WO2021197388A1 - 一种联邦学习中用户的索引方法及联邦学习装置 - Google Patents

一种联邦学习中用户的索引方法及联邦学习装置 Download PDF

Info

Publication number
WO2021197388A1
WO2021197388A1 PCT/CN2021/084610 CN2021084610W WO2021197388A1 WO 2021197388 A1 WO2021197388 A1 WO 2021197388A1 CN 2021084610 W CN2021084610 W CN 2021084610W WO 2021197388 A1 WO2021197388 A1 WO 2021197388A1
Authority
WO
WIPO (PCT)
Prior art keywords
multiple users
users
federated learning
new round
learning
Prior art date
Application number
PCT/CN2021/084610
Other languages
English (en)
French (fr)
Inventor
刘洋
于涵
陈天健
杨强
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 WO2021197388A1 publication Critical patent/WO2021197388A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • the present invention relates to the field of Fintech technology and artificial intelligence technology, in particular to a method for indexing users in federated learning and a federated learning device.
  • Federated Learning uses distributed training and encryption technology to ensure that users' private data are protected to the utmost extent, so as to enhance users' trust in artificial intelligence technology.
  • each participant contributes the encrypted data model to the alliance, jointly trains a federated model, and then opens the model for use by all participants.
  • how to improve the interaction efficiency between federated learning participants and federated learning devices is of great significance to improving the training efficiency of federated learning models.
  • the federated learning device mainly selects the participants of the federated learning through a random method.
  • the federated learning device uses a random algorithm to randomly index a corresponding number of users from users who meet the constraint conditions (the terminal device used must meet the conditions of charging and using non-metering network links such as WiFi) to participate in the federation. Learn.
  • the federated learning device can use a random algorithm to randomly index 8 users from the users meeting the constraint conditions as participants in the federated learning.
  • the federated learning device uses a random method to index the users participating in the federated learning, which lacks optimization considerations.
  • the users randomly indexed by the federated learning device are not optimized and filtered by the federated learning device. Therefore, it is easy for the indexed users to refuse to participate in the federation.
  • the possibility of learning is relatively high (that is, the indexed users are less suitable for federated learning), which leads to the need for repeated trials of federated learning devices to contact users to participate in federated learning, thereby reducing the risk of federated learning participants and federated learning devices.
  • Interactive efficiency is relatively high (that is, the indexed users are less suitable for federated learning), which leads to the need for repeated trials of federated learning devices to contact users to participate in federated learning, thereby reducing the risk of federated learning participants and federated learning devices.
  • the present invention provides an indexing method for users in federated learning and a federated learning device to solve the problem of low interaction efficiency between participants of federated learning and the federated learning device in the prior art.
  • the present invention provides a method for indexing users in federated learning, which includes:
  • the index value is used to characterize the suitability value of each of the multiple users participating in a new round of federated learning
  • each of the multiple users is calculated.
  • the index value of includes:
  • modeling is performed according to the respective feedback data of the multiple users, the confidence of the user portrait, and the number of times each of the multiple users participated in the federated learning modeling in the last time window, and the calculation is performed
  • the respective index values of the multiple users include:
  • the magnitude of the value corresponding to a time point is positively correlated with the time-sharing responsiveness of each of the multiple users at the time point; the time-sharing responsiveness is used to characterize the speed at which the multiple users each feedback accepting the federal study invitation ;
  • calculating the respective index values of the multiple users according to the first probability, the experience loss, and the frequency includes:
  • the largest value among the first probability, the experience loss, and the frequency is used as the respective index values of the multiple users.
  • using the largest value among the first probability, the experience loss, and the frequency as the respective index values of the multiple users includes:
  • the largest value among the first probability, the experience loss, and the frequency is used as the respective index values of the multiple users.
  • inviting users who meet preset conditions among the multiple users to participate in federated learning includes:
  • N users with index values smaller than the first preset index value among the plurality of users are screened out, and the N users participate in a new round of federated learning;
  • N users with index values greater than the second preset index value among the plurality of users are screened out , Invite the N users to participate in a new round of federated learning.
  • the method further includes:
  • Receiving feedback data of the N users updating the probability of the N users participating in a new round of federated learning invitations, the experience loss of the N users, and the frequency of interaction between the federated learning server and the N users, It is used to calculate the suitability value for each of the N users to participate in the next round of federated learning.
  • the present invention provides a federated learning device, and the federated learning device includes:
  • the obtaining unit is used to obtain feedback data of multiple users after accepting the invitation to the federated study in history, and obtain respective user portrait data of the multiple users;
  • the processing unit is used to determine the number of times each of the multiple users participated in the federated learning modeling in the last time window; according to the respective feedback data of the multiple users, the user portrait data, and the multiple users The number of times of participating in federated learning modeling in the last time window, calculating the respective index values of the multiple users; the index value is used to represent the suitability value of each of the multiple users participating in a new round of federated learning;
  • the inviting unit is configured to invite users who meet preset conditions among the multiple users to participate in federated learning according to the respective index values of the multiple users.
  • the processing unit is specifically configured to:
  • the processing unit is specifically configured to:
  • the magnitude of the value corresponding to a time point is positively correlated with the time-sharing responsiveness of each of the multiple users at the time point; the time-sharing responsiveness is used to characterize the speed at which the multiple users each feedback accepting the federal study invitation ;
  • the processing unit is specifically configured to:
  • the largest value among the first probability, the experience loss, and the frequency is used as the respective index values of the multiple users.
  • the processing unit is specifically configured to:
  • the largest value among the first probability, the experience loss, and the frequency is used as the respective index values of the multiple users.
  • the invitation unit is specifically used for:
  • N users with index values smaller than the first preset index value among the plurality of users are screened out, and the N users participate in a new round of federated learning;
  • N users with index values greater than the second preset index value among the plurality of users are screened out , Invite the N users to participate in a new round of federated learning.
  • the processing unit is also used for:
  • Receiving feedback data of the N users updating the probability of the N users participating in a new round of federated learning invitations, the experience loss of the N users, and the frequency of interaction between the federated learning server and the N users, It is used to calculate the suitability value for each of the N users to participate in the next round of federated learning.
  • the present invention provides a computer device, the computer device includes: at least one processor and a memory;
  • the memory stores one or more computer programs
  • the processor reads one or more computer programs stored in the memory, and executes the following method: obtain feedback data of a plurality of users after accepting a federal study invitation in the history of a plurality of users, and obtain respective user portrait data of the plurality of users; determine The number of times each of the multiple users participated in federated learning modeling in the last time window; according to the feedback data of the multiple users, the user profile data, and the multiple users each participated in the previous time window The number of times of federated learning modeling is to calculate the respective index values of the multiple users; the index value is used to characterize the suitability of each of the multiple users to participate in a new round of federated learning; according to the respective values of the multiple users Index value, inviting users who meet preset conditions among the multiple users to participate in federated learning.
  • the processor is specifically configured to:
  • the processor is specifically configured to:
  • the magnitude of the value corresponding to a time point is positively correlated with the time-sharing responsiveness of each of the multiple users at the time point; the time-sharing responsiveness is used to characterize the speed at which the multiple users each feedback accepting the federal study invitation ;
  • the processor is specifically configured to:
  • the largest value among the first probability, the experience loss, and the frequency is used as the respective index values of the multiple users.
  • the processor is specifically configured to:
  • the largest value among the first probability, the experience loss, and the frequency is used as the respective index values of the multiple users.
  • the processor is specifically configured to:
  • N users with index values smaller than the first preset index value among the plurality of users are screened out, and the N users participate in a new round of federated learning;
  • N users with index values greater than the second preset index value among the plurality of users are screened out , Invite the N users to participate in a new round of federated learning.
  • the processor is specifically configured to:
  • Receiving feedback data of the N users updating the probability of the N users participating in a new round of federated learning invitations, the experience loss of the N users, and the frequency of interaction between the federated learning server and the N users, It is used to calculate the suitability value for each of the N users to participate in the next round of federated learning.
  • the present invention provides a computer-readable storage medium, the computer-readable storage medium stores computer instructions, and when the computer instructions run on a computer device, the computer device can execute the first aspect described above. Or any one of the possible design methods of the first aspect mentioned above.
  • the present invention provides a computer program product.
  • the computer program product includes a computer program stored on a computer-readable storage medium.
  • the computer program includes program instructions.
  • the program instructions When the program instructions are executed by a computer device, the computer device executes the above The first aspect or any one of the possible design methods of the above-mentioned first aspect.
  • the indexing method of users in federated learning is that the index values of multiple users are feedback data after the federated learning device accepts the federated learning invitation according to the respective histories of the multiple users.
  • the user s respective user profile data and the number of times that multiple users participated in federated learning modeling in the last time window are calculated. Therefore, the index values of multiple users can be used for federated learning at different periods of time with multiple users.
  • the probability of invitation, the probability of multiple users participating in a new round of federated learning, and the satisfaction degree of multiple users participating in federated learning modeling in the last time window are related, which can improve the federated learning device from multiple users according to the index value.
  • the selected users who meet the preset conditions are suitable for federated learning, which reduces the possibility that the indexed users will refuse to participate in federated learning. There is no need for trial and error to contact users to participate in federated learning, which can effectively improve the participants of federated learning and federated learning.
  • the interactive efficiency of the device is not limited.
  • FIG. 1 is a schematic flowchart of a method for indexing users in federated learning according to an embodiment of the present invention
  • Figure 2 is a schematic structural diagram of a federated learning device provided by an embodiment of the present invention.
  • Fig. 3 is a schematic structural diagram of a computer device provided by an embodiment of the present invention.
  • first and second are used to distinguish different objects, rather than to describe a specific sequence.
  • the term “including” and any variations of them are intended to cover non-exclusive protection.
  • a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes unlisted steps or units, or optionally also includes Other steps or units inherent to these processes, methods, products or equipment.
  • "and/or” is merely an association relationship describing associated objects, which means that there can be three types of relationships.
  • a and/or B can mean that A exists alone, and A and B exist at the same time. , There are three cases of B alone.
  • the character "/" in the embodiment of the present invention generally indicates that the associated objects before and after are in an "or" relationship.
  • a plurality of may mean at least two, for example, it may be two, three, or more, which is not limited in the embodiment of the present invention.
  • the current federated learning device uses a random method to index users participating in federated learning, which lacks optimization considerations. It is easy to have the indexed users have low suitability for federated learning. This leads to the emergence of federated learning devices that require trial and error to contact users to participate in federation. The phenomenon of learning reduces the interaction efficiency between the participants of federated learning and the federated learning device.
  • the embodiment of the present invention provides a method for indexing users in federated learning, so as to improve the interaction efficiency between the participants of federated learning and the federated learning device.
  • the following specifically introduces the specific process of the federated learning device in the embodiment of the present invention for indexing users participating in federated learning.
  • FIG. 1 is a schematic flowchart of a method for indexing users in federated learning according to an embodiment of the present invention.
  • the method can be applied to a federated learning device.
  • the method flow includes:
  • the federated learning device can understand the probability distribution of multiple users accepting the federated learning invitation in different time periods by acquiring feedback data of the history of multiple users accepting the federated learning invitation.
  • the federated learning device invited user a to participate in the federated learning in the last round of federated learning.
  • user a can feedback data to the federated learning device after accepting the invitation of the last federated learning to inform the federated learning device that user a is at a certain time Point has accepted the invitation of the last round of federated learning.
  • the feedback data of user a can be expressed as: User a accepted the invitation of the last round of federated learning at 14:05.
  • the probability of the user accepting the federal study invitation between 8:00-9:00 is 10%
  • the probability of accepting the federal study invitation from 11:00-12:00 is 20%
  • the probability of accepting the federal study between 14:00-15:00 The probability of invitation is 55%
  • the probability of accepting the federal study invitation between 16:00-17:00 is 15%
  • the probability of accepting the federal study during the rest of the time period is 0.
  • the federated learning device can also obtain feedback data after multiple user history rejects the federated learning invitation, so as to understand the probability distribution of multiple users rejecting the federated learning invitation at different time periods.
  • the probability distribution method of each user accepting the federated learning invitation at different time periods will not be repeated here.
  • the federated learning device may also obtain user portrait data of multiple users.
  • the federated learning device may obtain user portrait data stored in the server, and/or obtain user portrait data stored in the terminal.
  • the user portrait data can include the respective basic attributes (such as age, gender, region, etc.), social attributes (such as occupation, income, etc.), behavior attributes (such as shopping preferences, viewing preferences, etc.), and psychological attributes of multiple users (such as focusing on cost performance, loving nature, etc.) data, and so on.
  • the terminal may be any device that can participate in federated learning, such as a tablet, a mobile phone, or a notebook computer, which is not specifically limited in the embodiment of the present invention.
  • the federated learning device can analyze and determine the future needs of multiple users by acquiring the user profile data of multiple users, so as to understand the respective future needs of multiple users, such as financial needs, product needs, entertainment Demand, etc.
  • the terminal can also be used for other purposes, such as playing videos and browsing web pages. Because when users use the terminal to participate in federated learning modeling, the terminal's operating speed and network loading speed will be temporarily affected by the transmission of the federated learning model parameters. Therefore, when the terminal is used to participate in federated learning modeling, it also runs other When using applications (such as video applications, etc.), the user's experience of using other applications of the terminal will be reduced to a certain extent. Therefore, in a certain time period in a certain time window, in order to improve the experience of using other applications of the terminal, the user may not participate in the federated learning modeling during this time period. Among them, the time window can be expressed as the length of time required for one round of federated learning or multiple rounds of federated learning.
  • the federated learning device determines the number of times that multiple users have participated in federated learning modeling in the last time window, so as to understand the degree of satisfaction of multiple users in participating in federated learning modeling in the last time window. For example, take user b among multiple users as an example. If the number of times of federated learning modeling in the last time window is 10, and user b uses the terminal to participate in federated learning modeling only once, then the satisfaction degree of user b participating in federated learning modeling in the last time window is 10% , That is, user b is less satisfied with participating in federated learning modeling in the last time window.
  • the execution sequence of the above S101 and S102 is not specifically limited in the embodiment of the present invention.
  • the federated learning device may first execute S101 and then execute S102, or execute S102 first, then execute S101, or execute both S101 and S101 at the same time. S102.
  • S103 Calculate the respective index values of the multiple users according to the respective feedback data of the multiple users, the user profile data, and the number of times each of the multiple users participated in federated learning modeling in the last time window;
  • the index value is used to characterize the suitability value of each of the multiple users participating in a new round of federated learning.
  • the federated learning device can determine the confidence of the user portraits of the multiple users, that is, determine the probability of each of the multiple users participating in a new round of federated learning.
  • the federated learning device can determine the future needs of multiple users through their respective user profile data. After that, the federated learning device can determine that multiple users will participate in a new round of federation based on their respective future needs. Probability of learning.
  • the federated learning determines the confidence of the respective user portraits of multiple users, it can be based on the respective feedback data of the multiple users, the confidence of the user portraits, and the number of times each of the multiple users participated in the federated learning modeling in the last time window Modeling is performed separately, and the respective index values of multiple users are calculated.
  • the federated learning device can understand the probability distribution of multiple users receiving the federated learning invitation at different time periods through the respective feedback data of multiple users, and the confidence level of the respective user portraits of multiple users can be used to determine The probability of multiple users participating in a new round of federated learning.
  • the index values of multiple users calculated by the federated learning device can be compared with the probability that multiple users will accept the invitation of federated learning at different times, the probability that multiple users will participate in a new round of federated learning, and the probability of multiple users.
  • the satisfaction degree of each participating in the federated learning modeling in the last time window is related, which helps to improve the suitability of the federated learning device to select users who meet the preset conditions from multiple users according to the index value and the federated learning.
  • the phenomenon that the federated learning device needs to test and contact the user to participate in the federated learning can effectively improve the interaction efficiency between the participants of the federated learning and the federated learning device.
  • the federated learning device performs modeling based on the feedback data of multiple users, which can be used to predict the first probability of multiple users participating in a new round of federated learning invitations in the current time period.
  • the value of the first probability at any time point in the time window of the new round of federated learning is positively correlated with the time-sharing responsiveness of multiple users at each time point, that is, the time-sharing responsiveness of multiple users at each time point is positively correlated.
  • the higher the time-sharing responsiveness the greater the first probability that multiple users will participate in a new round of federal learning invitations at the same time point.
  • the time-sharing responsiveness is used to characterize the speed at which multiple users receive the federal learning invitation.
  • the federated learning device can determine that user a participates in the first round of the new round of federated learning invitation at time point a.
  • a probability is greater than the first probability of participating in a new round of federal learning invitation at time point b, that is, the first probability of determining that user a participates in a new round of federal learning invitation in time period a is greater than the first probability of participating in a new round of federal learning invitation in time period b The first probability.
  • the federated learning device predicts the first probability that multiple users will participate in a new round of federated learning invitations in the current period by modeling based on the respective feedback data of multiple users, and can understand that multiple users are in different periods of time.
  • the probability of participating in a new round of federated learning so as to avoid the phenomenon that the federated learning device invites users to participate in federated learning during a time period when the probability of users participating in federated learning is low, and reduces the possibility of invited users to refuse to participate in federated learning. Therefore, the suitability of the indexed users to participate in the federated learning can be improved, and the interaction efficiency between the participants of the federated learning and the federated learning device can be effectively improved.
  • the federated learning device performs modeling according to the number of times that multiple users participated in federated learning modeling in the last time window, which can be used to predict the experience loss of multiple users; where the experience loss is used for Indicates the degree of satisfaction of multiple users in accepting the invitation to participate in a new round of federal learning in the current period. For example, taking user a among multiple users as an example, if the number of times of federated learning modeling in the last time window is 20, the number of times that user a participated in federated learning modeling in the last time window is 2, and they are in time respectively.
  • the federated learning device can determine that the total experience loss of user a participating in federated learning modeling in the last time window is ((20-2)/20)%, that is, 90%, that is, the degree of satisfaction is 10%, the experience loss of participating in federated learning modeling in time period e and time period f in the last time window is ((20-1)/20)%, which is 95%, which is the satisfaction degree 5%. Then, when the federated learning device can perform modeling according to the number of times that user a participated in federated learning modeling in the last time window, it can predict the time period e, time period f and other time periods of user a in multiple time windows in the future. Loss of experience over time.
  • the federated learning device performs modeling based on the number of times that multiple users participate in federated learning modeling in the last time window, predicts the experience loss of multiple users, and can understand that multiple users participate in different periods of time.
  • the degree of satisfaction of federated learning modeling can avoid the phenomenon that the federated learning device invites users to participate in federated learning modeling during the time period when the user's satisfaction with participating in federated learning modeling is low, and reduces the number of invited users who refuse to participate in federated learning.
  • the possibility of learning can further improve the suitability of the indexed users with the federated learning, and can effectively improve the interaction efficiency between the participants of federated learning and the federated learning device.
  • the federated learning device performs modeling based on the confidence of the user portrait, which can be used to predict how often the federated learning server needs to interact with multiple users in a new round of federated learning. For example, take user b among multiple users as an example. If the amount of user portrait data of user b is small or the user portrait data is divergent (for example, the amount of behavior attribute data is much larger than the amount of psychological attribute data), there may be low confidence in the user portrait of user b, resulting in a federated learning device In the future, it is predicted that the accuracy of user b using the terminal to participate in a new round of federated learning is not high.
  • the federated learning device when the federated learning device performs modeling according to the confidence level of the user portrait of user b, and determines that the probability of user b participating in a new round of federated learning is low, the federated learning device can determine that the federated learning server needs to interact with user b more frequently , In order to improve the accuracy of the subsequent prediction of the probability of user b participating in a new round of federated learning by the federated learning device.
  • the federated learning device predicts the frequency of interaction between the federated learning server and the multiple users in the new round of federated learning by modeling according to the confidence of the respective user portraits of multiple users, which can improve users
  • the confidence of the user portrait of the user with a small amount of portrait data or with a large divergence of user portrait data can increase the probability of the federated learning device predicting that users with a small amount of user portrait data or with a large divergence of user portrait data will participate in a new round of federated learning in the future.
  • Accuracy in turn, can avoid the unevenness of the chances of multiple users participating in federated learning due to factors of user portrait data.
  • the federated learning device obtains the first probability of multiple users participating in a new round of federated learning invitations in the current time period, the respective experience loss of multiple users and the federated learning server in the new round of federated learning obtained by the federated learning device according to the above predictions
  • the respective index values of multiple users can be improved, and the federated learning device can increase the relevance of the index value according to the index value.
  • Each user selects the suitability of the users participating in the federated learning and the federated learning, so as to avoid the phenomenon that the federated learning device needs to test and contact the user to participate in the federated learning, and reduce the possibility of the invited users to refuse to participate in the federated learning.
  • the interaction efficiency between the participants of the federated learning and the federated learning device is improved, and in addition, the communication burden of the federated learning device can be reduced.
  • the federated learning device is based on the first probability of multiple users participating in a new round of federated learning invitations in the current period, the respective experience losses of multiple users, and the federated in the new round of federated learning.
  • the learning server needs to interact with multiple users individually, and there can be multiple ways to calculate the respective index values of multiple users. for example:
  • the federated learning device can calculate the first probability of multiple users participating in a new round of federated learning invitations in the current period, the respective experience loss of multiple users, and the federated learning server needs to interact with multiple users in the new round of federated learning.
  • the corresponding average value of the three interaction frequencies, and then the calculated average value is used as the respective index value of multiple users.
  • the first probability of multiple users participating in a new round of federated learning invitations in the current period, the respective experience loss of multiple users, and the federated learning server in the new round of federated learning The average value of the interaction frequency between the three, as the respective index value of multiple users, can balance the respective index values of multiple users, and the first probability of multiple users participating in a new round of federal learning invitation in the current period
  • Probability, experience loss, and the frequency with which the federated learning server needs to interact with multiple users in the new round of federated learning is the lowest of the three as the respective index value of multiple users, resulting in multiple users’ respective index values and multiple
  • the phenomenon that the user’s relevance is low, which can improve the suitability of the federated learning device for selecting users to participate in federated learning from multiple users
  • the federated learning device can, according to a preset strategy, assign the first probability of multiple users participating in a new round of federated learning invitations in the current period, the respective experience losses of multiple users, and the federated learning server in the new round of federated learning.
  • the maximum value of the frequency of interaction with multiple users is used as the index value of each of the multiple users.
  • the federated learning device may compare the calculated average value with a preset threshold value to determine whether the average value is greater than or equal to the preset threshold value; if it is determined that the average value is greater than or equal to the preset threshold value, the first probability and the experience The maximum value of the loss and the frequency is used as the index value of each of the multiple users. Otherwise, the average value is used as the respective index value of multiple users.
  • the first probability of multiple users participating in a new round of federated learning invitations in the current period, the respective experience loss of multiple users, and the number of federated learning servers in the new round of federated learning are calculated through calculations.
  • the average of the three interaction frequencies of each user is greater than or equal to the preset threshold, the maximum value among the three is used as the index value of multiple users, or the average value between the three is less than
  • the threshold is preset, the average value is used as the index value of multiple users, which can improve the relevance of the index value of multiple users to multiple users, thereby improving the federated learning device to select from multiple users based on the index value.
  • the suitability of users participating in federated learning and federated learning avoids the phenomenon that federated learning devices need to trial and contact users to participate in federated learning. There is no need for trial and error to contact users to participate in federated learning, which can effectively improve federated learning participants and federated learning devices. Interactive efficiency.
  • the federated learning device may invite users from the multiple users who meet the preset conditions to participate in federated learning according to different preset conditions. For example:
  • Example 1 If the preset condition is to preferentially mobilize users whose frequency of federal learning invitations is lower than a preset threshold to participate in a new round of federated learning, then the federated learning device can filter out the plurality of users whose index value is less than the first N users with a preset index value are invited to participate in a new round of federated learning.
  • the federated learning device may sort the respective index values of multiple users in ascending order, and invite the first N users to participate in a new round of federated learning based on the sorting, or may combine the respective index values of multiple users Sorting is performed in descending order, and N users are invited to participate in a new round of federated learning based on the sorting, wherein each index value of the N users is less than the first preset index value.
  • federated learning can improve the accuracy that the multiple users selected are users with low frequency of participating in federated learning, thereby improving the federated learning device to select users participating in federated learning from multiple users based on the index value and federated learning
  • the suitability of the federated learning device avoids the phenomenon that the federated learning device needs to trial and contact the user to participate in the federated learning. There is no need for trial and error to contact the user to participate in the federated learning, which can effectively improve the interaction efficiency between the federated learning participant and the federated learning device.
  • Example 2 If the preset condition is to preferentially mobilize users whose frequency of federal learning invitations is higher than or equal to the preset threshold to participate in a new round of federated learning, then the federated learning device can filter out the indexes among the multiple users N users with a value greater than the second preset index value are invited to participate in a new round of federated learning.
  • the federated learning device can sort the respective index values of multiple users in ascending order, and invite N users to participate in a new round of federated learning based on the sorting, or it can sort the respective index values of multiple users Sorting is performed in descending order, and based on the sorting, the first N users are invited to participate in a new round of federated learning, where the respective index values of the N users are all greater than the second preset index value.
  • federated learning can improve the accuracy that the multiple users selected are users with high frequency of participating in federated learning, thereby improving the federated learning device to select users participating in federated learning from multiple users based on the index value and federated learning
  • the suitability of the federated learning device eliminates the need for trial and error to contact users to participate in federated learning, avoiding the phenomenon that the federated learning device needs to trial and contact users to participate in federated learning, and can effectively improve the interaction efficiency between federated learning participants and the federated learning device.
  • the federated learning device may, after inviting N users among multiple users to participate in a new round of federated learning, may receive feedback data of N users, and update the N users to participate in the new round according to the feedback data of N users.
  • the probability of a round of federated learning invitation, the experience loss of N users, and the frequency of interaction between the federated learning server and N users are used to calculate the suitability of each of the N users to participate in the next round of federated learning, so as to increase the participation of N users.
  • the suitability of the next round of federated learning can effectively improve the interaction efficiency between the participants of the N users in the next round of federated learning and the federated learning device.
  • first preset index value and the second preset index value may be the same or different, which is not specifically limited in the embodiment of the present invention.
  • N can be set by the system administrator of the federated learning device, or can be determined by the upper limit of the preset index value, or the lower limit of the preset index value, which is not specifically limited in the embodiment of the present invention.
  • the index values of multiple users are feedback data after the federated learning device accepts the federated learning invitation according to the respective history of the multiple users.
  • the user profile data of each user and the number of times that multiple users participated in federated learning modeling in the last time window are calculated. Therefore, the index values of multiple users can be compared with multiple users in different periods of time.
  • the probability of learning invitation, the probability of multiple users participating in a new round of federated learning, and the satisfaction degree of multiple users participating in federated learning modeling in the previous time window are related, which can improve the federated learning device from multiple users according to the index value.
  • the suitability of the selected users who meet the preset conditions with the federated learning reduces the possibility that the indexed users will refuse to participate in the federated learning. There is no need for trial and error to contact users to participate in the federated learning, which can effectively improve the participants and federated federated learning.
  • the interactive efficiency of the learning device reduces the possibility that the indexed users will refuse to participate in the federated learning.
  • FIG. 2 is a schematic structural diagram of a federated learning device according to an embodiment of the present invention.
  • the federated learning device 200 includes:
  • the acquiring unit 201 is configured to acquire historical feedback data of multiple users after accepting a federal study invitation, and acquire respective user portrait data of the multiple users;
  • the processing unit 202 is configured to determine the number of times each of the multiple users participated in the federated learning modeling in the last time window; according to the respective feedback data of the multiple users, the user portrait data, and the respective multiple users The number of times of participating in federated learning modeling in the last time window, calculating the respective index values of the multiple users; the index value is used to represent the suitability value of each of the multiple users participating in a new round of federated learning;
  • the inviting unit 203 is configured to invite users who meet preset conditions among the multiple users to participate in federated learning according to respective index values of the multiple users.
  • processing unit 202 is specifically configured to:
  • processing unit 202 is specifically configured to:
  • the magnitude of the value corresponding to a time point is positively correlated with the time-sharing responsiveness of each of the multiple users at the time point; the time-sharing responsiveness is used to characterize the speed at which the multiple users each feedback accepting the federal study invitation ;
  • processing unit 202 is specifically configured to:
  • the largest value among the first probability, the experience loss, and the frequency is used as the respective index values of the multiple users.
  • processing unit 202 is specifically configured to:
  • the largest value among the first probability, the experience loss, and the frequency is used as the respective index values of the multiple users.
  • the inviting unit 203 is specifically configured to:
  • N users with index values smaller than the first preset index value among the plurality of users are screened out, and the N users participate in a new round of federated learning;
  • N users with index values greater than the second preset index value among the plurality of users are screened out , Invite the N users to participate in a new round of federated learning.
  • processing unit 202 is further configured to:
  • Receiving feedback data of the N users updating the probability of the N users participating in a new round of federated learning invitations, the experience loss of the N users, and the frequency of interaction between the federated learning server and the N users, It is used to calculate the suitability value for each of the N users to participate in the next round of federated learning.
  • the federated learning device 200 in the embodiment of the present invention and the indexing method for users in federated learning shown in FIG. 1 are inventions based on the same concept.
  • the implementation process of the federated learning device 200 in this embodiment can be clearly understood, so for the sake of brevity of the description, it will not be repeated here.
  • FIG. 3 is a schematic structural diagram of a computer device according to an embodiment of the present invention.
  • the computer device 300 includes: a memory 301 and at least one processor 302. Wherein, the memory 301 stores one or more computer programs; when the one or more computer programs stored in the memory 301 are executed by the at least one processor 302, the computer device 300 is caused to execute the user in the federated learning. The steps of the indexing method.
  • the memory 301 may include a high-speed random access memory, and may also include a non-volatile memory, such as a magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices, etc., which is not limited in the embodiment of the present invention.
  • a non-volatile memory such as a magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices, etc., which is not limited in the embodiment of the present invention.
  • the processor 302 may be a general-purpose processor (central processing unit, CPU), or ASIC, or FPGA, or may be one or more integrated circuits for controlling program execution.
  • CPU central processing unit
  • ASIC application specific integrated circuit
  • FPGA field-programmable gate array
  • the memory 301 and the processor 302 may be implemented on the same chip. In other embodiments, they may also be implemented on separate chips, which is not limited in the embodiment of the present invention.
  • the present invention also provides a computer-readable storage medium that stores computer instructions.
  • the computer instructions When the computer instructions are executed by a computer device, the computer device can execute the foregoing The steps of the user index method in federated learning.
  • the present invention also provides a computer program product.
  • the computer program product includes a computer program stored on a computer-readable storage medium.
  • the computer program includes program instructions.
  • the program instructions When the program instructions are executed by a computer device, the computer program The computer device executes the steps of the user indexing method in the federated learning described above.
  • These computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operation steps are executed on the computer or other programmable equipment to produce computer-implemented processing, so as to execute on the computer or other programmable equipment.
  • the instructions provide steps for implementing the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.
  • These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device.
  • the device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种联邦学习中参与者权重的设置方法及装置，该方法适用于包括一个协调者和多个参与者的联邦学习；其中方法为：所述协调者将根据多个参与者上报的本地模型参数得到的联邦模型参数发送至所述多个参与者；所述协调者获取所述多个参与者反馈的模型性能测试结果；所述协调者至少根据所述多个模型性能测试结果，设置所述多个参与者的权重值，参与者的权重值用于表征参与者在后续联邦学习中的贡献度。上述方法应用于金融科技(Fintech)时，有效激励了参与者汇报真实模型性能测试结果，同时削弱了虚报模型性能测试结果对整个联邦学习模型的影响。

Description

一种联邦学习中用户的索引方法及联邦学习装置
相关申请的交叉引用
本申请要求在2020年03月31日提交中国专利局、申请号为202010244824.9、申请名称为“一种联邦学习中用户的索引方法及联邦学习装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及金融科技(Fintech)技术领域和人工智能技术领域，尤指一种联邦学习中用户的索引方法及联邦学习装置。
背景技术
作为一种新型机器学习理念，联邦学习通过分布式训练及加密技术确保用户隐私数据得到最大限度的保护，以提升用户对人工智能技术的信任。在联邦学习机制下，各参与方把加密后的数据模型贡献给联盟，联合训练一个联邦模型，再开放这个模型给各参与方使用。在这个联邦学习的训练过程中，如何提升联邦学习的参与者与联邦学习装置的交互效率，对于提升联邦学习的模型训练效率有重要意义。
目前的联邦学习中，联邦学习装置对于联邦学习的参与者主要是通过随机方法进行选择的。比如，联邦学习装置采用随机算法从满足约束条件下的用户(使用的终端设备必须满足在充电状态且使用非流量计价网络链接如WiFi等的条件的用户)中随机索引相应数目的用户来参与联邦学习。例如，若联邦学习需要选择8个参与者，那么联邦学习装置可以采用随机算法从满足约束条件下的用户中随机索引出8个用户作为联邦学习的参与者。
然而，联邦学习装置采用随机方法来索引参与联邦学习的用户，缺乏优化考虑，换言之，联邦学习装置随机索引出的用户不是联邦学习装置优化筛选出来的，因此，容易存在索引出的用户拒绝参与联邦学习的可能性比较大(即索引出的用户与联邦学习的适合度低)，而导致联邦学习装置需要反复试验联系用户参与联邦学习的现象，从而降低了联邦学习的参与者与联邦学习装置的交互效率。
发明内容
本发明提供一种联邦学习中用户的索引方法及联邦学习装置，用以解决现有技术中存在联邦学习的参与者与联邦学习装置的交互效率低的问题。
为实现上述目的，第一方面，本发明提供一种联邦学习中用户的索引方法，该方法包括：
获取多个用户历史接受联邦学习邀请后的反馈数据，以及获取所述多个用户各自的用户画像数据；
确定所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数；
根据所述多个用户各自的反馈数据、所述用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值；所述索引值用于表征所述多个用户各自参与新一轮联邦学习的适合度值；
根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习。
在一种可能的设计中，根据所述多个用户各自的反馈数据、用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值，包括：
根据所述多个用户各自的用户画像数据，确定所述多个用户各自的用户画像置信度；
根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值。
在一种可能的设计中，根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值，包括：
根据所述反馈数据进行建模，预测所述多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率；其中，所述第一概率在新一轮联邦学习的时间窗口中的任意一个时间点对应的数值大小，与所述多个用户各自在所述时间点的分时响应度正相关；所述分时响应度用于表征所述多个用户各自反馈接受联邦学习邀请的速度；
根据所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，预测所述多个用户各自的体验损失；其中，所述体验损失用于表示所述多个用户各自当前时段接受参与新一轮联邦学习邀请行为的满意程度；
根据所述用户画像置信度进行建模，预测在新一轮的联邦学习中联邦学习服务器需与所述多个用户各自互动的频率；
根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值。
在一种可能的设计中，根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值，包括：
计算所述第一概率、所述体验损失和所述频率三者之间对应的平均值；将所述平均值作为所述多个用户各自的索引值；或者，
根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
在一种可能的设计中，根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值，包括：
将所述平均值与预设阈值进行比较，确定所述平均值是否大于或者等于所述预设阈值；
若确定所述平均值大于或者等于所述预设阈值，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
在一种可能的设计中，根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习，包括：
若确定优先动员被联邦学习邀请频率低于预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值小于第一预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习；
若确定优先动员被联邦学习邀请频率高于或者等于所述预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值大于第二预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习。
在一种可能的设计中，邀请所述N个用户参与新一轮的联邦学习之后，所述方法还包括：
接收所述N个用户的反馈数据，更新所述N个用户参与新一轮联邦学习邀请的概率、所述N个用户的体验损失和所述联邦学习服务器与所述N个用户互动的频率，用于计算所述N个用户各自参与下一轮联邦学习的适合度值。
第二方面，本发明提供一种联邦学习装置，所述联邦学习装置包括：
获取单元，用于获取多个用户历史接受联邦学习邀请后的反馈数据，以及获取所述多个用户各自的用户画像数据；
处理单元，用于确定所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数；根据所述多个用户各自的反馈数据、所述用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值；所述索引值用于表征所述多个用户各自参与新一轮联邦学习的适合度值；
邀请单元，用于根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习。
在一种可能的设计中，所述处理单元具体用于：
根据所述多个用户各自的用户画像数据，确定所述多个用户各自的用户画像置信度；
根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值。
在一种可能的设计中，所述处理单元具体用于：
根据所述反馈数据进行建模，预测所述多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率；其中，所述第一概率在新一轮联邦学习的时间窗口中的任意一个时间点对应的数值大小，与所述多个用户各自在所述时间点的分时响应度正相关；所述分时响应度用于表征所述多个用户各自反馈接受联邦学习邀请的速度；
根据所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，预测所述多个用户各自的体验损失；其中，所述体验损失用于表示所述多个用户各自当前时段接受参与新一轮联邦学习邀请行为的满意程度；
根据所述用户画像置信度进行建模，预测在新一轮的联邦学习中联邦学习服务器需与所述多个用户各自互动的频率；
根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值。
在一种可能的设计中，所述处理单元具体用于：
计算所述第一概率、所述体验损失和所述频率三者之间对应的平均值；将所述平均值作为所述多个用户各自的索引值；或者，
根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
在一种可能的设计中，所述处理单元具体用于：
将所述平均值与预设阈值进行比较，确定所述平均值是否大于或者等于所述预设阈值；
若确定所述平均值大于或者等于所述预设阈值，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
在一种可能的设计中，所述邀请单元具体用于：
若确定优先动员被联邦学习邀请频率低于预设阈值的用户参与新一轮的联邦学习，则 筛选出所述多个用户中索引值小于第一预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习；
若确定优先动员被联邦学习邀请频率高于或者等于所述预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值大于第二预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习。
在一种可能的设计中，所述处理单元还用于：
接收所述N个用户的反馈数据，更新所述N个用户参与新一轮联邦学习邀请的概率、所述N个用户的体验损失和所述联邦学习服务器与所述N个用户互动的频率，用于计算所述N个用户各自参与下一轮联邦学习的适合度值。
第三方面，本发明提供一种计算机设备，所述计算机设备包括：至少一个处理器和存储器；
所述存储器存储一个或多个计算机程序；
所述处理器读取所述存储器存储的一个或多个计算机程序，执行以下方法：获取多个用户历史接受联邦学习邀请后的反馈数据，以及获取所述多个用户各自的用户画像数据；确定所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数；根据所述多个用户各自的反馈数据、所述用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值；所述索引值用于表征所述多个用户各自参与新一轮联邦学习的适合度值；根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习。
可选地，所述处理器具体用于：
根据所述多个用户各自的用户画像数据，确定所述多个用户各自的用户画像置信度；
根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值。
可选地，所述处理器具体用于：
根据所述反馈数据进行建模，预测所述多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率；其中，所述第一概率在新一轮联邦学习的时间窗口中的任意一个时间点对应的数值大小，与所述多个用户各自在所述时间点的分时响应度正相关；所述分时响应度用于表征所述多个用户各自反馈接受联邦学习邀请的速度；
根据所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，预测所述多个用户各自的体验损失；其中，所述体验损失用于表示所述多个用户各自当前时段接受参与新一轮联邦学习邀请行为的满意程度；
根据所述用户画像置信度进行建模，预测在新一轮的联邦学习中联邦学习服务器需与所述多个用户各自互动的频率；
根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值。
可选地，所述处理器具体用于：
计算所述第一概率、所述体验损失和所述频率三者之间对应的平均值；将所述平均值作为所述多个用户各自的索引值；或者，
根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
可选地，所述处理器具体用于：
将所述平均值与预设阈值进行比较，确定所述平均值是否大于或者等于所述预设阈值；
若确定所述平均值大于或者等于所述预设阈值，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
可选地，所述处理器具体用于：
若确定优先动员被联邦学习邀请频率低于预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值小于第一预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习；
若确定优先动员被联邦学习邀请频率高于或者等于所述预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值大于第二预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习。
可选地，所述处理器具体用于：
接收所述N个用户的反馈数据，更新所述N个用户参与新一轮联邦学习邀请的概率、所述N个用户的体验损失和所述联邦学习服务器与所述N个用户互动的频率，用于计算所述N个用户各自参与下一轮联邦学习的适合度值。
第四方面，本发明提供一种计算机可读存储介质，所述计算机可读存储介质存储有计算机指令，当所述计算机指令在计算机设备上运行时，使得所述计算机设备能够执行上述第一方面或上述第一方面的任意一种可能的设计的方法。
第五方面，本发明提供一种计算机程序产品，该计算机程序产品包括存储在计算机可读存储介质上的计算机程序，计算机程序包括程序指令，当程序指令被计算机设备执行时，使计算机设备执行上述第一方面或上述第一方面的任意一种可能的设计的方法。
本发明有益效果如下：
本发明提供的一种联邦学习中用户的索引方法，与现有技术相比，由于多个用户各自的索引值是联邦学习装置根据多个用户各自历史接受联邦学习邀请后的反馈数据、多个用户各自的用户画像数据和多个用户各自在上一个时间窗口内参与联邦学习建模的次数计算得到的，因此，多个用户各自的索引值，可以与多个用户各自在不同时段接受联邦学习邀请的概率、多个用户各自参与新一轮联邦学习的概率、多个用户各自在上一个时间窗口参与联邦学习建模的满意程度相关，从而可以提高联邦学习装置根据索引值从多个用户中选出的满足预设条件的用户与联邦学习的适合度，降低了索引出的用户拒绝参与联邦学习的可能性，无需反复试验联系用户参与联邦学习，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
附图说明
图1为本发明实施例提供的一种联邦学习中用户的索引方法的流程示意图；
图2为本发明实施例提供的一种联邦学习装置的结构示意图；
图3为本发明实施例提供的一种计算机设备的结构示意图。
具体实施方式
为了使本发明的目的、技术方案和优点更加清楚，下面将结合附图对本发明作进一步地详细描述，显然，所描述的实施例仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其 它实施例，都属于本发明保护的范围。
附图中各部件的形状和大小不反映真实比例，目的只是示意说明本发明内容。
本发明实施例中，“第一”、“第二”是用于区别不同对象，而非用于描述特定顺序。此外，术语“包括”以及它们任何变形，意图在于覆盖不排他的保护。例如包含了一系列步骤或单元的过程、方法、系统、产品或设备没有限定于已列出的步骤或单元，而是可选地还包括没有列出的步骤或单元，或可选地还包括对于这些过程、方法、产品或设备固有的其它步骤或单元。
本发明实施例中，“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本发明实施例中的字符“/”，一般表示前后关联对象是一种“或”的关系。
本发明实施例中，“多个”可以表示至少两个，例如可以是两个、三个或者更多，本发明实施例不限定。
由前述内容可知，目前联邦学习装置采用随机方法来索引参与联邦学习的用户，缺乏优化考虑，容易存在索引出的用户与联邦学习的适合度低，导致出现联邦学习装置需要反复试验联系用户参与联邦学习的现象，降低了联邦学习的参与者与联邦学习装置的交互效率。为了解决该问题，本发明实施例提供了一种联邦学习中用户的索引方法，以提高联邦学习的参与者与联邦学习装置的交互效率。
下面具体介绍本发明实施例中联邦学习装置索引参与联邦学习的用户的具体过程。
示例性的，请参考图1所示，为本发明实施例提供的一种联邦学习中用户的索引方法的流程示意图。其中，该方法可以应用于联邦学习装置。如图1所示，该方法流程包括：
S101、获取多个用户历史接受联邦学习邀请后的反馈数据，以及获取所述多个用户各自的用户画像数据。
通常，由于工作时间或者作息时间等因素的不同，不同的用户使用终端的时间段一般也不同，因此，用户满足参与联邦学习的基本条件(如使用的终端必须满足在充电状态且使用非流量计价网络链接如WiFi等条件)在不同时段上的概率分布也有所区别，换言之，多个用户在不同时间段接受联邦学习邀请的概率分布也有所区别。
在本发明实施例中，联邦学习装置通过获取多个用户历史接受联邦学习邀请的反馈数据，可以了解多个用户在不同时间段接受联邦学习邀请的概率分布。
比如，以多个用户中的用户a为例。联邦学习装置在上一轮联邦学习中邀请了用户a参与联邦学习，之后，用户a接受上一轮联邦学习邀请后可以向联邦学习装置反馈数据，用于告知联邦学习装置用户a在某个时间点接受了上一轮联邦学习的邀请，例如，用户a的反馈数据可以表示为：用户a在14:05接受了上一轮联邦学习的邀请。示例性的，若历史联邦学习的轮回次数为20，用户a在8：00-9：00之间参与第11轮、第16轮联邦学习，在11：00-12：00之间参与第12轮至第15轮联邦学习，在14：00-15：00之间参与第1轮至第10轮联邦学习和第20轮(即上一轮)联邦学习，在16：00-17：00之间参与第17轮至第19轮联邦学习。那么，用户在8：00-9：00接受联邦学习邀请的概率为10％，在11：00-12：00接受联邦学习邀请的概率为20％，在14：00-15：00接受联邦学习邀请的概率为55％，在16：00-17：00接受联邦学习邀请的概率为15％，在其余时间段接受联邦学习的概率为0。
当然，联邦学习装置还可以通过获取多个用户历史拒绝联邦学习邀请后的反馈数据， 以便了解多个用户在不同时段拒绝联邦学习邀请的概率分布，其具体实现方式可以参见上述联邦学习装置确定多个用户在不同时段接受联邦学习邀请的概率分布的方式，在此不再赘述。
可选地，联邦学习装置还可以获取多个用户各自的用户画像数据。比如，联邦学习装置可以获取服务器存储的用户的画像数据，和/或，获取终端中存储的用户画像数据。其中，用户画像数据可以包括多个用户各自的基本属性(如年龄、性别、地域等)、社会属性(如职业、收入等)、行为属性(如购物偏好、观影偏好等)、心理属性(如注重性价比、喜爱自然等)数据，等等。终端可以为平板、手机、笔记本电脑等任何可以参与联邦学习的设备，本发明实施例不做具体限定。
在本发明实施例中，联邦学习装置通过获取多个用户各自的用户画像数据，可以分析确定多个用户各自未来的需求，以便了解多个用户各自未来的需求，如理财需求、产品需求、娱乐需求等。
S102、确定所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数。
通常，终端除了可以用于参与联邦学习之外，还可以有其他用途，如播放视频、浏览网页等。由于用户使用终端参与联邦学习建模时，终端的运行速度和网络加载速度等会暂时受到联邦学习模型参数的传输的影响而降低，因此，终端被使用参与联邦学习建模的同时，还运行其他应用(如视频应用等)时，会在一定程度上降低用户使用该终端的其他应用的体验。因此，在某个时间窗口中的某个时间段，用户为了提高使用该终端的其他应用的体验，可能会在该时间段中不参与联邦学习建模。其中，时间窗口可以表示为一轮联邦学习或者多轮联邦学习所需的时长。
在本发明实施例中，联邦学习装置通过确定多个用户各自在上一个时间窗口内参与联邦学习建模的次数，可以了解多个用户各自在上一个时间窗口参与联邦学习建模的满意程度。比如，以多个用户中的用户b为例。若上一个时间窗口联邦学习建模的次数为10次，而用户b使用终端参与联邦学习建模的次数只有一次，那么，用户b在上一个时间窗口参与联邦学习建模的满意程度为10％，即用户b在上一个时间窗口参与联邦学习建模的满意程度较低。
需要说明的是，上述S101和S102的执行顺序，本发明实施例不作具体限定，比如，联邦学习装置可以先执行S101，后执行S102，或者，先执行S102，后执行S101，或者同时执行S101和S102。
S103、根据所述多个用户各自的反馈数据、所述用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值；所述索引值用于表征所述多个用户各自参与新一轮联邦学习的适合度值。
可选地，联邦学习装置获取多个用户各自的用户画像数据后，可以确定多个用户各自的用户画像置信度，即确定多个用户各自参与新一轮联邦学习的概率。换言之，联邦学习装置可以通过多个用户各自的用户画像数据，确定多个用户各自未来的需求，之后，联邦学习装置可以根据多个用户各自未来的需求，确定多个用户各自参与新一轮联邦学习的概率。
可选地，联邦学习确定多个用户各自的用户画像置信度后，可以根据多个用户各自的反馈数据、用户画像置信度和多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到多个用户各自的索引值。在本发明实施例中，由于联邦学习装置 通过多个用户各自的反馈数据，可以了解多个用户各自在不同时段接受联邦学习邀请的概率分布，通过多个用户各自的用户画像置信度，可以确定多个用户各自参与新一轮联邦学习的概率，通过多个用户各自在上一个时间窗口内参与联邦学习建模的次数，可以了解多个用户各自在上一个时间窗口参与联邦学习建模的满意程度，因此，联邦学习装置计算得到的多个用户各自的索引值，可以与多个用户各自在不同时段接受联邦学习邀请的概率、多个用户各自参与新一轮联邦学习的概率、多个用户各自在上一个时间窗口参与联邦学习建模的满意程度相关，有助于提高联邦学习装置后续根据索引值从多个用户中选出满足预设条件的用户与联邦学习的适合度，从而可以避免出现联邦学习装置需要反复试验联系用户参与联邦学习的现象，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
在具体的实现过程中，联邦学习装置根据多个用户各自的反馈数据进行建模，可以用于预测多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率。其中，第一概率在新一轮联邦学习的时间窗口中的任意一个时间点对应的数值大小，与多个用户各自在时间点的分时响应度正相关，即多个用户各自在时间点的分时响应度越高，多个用户各自在相同时间点参与新一轮联邦学习邀请的第一概率越大，其中，分时响应度用于表征多个用户各自反馈接受联邦学习邀请的速度。比如，以多个用户中的用户a为例，若用户a在新一轮联邦学习的时间窗口中的时间点a(位于时间段a内)对应的分时响应度为4，用户a在新一轮联邦学习的时间窗口中的时间点b(位于时间段b内)对应的分时响应度为2，那么，联邦学习装置可以确定用户a在时间点a参与新一轮联邦学习邀请的第一概率大于在时间点b参与新一轮联邦学习邀请的第一概率，即确定用户a在时间段a参与新一轮联邦学习邀请的第一概率大于在时间段b参与新一轮联邦学习邀请的第一概率。
在本发明实施中，联邦学习装置通过根据多个用户各自的反馈数据进行建模，预测多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率，可以了解多个用户在不同时段参与新一轮联邦学习的概率，从而可以避免出现联邦学习装置在用户参与联邦学习的概率较低的时间段内邀请用户参与联邦学习的现象，降低了被邀请的用户拒绝参与联邦学习的可能性，进而可以提高索引出的用户参与联邦学习的适合度，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
在具体的实现过程中，联邦学习装置根据多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，可以用于预测多个用户各自的体验损失；其中，体验损失用于表示多个用户各自当前时段接受参与新一轮联邦学习邀请行为的满意程度。比如，以多个用户中的用户a为例，若上一个时间窗口内联邦学习建模的次数为20，用户a在上一个时间窗口内参与联邦学习建模的次数为2，且分别处于时间段e和时间段f内，联邦学习装置可以确定用户a在上一个时间窗口参与联邦学习建模总的体验损失为((20-2)/20)％，即90％，也即满意程度为10％，在上一个时间窗口中的时间段e和时间段f参与联邦学习建模的体验损失均为体验损失为((20-1)/20)％，即95％，也即满意程度为5％。那么，当联邦学习装置可以根据用户a在上一个时间窗口内参与联邦学习建模的次数进行建模时，可以预测用户a在未来的多个时间窗口中的时间段e、时间段f以及其它时间段的体验损失。
在本发明实施中，联邦学习装置通过根据多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，预测多个用户各自的体验损失，可以了解多个用户在不同时段参与联邦学习建模的满意程度，从而可以避免出现联邦学习装置在用户参与联邦学习建模的满意程度较低的时间段内邀请用户参与联邦学习建模的现象，降低了被邀请的用户拒绝 参与联邦学习的可能性，进而可以提高索引出的用户与联邦学习的适合度，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
在具体的实现过程中，联邦学习装置根据用户画像置信度进行建模，可以用于预测在新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率。比如，以多个用户中的用户b为例。若用户b的用户画像数据量较少或者用户画像数据分歧较大(例如行为属性数据量远大于心理属性数据量)，则可能会存在用户b的用户画像置信度不高，而导致联邦学习装置未来预测用户b使用终端参与新一轮联邦学习的准确性不高的现象。因此，当联邦学习装置根据用户b的用户画像置信度进行建模，确定用户b参与新一轮联邦学习的概率较低时，联邦学习装置可以确定联邦学习服务器需与用户b互动的频率较高，以提高联邦学习装置后续预测用户b参与新一轮联邦学习的概率的准确性。
在本发明实施例中，联邦学习装置通过根据多个用户各自的用户画像置信度进行建模，预测在新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率，可以提高用户画像数据量少或者用户画像数据分歧较大的用户的用户画像置信度，从而可以提高联邦学习装置未来预测用户画像数据量少或者用户画像数据分歧较大的用户参与新一轮联邦学习的概率的准确性，进而可以避免出现由于用户画像数据的因素而导致对多个用户各自参与联邦学习机会的不均的现象。
因此，当联邦学习装置通过根据上述预测得到的多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率、多个用户各自的体验损失和新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率，来计算多个用户各自的索引值时，可以提高多个用户各自的索引值与多个用户之间的关联性，从而可以提高联邦学习装置根据索引值从多个用户选出参与联邦学习的用户与联邦学习的适合度，从而可以避免出现联邦学习装置需要反复试验联系用户参与联邦学习的现象，降低了被邀请的用户拒绝参与联邦学习的可能性，可以有效提高联邦学习的参与者与联邦学习装置的交互效率，另外，还可以降低联邦学习装置的通讯负担。
可选地，在具体的实现过程中，联邦学习装置根据多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率、多个用户各自的体验损失和新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率，计算多个用户各自的索引值的方式可以有多种。比如：
方式1，联邦学习装置可以计算多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率、多个用户各自的体验损失和新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率三者之间对应的平均值，之后，将计算得到的平均值作为多个用户各自的索引值。
在方式1中，通过将多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率、多个用户各自的体验损失和新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率三者之间对应的平均值，作为多个用户各自的索引值，可以平衡多个用户各自的索引值，与多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率、多个用户各自的体验损失和新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率三者之间的关系，避免出现将当前时段参与新一轮联邦学习邀请的第一概率、体验损失和新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率三者中最低的作为多个用户各自的索引值，而导致多个用户各自的索引值与多个用户的关联性较低的现象，从而可以提高联邦学习装置根据索引值从多个用户选出参与联邦学习的用户与联邦学习的适合度，避 免了出现联邦学习装置需要反复试验联系用户参与联邦学习的现象，无需反复试验联系用户参与联邦学习，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
方式2，联邦学习装置可以根据预设策略，将多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率、多个用户各自的体验损失和新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率的最大数值作为多个用户各自的索引值。比如，联邦学习装置可以计算得到的平均值与预设阈值进行比较，确定平均值是否大于或者等于预设阈值；若确定平均值大于或者等于预设阈值，将所述第一概率、所述体验损失和所述频率中的最大数值作为多个用户各自的索引值。否则，将所述平均值作为多个用户各自的索引值。
在方式2中，通过在计算得到的多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率、多个用户各自的体验损失和新一轮的联邦学习中联邦学习服务器需与多个用户各自互动的频率的三者之间的平均值大于或者等于预设阈值时，将三者之间的最大数值作为多个用户各自的索引值，或者，在三者之间的平均值小于预设阈值时，将该平均值作为多个用户各自的索引值，可以提高多个用户各自的索引值与多个用户的关联性，从而可以提高联邦学习装置根据索引值从多个用户选出参与联邦学习的用户与联邦学习的适合度，避免了出现联邦学习装置需要反复试验联系用户参与联邦学习的现象，无需反复试验联系用户参与联邦学习，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
S104、根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习。
可选地，联邦学习装置计算得到多个用户各自的索引值后，可以根据不同的预设条件，邀请所述多个用户中满足预设条件的用户参与联邦学习。比如：
示例1，若预设条件为优先动员被联邦学习邀请频率低于预设阈值的用户参与新一轮的联邦学习，那么，联邦学习装置则可以筛选出所述多个用户中索引值小于第一预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习。比如，联邦学习装置可以将多个用户各自的索引值按照从小到大的顺序进行排序，基于该排序邀请前N个用户参与新一轮的联邦学习，或者，可以将多个用户各自的索引值按照从大到小的顺序进行排序，基于该排序邀请后N个用户参与新一轮的联邦学习，其中，N个用户各自的索引值均小于第一预设索引值。
在示例1中，联邦学习可以提高筛选出的多个用户为参与联邦学习频率低的用户的准确性，从而可以提高联邦学习装置根据索引值从多个用户选出参与联邦学习的用户与联邦学习的适合度，避免了出现联邦学习装置需要反复试验联系用户参与联邦学习的现象，无需反复试验联系用户参与联邦学习，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
示例2，若预设条件为优先动员被联邦学习邀请频率高于或者等于所述预设阈值的用户参与新一轮的联邦学习，那么，联邦学习装置则可以筛选出所述多个用户中索引值大于第二预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习。比如，联邦学习装置可以将多个用户各自的索引值按照从小到大的顺序进行排序，基于该排序邀请后N个用户参与新一轮的联邦学习，或者，可以将多个用户各自的索引值按照从大到小的顺序进行排序，基于该排序邀请前N个用户参与新一轮的联邦学习，其中，N个用户各自的索引值均大于第二预设索引值。
在示例2中，联邦学习可以提高筛选出的多个用户为参与联邦学习频率高的用户的 准确性，从而可以提高联邦学习装置根据索引值从多个用户选出参与联邦学习的用户与联邦学习的适合度，无需反复试验联系用户参与联邦学习，避免了出现联邦学习装置需要反复试验联系用户参与联邦学习的现象，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
可选地，联邦学习装置可以在邀请多个用户中N个用户参与新一轮的联邦学习之后，可以接收N个用户的反馈数据，并根据N个用户的反馈数据更新N个用户参与新一轮联邦学习邀请的概率、N个用户的体验损失和联邦学习服务器与N个用户互动的频率，用于计算N个用户各自参与下一轮联邦学习的适合度值，以提高N个用户各自参与下一轮联邦学习的适合度，可以有效提高下一轮联邦学习的N个用户中的参与者与联邦学习装置的交互效率。
需要说明的是，上述第一预设索引值和第二预设索引值可以相同，也可以不同，本发明实施例不做具体限定。
需要说明的是，上述N的数值可由联邦学习装置的系统管理员设定，也可由预设索引值上限，或者预设索引值下限决定，本发明实施例不做具体限定。
通过以上描述可知，本发明实施例提供的技术方案中，与现有技术相比，由于多个用户各自的索引值是联邦学习装置根据多个用户各自历史接受联邦学习邀请后的反馈数据、多个用户各自的用户画像数据和多个用户各自在上一个时间窗口内参与联邦学习建模的次数计算得到的，因此，多个用户各自的索引值，可以与多个用户各自在不同时段接受联邦学习邀请的概率、多个用户各自参与新一轮联邦学习的概率、多个用户各自在上一个时间窗口参与联邦学习建模的满意程度相关，从而可以提高联邦学习装置根据索引值从多个用户中选出的满足预设条件的用户与联邦学习的适合度，降低了索引出的用户拒绝参与联邦学习的可能性，无需反复试验联系用户参与联邦学习，可以有效提高联邦学习的参与者与联邦学习装置的交互效率。
基于同一发明构思下，本发明还提供了一种联邦学习装置。请参考图2所示，为本发明实施例提供的一种联邦学习装置的结构示意图。
如图2所示，联邦学习装置200包括：
获取单元201，用于获取多个用户历史接受联邦学习邀请后的反馈数据，以及获取所述多个用户各自的用户画像数据；
处理单元202，用于确定所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数；根据所述多个用户各自的反馈数据、所述用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值；所述索引值用于表征所述多个用户各自参与新一轮联邦学习的适合度值；
邀请单元203，用于根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习。
在一种可能的设计中，所述处理单元202具体用于：
根据所述多个用户各自的用户画像数据，确定所述多个用户各自的用户画像置信度；
根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值。
在一种可能的设计中，所述处理单元202具体用于：
根据所述反馈数据进行建模，预测所述多个用户各自在当前时段参与新一轮联邦学习 邀请的第一概率；其中，所述第一概率在新一轮联邦学习的时间窗口中的任意一个时间点对应的数值大小，与所述多个用户各自在所述时间点的分时响应度正相关；所述分时响应度用于表征所述多个用户各自反馈接受联邦学习邀请的速度；
根据所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，预测所述多个用户各自的体验损失；其中，所述体验损失用于表示所述多个用户各自当前时段接受参与新一轮联邦学习邀请行为的满意程度；
根据所述用户画像置信度进行建模，预测在新一轮的联邦学习中联邦学习服务器需与所述多个用户各自互动的频率；
根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值。
在一种可能的设计中，所述处理单元202具体用于：
计算所述第一概率、所述体验损失和所述频率三者之间对应的平均值；将所述平均值作为所述多个用户各自的索引值；或者，
根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
在一种可能的设计中，所述处理单元202具体用于：
将所述平均值与预设阈值进行比较，确定所述平均值是否大于或者等于所述预设阈值；
若确定所述平均值大于或者等于所述预设阈值，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
在一种可能的设计中，所述邀请单元203具体用于：
若确定优先动员被联邦学习邀请频率低于预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值小于第一预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习；
若确定优先动员被联邦学习邀请频率高于或者等于所述预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值大于第二预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习。
在一种可能的设计中，所述处理单元202还用于：
接收所述N个用户的反馈数据，更新所述N个用户参与新一轮联邦学习邀请的概率、所述N个用户的体验损失和所述联邦学习服务器与所述N个用户互动的频率，用于计算所述N个用户各自参与下一轮联邦学习的适合度值。
本发明实施例中的联邦学习装置200与前述图1所示的联邦学习中用户的索引方法是基于同一构思下的发明，通过前述对联邦学习中用户的索引方法的详细描述，本领域技术人员可以清楚的了解本实施例中联邦学习装置200的实施过程，所以为了说明书的简洁，在此不再赘述。
基于同一发明构思下，本发明还提供了一种计算机设备。请参考图3所示，为本发明实施例提供的一种计算机设备的结构示意图。
如图3所示，计算机设备300包括：存储器301和至少一个处理器302。其中，所述存储器301存储一个或多个计算机程序；当所述存储器301存储的一个或多个计算机程序被所述至少一个处理器302执行时，使得所述计算机设备300执行上述联邦学习中用户的索引方法的步骤。
可选地，所述存储器301可以包括高速随机存取存储器，还可以包括非易失存储器， 例如磁盘存储器件、闪存器件或其他非易失性固态存储器件等，本发明实施例不作限定。
可选地，所述处理器302可以是通用的处理器(central processing unit，CPU)，或ASIC，或FPGA，也可以是一个或多个用于控制程序执行的集成电路。
在一些实施例中，所述存储器301和所述处理器302可以在同一芯片上实现，在另一些实施例中，它们也可以在独立的芯片上分别实现，本发明实施例不作限定。
基于同一发明构思下，本发明还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机指令，当所述计算机指令被计算机设备执行时，使所述计算机设备可以执行上述联邦学习中用户的索引方法的步骤。
基于同一发明构思下，本发明还提供了一种计算机程序产品，该计算机程序产品包括存储在计算机可读存储介质上的计算机程序，计算机程序包括程序指令，当程序指令被计算机设备执行时，使计算机设备执行上述联邦学习中用户的索引方法的步骤。
本发明是参照根据本发明实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
显然，本领域的技术人员可以对本发明进行各种改动和变型而不脱离本发明的精神和范围。这样，倘若本发明的这些修改和变型属于本发明权利要求及其等同技术的范围之内，则本发明也意图包含这些改动和变型在内。

Claims (20)

  1. 一种联邦学习中用户的索引方法，其特征在于，包括：
    获取多个用户历史接受联邦学习邀请后的反馈数据，以及获取所述多个用户各自的用户画像数据；
    确定所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数；
    根据所述多个用户各自的反馈数据、所述用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值；所述索引值用于表征所述多个用户各自参与新一轮联邦学习的适合度值；
    根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习。
  2. 如权利要求1所述的方法，其特征在于，根据所述多个用户各自的反馈数据、用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值，包括：
    根据所述多个用户各自的用户画像数据，确定所述多个用户各自的用户画像置信度；
    根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值。
  3. 如权利要求2所述的方法，其特征在于，根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值，包括：
    根据所述反馈数据进行建模，预测所述多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率；其中，所述第一概率在新一轮联邦学习的时间窗口中的任意一个时间点对应的数值大小，与所述多个用户各自在所述时间点的分时响应度正相关；所述分时响应度用于表征所述多个用户各自反馈接受联邦学习邀请的速度；
    根据所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，预测所述多个用户各自的体验损失；其中，所述体验损失用于表示所述多个用户各自当前时段接受参与新一轮联邦学习邀请行为的满意程度；
    根据所述用户画像置信度进行建模，预测在新一轮的联邦学习中联邦学习服务器需与所述多个用户各自互动的频率；
    根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值。
  4. 如权利要求3所述的方法，其特征在于，根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值，包括：
    计算所述第一概率、所述体验损失和所述频率三者之间对应的平均值；将所述平均值作为所述多个用户各自的索引值；或者，
    根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
  5. 如权利要求4所述的方法，其特征在于，根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值，包括：
    将所述平均值与预设阈值进行比较，确定所述平均值是否大于或者等于所述预设阈值；
    若确定所述平均值大于或者等于所述预设阈值，将所述第一概率、所述体验损失和所 述频率中的最大数值作为所述多个用户各自的索引值。
  6. 如权利要求1-5任一项所述的方法，其特征在于，根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习，包括：
    若确定优先动员被联邦学习邀请频率低于预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值小于第一预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习；
    若确定优先动员被联邦学习邀请频率高于或者等于所述预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值大于第二预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习。
  7. 如权利要求6所述的方法，其特征在于，邀请所述N个用户参与新一轮的联邦学习之后，所述方法还包括：
    接收所述N个用户的反馈数据，更新所述N个用户参与新一轮联邦学习邀请的概率、所述N个用户的体验损失和所述联邦学习服务器与所述N个用户互动的频率，用于计算所述N个用户各自参与下一轮联邦学习的适合度值。
  8. 一种联邦学习装置，其特征在于，包括：
    获取单元，用于获取多个用户历史接受联邦学习邀请后的反馈数据，以及获取所述多个用户各自的用户画像数据；
    处理单元，用于确定所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数；根据所述多个用户各自的反馈数据、所述用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值；所述索引值用于表征所述多个用户各自参与新一轮联邦学习的适合度值；
    邀请单元，用于根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习。
  9. 如权利要求8所述的装置，其特征在于，所述处理单元具体用于：
    根据所述多个用户各自的用户画像数据，确定所述多个用户各自的用户画像置信度；
    根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值。
  10. 如权利要求9所述的装置，其特征在于，所述处理单元具体用于：
    根据所述反馈数据进行建模，预测所述多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率；其中，所述第一概率在新一轮联邦学习的时间窗口中的任意一个时间点对应的数值大小，与所述多个用户各自在所述时间点的分时响应度正相关；所述分时响应度用于表征所述多个用户各自反馈接受联邦学习邀请的速度；
    根据所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，预测所述多个用户各自的体验损失；其中，所述体验损失用于表示所述多个用户各自当前时段接受参与新一轮联邦学习邀请行为的满意程度；
    根据所述用户画像置信度进行建模，预测在新一轮的联邦学习中联邦学习服务器需与所述多个用户各自互动的频率；
    根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值。
  11. 如权利要求10所述的装置，其特征在于，所述处理单元具体用于：
    计算所述第一概率、所述体验损失和所述频率三者之间对应的平均值；将所述平均值 作为所述多个用户各自的索引值；或者，
    根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
  12. 如权利要求11所述的装置，其特征在于，所述处理单元具体用于：
    将所述平均值与预设阈值进行比较，确定所述平均值是否大于或者等于所述预设阈值；
    若确定所述平均值大于或者等于所述预设阈值，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
  13. 如权利要求8-12任一项所述的装置，其特征在于，所述邀请单元具体用于：
    若确定优先动员被联邦学习邀请频率低于预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值小于第一预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习；
    若确定优先动员被联邦学习邀请频率高于或者等于所述预设阈值的用户参与新一轮的联邦学习，则筛选出所述多个用户中索引值大于第二预设索引值的N个用户，邀请所述N个用户参与新一轮的联邦学习。
  14. 如权利要求13所述的装置，其特征在于，所述处理单元还用于：
    接收所述N个用户的反馈数据，更新所述N个用户参与新一轮联邦学习邀请的概率、所述N个用户的体验损失和所述联邦学习服务器与所述N个用户互动的频率，用于计算所述N个用户各自参与下一轮联邦学习的适合度值。
  15. 一种计算机设备，其特征在于，所述计算机设备包括至少一个处理器和存储器；
    所述存储器存储一个或多个计算机程序；
    所述处理器读取所述存储器存储的一个或多个计算机程序，执行以下方法：获取多个用户历史接受联邦学习邀请后的反馈数据，以及获取所述多个用户各自的用户画像数据；确定所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数；根据所述多个用户各自的反馈数据、所述用户画像数据和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数，计算所述多个用户各自的索引值；所述索引值用于表征所述多个用户各自参与新一轮联邦学习的适合度值；根据所述多个用户各自的索引值，邀请所述多个用户中满足预设条件的用户参与联邦学习。
  16. 如权利要求15所述的计算机设备，其特征在于，所述处理器具体用于：
    根据所述多个用户各自的用户画像数据，确定所述多个用户各自的用户画像置信度；
    根据所述多个用户各自的反馈数据、用户画像置信度和所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数分别进行建模，计算得到所述多个用户各自的索引值。
  17. 如权利要求16所述的计算机设备，其特征在于，所述处理器具体用于：
    根据所述反馈数据进行建模，预测所述多个用户各自在当前时段参与新一轮联邦学习邀请的第一概率；其中，所述第一概率在新一轮联邦学习的时间窗口中的任意一个时间点对应的数值大小，与所述多个用户各自在所述时间点的分时响应度正相关；所述分时响应度用于表征所述多个用户各自反馈接受联邦学习邀请的速度；
    根据所述多个用户各自在上一个时间窗口内参与联邦学习建模的次数进行建模，预测所述多个用户各自的体验损失；其中，所述体验损失用于表示所述多个用户各自当前时段接受参与新一轮联邦学习邀请行为的满意程度；
    根据所述用户画像置信度进行建模，预测在新一轮的联邦学习中联邦学习服务器需与 所述多个用户各自互动的频率；
    根据所述第一概率、所述体验损失和所述频率，计算所述多个用户各自的索引值。
  18. 如权利要求17所述的计算机设备，其特征在于，所述处理器具体用于：
    计算所述第一概率、所述体验损失和所述频率三者之间对应的平均值；将所述平均值作为所述多个用户各自的索引值；或者，
    根据预设策略，将所述第一概率、所述体验损失和所述频率中的最大数值作为所述多个用户各自的索引值。
  19. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有计算机指令，当所述计算机指令在计算机设备上运行时，使得所述计算机设备执行如权利要求1-7任一项所述的方法的步骤。
  20. 一种计算机程序产品，其特征在于，该计算机程序产品包括存储在计算机可读存储介质上的计算机程序，计算机程序包括程序指令，当程序指令被计算机设备执行时，使计算机设备执行如权利要求1-7任一项所述的方法的步骤。
PCT/CN2021/084610 2020-03-31 2021-03-31 一种联邦学习中用户的索引方法及联邦学习装置 WO2021197388A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010244824.9A CN111428885B (zh) 2020-03-31 2020-03-31 一种联邦学习中用户的索引方法及联邦学习装置
CN202010244824.9 2020-03-31

Publications (1)

Publication Number Publication Date
WO2021197388A1 true WO2021197388A1 (zh) 2021-10-07

Family

ID=71550052

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/084610 WO2021197388A1 (zh) 2020-03-31 2021-03-31 一种联邦学习中用户的索引方法及联邦学习装置

Country Status (2)

Country Link
CN (1) CN111428885B (zh)
WO (1) WO2021197388A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111428885B (zh) * 2020-03-31 2021-06-04 深圳前海微众银行股份有限公司 一种联邦学习中用户的索引方法及联邦学习装置
CN112508205A (zh) * 2020-12-04 2021-03-16 中国科学院深圳先进技术研究院 一种联邦学习调度方法、装置及系统
CN116567702A (zh) * 2022-01-26 2023-08-08 展讯通信（上海）有限公司 一种用户设备选择方法、装置、芯片及模组设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180373988A1 (en) * 2017-06-27 2018-12-27 Hcl Technologies Limited System and method for tuning and deploying an analytical model over a target eco-system
CN110443063A (zh) * 2019-06-26 2019-11-12 电子科技大学 自适性保护隐私的联邦深度学习的方法
CN110572253A (zh) * 2019-09-16 2019-12-13 济南大学 一种联邦学习训练数据隐私性增强方法及系统
CN110598870A (zh) * 2019-09-02 2019-12-20 深圳前海微众银行股份有限公司 一种联邦学习方法及装置
CN110610242A (zh) * 2019-09-02 2019-12-24 深圳前海微众银行股份有限公司 一种联邦学习中参与者权重的设置方法及装置
CN111428885A (zh) * 2020-03-31 2020-07-17 深圳前海微众银行股份有限公司 一种联邦学习中用户的索引方法及联邦学习装置

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11475350B2 (en) * 2018-01-22 2022-10-18 Google Llc Training user-level differentially private machine-learned models
CN109300050A (zh) * 2018-08-31 2019-02-01 平安科技（深圳）有限公司 基于用户画像的保险推送方法、装置及存储介质
CN110245510B (zh) * 2019-06-19 2021-12-07 北京百度网讯科技有限公司 用于预测信息的方法和装置
CN110297848B (zh) * 2019-07-09 2024-02-23 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
CN110908893A (zh) * 2019-10-08 2020-03-24 深圳逻辑汇科技有限公司 联邦学习的沙盒机制

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180373988A1 (en) * 2017-06-27 2018-12-27 Hcl Technologies Limited System and method for tuning and deploying an analytical model over a target eco-system
CN110443063A (zh) * 2019-06-26 2019-11-12 电子科技大学 自适性保护隐私的联邦深度学习的方法
CN110598870A (zh) * 2019-09-02 2019-12-20 深圳前海微众银行股份有限公司 一种联邦学习方法及装置
CN110610242A (zh) * 2019-09-02 2019-12-24 深圳前海微众银行股份有限公司 一种联邦学习中参与者权重的设置方法及装置
CN110572253A (zh) * 2019-09-16 2019-12-13 济南大学 一种联邦学习训练数据隐私性增强方法及系统
CN111428885A (zh) * 2020-03-31 2020-07-17 深圳前海微众银行股份有限公司 一种联邦学习中用户的索引方法及联邦学习装置

Also Published As

Publication number Publication date
CN111428885B (zh) 2021-06-04
CN111428885A (zh) 2020-07-17

Similar Documents

Publication Publication Date Title
WO2021197388A1 (zh) 一种联邦学习中用户的索引方法及联邦学习装置
Apostolopoulos et al. Cognitive data offloading in mobile edge computing for internet of things
CN110266745B (zh) 基于深度网络的信息流推荐方法、装置、设备及存储介质
CN110417607B (zh) 一种流量预测方法、装置及设备
Brandão et al. A biased random‐key genetic algorithm for scheduling heterogeneous multi‐round systems
Zhao et al. Emergency transshipment in decentralized dealer networks: When to send and accept transshipment requests
Abraham et al. Reputation formation in economic transactions
Kasthurirathna et al. Emergence of scale-free characteristics in socio-ecological systems with bounded rationality
CN110457577A (zh) 数据处理方法、装置、设备和计算机存储介质
Tan et al. ECRModel: An elastic collision-based rumor-propagation model in online social networks
Tang et al. Budget‐constraint stochastic task scheduling on heterogeneous cloud systems
Di Tollo et al. An experimental study of adaptive control for evolutionary algorithms
Wang et al. Bad-scenario-set robust scheduling for a job shop to hedge against processing time uncertainty
CN109246178B (zh) 一种在线题目发放方法、装置、服务器及存储介质
CN110427160A (zh) 归置组分布的均衡方法及装置
Raj et al. A firefly inspired game dissemination and QoS-based priority pricing strategy for online social network games
CN111652673B (zh) 智能推荐方法、装置、服务器和存储介质
CN112700003A (zh) 网络结构搜索方法、装置、设备、存储介质及程序产品
CN104967643A (zh) 一种获取转移数据的方法及装置
Ma et al. Dynamic pricing in the presence of participation-dependent social learning
CN106612296A (zh) 一种用户设备连接请求分配的方法及装置
WO2021103508A1 (zh) 种子用户的选取方法、装置、设备及存储介质
Polevoy et al. Nash equilibria in shared effort games
CN108512842A (zh) 一种面向移动终端的网络应用账号共享系统
Singh et al. Apply AHP for resource allocation problem in cloud

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: 21781231

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 16/01/2023)

122 Ep: pct application non-entry in european phase

Ref document number: 21781231

Country of ref document: EP

Kind code of ref document: A1