and the second weight value by using a preset intermediate parameter formula, And sending the gradient auxiliary variable to a second device, where the second device is used to calculate a second gradient corresponding to the gradient auxiliary variable;
  • the third calculation unit is configured to calculate a first gradient based on the gradient auxiliary variable, and send the first gradient to a preset federated server, where the preset federated server is used to calculate the first gradient based on the first gradient.
  • An iterative update unit configured to receive the total gradient of the federation model fed back by the federation server, and to iteratively update the local training model of the first device based on the total gradient of the federation model to obtain the logistic regression model .
  • the extraction module includes:
  • An extraction unit configured to receive the uploaded data and extract the sample ID in the uploaded data
  • the query unit is configured to query the corresponding target recall set in the preset recall set storage database based on the sample ID.
  • the personalized recommendation device based on federated learning includes:
  • a second federated learning module configured to perform federated learning for the second device associated with the first device to obtain a federated recall algorithm model
  • the storage module is used to obtain the sample upload data and input the sample upload data into the federated recall algorithm model to obtain a target recall set set, and store the target recall set set in the preset recall set storage In the database.
  • the specific implementation of the personalized recommendation device based on federated learning in this application is basically the same as the above-mentioned embodiments of the personalized recommendation method based on federated learning, and will not be repeated here.
  • the embodiment of the present application provides a medium, the medium is a readable storage medium, and the medium stores one or more programs, and the one or more programs may also be executed by one or more processors to It is used to implement the steps of any one of the above-mentioned personalized recommendation methods based on federated learning.

Abstract

L'invention concerne un procédé, un appareil et un dispositif de recommandation personnalisée basés sur un apprentissage fédéré, et un support. Le procédé de recommandation personnalisée basé sur un apprentissage fédéré comprend les étapes consistant : à recevoir des données téléchargées, et à extraire un ensemble de rappel cible correspondant aux données téléchargées d'une base de données de stockage d'ensembles de rappel prédéfinis (S10) ; à acquérir des données de prédiction correspondant à la fois aux données téléchargées et à l'ensemble de rappel cible, et à entrer les données de prédiction dans un modèle de recommandation personnalisée acquis sur la base de l'apprentissage fédéré et à acquérir un résultat de sortie de modèle (S20) ; et à filtrer le résultat de sortie de modèle pour obtenir un résultat de recommandation personnalisée (S30).
PCT/CN2020/135030 2019-12-20 2020-12-09 Procédé, appareil et dispositif de recommandation personnalisée basées sur un apprentissage fédéré, et support WO2021121106A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911326853.3A CN111079022B (zh) 2019-12-20 2019-12-20 基于联邦学习的个性化推荐方法、装置、设备及介质
CN201911326853.3 2019-12-20

Publications (1)

Publication Number Publication Date
WO2021121106A1 true WO2021121106A1 (fr) 2021-06-24

Family

ID=70316309

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/135030 WO2021121106A1 (fr) 2019-12-20 2020-12-09 Procédé, appareil et dispositif de recommandation personnalisée basées sur un apprentissage fédéré, et support

Country Status (2)

Country Link
CN (1) CN111079022B (fr)
WO (1) WO2021121106A1 (fr)

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113516255A (zh) * 2021-07-28 2021-10-19 深圳前海微众银行股份有限公司 联邦学习建模优化方法、设备、可读存储介质及程序产品
CN113569151A (zh) * 2021-09-18 2021-10-29 平安科技（深圳）有限公司 基于人工智能的数据推荐方法、装置、设备及介质
CN113627496A (zh) * 2021-07-27 2021-11-09 交控科技股份有限公司 道岔转辙机故障预测方法、装置、电子设备和可读存储介质
CN113642707A (zh) * 2021-08-12 2021-11-12 深圳平安智汇企业信息管理有限公司 基于联邦学习的模型训练方法、装置、设备及存储介质
CN113722738A (zh) * 2021-09-02 2021-11-30 脸萌有限公司 数据保护方法、装置、介质及电子设备
CN113762532A (zh) * 2021-06-28 2021-12-07 京东科技控股股份有限公司 联邦学习模型的训练方法、装置、电子设备和存储介质
CN113902473A (zh) * 2021-09-29 2022-01-07 支付宝(杭州)信息技术有限公司 业务预测系统的训练方法及装置
CN114186263A (zh) * 2021-12-17 2022-03-15 大连理工大学 一种基于纵向联邦学习的数据回归方法及电子装置
CN114363176A (zh) * 2021-12-20 2022-04-15 中山大学 一种基于联邦学习的网络辨别方法、装置、终端及介质
CN114676851A (zh) * 2022-04-08 2022-06-28 中国科学技术大学 召回和排序模型的联合训练方法、设备和存储介质
CN114741611A (zh) * 2022-06-08 2022-07-12 杭州金智塔科技有限公司 联邦推荐模型训练方法以及系统
CN114863499A (zh) * 2022-06-30 2022-08-05 广州脉泽科技有限公司 一种基于联邦学习的指静脉与掌静脉识别方法
CN114969503A (zh) * 2022-03-30 2022-08-30 贵州大学 一种基于联邦学习的多元数据用户画像实现方法
CN114996317A (zh) * 2022-07-05 2022-09-02 中国电信股份有限公司 基于纵向联邦学习的异步优化方法、装置及存储介质
CN115169576A (zh) * 2022-06-24 2022-10-11 上海富数科技有限公司广州分公司 基于联邦学习的模型训练方法、装置及电子设备
CN116629374A (zh) * 2023-04-13 2023-08-22 南京大学 一种针对异构特征空间学件的查搜与复用方法
CN117076783A (zh) * 2023-10-16 2023-11-17 广东省科技基础条件平台中心 基于数据分析的科技信息推荐方法、装置、介质及设备
CN117421486A (zh) * 2023-12-18 2024-01-19 杭州金智塔科技有限公司 基于球树算法和联邦学习的推荐模型更新系统及方法
CN117592042A (zh) * 2024-01-17 2024-02-23 杭州海康威视数字技术股份有限公司 面向联邦推荐系统的隐私泄露检测方法及装置

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111079022B (zh) * 2019-12-20 2023-10-03 深圳前海微众银行股份有限公司 基于联邦学习的个性化推荐方法、装置、设备及介质
CN111291417B (zh) * 2020-05-09 2020-08-28 支付宝(杭州)信息技术有限公司 保护数据隐私的多方联合训练对象推荐模型的方法及装置
CN111859136B (zh) * 2020-07-23 2024-03-15 深圳前海微众银行股份有限公司 个性化推荐方法、装置、设备及可读存储介质
CN113781134A (zh) * 2020-07-28 2021-12-10 北京沃东天骏信息技术有限公司 物品推荐方法、装置及计算机可读存储介质
CN111897800B (zh) * 2020-08-05 2021-06-22 全球能源互联网研究院有限公司 一种基于联邦学习的电动汽车充电设施推荐方法及系统
CN112287231B (zh) * 2020-11-05 2024-04-05 深圳大学 一种联邦推荐梯度获取方法、装置、智能终端及存储介质
CN112449009B (zh) * 2020-11-12 2023-01-10 深圳大学 一种基于svd的联邦学习推荐系统通信压缩方法及装置
CN112836830B (zh) * 2021-02-01 2022-05-06 广西师范大学 一种联邦梯度提升决策树投票并行训练方法
CN112966182B (zh) * 2021-03-09 2024-02-09 中国民航信息网络股份有限公司 一种项目推荐方法及相关设备
CN113313268A (zh) * 2021-06-11 2021-08-27 杭州煋辰数智科技有限公司 一种基于联邦学习预测方法、装置、存储介质及遥感设备
CN113656681B (zh) * 2021-07-08 2023-08-11 北京奇艺世纪科技有限公司 一种对象评价方法、装置、设备及存储介质
CN113704555B (zh) * 2021-07-16 2023-11-07 杭州医康慧联科技股份有限公司 基于医疗方向联邦学习的特征管理方法
CN113722537B (zh) * 2021-08-11 2024-04-26 北京奇艺世纪科技有限公司 短视频排序及模型训练方法、装置、电子设备和存储介质
CN114969486B (zh) * 2022-08-02 2022-11-04 平安科技（深圳）有限公司 语料推荐方法、装置、设备及存储介质
CN115238065B (zh) * 2022-09-22 2022-12-20 太极计算机股份有限公司 一种基于联邦学习的公文智能推荐方法
CN117493662A (zh) * 2023-10-09 2024-02-02 上海及未科技有限公司 一种基于联邦学习的个性化推荐方法和系统
CN117454185B (zh) * 2023-12-22 2024-03-12 深圳市移卡科技有限公司 联邦模型训练方法、装置、计算机设备和存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110110229A (zh) * 2019-04-25 2019-08-09 深圳前海微众银行股份有限公司 一种信息推荐方法及装置
CN110297848A (zh) * 2019-07-09 2019-10-01 深圳前海微众银行股份有限公司 基于联邦学习的推荐模型训练方法、终端及存储介质
US20190303838A1 (en) * 2018-03-30 2019-10-03 Atlassian Pty Ltd Using a productivity index and collaboration index for validation of recommendation models in federated collaboration systems
CN110458663A (zh) * 2019-08-06 2019-11-15 上海新共赢信息科技有限公司 一种车辆推荐方法、装置、设备及存储介质
CN111079022A (zh) * 2019-12-20 2020-04-28 深圳前海微众银行股份有限公司 基于联邦学习的个性化推荐方法、装置、设备及介质

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10102478B2 (en) * 2015-06-26 2018-10-16 Conduent Business Services, Inc. Distributed and privacy-preserving prediction method
CN109754105B (zh) * 2017-11-07 2024-01-05 华为技术有限公司 一种预测方法及终端、服务器
US20190182059A1 (en) * 2017-12-12 2019-06-13 Facebook, Inc. Utilizing machine learning from exposed and non-exposed user recall to improve digital content distribution
CN110134868B (zh) * 2019-05-14 2023-04-07 辽宁工程技术大学 一种基于用户偏好异构性分析的推荐方法
CN110443063B (zh) * 2019-06-26 2023-03-28 电子科技大学 自适性保护隐私的联邦深度学习的方法
CN110428058B (zh) * 2019-08-08 2024-04-26 深圳前海微众银行股份有限公司 联邦学习模型训练方法、装置、终端设备及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190303838A1 (en) * 2018-03-30 2019-10-03 Atlassian Pty Ltd Using a productivity index and collaboration index for validation of recommendation models in federated collaboration systems
CN110110229A (zh) * 2019-04-25 2019-08-09 深圳前海微众银行股份有限公司 一种信息推荐方法及装置
CN110297848A (zh) * 2019-07-09 2019-10-01 深圳前海微众银行股份有限公司 基于联邦学习的推荐模型训练方法、终端及存储介质
CN110458663A (zh) * 2019-08-06 2019-11-15 上海新共赢信息科技有限公司 一种车辆推荐方法、装置、设备及存储介质
CN111079022A (zh) * 2019-12-20 2020-04-28 深圳前海微众银行股份有限公司 基于联邦学习的个性化推荐方法、装置、设备及介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
FEI CHEN, MI LUO, ZHENHUA DONG, ZHENGUO LI, XIUQIANG HE: "Federated Meta-Learning with Fast Convergence and Efficient Communication", COMPUTER SCIENCE MACHINE LEARNING (ARXIV), 14 December 2019 (2019-12-14), pages 1 - 14, XP081556998 *

Cited By (34)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113762532B (zh) * 2021-06-28 2024-04-05 京东科技控股股份有限公司 联邦学习模型的训练方法、装置、电子设备和存储介质
CN113762532A (zh) * 2021-06-28 2021-12-07 京东科技控股股份有限公司 联邦学习模型的训练方法、装置、电子设备和存储介质
CN113627496A (zh) * 2021-07-27 2021-11-09 交控科技股份有限公司 道岔转辙机故障预测方法、装置、电子设备和可读存储介质
CN113516255A (zh) * 2021-07-28 2021-10-19 深圳前海微众银行股份有限公司 联邦学习建模优化方法、设备、可读存储介质及程序产品
CN113642707A (zh) * 2021-08-12 2021-11-12 深圳平安智汇企业信息管理有限公司 基于联邦学习的模型训练方法、装置、设备及存储介质
CN113642707B (zh) * 2021-08-12 2023-08-18 深圳平安智汇企业信息管理有限公司 基于联邦学习的模型训练方法、装置、设备及存储介质
CN113722738A (zh) * 2021-09-02 2021-11-30 脸萌有限公司 数据保护方法、装置、介质及电子设备
CN113722738B (zh) * 2021-09-02 2023-08-08 脸萌有限公司 数据保护方法、装置、介质及电子设备
CN113569151A (zh) * 2021-09-18 2021-10-29 平安科技（深圳）有限公司 基于人工智能的数据推荐方法、装置、设备及介质
CN113902473A (zh) * 2021-09-29 2022-01-07 支付宝(杭州)信息技术有限公司 业务预测系统的训练方法及装置
CN114186263A (zh) * 2021-12-17 2022-03-15 大连理工大学 一种基于纵向联邦学习的数据回归方法及电子装置
CN114186263B (zh) * 2021-12-17 2024-05-03 大连理工大学 一种基于纵向联邦学习的数据回归方法及电子装置
CN114363176A (zh) * 2021-12-20 2022-04-15 中山大学 一种基于联邦学习的网络辨别方法、装置、终端及介质
CN114363176B (zh) * 2021-12-20 2023-08-08 中山大学 一种基于联邦学习的网络辨别方法、装置、终端及介质
CN114969503A (zh) * 2022-03-30 2022-08-30 贵州大学 一种基于联邦学习的多元数据用户画像实现方法
CN114969503B (zh) * 2022-03-30 2024-04-02 贵州大学 一种基于联邦学习的多元数据用户画像实现方法
CN114676851B (zh) * 2022-04-08 2024-03-29 中国科学技术大学 召回和排序模型的联合训练方法、设备和存储介质
CN114676851A (zh) * 2022-04-08 2022-06-28 中国科学技术大学 召回和排序模型的联合训练方法、设备和存储介质
CN114741611A (zh) * 2022-06-08 2022-07-12 杭州金智塔科技有限公司 联邦推荐模型训练方法以及系统
CN114741611B (zh) * 2022-06-08 2022-10-14 杭州金智塔科技有限公司 联邦推荐模型训练方法以及系统
CN115169576A (zh) * 2022-06-24 2022-10-11 上海富数科技有限公司广州分公司 基于联邦学习的模型训练方法、装置及电子设备
CN115169576B (zh) * 2022-06-24 2024-02-09 上海富数科技有限公司 基于联邦学习的模型训练方法、装置及电子设备
CN114863499B (zh) * 2022-06-30 2022-12-13 广州脉泽科技有限公司 一种基于联邦学习的指静脉与掌静脉识别方法
CN114863499A (zh) * 2022-06-30 2022-08-05 广州脉泽科技有限公司 一种基于联邦学习的指静脉与掌静脉识别方法
CN114996317B (zh) * 2022-07-05 2024-02-23 中国电信股份有限公司 基于纵向联邦学习的异步优化方法、装置及存储介质
CN114996317A (zh) * 2022-07-05 2022-09-02 中国电信股份有限公司 基于纵向联邦学习的异步优化方法、装置及存储介质
CN116629374B (zh) * 2023-04-13 2023-12-26 南京大学 一种针对异构特征空间学件的查搜与复用方法
CN116629374A (zh) * 2023-04-13 2023-08-22 南京大学 一种针对异构特征空间学件的查搜与复用方法
CN117076783B (zh) * 2023-10-16 2023-12-26 广东省科技基础条件平台中心 基于数据分析的科技信息推荐方法、装置、介质及设备
CN117076783A (zh) * 2023-10-16 2023-11-17 广东省科技基础条件平台中心 基于数据分析的科技信息推荐方法、装置、介质及设备
CN117421486A (zh) * 2023-12-18 2024-01-19 杭州金智塔科技有限公司 基于球树算法和联邦学习的推荐模型更新系统及方法
CN117421486B (zh) * 2023-12-18 2024-03-19 杭州金智塔科技有限公司 基于球树算法和联邦学习的推荐模型更新系统及方法
CN117592042A (zh) * 2024-01-17 2024-02-23 杭州海康威视数字技术股份有限公司 面向联邦推荐系统的隐私泄露检测方法及装置
CN117592042B (zh) * 2024-01-17 2024-04-05 杭州海康威视数字技术股份有限公司 面向联邦推荐系统的隐私泄露检测方法及装置

Also Published As

Publication number Publication date
CN111079022A (zh) 2020-04-28
CN111079022B (zh) 2023-10-03

Similar Documents

Publication Publication Date Title
WO2021121106A1 (fr) Procédé, appareil et dispositif de recommandation personnalisée basées sur un apprentissage fédéré, et support
CN102043833B (zh) 一种基于查询词进行搜索的方法和搜索装置
JP5721818B2 (ja) 検索におけるモデル情報群の使用
JP5575902B2 (ja) クエリのセマンティックパターンに基づく情報検索
WO2021143267A1 (fr) Procédé de traitement de modèle de classification à grain fin basé sur la détection d'image, et dispositifs associés
CN101984422B (zh) 一种容错文本查询的方法和设备
CN101984420B (zh) 一种基于拆词处理进行图片搜索的方法与设备
CN110134800A (zh) 一种文档关系可视化处理方法及装置
WO2010022655A1 (fr) Procédé et système de recherche
CN103136213B (zh) 一种提供相关词的方法及装置
WO2011137125A1 (fr) Procédé, système et dispositif de requête basés sur une recherche verticale
WO2016045567A1 (fr) Procédé et dispositif d'analyse de données de page internet
CN105531701A (zh) 个性化趋势图像搜索建议
CN102456057B (zh) 基于网上交易平台的检索方法、装置和服务器
CN102236710A (zh) 一种用于在查询结果中展现新闻信息的方法与设备
WO2015081792A1 (fr) Procédé, dispositif et système de recherche étendue corrélative et personnalisée
WO2016078533A1 (fr) Procédé, appareil et dispositif de recherche et support de stockage informatique non volatil
CN102402589A (zh) 一种提供与搜索请求相关的参考搜索信息的方法与设备
Babu et al. Improving Quality of Content Based Image Retrieval with Graph Based Ranking
CN103049495A (zh) 用于提供与查询序列相对应的搜索建议的方法、装置与设备
CN110968801A (zh) 地产产品搜索方法、存储介质及电子设备
CN102609539A (zh) 一种搜索方法和系统
CN112825089A (zh) 文章推荐方法、装置、设备及存储介质
JP2011248762A (ja) 分類装置、コンテンツ検索システム、コンテンツ分類方法、コンテンツ検索方法及びプログラム
CN102760127A (zh) 基于扩展文本信息来确定资源类型的方法、装置及设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20903485

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20903485

Country of ref document: EP

Kind code of ref document: A1