A method and system for federated learning feature screening based on homomorphic encryption
Through the federated learning feature screening method based on homomorphic encryption, the problem that training participants who lack target data in vertical federated learning cannot complete local feature selection, achieving improvement in model performance and guaranteeing data privacy.
Patent Information
- Application Number
- CN202311354062.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-18
- Publication Date
- 2025-02-28
- Estimated Expiration
- 2043-10-18
AI Technical Summary
In vertical federated learning scenarios, training participants who lack target data cannot complete local feature selection alone, resulting in poor model performance.
A federated learning feature screening method based on homomorphic encryption is used to obtain the encryption target value of the second participant, and feature binning and information value calculation are performed to screen out features that meet the requirements.
While ensuring the privacy of target data of the second participant user, the first participant who lacks target data can complete local feature screening to improve model performance and performance.
Smart Images
Figure CN117278199B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of privacy computing technology, and in particular to a method and system for screening features of federated learning based on homomorphic encryption. Background Art
[0002] As a new paradigm of distributed machine learning applications with "data available but invisible, data fixed but model moving", the core idea of federated learning is to build a global model based on virtual fusion data by training distributed models among multiple data sources, that is, without exchanging local data, only exchanging the intermediate results of model parameters, so as to achieve a balance between data privacy protection and value sharing. According to the data feature space and sample ID space distribution of different participants, federated learning is divided into horizontal federated learning, vertical federated learning, and migration federated learning.
[0003] Similar to traditional centralized machine learning technology, the "feature combination of each training party" in federated learning is crucial to the performance of the global virtual model obtained. Appropriate feature combination enables the trained global virtual model to better capture data information and improve model expressiveness.
[0004] In the vertical federated learning scenario, training participants who "lack target data" "cannot complete local feature selection independently" and can only use all features for model training, resulting in poor model performance. Summary of the invention
[0005] In an embodiment of the present invention, a method and system for federated learning feature screening based on homomorphic encryption are provided to solve the problem that training participants who "lack target data" "cannot complete local feature selection independently" and can only use all features for model training, resulting in poor model performance.
[0006] To achieve the above-mentioned purpose, on the one hand, the present invention provides a method for selecting features of federated learning based on homomorphic encryption, the method comprising: S1, obtaining the encrypted target values of all users in the second participant; S2, binning the users corresponding to each feature of the first participant according to the feature binning method, and counting the number of users in each bin of each feature; S3, calculating the sum of the encrypted target values of all users in each bin of each feature according to the encrypted target values of all users; sending all the sum values corresponding to each feature to the second participant so that the second participant decrypts them, and counting the number of users with a target value of 1 in each bin of each feature; S4, obtaining the number of users with a target value of 1 in each bin of each feature; and calculating the information value of each feature according to the number of users in each bin of each feature and the number of users with a target value of 1 in each bin of each feature; S5, filtering out features that meet the requirements according to the information value of all features and preset feature evaluation criteria.
[0007] Optionally, S4 includes: obtaining the number of users whose target value is 1 in each bin in each feature; and calculating the number of users whose target value is 0 in each bin in each feature based on the number of users in each bin in each feature and the number of users whose target value is 1 in each bin in each feature; calculating the evidence weight of each bin in each feature based on the number of users whose target value is 0 in each bin in each feature and the number of users whose target value is 1 in each bin in each feature; calculating the information value of each bin in each feature based on the number of users whose target value is 0 in each bin in each feature, the number of users whose target value is 1 in each bin in each feature, and the evidence weight of each bin in each feature; calculating the information value of each feature based on the information value of all bins in each feature.
[0008] Optionally, the evidence weight for each bin of the current feature is calculated according to the following formula:
[0009]
[0010] in, is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature;
[0011] The information value of each bin of the current feature is calculated according to the following formula:
[0012]
[0013] in, is the information value of the i-th bin of the current feature; is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature.
[0014] Optionally, the feature binning method includes: chi-square binning and minimum entropy binning.
[0015] Optionally, the preset feature evaluation standard is: if the information value of the current feature is less than a first preset threshold, the current feature is judged to be a worthless feature; if the information value of the current feature is greater than or equal to the first preset threshold and less than the second preset threshold, the current feature is judged to be a weak value feature; if the information value of the current feature is greater than or equal to the second preset threshold and less than the third preset threshold, the current feature is judged to be a medium value feature; if the information value of the current feature is greater than or equal to the third preset threshold and less than the fourth preset threshold, the current feature is judged to be a strong value feature; if the information value of the current feature is greater than or equal to the fourth preset threshold, the current feature is judged to be an unreal feature.
[0016] On the other hand, the present invention provides a federated learning feature screening system based on homomorphic encryption, which includes: an acquisition unit, used to obtain the encrypted target values of all users in the second participant; a binning and counting unit, used to bin the users corresponding to each feature of the first participant according to the feature binning method, and count the number of users in each bin of each feature; a summing and sending unit, used to calculate the sum of the encrypted target values of all users in each bin of each feature according to the encrypted target values of all users; send all the sum values corresponding to each feature to the second participant so that the second participant decrypts them, and counts the number of users whose target value is 1 in each bin of each feature; a calculation unit, used to obtain the number of users whose target value is 1 in each bin of each feature; and calculate the information value of each feature based on the number of users in each bin of each feature and the number of users whose target value is 1 in each bin of each feature; a screening unit, used to screen out features that meet the requirements according to the information value of all features and preset feature evaluation criteria.
[0017] Optionally, the calculation unit includes: a quantity calculation subunit, used to obtain the number of users whose target value is 1 in each subbox of each feature; and calculate the number of users whose target value is 0 in each subbox of each feature according to the number of users in each subbox of each feature and the number of users whose target value is 1 in each subbox of each feature; an evidence weight calculation subunit, used to calculate the evidence weight of each subbox in each feature according to the number of users whose target value is 0 in each subbox of each feature and the number of users whose target value is 1 in each subbox of each feature; an information value calculation subunit, used to calculate the information value of each subbox in each feature according to the number of users whose target value is 0 in each subbox of each feature, the number of users whose target value is 1 in each subbox of each feature, and the evidence weight of each subbox in each feature; and a summation subunit, used to calculate the information value of each feature according to the information values of all subboxes in each feature.
[0018] Optionally, the evidence weight for each bin of the current feature is calculated according to the following formula:
[0019]
[0020] in, is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature;
[0021] The information value of each bin of the current feature is calculated according to the following formula:
[0022]
[0023] in, is the information value of the i-th bin of the current feature; is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature.
[0024] Optionally, the feature binning method includes: chi-square binning and minimum entropy binning.
[0025] Optionally, the preset feature evaluation standard is: if the information value of the current feature is less than a first preset threshold, the current feature is judged to be a worthless feature; if the information value of the current feature is greater than or equal to the first preset threshold and less than the second preset threshold, the current feature is judged to be a weak value feature; if the information value of the current feature is greater than or equal to the second preset threshold and less than the third preset threshold, the current feature is judged to be a medium value feature; if the information value of the current feature is greater than or equal to the third preset threshold and less than the fourth preset threshold, the current feature is judged to be a strong value feature; if the information value of the current feature is greater than or equal to the fourth preset threshold, the current feature is judged to be an unreal feature.
[0026] Beneficial effects of the present invention:
[0027] The present invention provides a method and system for selecting features of federated learning based on homomorphic encryption, wherein the method comprises: S1, obtaining the encrypted target values of all users in the second participant; S2, binning the users corresponding to each feature of the first participant according to the feature binning method, and counting the number of users in each bin of each feature; S3, calculating the sum of the encrypted target values of all users in each bin of each feature according to the encrypted target values of all users; sending all the sum values corresponding to each feature to the second participant so that the second participant decrypts and counts the number of users with a target value of 1 in each bin of each feature; S4, obtaining the number of users with a target value of 1 in each bin of each feature; and calculating the information value of each feature according to the number of users in each bin of each feature and the number of users with a target value of 1 in each bin of each feature; S5, filtering out the features that meet the requirements according to the information value of all features and the preset feature evaluation criteria. Through the above method, while ensuring the privacy of the target data of the second participant user, the first participant who lacks the target data can use the target data of the second participant to complete the local feature screening, and screen out the features that meet the requirements to participate in the model training, thereby improving the performance and expression of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 It is a flowchart of a method for selecting features of federated learning based on homomorphic encryption provided by an embodiment of the present invention;
[0029] Figure 2 is a flow chart of calculating the information value of each feature provided by an embodiment of the present invention;
[0030] Figure 3 It is a structural schematic diagram of a federated learning feature screening system based on homomorphic encryption provided by an embodiment of the present invention;
[0031] Figure 4 It is a schematic diagram of the structure of a computing unit provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0032] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0033] In the vertical federated learning scenario, training participants who "lack target data" "cannot complete local feature selection independently" and can only use all features for model training, resulting in poor model performance.
[0034] In view of the above problems, the present invention provides a method for selecting features of federated learning based on homomorphic encryption. Figure 1 is a flowchart of a method for selecting features of federated learning based on homomorphic encryption provided by an embodiment of the present invention, such as Figure 1 As shown, the method includes:
[0035] S1. Obtain the encrypted target values of all users in the second party;
[0036] Specifically, taking the loan amount prediction scenario of the vertical federated learning regression model as an example, it is assumed that the first participant is the tax management agency P1 that lacks target data; the second participant is the lending bank P2 that has target data; the sample data of the first participant is composed of: dataP1 = {salary x1, accumulated provident fund x2, job stability coefficient x3, age x4}; the sample data of the second participant is composed of: dataP2 = {accumulated assets x5, repayment label value y}; the target data is the repayment label value y; the specific data values are shown in Table 1 and Table 2. Here, the users in Table 1 and Table 2 are the same.
[0037] Table 1 Sample data of tax administration agency P1
[0038] User ID <![CDATA[Salary x1]]> <![CDATA[Accumulated provident fund x2]]> <![CDATA[Working stability coefficient x3]]> <![CDATA[Age x4]]> N000001 39299 13473 0.976575052 25 N000002 23297 73091 0.317665274 28 …… …… …… …… …… N100000 17166 106008 0.422950113 30
[0039] Table 2 Sample data of lending bank P2
[0040] User ID <![CDATA[Accumulated value of assets x5]]> Repayment tag value Y N000001 135172 1 N000002 50109 0 …… …… …… N100000 78183 1
[0041] First, the second party, the lending bank P2, uses the public key pk and the additive homomorphic encryption algorithm to encrypt all user IDs = {id1, id2, ..., id 100000}The corresponding repayment label value Y={y1,y2,…,y 100000}, the encrypted target value of the processed data set (i.e., all users in the second participant) is:
[0042]
[0043] The first participant obtains the encrypted target value dataCipher of all users in the second participant.
[0044] S2. bin the users corresponding to each feature of the first participant according to the feature binning method, and count the number of users in each bin for each feature;
[0045] Specifically, the feature binning methods include: chi-square binning, minimum entropy binning, etc.
[0046] Select appropriate feature binning method for each feature (four features) of the first participant corresponding to user ID = {id i}i∈100000 Perform binning and count the number of users in each bin for each feature.
[0047] Taking the feature salary x1 as an example, the results of the labeling and binning are: That is, there are n1 bins in total.
[0048] in, The number of users in the bin for b1;
[0049] The number of users in the bin for b2;
[0050] …
[0051] for The number of users in the bin;
[0052] Special note: for all user IDs = {id i} i∈100000 , can only and must belong to one of the bins of this feature.
[0053] S3. Calculate the sum of the encrypted target values of all users in each sub-box in each feature according to the encrypted target values of all users; send all the sum values corresponding to each feature to the second participant so that the second participant decrypts them, and count the number of users whose target value in each sub-box in each feature is 1;
[0054] Specifically, based on the encrypted target values of all users in the second participant, the encrypted target values (repayment tag value ciphertext) and values of all users in each bin of each feature are calculated respectively;
[0055] Similarly, taking the feature salary x1 as an example, based on the above binning results, we have:
[0056] The sum of the ciphertexts of all users’ repayment label values in bin b1 is
[0057] The sum of the ciphertexts of all users’ repayment label values in bin b2 is
[0058] …
[0059] Binning The sum of the ciphertexts of all users’ repayment tag values is
[0060] Send the sum sequence of encrypted target values (repayment tag value ciphertext) of all users in each bin in the feature salary x1 to the second participant
[0061] Similarly, the sum sequence of encrypted target values of all users in each subbox in the feature accumulated provident fund x2 is sent to the second participant; the sum sequence of encrypted target values of all users in each subbox in the feature work stability coefficient x3 is sent to the second participant; the sum sequence of encrypted target values of all users in each subbox in the feature age x4 is sent to the second participant.
[0062] The second participant decrypts the received encrypted target value sum sequence of all users in each subbox of each feature, and obtains the set of users whose target value (repayment label value) is 1 in each subbox of each feature; it should be noted that the repayment label value has only two values: 1 and 0.
[0063] Taking the characteristic salary x1 as an example, based on the private key sk and the homomorphic decryption algorithm, the sum sequence of the repayment label values of all users in each bin corresponding to the characteristic salary x1 (encrypted target value and value sequence) Decrypt them separately to obtain the number of users whose repayment label value (target value) is 1 in each bin corresponding to the feature salary x1:
[0064] Similarly, the statistical feature accumulated provident fund x2 corresponds to the number of users whose repayment label value (target value) is 1 in each sub-box: n1_x2; the statistical feature work stability coefficient x3 corresponds to the number of users whose repayment label value (target value) is 1 in each sub-box: n1_x3; the statistical feature age x4 corresponds to the number of users whose repayment label value (target value) is 1 in each sub-box: n1_x4.
[0065] S4. Obtain the number of users whose target value is 1 in each bin of each feature; and calculate the information value of each feature according to the number of users in each bin of each feature and the number of users whose target value is 1 in each bin of each feature;
[0066] Figure 2 is a flow chart of calculating the information value of each feature provided by an embodiment of the present invention; Figure 2 As shown, the S4 includes:
[0067] S41, obtaining the number of users whose target value is 1 in each bin in each feature; and calculating the number of users whose target value is 0 in each bin in each feature according to the number of users in each bin in each feature and the number of users whose target value is 1 in each bin in each feature;
[0068] That is, the second participant sends the set of the number of users whose repayment label value is 1 in each bin in each feature to the first participant;
[0069] Taking the feature salary x1 as an example, the number of users whose target value is 0 in each bin of the current feature is:
[0070]
[0071] Similarly, the number of users whose target value is 0 in each bin in each feature is obtained.
[0072] S42, calculating the evidence weight of each bin in each feature according to the number of users with a target value of 0 in each bin in each feature and the number of users with a target value of 1 in each bin in each feature;
[0073] Specifically, the evidence weight of each bin of the current feature is calculated according to the following formula:
[0074]
[0075] in, is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature.
[0076] S43, calculating the information value of each bin in each feature according to the number of users with a target value of 0 in each bin in each feature, the number of users with a target value of 1 in each bin in each feature, and the evidence weight of each bin in each feature;
[0077] The information value of each bin of the current feature is calculated according to the following formula:
[0078]
[0079] in, The information value of the i-th bin of the current feature; is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature.
[0080] S44. Calculate the information value of each feature based on the information value of all bins in each feature.
[0081] The information value of the current feature
[0082] According to the above method, the information value of each feature is obtained.
[0083] S5. Filter out the features that meet the requirements based on the information value of all features and the preset feature evaluation criteria.
[0084] The preset feature evaluation criteria are:
[0085] If the information value of the current feature is less than the first preset threshold, the current feature is judged to be a worthless feature;
[0086] If the information value of the current feature is greater than or equal to the first preset threshold and less than the second preset threshold, the current feature is judged to be a weak value feature;
[0087] If the information value of the current feature is greater than or equal to the second preset threshold and less than the third preset threshold, the current feature is judged to be a medium value feature;
[0088] If the information value of the current feature is greater than or equal to the third preset threshold and less than the fourth preset threshold, the current feature is judged to be a high-value feature;
[0089] If the information value of the current feature is greater than or equal to the fourth preset threshold, the current feature is determined to be a false feature.
[0090] For example: if it is required to filter out false features, the information value of the corresponding required features is greater than or equal to the fourth preset threshold.
[0091] Specifically, in the present invention, the first preset threshold, the second preset threshold, the third preset threshold, and the fourth preset threshold are 0.02, 0.1, 0.3, and 0.5, respectively. It should be noted that the present invention does not limit the specific numerical values of the first preset threshold, the second preset threshold, the third preset threshold, and the fourth preset threshold.
[0092] The method of the present invention can use the target data of the second participant while ensuring the data privacy of the second participant, so that the first participant can use the target data of the second participant to complete local feature screening, screen out features that meet the requirements to participate in model training, and thus improve the performance and expression of the model.
[0093] Figure 3 is a structural diagram of a federated learning feature screening system based on homomorphic encryption provided by an embodiment of the present invention; Figure 3 As shown, the system includes:
[0094] An acquisition unit 201 is used to acquire the encrypted target values of all users in the second party;
[0095] A binning and counting unit 202, configured to bin the users corresponding to each feature of the first participant according to a feature binning method, and count the number of users in each bin of each feature;
[0096] Specifically, the feature binning methods include: chi-square binning, minimum entropy binning, etc.
[0097] A summing and sending unit 203 is used to calculate the sum of the encrypted target values of all users in each sub-box in each feature according to the encrypted target values of all users; send all the sum values corresponding to each feature to the second participant so that the second participant decrypts them, and count the number of users whose target value in each sub-box in each feature is 1;
[0098] A calculation unit 204 is used to obtain the number of users whose target value is 1 in each bin of each feature; and calculate the information value of each feature according to the number of users in each bin of each feature and the number of users whose target value is 1 in each bin of each feature;
[0099] Figure 4 is a schematic diagram of the structure of a computing unit provided in an embodiment of the present invention, such as Figure 4 As shown, the calculation unit 204 includes:
[0100] The quantity calculation subunit 2041 is used to obtain the number of users whose target value is 1 in each bin of each feature; and calculate the number of users whose target value is 0 in each bin of each feature according to the number of users in each bin of each feature and the number of users whose target value is 1 in each bin of each feature;
[0101] The evidence weight calculation subunit 2042 is used to calculate the evidence weight of each bin in each feature according to the number of users with a target value of 0 in each bin in each feature and the number of users with a target value of 1 in each bin in each feature;
[0102] The evidence weight of each bin of the current feature is calculated according to the following formula:
[0103]
[0104] in, is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature;
[0105] The information value calculation subunit 2043 is used to calculate the information value of each bin in each feature according to the number of users with a target value of 0 in each bin in each feature, the number of users with a target value of 1 in each bin in each feature, and the evidence weight of each bin in each feature;
[0106] The information value of each bin of the current feature is calculated according to the following formula:
[0107]
[0108] in, The information value of the i-th bin of the current feature; is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature.
[0109] The summing subunit 2044 is used to calculate the information value of each feature according to the information value of all bins in each feature.
[0110] The screening unit 205 is used to screen out the features that meet the requirements according to the information value of all the features and the preset feature evaluation criteria.
[0111] The preset feature evaluation criteria are:
[0112] If the information value of the current feature is less than the first preset threshold, the current feature is judged to be a worthless feature;
[0113] If the information value of the current feature is greater than or equal to the first preset threshold and less than the second preset threshold, the current feature is judged to be a weak value feature;
[0114] If the information value of the current feature is greater than or equal to the second preset threshold and less than the third preset threshold, the current feature is judged to be a medium value feature;
[0115] If the information value of the current feature is greater than or equal to the third preset threshold and less than the fourth preset threshold, the current feature is judged to be a high-value feature;
[0116] If the information value of the current feature is greater than or equal to the fourth preset threshold, the current feature is determined to be a false feature.
[0117] Beneficial effects of the present invention:
[0118] The present invention provides a method and system for selecting features of federated learning based on homomorphic encryption, wherein the method comprises: S1, obtaining the encrypted target values of all users in the second participant; S2, binning the users corresponding to each feature of the first participant according to the feature binning method, and counting the number of users in each bin of each feature; S3, calculating the sum of the encrypted target values of all users in each bin of each feature according to the encrypted target values of all users; sending all the sum values corresponding to each feature to the second participant so that the second participant decrypts and counts the number of users with a target value of 1 in each bin of each feature; S4, obtaining the number of users with a target value of 1 in each bin of each feature; and calculating the information value of each feature according to the number of users in each bin of each feature and the number of users with a target value of 1 in each bin of each feature; S5, filtering out the features that meet the requirements according to the information value of all features and the preset feature evaluation criteria. Through the above method, while ensuring the privacy of the target data of the second participant user, the first participant who lacks the target data can use the target data of the second participant to complete the local feature screening, and screen out the features that meet the requirements to participate in the model training, thereby improving the performance and expression of the model.
[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for selecting features of federated learning based on homomorphic encryption, characterized in that: include: S1. The first participant obtains the encrypted target values of all users in the second participant; S2. The first participant bins the users corresponding to each feature of the first participant according to the feature binning method, and counts the number of users in each bin for each feature; S3. The first participant calculates the sum of the encrypted target values of all users in each sub-box in each feature according to the encrypted target values of all users; sends all the sum values corresponding to each feature to the second participant so that the second participant decrypts the sum values and counts the number of users whose target value in each sub-box in each feature is 1; S4. The first participant obtains the number of users whose target value is 1 in each sub-box of each feature; and calculates the information value of each feature based on the number of users in each sub-box of each feature and the number of users whose target value is 1 in each sub-box of each feature; S5. The first participant selects the features that meet the requirements based on the information value of all features and the preset feature evaluation criteria; The preset feature evaluation criteria are: If the information value of the current feature is less than the first preset threshold, the current feature is judged to be a worthless feature; If the information value of the current feature is greater than or equal to the first preset threshold and less than the second preset threshold, the current feature is judged to be a weak value feature; If the information value of the current feature is greater than or equal to the second preset threshold and less than the third preset threshold, the current feature is judged to be a medium value feature; If the information value of the current feature is greater than or equal to the third preset threshold and less than the fourth preset threshold, the current feature is judged to be a high-value feature; If the information value of the current feature is greater than or equal to the fourth preset threshold, the current feature is judged to be a false feature; The first participant is a characteristic variable provider; The second participant is a target variable provider.
2. The method according to claim 1, characterized in that The S4 includes: Get the number of users whose target value is 1 in each bin of each feature; and calculate the number of users whose target value is 0 in each bin of each feature based on the number of users in each bin of each feature and the number of users whose target value is 1 in each bin of each feature; The evidence weight of each bin in each feature is calculated based on the number of users with a target value of 0 in each bin in each feature and the number of users with a target value of 1 in each bin in each feature; The information value of each bin in each feature is calculated based on the number of users with a target value of 0 in each bin in each feature, the number of users with a target value of 1 in each bin in each feature, and the evidence weight of each bin in each feature; The information value of each feature is calculated based on the information value of all bins in each feature.
3. The method according to claim 2, characterized in that: The evidence weight of each bin of the current feature is calculated according to the following formula: in, is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature; The information value of each bin of the current feature is calculated according to the following formula: in, The information value of the i-th bin of the current feature; is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature.
4. The method according to claim 1, characterized in that: The feature binning methods include: chi-square binning and minimum entropy binning.
5. A federated learning feature screening system based on homomorphic encryption, characterized in that: include: An acquisition unit, used by the first participant to acquire the encrypted target values of all users of the second participant; A binning and statistics unit, used for the first participant to bin the users corresponding to each feature of the first participant according to the feature binning method, and to count the number of users in each bin of each feature; A summing and sending unit is used for the first participant to calculate the sum of the encrypted target values of all users in each subbox in each feature according to the encrypted target values of all users; send all the sum values corresponding to each feature to the second participant so that the second participant decrypts and counts the number of users whose target value is 1 in each subbox in each feature; A calculation unit, used by the first participant to obtain the number of users whose target value is 1 in each bin in each feature; The information value of each feature is calculated based on the number of users in each bin of each feature and the number of users with a target value of 1 in each bin of each feature; A screening unit, used by the first participant to screen out features that meet the requirements according to the information value of all features and preset feature evaluation criteria; The preset feature evaluation criteria are: If the information value of the current feature is less than the first preset threshold, the current feature is judged to be a worthless feature; If the information value of the current feature is greater than or equal to the first preset threshold and less than the second preset threshold, the current feature is judged to be a weak value feature; If the information value of the current feature is greater than or equal to the second preset threshold and less than the third preset threshold, the current feature is judged to be a medium value feature; If the information value of the current feature is greater than or equal to the third preset threshold and less than the fourth preset threshold, the current feature is judged to be a high-value feature; If the information value of the current feature is greater than or equal to the fourth preset threshold, the current feature is judged to be a false feature; The first participant is a characteristic variable provider; The second participant is a target variable provider.
6. The system according to claim 5, characterized in that The computing unit comprises: The quantity calculation subunit is used to obtain the number of users whose target value is 1 in each bin of each feature; and calculate the number of users whose target value is 0 in each bin of each feature according to the number of users in each bin of each feature and the number of users whose target value is 1 in each bin of each feature; The evidence weight calculation subunit is used to calculate the evidence weight of each bin in each feature according to the number of users with a target value of 0 in each bin in each feature and the number of users with a target value of 1 in each bin in each feature; An information value calculation subunit, used to calculate the information value of each bin in each feature according to the number of users with a target value of 0 in each bin in each feature, the number of users with a target value of 1 in each bin in each feature, and the evidence weight of each bin in each feature; The summation subunit is used to calculate the information value of each feature based on the information value of all bins in each feature.
7. The system according to claim 6, characterized in that: The evidence weight of each bin of the current feature is calculated according to the following formula: in, is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature; The information value of each bin of the current feature is calculated according to the following formula: in, The information value of the i-th bin of the current feature; is the number of users whose target value is 0 in the i-th bin of the current feature, is the number of users whose target value is 1 in the i-th bin of the current feature, n1 is the number of bins of the current feature, is the evidence weight of the i-th bin of the current feature.
8. The system according to claim 5, characterized in that: The feature binning methods include: chi-square binning and minimum entropy binning.
Citation Information
Patent Citations
Data processing method and device, equipment and storage medium
CN114547635A