Federal credit evaluation method and system

Abstract

The invention discloses a federal credit evaluation method and system. The method comprises the steps of joint feature binning, joint modeling, joint calculation of credit scores and rating according to the credit scores, according to the method, a bank cooperates with other banks, financial institutions or credit investigation institutions, credit information is shared, joint modeling is carried out to jointly calculate the credit scores, the performance on internal data is good, and the credit scores can be graded according to the credit scores. And the credit risk of customers of other banks, financial institutions or credit investigation institutions can be accurately predicted.

Description

Technical Field

[0001] The present invention relates to the field of enterprise credit technology, and in particular to a federal credit evaluation method and system. Background Art

[0002] As the market and economic environment changes, the original credit scoring model may no longer be applicable and needs to be regularly updated and maintained to ensure the accuracy and effectiveness of the model. The risk control model established independently by the bank may perform well on internal data, but it lacks generalization capabilities when facing external data and cannot accurately predict the credit risks of customers of other banks, financial institutions or credit reporting agencies.

[0003] To address the above problems, no effective solution has been proposed yet. Summary of the invention

[0004] A federal credit evaluation method and system are provided in the embodiments of the present invention to solve the problem that the risk control model independently established by the bank in the prior art may perform well on internal data, but has insufficient generalization ability when facing external data and cannot accurately predict the credit risk of customers of other banks, financial institutions or credit reporting agencies.

[0005] To achieve the above-mentioned purpose, on the one hand, the present invention provides a federal credit evaluation method, wherein the method comprises: S1, the initiator bins each initiation feature of all users, obtains the evidence weight of each bin in each initiation feature, and takes the evidence weight of all initiation features of all users as the initiator data set; the participant obtains the encrypted default value of all users of the initiator, and bins each participating feature of all users with the initiator according to the feature binning method, obtains the evidence weight of each bin in each participating feature, and takes the evidence weight of all participating features of all users as the participant data set; wherein all users of the initiator and the participant are the same; S2, according to the default value of all users of the initiator, the initiator data set and the initiator model The initial weights of the participant data set and the initial weights of the participant model are jointly used for model training to update the initial weights of the initiator model to obtain the final weights, and the initial weights of the participant model are updated to obtain the final weights; S3, the initiator sends the obtained initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participant; the initiator and the participant jointly calculate the final credit score of each user based on the initiator data set, the participant data set, the final weight of the initiator model, the final weight of the participant model, the initial credit score, ratio doubling change value, factor, bad sample probability and good sample probability ratio; S4, judge the credit level of the current user according to the final credit score of the current user and the preset credit score range.

[0006] Optionally, the S2 includes: S21, the initiator generates a public-private key pair, and sends the public key to the participant; S22, the initiator's data set is input into the initiator's model for training, and the participant's data set is input into the participant's model for training; the initiator obtains a first calculation value based on the initiator's data set and the initial weight of the initiator's model, and the participant obtains a second calculation value based on the participant's data set and the initial weight of the participant's model, and sends the second calculation value to the initiator; the initiator calculates the predicted values ​​of all users based on the first calculation value and the second calculation value; S23, the initiator calculates the residual based on the predicted values ​​and default values ​​of all users, and encrypts the residual based on the public key to obtain an encrypted residual; the initiator sends the encrypted residual to the participant; S24, the initiator calculates the gradient of the initiator's model based on the residual and the initiator's data set; the participant calculates the gradient based on the encrypted residual, the participant's data set, the random number randomly generated by the participant, and the public key The participant sends the gradient blinding result to the initiator so that the initiator decrypts the gradient blinding result according to the private key; the initiator sends the decrypted gradient blinding result to the participant so that the participant obtains the gradient of the participant model according to the decrypted gradient blinding result and the random number; S25, the initiator updates the initial weight of the initiator model according to the initial weight of the initiator model and the gradient of the initiator model, and the participant updates the initial weight of the participant model according to the initial weight of the participant model and the gradient of the participant model; S26, the initiator calculates the loss value according to the predicted value and the default value of all users, and determines whether the loss value is less than the preset loss value. If not, repeat S22 to S26. Otherwise, the initiator model and the participant model are no longer repeatedly trained, and the initial weight of the initiator model obtained by the latest update is used as the final weight of the initiator model, and the initial weight of the participant model obtained by the latest update is used as the final weight of the participant model.

[0007] Optionally, S3 includes: S31, the initiator sends the acquired initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participants; S32, the initiator and the participants calculate the final factor based on the ratio doubling change value and the initial factor; S33, the initiator calculates a third calculation value based on the final weight of the initiator's data set and the initiator's model, and the participants calculate a fourth calculation value based on the final weight of the participant's data set and the participant's model, and send the fourth calculation value to the initiator; the initiator calculates the credit normalization value of all users based on the third calculation value and the fourth calculation value; S34, the initiator calculates the final credit score of the current user based on the credit normalization value, initial credit score, bad sample probability and good sample probability ratio, and final factor of the current user.

[0008] Optionally, S34 includes: if the credit normalization value of the current user is greater than 0 and less than 1, then calculating a benchmark credit score based on the current user's initial credit score, the ratio of the bad sample probability to the good sample probability, and the final factor; calculating the current user's final credit score based on the current user's benchmark credit score, the final factor and the credit normalization value; if the current user's credit normalization value is less than 0 or greater than 1, then the current user's final credit score is -1, and S1 to S3 need to be repeated.

[0009] Optionally, the feature binning method includes: equal-interval binning, equal-frequency binning, and chi-square binning.

[0010] On the other hand, the present invention provides a federal credit evaluation system, which includes: a feature binning unit, which is used by the initiator to bin each initiation feature of all users, obtain the evidence weight of each bin in each initiation feature, and use the evidence weight of all initiation features of all users as the initiator data set; the participant obtains the encrypted default values ​​of all users of the initiator, and bins each participating feature of all users with the initiator according to the feature binning method, obtains the evidence weight of each bin in each participating feature, and uses the evidence weight of all participating features of all users as the participant data set; wherein all users of the initiator and the participant are the same; a model training unit, which is used to train the model according to the default values ​​of all users of the initiator, the initiator data set and the initial weight of the initiator model. The initial weights of the participant data set and the participant model are jointly used for model training to update the initial weights of the initiator model to obtain the final weights, and the initial weights of the participant model are updated to obtain the final weights; the score calculation unit is used for the initiator to send the obtained initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participant; the initiator and the participant jointly calculate the final credit score of each user based on the initiator data set, the participant data set, the final weight of the initiator model, the final weight of the participant model, the initial credit score, the ratio doubling change value, the factor, the bad sample probability and the good sample probability ratio; the credit rating unit is used to judge the credit level of the current user according to the final credit score of the current user and the preset credit score range.

[0011] Optionally, the model training unit includes: a sending subunit, which is used by the initiator to generate a public-private key pair and send the public key to the participating parties; a predicted value calculation subunit, which is used by the initiator to input the initiator's data set into the initiator's model for training, and the participant's data set into the participant's model for training; the initiator obtains a first calculated value based on the initiator's data set and the initial weight of the initiator's model, and the participant obtains a second calculated value based on the participant's data set and the initial weight of the participant's model, and sends the second calculated value to the initiator; the initiator calculates the predicted values ​​of all users based on the first calculated value and the second calculated value; a residual calculation subunit, which is used by the initiator to calculate the residual based on the predicted values ​​and default values ​​of all users, and encrypt the residual based on the public key to obtain an encrypted residual; the initiator sends the encrypted residual to the participating parties; a gradient calculation subunit, which is used by the initiator to calculate the gradient of the initiator's model based on the residual and the initiator's data set; the participating party obtains a gradient blinding result based on the encrypted residual, the participant's data set, a random number randomly generated by the participant and the public key; the participating party The blinded gradient result is sent to the initiator so that the initiator decrypts the blinded gradient result according to the private key; the initiator sends the decrypted blinded gradient result to the participant so that the participant obtains the gradient of the participant model according to the decrypted blinded gradient result and the random number; the weight updating subunit is used for the initiator to update the initial weight of the initiator model according to the initial weight of the initiator model and the gradient of the initiator model, and the participant updates the initial weight of the participant model according to the initial weight of the participant model and the gradient of the participant model; the repeated training subunit is used for the initiator to calculate the loss value according to the predicted values ​​and default values ​​of all users, and judge whether the loss value is less than the preset loss value. If not, repeat the predicted value calculation subunit, the residual calculation subunit, the gradient calculation subunit, the weight updating subunit, and the repeated training subunit. Otherwise, the initiator model and the participant model are no longer repeatedly trained, and the initial weight of the initiator model obtained by the latest update is used as the final weight of the initiator model, and the initial weight of the participant model obtained by the latest update is used as the final weight of the participant model.

[0012] Optionally, the score calculation unit includes: an acquisition subunit, which is used by the initiator to send the acquired initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participants; a final factor calculation subunit, which is used by the initiator and the participants to calculate the final factor based on the ratio doubling change value and the initial factor; a credit normalization value calculation subunit, which is used by the initiator to calculate the third calculation value based on the final weight of the initiator's data set and the initiator's model, and the participants to calculate the fourth calculation value based on the final weight of the participant's data set and the participant's model, and send the fourth calculation value to the initiator; the initiator calculates the credit normalization value of all users based on the third calculation value and the fourth calculation value; a final credit score calculation subunit, which is used by the initiator to calculate the final credit score of the current user based on the credit normalization value, initial credit score, bad sample probability and good sample probability ratio, and final factor of the current user.

[0013] Optionally, the final credit score calculation subunit is further used to: if the credit normalization value of the current user is greater than 0 and less than 1, calculate a baseline credit score based on the current user's initial credit score, the ratio of the bad sample probability to the good sample probability, and the final factor; calculate the current user's final credit score based on the current user's baseline credit score, the final factor and the credit normalization value; if the credit normalization value of the current user is less than 0 or greater than 1, the current user's final credit score is -1, and the feature binning unit, the model training unit, and the score calculation unit need to be repeated.

[0014] Optionally, the feature binning method includes: equal-interval binning, equal-frequency binning, and chi-square binning.

[0015] Beneficial effects of the present invention:

[0016] The present invention provides a federal credit evaluation method and system, wherein the method includes joint feature binning, joint modeling, and joint calculation of credit scores, and rating is performed based on the credit scores. The method of the present invention allows a bank to cooperate with other banks, financial institutions, or credit reporting agency users to share credit information and perform joint modeling to jointly calculate credit scores. It performs well on both internal and external data, and can accurately predict the credit risk issues of other banks, financial institutions, or credit reporting agency customers. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of a federal credit evaluation method provided by an embodiment of the present invention;

[0018] Figure 2 is a flowchart of joint modeling provided by an embodiment of the present invention;

[0019] Figure 3is a flow chart of jointly calculating a final credit score provided by an embodiment of the present invention;

[0020] Figure 4 is a structural diagram of a federal credit evaluation system provided by an embodiment of the present invention;

[0021] Figure 5 is a schematic diagram of the structure of a model training unit provided in an embodiment of the present invention;

[0022] Figure 6 It is a schematic diagram of the structure of a score calculation unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] 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.

[0024] The risk control model established independently by a bank may perform well on internal data, but it lacks generalization capabilities when faced with external data and cannot accurately predict the credit risks of customers of other banks, financial institutions or credit reporting agencies.

[0025] In order to solve the above problems, the present invention provides a federal credit evaluation method and system. Figure 1 is a flow chart of a federal credit evaluation method provided by an embodiment of the present invention. Figure 1 As shown, the method includes:

[0026] S1. The initiator bins each initiation feature of all users, obtains the evidence weight of each bin in each initiation feature, and uses the evidence weight of all initiation features of all users as the initiator data set; the participants obtain the encrypted default values ​​of all users of the initiator, and bin each participation feature of all users with the initiator according to the feature binning method, obtain the evidence weight of each bin in each participation feature, and use the evidence weight of all participation features of all users as the participant data set; wherein, all users of the initiator and the participants are the same;

[0027] Assume that the bank is the initiator P0, the credit reporting agency is the participant P1, and the IDs of all users in P0 and P1 are their ID numbers;

[0028] All the initiation features of the initiator P0 include: Month: represents the month of the year, Name: represents a person's name, Age: represents the person's age, SSN: represents a person's social security number, Occupation: represents the person's occupation, Annual_Income: represents the person's annual income, Monthly_Inhand_Salary: represents a person's monthly basic salary, Num_Bank_Accounts: represents the number of bank accounts a person has, Num_Credit_Card: represents the number of credit cards, Interest_Rate: represents the interest rate of the credit card, Num_of_Loan: represents the number of loans from the bank, Type_of_Loan: represents a person's loan type, Delay_from_due_date: represents the average number of days delayed from the payment date, Num_of_Delayed_Payment: represents the average number of delayed payments by a person, Changed_Cr edit_Limit: represents the percentage change in the credit card limit, Num_Credit_Inquiries: represents the number of credit card inquiries, Credit_Mix: represents the classification of the credit portfolio, Outstanding_Debt: represents the remaining debt to be paid (in US dollars), Credit_Utilization_Ratio: represents the utilization ratio of credit cards, Credit_History_Age: represents the age of the individual's credit history, Payment_of_Min_Amount: represents whether the person has only paid the minimum amount, Total_EMI_per_month: represents the monthly EMI payment (in US dollars), Amount_invested_monthly: represents the amount invested by the customer every month (in US dollars), Payment_Behaviour: represents the payment behavior of the customer (in US dollars), Monthly_Balance: represents the monthly balance amount of the customer (in US dollars).

[0029] The initiator P0 also includes the default value y of all users, where y is 1, indicating default, and y is 0, indicating no default.

[0030] All participating characteristics of Party 1 include: number of loans / credit cards / quasi-credit cards, month of first issuance, summary of overdue information, summary of outstanding loans, summary of credit cards / quasi-credit cards that have not been closed, summary of external guarantee information, repayment record of each loan / credit card / quasi-credit card of the borrower within 2 years and overdue record within 5 years, account status, five-level classification, balance and remaining repayment period, actual repayment due this month, current overdue period and amount, outstanding principal at different overdue stages, borrower's tax arrears record within 5 years, civil judgment record, compulsory execution record, administrative penalty record and telecommunications arrears record.

[0031] The initiator P0 and the participant P1 each collect data corresponding to local features, perform cleaning, outlier detection processing, and standardization. The initiator P0 and the participant P1 align samples based on ID to obtain samples shared by both parties (that is, all users of both parties are the same). Among them, the total number of positive samples is Good, and the total number of negative samples is Bad;

[0032] The initiator divides each initiation feature of all users into boxes and obtains the evidence weight of each box in each initiation feature.

[0033] Take the current launch characteristics of all users in the initiator as an example:

[0034] Assume that there are 100 users in total, which are divided into 10 boxes, and the number of users in each box is 11, 10, 8, 7, 5, 15, 8, 13, 12, and 11 respectively. The number of defaulting users in the first box is a, and the number of non-defaulting users is b. Then the evidence weight of the first box is: woe = ln((a / Bad) / (b / Good)), and the information value of the first box is: IV = ((a / Bad)-(b / Good))*woe. The information value of all boxes is summed to obtain the information value of the current initiating feature. If the information value of the current initiating feature is less than the preset threshold (set to 0.02 in the present invention), the current initiating feature is re-boxed until the information value of the current initiating feature is greater than or equal to the preset threshold. It should be noted that the users in each box can be continuous users or discontinuous users.

[0035] If the current user ID has 30 initiating features, then the current user's woe values ​​are 30, which are the woe values ​​corresponding to the bins in each initiating feature. The woe values ​​under all initiating features of all users are taken as the initiator data set woe0.

[0036] The participating party obtains the encrypted default values ​​of all users of the initiator, and jointly with the initiator bins each participating feature of all users according to the feature binning method to obtain the evidence weight of each bin in each participating feature.

[0037] Among them, the feature binning methods include: equal-distance binning, equal-frequency binning, and chi-square binning.

[0038] The following example uses equal-frequency binning to bin the current participation characteristics of all users in the participant:

[0039] The initiator P0 performs homomorphic encryption on y and 1-y of all users to obtain E(y) and E(1-y). The initiator P0 sends E(y) and E(1-y) to the participant P1. The participant P1 receives E(y) and E(1-y). The participant P1 performs equal-frequency binning on the current participating feature to obtain the binning result, that is, the binning range of the current participating feature and the encrypted E(y) and E(1-y) corresponding to each bin. The participant P1 calculates the ciphertext and sum(E(y)) and sum(E(1-y)) of E(y) and E(1-y) in each bin, and sends each bin number, the corresponding sum(E(y)) and sum(E(1-y)) to the initiator P0. The initiator P0 receives the bin number of the current participating feature of the participant P1, the corresponding sum(E(y)) and sum(E(1-y)) ), decrypt sum(E(y)) and sum(E(1-y)) to obtain the total number of positive samples sum(y) and the total number of negative samples sum(1-y) in each bin. The initiator P0 calculates the woe value and IV value in each bin of the current participating feature of the participant P1. Taking the woe value and IV value in the first bin of the current participating feature as an example: woe=ln(sum(1-y) / Bad / (sum(y) / Good)), IV=(sum(1-y) / Bad-sum(y) / Good)*woe. The information value of all bins is summed to obtain the information value of the current participating feature. If the information value of the current participating feature is less than the preset threshold (set to 0.02 in the present invention), the current participating feature is re-binned until the information value of the current participating feature is greater than or equal to the preset threshold.

[0040] If the current user ID has 20 participating features, then the current user's woe values ​​are 20, which are the woe values ​​corresponding to the bins in each participating feature. The woe values ​​of all participating features of all users are taken as the participant data set woe1.

[0041] During the feature binning process, there is no plaintext data transmission, the participants do not know the initiator's y value, and the initiator does not know the value of each participating feature of the participants, thereby completing the feature binning calculation and woe conversion under the condition of security and privacy protection.

[0042] S2. Perform model training based on the default values ​​of all users of the initiator, the initiator data set and the initial weight of the initiator model, and the participant data set and the initial weight of the participant model, so as to update the initial weight of the initiator model to obtain the final weight, and update the initial weight of the participant model to obtain the final weight;

[0043] Figure 2 is a flowchart of joint modeling provided by an embodiment of the present invention, such as Figure 2 As shown, the S2 includes:

[0044] S21. The initiator generates a public-private key pair and sends the public key to the participants;

[0045] The initiator P0 generates a public-private key pair (sk, pk), and sends the public key to the participant P1.

[0046] S22, the initiator data set is input into the initiator model for training, and the participant data set is input into the participant model for training; the initiator obtains a first calculation value based on the initiator data set and the initial weight of the initiator model, and the participant obtains a second calculation value based on the participant data set and the initial weight of the participant model, and sends the second calculation value to the initiator; the initiator obtains the predicted values ​​of all users based on the first calculation value and the second calculation value;

[0047] Specifically, the initiator P0 calculates the first calculated value = the initiator data set * the initial weight of the initiator model = woe0*w0; the participant P1 calculates the second calculated value = the participant data set * the initial weight of the participant model = woe1*w1, and sends it to the initiator P0. The initiator P0 calculates the predicted value y′=1 / (1+exp(-(h))) of all users, h=woe0*w0+woe1*w1.

[0048] S23, the initiator calculates the residual according to the predicted values ​​and default values ​​of all users, and encrypts the residual according to the public key to obtain the encrypted residual; the initiator sends the encrypted residual to the participating parties;

[0049] Specifically, the initiator P0 calculates the residual d=y′-y, encrypts it with the public key pk to obtain the encrypted residual E(d), and sends it to the participant P1.

[0050] S24, the initiator calculates the gradient of the initiator's model according to the residual and the initiator's data set; the participant calculates the gradient blinding result according to the encrypted residual, the participant's data set, the random number randomly generated by the participant and the public key; the participant sends the gradient blinding result to the initiator, so that the initiator decrypts the gradient blinding result according to the private key; the initiator sends the decrypted gradient blinding result to the participant, so that the participant obtains the gradient of the participant's model according to the decrypted gradient blinding result and the random number;

[0051] Specifically, the initiator P0 calculates the gradient g0=d*woe0 of the initiator model; the participant P1 calculates the ciphertext gradient E(g01)=E(d)*woe1 of the participant model, and the participant P1 generates a random number R. The participant P1 obtains the gradient blinding result E(g01+R)=E(d*woe1+R) based on the ciphertext gradient, random number and public key, and the participant P1 sends the gradient blinding result E(g01+R) to the initiator P0. The initiator P0 uses the private key sk to decrypt E(g01+R) and obtain d*woe1+R. The initiator P0 sends d*woe1+R to the participant P1, and the participant P1 removes the random number R to obtain the gradient g1=d*woe1 of the participant model.

[0052] S25, the initiator updates the initial weight of the initiator model according to the initial weight of the initiator model and the gradient of the initiator model, and the participant updates the initial weight of the participant model according to the initial weight of the participant model and the gradient of the participant model;

[0053] Specifically, the initial weight of the initiator model is updated to = w0–alfa*g0, and the initial weight of the participant model is updated to = w1–alfa*g1, where alfa is the learning rate, which is generally set in the range of (0,1).

[0054] S26. The initiator calculates the loss value according to the predicted values ​​and default values ​​of all users, and determines whether the loss value is less than the preset loss value. If not, repeat S22 to S26. Otherwise, the initiator model and the participant model are not repeatedly trained, and the initial weight of the initiator model obtained by the latest update is used as the final weight of the initiator model, and the initial weight of the participant model obtained by the latest update is used as the final weight of the participant model.

[0055] Specifically, the initiator P0 calculates the loss value Loss = log2-1 / 2*y*y′+1 / 8*y′^2; the initiator P0 determines whether the loss value is less than the preset loss value. If so, the initiator model and the participant model converge, and the initial weight of the initiator model finally updated is used as the final weight w00 of the initiator model, and the initial weight of the participant model finally updated is used as the final weight w11 of the participant model. Otherwise, repeat S22 to S26.

[0056] S3. The initiator sends the obtained initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participant; the initiator and the participant jointly calculate the final credit score of each user based on the initiator's data set, the participant's data set, the final weight of the initiator's model, the final weight of the participant's model, the initial credit score, ratio doubling change value, factor, bad sample probability and good sample probability ratio;

[0057] Figure 3 is a flowchart of jointly calculating the final credit score provided by an embodiment of the present invention, such as Figure 3 As shown, the S3 includes:

[0058] S31. The initiator sends the obtained initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participating parties;

[0059] Specifically, the initiator P0 directly obtains the parameters input on the page, namely, the initial credit score init_score, the ratio doubling change value pdo, the initial factor fact_base, the bad sample probability and the good sample probability ratio odds, and sends these parameters to the participant P1.

[0060] S32. The initiator and the participant calculate the final factor based on the ratio doubling change value and the initial factor;

[0061] Specifically, the initiator P0 and the participant P1 both calculate the final factor factor=pdo / math.log10(fact_base).

[0062] S33: The initiator obtains a third calculated value based on the final weight of the initiator's data set and the initiator's model, and the participant obtains a fourth calculated value based on the final weight of the participant's data set and the participant's model, and sends the fourth calculated value to the initiator; the initiator obtains the credit normalization value of all users based on the third calculated value and the fourth calculated value;

[0063] Specifically, the initiator P0 calculates the third calculated value y′0=w00*woe0; the participant P1 calculates the fourth calculated value y′1=w11*woe1 and sends it to the initiator P0. The initiator P0 calculates the credit normalization value S=1 / (1+exp(-(y′0+y′1))) of all users.

[0064] S34. The initiator calculates the final credit score of the current user based on the normalized credit value of the current user, the initial credit score, the ratio of the bad sample probability to the good sample probability, and the final factor.

[0065] Specifically, the S34 includes:

[0066] If the normalized credit value of the current user is greater than 0 and less than 1, the base credit score is calculated based on the initial credit score of the current user, the ratio of the bad sample probability to the good sample probability, and the final factor; the final credit score of the current user is calculated based on the base credit score of the current user, the final factor, and the normalized credit value;

[0067] Specifically, the base credit score of the current user: base_score=init_score-factor*math.log10(odds); the final credit score of the current user: scores=base_score+factor*S.

[0068] If the normalized credit value of the current user is less than 0 or greater than 1, the final credit score of the current user is -1, and S1 to S3 need to be repeated.

[0069] If the normalized credit value of the current user is close to 0, the final credit score of the current user is high and the credit rating of the current user is excellent. If the normalized credit value of the current user is close to 1, the final credit score of the current user is low.

[0070] S4. Determine the credit rating of the current user based on the final credit score of the current user and the preset credit score range.

[0071] In a preferred embodiment, the credit rating can be divided into four levels: excellent, good, qualified, and poor; each level has a corresponding preset credit score range; the credit rating of the current user can be determined based on the current user's final credit score and the preset credit score range.

[0072] It should be noted that the present invention does not limit the preset credit score range and credit level.

[0073] Figure 4 is a schematic diagram of the structure of a federal credit evaluation system provided by an embodiment of the present invention. Figure 4 As shown, the system includes:

[0074] The feature binning unit 201 is used for the initiator to bin each initiation feature of all users, obtain the evidence weight of each bin in each initiation feature, and use the evidence weight of all initiation features of all users as the initiator data set; the participant obtains the encrypted default value of all users of the initiator, and bins each participating feature of all users with the initiator according to the feature binning method, obtains the evidence weight of each bin in each participating feature, and uses the evidence weight of all participating features of all users as the participant data set; wherein all users of the initiator and the participant are the same;

[0075] The feature binning methods include: equal-distance binning, equal-frequency binning, and chi-square binning.

[0076] The model training unit 202 is used to jointly perform model training based on the default values ​​of all users of the initiator, the initiator data set and the initial weight of the initiator model, the participant data set and the initial weight of the participant model, so as to update the initial weight of the initiator model to obtain the final weight, and update the initial weight of the participant model to obtain the final weight;

[0077] Figure 5 is a schematic diagram of the structure of the model training unit provided in an embodiment of the present invention, such as Figure 5 As shown, the model training unit 202 includes:

[0078] The sending subunit 2021 is used for the initiator to generate a public-private key pair and send the public key to the participating parties;

[0079] The prediction value calculation subunit 2022 is used for inputting the initiator data set into the initiator model for training, and the participant data set into the participant model for training; the initiator obtains a first calculation value according to the initial weight calculation of the initiator data set and the initiator model, and the participant obtains a second calculation value according to the initial weight calculation of the participant data set and the participant model, and sends the second calculation value to the initiator; the initiator obtains the prediction value of all users according to the first calculation value and the second calculation value;

[0080] The residual calculation subunit 2023 is used for the initiator to calculate the residual according to the predicted values ​​and default values ​​of all users, and encrypt the residual according to the public key to obtain the encrypted residual; the initiator sends the encrypted residual to the participants;

[0081] The gradient calculation subunit 2024 is used for the initiator to calculate the gradient of the initiator's model according to the residual and the initiator's data set; the participant calculates the gradient blinding result according to the encrypted residual, the participant's data set, the random number randomly generated by the participant and the public key; the participant sends the gradient blinding result to the initiator so that the initiator decrypts the gradient blinding result according to the private key; the initiator sends the decrypted gradient blinding result to the participant so that the participant obtains the gradient of the participant's model according to the decrypted gradient blinding result and the random number;

[0082] The weight updating subunit 2025 is used for the initiator to update the initial weight of the initiator model according to the initial weight of the initiator model and the gradient of the initiator model, and the participant to update the initial weight of the participant model according to the initial weight of the participant model and the gradient of the participant model;

[0083] The repeated training subunit 2026 is used for the initiator to calculate the loss value according to the predicted values ​​and default values ​​of all users, and determine whether the loss value is less than the preset loss value. If not, the predicted value calculation subunit 2022, the residual calculation subunit 2023, the gradient calculation subunit 2024, the weight update subunit 2025 and the repeated training subunit 2026 are repeated. Otherwise, the initiator model and the participant model are no longer repeatedly trained, and the initial weight of the initiator model obtained by the last update is used as the final weight of the initiator model, and the initial weight of the participant model obtained by the last update is used as the final weight of the participant model.

[0084] The score calculation unit 203 is used for the initiator to send the acquired initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participant; the initiator and the participant jointly calculate the final credit score of each user based on the initiator data set, the participant data set, the final weight of the initiator model, the final weight of the participant model, the initial credit score, ratio doubling change value, factor, bad sample probability and good sample probability ratio;

[0085] Figure 6 is a schematic diagram of the structure of a score calculation unit provided in an embodiment of the present invention; Figure 6 As shown, the score calculation unit 203 includes:

[0086] The acquisition subunit 2031 is used for the initiator to send the acquired initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participating party;

[0087] The final factor calculation subunit 2032 is used for the initiator and the participant to calculate the final factor according to the ratio doubling change value and the initial factor;

[0088] The credit normalization value calculation subunit 2033 is used for the initiator to calculate the third calculation value according to the final weight of the initiator data set and the initiator model, and the participant to calculate the fourth calculation value according to the final weight of the participant data set and the participant model, and send the fourth calculation value to the initiator; the initiator calculates the credit normalization value of all users according to the third calculation value and the fourth calculation value;

[0089] The final credit score calculation subunit 2034 is used by the initiator to calculate the final credit score of the current user according to the normalized credit value of the current user, the initial credit score, the ratio of the bad sample probability to the good sample probability, and the final factor.

[0090] The final credit score calculation subunit 2034 is further used to:

[0091] If the normalized credit value of the current user is greater than 0 and less than 1, the base credit score is calculated based on the initial credit score of the current user, the ratio of the bad sample probability to the good sample probability, and the final factor; the final credit score of the current user is calculated based on the base credit score of the current user, the final factor, and the normalized credit value;

[0092] If the normalized credit value of the current user is less than 0 or greater than 1, the final credit score of the current user is -1, and the feature binning unit, model training unit, and score calculation unit need to be repeated.

[0093] The credit rating unit 204 is used to determine the credit level of the current user according to the final credit score of the current user and a preset credit score range.

[0094] The system of the present invention corresponds to the above-mentioned method, and the specific embodiments of the system will not be repeated here.

[0095] Beneficial effects of the present invention:

[0096] The present invention provides a federal credit evaluation method and system, wherein the method includes joint feature binning, joint modeling, and joint calculation of credit scores, and rating is performed based on the credit scores. The method of the present invention allows a bank to cooperate with other banks, financial institutions, or credit reporting agency users to share credit information and perform joint modeling to jointly calculate credit scores. It performs well on both internal and external data, and can accurately predict the credit risk issues of other banks, financial institutions, or credit reporting agency customers.

[0097] 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 federal credit evaluation method, characterized in that: include: S1. The initiator bins each initiation feature of all users, obtains the evidence weight of each bin in each initiation feature, and uses the evidence weight of all initiation features of all users as the initiator data set; the participants obtain the encrypted default values ​​of all users of the initiator, and bin each participation feature of all users with the initiator according to the feature binning method, obtain the evidence weight of each bin in each participation feature, and use the evidence weight of all participation features of all users as the participant data set; wherein, all users of the initiator and the participants are the same; S2. Perform model training based on the default values ​​of all users of the initiator, the initiator data set and the initial weight of the initiator model, and the participant data set and the initial weight of the participant model, so as to update the initial weight of the initiator model to obtain the final weight, and update the initial weight of the participant model to obtain the final weight; S3. The initiator sends the obtained initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participant; the initiator and the participant jointly calculate the final credit score of each user based on the initiator's data set, the participant's data set, the final weight of the initiator's model, the final weight of the participant's model, the initial credit score, ratio doubling change value, factor, bad sample probability and good sample probability ratio; S4. Determine the credit rating of the current user based on the final credit score of the current user and the preset credit score range.

2. The method according to claim 1, characterized in that The S2 includes: S21. The initiator generates a public-private key pair and sends the public key to the participants; S22, the initiator data set is input into the initiator model for training, and the participant data set is input into the participant model for training; the initiator obtains a first calculation value based on the initiator data set and the initial weight of the initiator model, and the participant obtains a second calculation value based on the participant data set and the initial weight of the participant model, and sends the second calculation value to the initiator; the initiator obtains the predicted values ​​of all users based on the first calculation value and the second calculation value; S23, the initiator calculates the residual according to the predicted values ​​and default values ​​of all users, and encrypts the residual according to the public key to obtain the encrypted residual; the initiator sends the encrypted residual to the participating parties; S24, the initiator calculates the gradient of the initiator's model according to the residual and the initiator's data set; the participant calculates the gradient blinding result according to the encrypted residual, the participant's data set, the random number randomly generated by the participant and the public key; the participant sends the gradient blinding result to the initiator, so that the initiator decrypts the gradient blinding result according to the private key; the initiator sends the decrypted gradient blinding result to the participant, so that the participant obtains the gradient of the participant's model according to the decrypted gradient blinding result and the random number; S25, the initiator updates the initial weight of the initiator model according to the initial weight of the initiator model and the gradient of the initiator model, and the participant updates the initial weight of the participant model according to the initial weight of the participant model and the gradient of the participant model; S26. The initiator calculates the loss value according to the predicted values ​​and default values ​​of all users, and determines whether the loss value is less than the preset loss value. If not, repeat S22 to S26. Otherwise, the initiator model and the participant model are not repeatedly trained, and the initial weight of the initiator model obtained by the latest update is used as the final weight of the initiator model, and the initial weight of the participant model obtained by the latest update is used as the final weight of the participant model.

3. The method according to claim 1, characterized in that The S3 includes: S31. The initiator sends the obtained initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participating parties; S32. The initiator and the participant calculate the final factor based on the ratio doubling change value and the initial factor; S33: The initiator obtains a third calculated value based on the final weight of the initiator's data set and the initiator's model, and the participant obtains a fourth calculated value based on the final weight of the participant's data set and the participant's model, and sends the fourth calculated value to the initiator; the initiator obtains the credit normalization value of all users based on the third calculated value and the fourth calculated value; S34. The initiator calculates the final credit score of the current user based on the normalized credit value of the current user, the initial credit score, the ratio of the bad sample probability to the good sample probability, and the final factor.

4. The method according to claim 3, characterized in that The S34 includes: If the normalized credit value of the current user is greater than 0 and less than 1, the base credit score is calculated based on the initial credit score of the current user, the ratio of the bad sample probability to the good sample probability, and the final factor; the final credit score of the current user is calculated based on the base credit score of the current user, the final factor, and the normalized credit value; If the normalized credit value of the current user is less than 0 or greater than 1, the final credit score of the current user is -1, and S1 to S3 need to be repeated.

5. The method according to claim 1, characterized in that: The feature binning methods include: equal-distance binning, equal-frequency binning, and chi-square binning.

6. A federal credit rating system, characterized in that: include: The feature binning unit is used for the initiator to bin each initiation feature of all users, obtain the evidence weight of each bin in each initiation feature, and use the evidence weight of all initiation features of all users as the initiator data set; the participant obtains the encrypted default value of all users of the initiator, and bins each participating feature of all users with the initiator according to the feature binning method, obtains the evidence weight of each bin in each participating feature, and uses the evidence weight of all participating features of all users as the participant data set; wherein, all users of the initiator and the participant are the same; A model training unit, used for jointly performing model training according to the default values ​​of all users of the initiator, the initiator data set and the initial weight of the initiator model, the participant data set and the initial weight of the participant model, so as to update the initial weight of the initiator model to obtain the final weight, and update the initial weight of the participant model to obtain the final weight; The score calculation unit is used for the initiator to send the obtained initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participant; the initiator and the participant jointly calculate the final credit score of each user based on the initiator's data set, the participant's data set, the final weight of the initiator's model, the final weight of the participant's model, the initial credit score, ratio doubling change value, factor, bad sample probability and good sample probability ratio; The credit rating unit is used to determine the credit level of the current user based on the final credit score of the current user and the preset credit score range.

7. The system according to claim 6, characterized in that The model training unit comprises: The sending subunit is used for the initiator to generate a public-private key pair and send the public key to the participating parties; The prediction value calculation subunit is used for inputting the initiator data set into the initiator model for training, and the participant data set into the participant model for training; the initiator obtains a first calculation value according to the initial weight of the initiator data set and the initiator model, and the participant obtains a second calculation value according to the initial weight of the participant data set and the participant model, and sends the second calculation value to the initiator; the initiator obtains the prediction value of all users according to the first calculation value and the second calculation value; The residual calculation subunit is used for the initiator to calculate the residual according to the predicted values ​​and default values ​​of all users, and encrypt the residual according to the public key to obtain the encrypted residual; the initiator sends the encrypted residual to the participating parties; The gradient calculation subunit is used for the initiator to calculate the gradient of the initiator's model according to the residual and the initiator's data set; the participant calculates the gradient blinding result according to the encrypted residual, the participant's data set, the random number randomly generated by the participant and the public key; the participant sends the gradient blinding result to the initiator so that the initiator decrypts the gradient blinding result according to the private key; the initiator sends the decrypted gradient blinding result to the participant so that the participant obtains the gradient of the participant's model according to the decrypted gradient blinding result and the random number; A weight updating subunit, used for the initiator to update the initial weight of the initiator model according to the initial weight of the initiator model and the gradient of the initiator model, and for the participant to update the initial weight of the participant model according to the initial weight of the participant model and the gradient of the participant model; The repeated training subunit is used for the initiator to calculate the loss value according to the predicted values ​​and default values ​​of all users, and determine whether the loss value is less than the preset loss value. If not, the predicted value calculation subunit, the residual calculation subunit, the gradient calculation subunit, the weight update subunit and the repeated training subunit are repeated. Otherwise, the initiator model and the participant model are no longer repeatedly trained, and the initial weight of the initiator model obtained by the last update is used as the final weight of the initiator model, and the initial weight of the participant model obtained by the last update is used as the final weight of the participant model.

8. The system according to claim 6, characterized in that The score calculation unit comprises: The acquisition subunit is used for the initiator to send the acquired initial credit score, ratio doubling change value, initial factor, bad sample probability and good sample probability ratio to the participating party; The final factor calculation subunit is used by the initiator and the participant to calculate the final factor based on the ratio doubling change value and the initial factor; The credit normalization value calculation subunit is used for the initiator to calculate the third calculation value according to the final weight of the initiator data set and the initiator model, and the participant to calculate the fourth calculation value according to the final weight of the participant data set and the participant model, and send the fourth calculation value to the initiator; the initiator calculates the credit normalization value of all users according to the third calculation value and the fourth calculation value; The final credit score calculation subunit is used by the initiator to calculate the final credit score of the current user based on the normalized credit value of the current user, the initial credit score, the ratio of the bad sample probability to the good sample probability, and the final factor.

9. The system according to claim 8, characterized in that The final credit score calculation subunit is further used for: If the normalized credit value of the current user is greater than 0 and less than 1, the benchmark credit score is calculated based on the current user's initial credit score, the ratio of the bad sample probability to the good sample probability, and the final factor; The final credit score of the current user is calculated based on the current user's baseline credit score, final factor and credit normalization value; If the normalized credit value of the current user is less than 0 or greater than 1, the final credit score of the current user is -1, and the feature binning unit, model training unit, and score calculation unit need to be repeated.

10. The system according to claim 6, characterized in that: The feature binning methods include: equal-distance binning, equal-frequency binning, and chi-square binning.

Images