Credit risk control prediction model training method, credit risk control prediction method and system

Abstract

The application provides a credit risk control prediction model training method, a credit risk control prediction method and a system, and relates to the technical field of artificial intelligence. The training method comprises the following steps: inputting the data in the respective training sets of the two credit institutions into respective local learning models, obtaining the hidden features of the data in the respective training sets, and using the hidden features to complete the data in the training sets; inputting the data into the local prediction model; comparing the model output result with the corresponding risk label to obtain a loss value and an alignment loss value; sending the loss value of one party to the other party to obtain a reconstruction loss value and a comprehensive gradient, which is equivalent to using the user data of the self and the user data of the other party to train the local prediction model, thereby avoiding the data island problem; updating the parameters of the local prediction model according to the comprehensive gradient until the loss value and the reconstruction loss value of each party converge to the corresponding preset accuracy, and stopping the training. The local prediction model is used as a credit risk control prediction model to ensure the safety of loan funds.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for training a credit risk control prediction model, a credit risk control prediction method, and a system. Background Technology

[0002] In credit business, to ensure the safety of loan funds, it is necessary to assess and predict the borrower's credit risk before granting credit to determine whether the lending activity can continue.

[0003] In general, to ensure user privacy and data security, credit institutions can only use private data to build risk control models. However, data is limited and there are data silos. The risk control models built may not be able to fully depict the financial risk profile of customers, thus affecting the corresponding decisions.

[0004] In existing methods, multiple credit institutions jointly select a trusted third party to achieve data cooperation, transmit private data to the third party's server, use the shared data to jointly train a risk control model and implement it, and use the above model and user data that want to apply for credit to make predictions.

[0005] However, the above training assumes that the user data uploaded by multiple credit institutions has the same format. In reality, the data sources of various institutions are quite complex, and there is no unified standard for the data format. Therefore, the above method is difficult to apply in actual use. Summary of the Invention

[0006] The problem this invention aims to solve is that user data held by multiple credit institutions are in different formats, making it difficult to use them together and resulting in data silos.

[0007] To address the aforementioned problems, firstly, this invention provides a method for training a credit risk control prediction model.

[0008] The data from the training sets of both credit institutions are input into their respective local learning models to obtain the hidden features of the data in their respective training sets;

[0009] Based on the hidden features, the data in the training set is completed to make the format of the data in each training set consistent;

[0010] Each of the data in the completed training set is input into its respective local prediction model. The output of the local prediction model is compared with the risk label of each data point to obtain the respective loss value and the common aligned loss value, wherein the risk label represents the user's credit risk level.

[0011] The loss value of one credit institution is sent to the other credit institution, and combined with the aligned loss value, the respective reconstruction loss value and the respective comprehensive gradient are obtained;

[0012] Based on the comprehensive gradient, the parameters of the local prediction model are updated until the reconstruction loss value converges to a preset accuracy. Then, the training of the local prediction model is stopped, and the trained local prediction model is used as the credit risk control prediction model corresponding to the credit institution.

[0013] Optionally, the step of inputting the data from each of the completed training sets into their respective local prediction models, and comparing the output of the local prediction models with the risk labels of each data point to obtain their respective loss values ​​and a common alignment loss value includes:

[0014] Each of the supplemented data is input into its respective local prediction model. Based on the output of the local prediction model and the risk label corresponding to the supplemented data, the loss value of each credit institution is obtained.

[0015] The data containing the hidden features of both parties is input into the local prediction model. Based on the output of the local prediction model and the risk label corresponding to the data containing the hidden features of both parties, a common alignment loss value is obtained.

[0016] Optionally, the step of sending the loss value of one credit institution to the other credit institution and combining it with the aligned loss value to obtain their respective reconstructed loss values ​​and their respective comprehensive gradients includes:

[0017] The loss value of one credit institution is sent to the other credit institution, and the loss values ​​of both parties and the alignment loss value are added together to obtain the reconstruction loss value;

[0018] Based on the reconstruction loss value, determine the aligned gradient and the independent gradient;

[0019] The individual gradient is added to the aligned gradient to obtain the combined gradient.

[0020] Optionally, the step of sending the loss value of one credit institution to the other credit institution, and adding the loss values ​​of both parties and the aligned loss value to obtain the reconstructed loss value includes:

[0021] The loss value of one credit institution is sent to the other credit institution. The loss values ​​of both parties are added together, and then the product of the aligned loss value and the preset weight is added to obtain the reconstructed loss value.

[0022] Optionally, updating the parameters of the local prediction model based on the comprehensive gradient until the reconstructed loss value converges to a preset accuracy, stopping the training of the local prediction model, and using the trained local prediction model as the credit risk control prediction model corresponding to the credit institution includes:

[0023] Based on the comprehensive gradient, the parameters of the local prediction model are updated until their respective loss values, the common alignment loss value, and their respective reconstruction loss values ​​all converge to the corresponding preset accuracy, at which point the training of the local prediction model is stopped.

[0024] Optionally, before inputting the data from the respective training sets of both credit institutions into their respective local learning models to obtain the hidden features of the data in their respective training sets, the method further includes:

[0025] Preprocess user data, add the risk label to the credit risk of each user, and construct sample data;

[0026] The sample data is differentially processed to obtain the training set.

[0027] Optionally, the preprocessing of user data, adding the risk label to the credit risk of each user, and constructing sample data includes:

[0028] Extract the data features from the user data, and add the risk label to the credit risk of each user based on the credit records in the user data to construct the sample data. The data features include user identity information, income level, assets and liabilities, tax payment status, business abnormalities and historical credit. The risk label of each user corresponds to the user data of each user.

[0029] Examine the data type in the sample data, wherein the data type includes string type and numeric type;

[0030] When the data in the sample data is of the string type, the string type data is converted into a floating-point number type, and any missing data after conversion is filled in;

[0031] When the data in the sample data is of the numerical type, the data of the numerical type is standardized.

[0032] Optionally, the step of performing differential processing on the sample data to obtain the training set includes:

[0033] When the data column corresponding to the data feature in the sample data is a continuous data column, independent random noise following a Gaussian mechanism is added to each data item in the data column, wherein the data corresponding to each data feature in the sample data constitutes a data column;

[0034] When the data column corresponding to the data feature in the sample data is a discrete data column, an exponential mechanism is used to perform difference processing on each data item in the data column.

[0035] The data in the processed data column is used to replace the original data in the sample data to obtain the training set.

[0036] Secondly, the present invention also provides a credit risk control prediction method, comprising:

[0037] The user data of users who apply for loans is input into a local learning model to obtain the hidden features of the user data;

[0038] Based on the user data, the hidden features, the local learning model, and the credit risk control prediction model training method described above, the predicted risk label corresponding to the user data is obtained.

[0039] Assess the user's credit risk based on the predicted risk label.

[0040] Thirdly, the present invention also provides a credit risk control prediction system, comprising:

[0041] The learning module is used to input user data of users applying for loans into a local learning model to obtain the hidden features of the user data;

[0042] The prediction module is used to obtain the predicted risk label corresponding to the user data based on the user data, the hidden features, the local learning model, and the credit risk control prediction model obtained by the above-mentioned credit risk control prediction model training method; it is also used to assess the credit risk of the user based on the predicted risk label.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] This invention provides a credit risk control prediction model training method, credit risk control prediction method, and system. The method involves inputting data from the training sets of two credit institutions into their respective local learning models to obtain hidden features from their training sets, thereby completing the data and ensuring consistency in the format of the training sets. This uniformly formatted training set data is then input into the initial, identical local prediction models. The output of each local prediction model is compared with the risk label of each data point to obtain the loss value of each local prediction model and a common alignment loss value. The loss value of one credit institution is then sent to the other. Since the loss value is an intermediate analysis result and does not involve specific user data, it avoids data leakage and ensures data security. Combined with the alignment loss value, the reconstructed loss value and the comprehensive gradient of each credit institution are obtained. By utilizing the combined gradient, the parameters of the local prediction model are updated. The loss values ​​are exchanged, allowing for multiple or deeper parameter updates to the local prediction model using both loss values. This is equivalent to training the local prediction model using its own user data while simultaneously training it using the user data of another entity. Combining the data from both parties increases the training speed of the local prediction model and allows for a suitable reduction in the size of the required training set, avoiding data silos. Training continues until both the individual loss values ​​and the reconstructed loss values ​​converge to the corresponding preset accuracy. At this point, the local prediction model training stops. Multiple credit institutions can use the trained local prediction model as a credit risk control prediction model, resulting in a wider range of applications. Using the locally learned model and the trained credit risk control prediction model to predict the credit risk of loan users yields more accurate prediction results, ensuring the security of loan funds. Attached Figure Description

[0045] Figure 1 A flowchart illustrating a credit risk control prediction model training method according to an embodiment of the present invention is shown.

[0046] Figure 2 This invention illustrates a schematic diagram of the practical application process of a credit risk control prediction model according to an embodiment of the present invention.

[0047] Figure 3 A flowchart illustrating a credit risk control prediction method according to an embodiment of the present invention is shown;

[0048] Figure 4 A schematic diagram of a credit risk control prediction system according to an embodiment of the present invention is shown. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0051] In the description of this specification, references to terms such as "embodiment," "one embodiment," and "one implementation" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or implementation is included in at least one embodiment or implementation of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or implementation. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or implementations.

[0052] Figure 1 The diagram illustrates a flowchart of a credit risk control prediction model training method according to an embodiment of the present invention. The credit risk control prediction model training method includes:

[0053] S1: Input the data from the training sets of both credit institutions into their respective local learning models to obtain the hidden features of the data in their respective training sets.

[0054] Specifically, assuming the two lending institutions are Party A and Party B, as follows: Figure 2 To jointly achieve credit risk control, the first step is to build a secure learning model based on data features. Then, the learning model is sent to both parties as their local learning models. The learning model can be a conventional fully connected neural network.

[0055] Suppose the dataset belonging to party A is The dataset belonging to Party B is in It is a feature set of two datasets. This is the risk label for the corresponding data row, N A N B ∈N represent the sizes of the datasets from both sides. Clearly, D A D B These data sources are owned separately by Party A and Party B, and cannot be directly exposed to each other. Furthermore, to achieve alignment between different data sources, we also need to consider the existence of a finite-sized overlapping sample set, based on the actual situation. Where N AB This refers to the size of the overlapping set. A's local learning model is a neural network. The local learning model of Party B is a neural network. The task of the two neural networks is to predict the hidden features of each other, respectively. and For example, suppose Party A has the following characteristics: age, gender, credit rating, and savings amount; and Party B has the following characteristics: credit rating, savings amount, work location, and years of employment. The output is the work location and years of service. The output is age and gender.

[0056] S2: Based on the hidden features, complete the data in the training sets to ensure that the format of the data in each training set is consistent. The supplemented data with missing hidden features has no practical significance; it is mainly to maintain the consistency of the data format between the two sets, facilitating the training of the prediction model.

[0057] S3: Input the data from the completed training sets into their respective local prediction models, compare the output of the local prediction models with the risk labels of each data point, and obtain their respective loss values ​​and a common aligned loss value, wherein the risk label represents the user's credit risk level.

[0058] Specifically, the objective functions of A's local prediction model and B's local prediction model can be written as follows: as well as Where θ A ,θ B These are the training parameter sets for the two networks, l1 is the loss function for the output layer (we choose the cross-entropy loss function), and φ is the local prediction model function for hidden features, whose output is the predicted risk label. In addition, we also need to minimize the alignment loss between A and B. The alignment loss function l2 is the square of the Frobenius norm.

[0059] Neural Networks and objective function L A In form, the objective function should be the risk label y A The function calculates the loss based on Party A's risk label and the predicted risk label. In this scheme, predicting the risk label requires the analyzed hidden features u. A And hidden features u A During the analysis process, hidden features u B Therefore, I calculated that the final objective function is the risk label y. A Hidden features of Party B B It takes the form of a function, but in essence, it analyzes the risk label y. A and hidden features uA Additionally, although the text above mentions using the other party's data (e.g., hidden features), B However, in actual training, the parameters (intermediate results) are obtained from the analysis of the other party's data, rather than directly using the other party's data, so there is no data leakage involved.

[0060] S4: Send the loss value of one credit institution to the other credit institution, and combine it with the alignment loss value to obtain their respective reconstruction loss values ​​and their respective comprehensive gradients.

[0061] Specifically, in order to combine the training and analysis results from both parties, such as Figure 2 This allows both parties to exchange intermediate analysis results, accelerating the training process. Essentially, two sets of data are simultaneously training the same prediction model, leveraging each other's intermediate analysis data to significantly reduce training time and the amount of data required. The reconstructed loss value can be the sum of the individual loss values ​​from both parties. Subsequently, this is equivalent to jointly training a local prediction model using user data from both parties without exposing the data directly.

[0062] S5: Update the parameters of the local prediction model according to the comprehensive gradient until the reconstruction loss value converges to the preset accuracy, stop the training of the local prediction model, and use the trained local prediction model as the credit risk control prediction model corresponding to the credit institution.

[0063] Specifically, the gradient represents the magnitude and direction of the contribution of the model parameters to the loss function in the current state, that is, the direction and magnitude of parameter updates. Using the gradient descent algorithm, we can adjust the model parameters based on this gradient, causing the loss function to gradually decrease and the model performance to gradually improve.

[0064] In this embodiment, data from the training sets of both credit institutions are input into their respective local learning models to obtain hidden features, thus completing the training sets and ensuring consistency in the format of the training data. The uniformly formatted training data is then input into the same initial local prediction model. The output of each local prediction model is compared with the risk label of each data point to obtain the loss value of each model and a common alignment loss value. The loss value of one credit institution is sent to the other. Since the loss value is an intermediate analysis result and does not involve specific user data, it avoids data leakage and ensures data security. Combined with the alignment loss value, the respective reconstruction loss value and their respective comprehensive gradient are obtained. The comprehensive gradient is then used to update the local prediction model. The parameters of the test model are exchanged, and the loss values ​​are shared. This allows for multiple or deeper parameter updates to the local prediction model using both user data. Essentially, the local prediction model is trained using both user data and the other party's data simultaneously. This combined approach increases the training speed of the local prediction model and allows for a reduction in the required training set size, avoiding data silos. Training continues until both the individual loss values ​​and the reconstructed loss values ​​converge to the corresponding preset accuracy. Multiple credit institutions can then use the trained local prediction model as a credit risk control prediction model, broadening its applicability. Using both the local learning model and the trained credit risk control prediction model to predict loan users' credit risk yields more accurate results, ensuring the security of loan funds.

[0065] In one embodiment of the present invention, the step of inputting the data from the respective completed training sets into their respective local prediction models, and comparing the output of the local prediction models with the risk labels of each data point to obtain their respective loss values ​​and a common alignment loss value includes:

[0066] Each of the supplemented data is input into its respective local prediction model. Based on the output of the local prediction model and the risk label corresponding to the supplemented data, the loss value of each credit institution is obtained.

[0067] The data containing the hidden features of both parties is input into the local prediction model. Based on the output of the local prediction model and the risk label corresponding to the data containing the hidden features of both parties, a common alignment loss value is obtained.

[0068] In this embodiment, parties A and B respectively calculate the corresponding loss value L. A and L B The results are then sent to the other party, and both Party A and Party B calculate the alignment loss value L based on the secret sharing protocol.AB .

[0069] In one embodiment of the present invention, the step of sending the loss value of one credit institution to the other credit institution and combining it with the aligned loss value to obtain their respective reconstructed loss values ​​and their respective comprehensive gradients includes:

[0070] The loss value from one credit institution is sent to the other credit institution. The loss values ​​from both parties and the aligned loss value are added together to obtain the reconstruction loss value, i.e., the reconstruction loss value L = L. A +L B +L AB .

[0071] Based on the reconstruction loss value, the aligned gradient and the independent gradient are determined. The aligned gradients obtained from the reconstruction loss value analysis are as follows: and The gradient alone is and

[0072] The individual gradient is added to the aligned gradient to obtain the combined gradient. The combined gradient of side A is... The overall gradient of side B is

[0073] In this embodiment, the step of sending the loss value of one credit institution to the other credit institution, and adding the loss values ​​of both parties and the aligned loss value to obtain the reconstructed loss value includes:

[0074] The loss value from one credit institution is sent to the other credit institution. The loss values ​​from both institutions are added together, and then the product of the aligned loss value and a preset weight is added to obtain the reconstructed loss value. Specifically, the reconstructed loss value L = L A +L B +αL AB Wherein, the weight α is a hyperparameter of the training process. Since the distribution of the original user data of each participant may differ to some extent, and L... A L B and L AB The choice of different loss methods (one being cross-entropy loss, the other Frobenius norm) may lead to significant differences in the calculated magnitudes. To avoid the problem of slow parameter updates caused by such differences, a weight α is used for balancing. The value of the weight α can be selected by pre-training with a small batch of data and then adjusting it based on the actual loss value.

[0075] In one embodiment of the present invention, updating the parameters of the local prediction model according to the comprehensive gradient until the reconstructed loss value converges to a preset accuracy, stopping the training of the local prediction model, and using the trained local prediction model as the credit risk control prediction model corresponding to the credit institution includes:

[0076] Based on the comprehensive gradient, the parameters of the local prediction model are updated until their respective loss values, the common alignment loss value, and their respective reconstruction loss values ​​all converge to the corresponding preset accuracy, at which point the training of the local prediction model is stopped.

[0077] Specifically, when checking the loss value, in addition to checking whether the reconstruction loss value L converges, the individual loss values ​​L can also be checked. A and L B and alignment loss value L AB Whether each model converges to a preset accuracy; this design ensures that each local prediction model can converge to a certain limit, resulting in more accurate predictions from the trained models.

[0078] During training, collaborative computation is performed based on a secret sharing protocol. Any information received by a participant from other participants is only partially encrypted data or partial intermediate analysis results, without access to the other party's private user data. This provides a high level of security for information or knowledge sharing among multiple parties, thereby protecting the privacy of each party's data and the security of the model. This method does not require large amounts of computing resources or large-scale parallel capabilities, significantly improving computational efficiency without any loss of accuracy.

[0079] In one embodiment of the present invention, before inputting the data from the respective training sets of both credit institutions into their respective local learning models to obtain the hidden features of the data in their respective training sets, the method further includes:

[0080] User data is preprocessed, and the risk label is added to the credit risk of each user to construct sample data.

[0081] Specifically, each credit institution or financial institution needs to collect user data as training data and perform preliminary preprocessing. Specifically, each participant constructs features for the sample data based on user information and user behavior records. Each user data point can include dimensions such as credit history, income level, assets and liabilities, tax payment status, and operational anomalies. Furthermore, based on actual credit granting records and business logic, the sample data is evaluated to generate labels for the dataset, where +1 represents low risk and credit approval is possible, while -1 represents the opposite.

[0082] The sample data is differentially processed to obtain the training set.

[0083] Specifically, for example, using Gaussian or exponential mechanisms to perform differencing on the sample data is equivalent to transforming each data point. Each user data point in the training dataset is already different from the original data. Even if the data is leaked during the establishment of the prediction model, other institutions that obtain the leaked data will not be able to know the true value of the corresponding user data, thus protecting user privacy.

[0084] In one embodiment of the present invention, the preprocessing of user data, adding the risk label to the credit risk of each user, and constructing sample data includes:

[0085] Extract the data features from the user data, and add the risk label to the credit risk of each user based on the credit records in the user data to construct the sample data. The data features include user identity information, income level, assets and liabilities, tax payment status, business abnormalities and historical credit. The risk label of each user corresponds to the user data of each user.

[0086] Examine the data type in the sample data, wherein the data type includes string type and numeric type.

[0087] When the data in the sample data is of the string type, the string type data is converted into a floating-point number type, and any missing data after conversion is filled in.

[0088] When the data in the sample data is of the numerical type, the data of the numerical type is standardized.

[0089] Specifically, after the dataset is constructed, data processing needs to be performed according to a certain paradigm. First, the data types in the dataset are examined, and numerical values ​​stored in string form are normalized by converting string data into floating-point numbers. Second, Lagrange interpolation is used to fill in missing data along the column dimension. Finally, the data columns of numerical types are observed, and for those with large ranges, mean-variance standardization is used to scale the entire column of data.

[0090] In this embodiment, user data from diverse sources is standardized, enabling it to participate in subsequent calculations with a similar distribution. Furthermore, preprocessed data can prevent anomalies during subsequent training of the local prediction model, reduce neuron oversaturation, and accelerate the convergence speed of the entire training process.

[0091] In one embodiment of the present invention, the step of performing differential processing on the sample data to obtain the training set includes:

[0092] When the data column corresponding to the data feature in the sample data is a continuous data column, independent random noise following a Gaussian mechanism is added to each data item in the data column, wherein the data corresponding to each data feature in the sample data constitutes a data column.

[0093] Specifically, for continuous data columns, independent random noise following a Gaussian distribution is added to each data item. Where ∈ represents the privacy budget, which is typically the reciprocal of the total sample size; δ represents the relaxation term, set to 10. -5 Δf is the sensitivity of the dataset, which can be set to 1 when the range of variation is small. After adding the difference term with the Gaussian mechanism, the continuous data column satisfies the relaxed (∈,δ)-DP mechanism.

[0094] When the data column corresponding to the data feature in the sample data is a discrete data column, an exponential mechanism is used to perform difference processing on each data item in the data column.

[0095] Specifically, for discrete data columns, let the discrete dataset D = {R1, R2, ..., R...} n}, where R i Let i = 1, 2, ..., n be the elements in the dataset. Construct a scoring function q(D, R). i The scoring function (D) can be constructed using methods such as frequency statistics and sum calculation. For a discrete dataset D, the scoring function constructs the relationship between each discrete point and the entire set. Furthermore, using... Output corresponding to R i The probability results are given here, assuming there are a total of 10 discrete points in the set, namely R1, R2, ..., R... 10 The function of M(R1) is to map the output R1 to R1, R2, ..., R 10 The probability is calculated, and then the one with the highest probability, R, is selected. i As the result of R1 differencing, other M(Ri) are similar, where ∈ and Δq represent the privacy budget and sensitivity, respectively, and the value with the highest probability is selected as the result of differencing. After differencing using the exponential mechanism, the discrete data column satisfies the ∈-DP mechanism.

[0096] The data in the processed data column is used to replace the original data in the sample data to obtain the training set.

[0097] In this embodiment, after adding random noise (i.e., difference terms) or performing function transformation on the data, each user's data in the training set is different from the original data. Even if leaked, the true value of the corresponding user's data cannot be known, thus protecting user privacy. At the same time, under the condition that the sample data scale is large enough, according to the differential privacy mechanism and the independent and identically distributed characteristics, the overall statistical characteristics of the samples have not changed. Continuing to participate in subsequent neural network calculations and training of local prediction models can still ensure that the information of the original samples is not lost, thus ensuring the effectiveness of the training results.

[0098] Figure 3 The diagram illustrates a flowchart of a credit risk control prediction method according to an embodiment of the present invention. The credit risk control prediction method includes:

[0099] S10: Input the user data of users who apply for credit into the local learning model to obtain the hidden features of the user data.

[0100] S20: Based on the user data, the hidden features, and the credit risk control prediction model obtained by the above-mentioned credit risk control prediction model training method, obtain the predicted risk label corresponding to the user data;

[0101] S30: Assess the user's credit risk based on the predicted risk label.

[0102] In this embodiment, as Figure 2 After the local prediction model is trained, the participating financial institutions save the model framework and parameters locally as a credit risk control prediction model. In subsequent scenarios where customers need to borrow money, customer data can be collected and transformed according to the preprocessing methods described above. This data is then input into the local learning model and the credit risk control prediction model for prediction. Based on the specific risk labels output by the model, the potential risks in this credit transaction are determined, ensuring the safety of loan funds.

[0103] Figure 4 This diagram illustrates the structure of a credit risk control prediction system according to an embodiment of the present invention. The credit risk control prediction system includes:

[0104] The learning module 10 is used to input user data of users who apply for credit into the local learning model to obtain the hidden features of the user data.

[0105] The prediction module 20 is used to obtain the predicted risk label corresponding to the user data based on the user data, the hidden features, the local learning model, and the credit risk control prediction model obtained by the above-mentioned credit risk control prediction model training method; it is also used to assess the credit risk of the user based on the predicted risk label.

[0106] In this embodiment, after the local prediction model is trained, the participating financial institutions save the model framework and parameters locally. In subsequent scenarios where customers need to borrow money, customer data can be collected and transformed according to the preprocessing method described above. Then, the learning module 10 inputs the user data of the user applying for credit into the local learning model to obtain the hidden features of the user data. The prediction module 20, based on the user data, the hidden features, the local learning model, and the credit risk control prediction model trained using the aforementioned method, obtains the predicted risk label corresponding to the user data. Based on the predicted risk label, it assesses the user's credit risk, determines the potential risks in this credit transaction, and ensures the safety of loan funds.

[0107] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for training a credit risk control prediction model, characterized in that, include: The data from the training sets of both credit institutions are input into their respective local learning models to obtain the hidden features of the data in their respective training sets; Based on the hidden features, the data in the training set is completed to make the format of the data in each training set consistent; Each of the data in the completed training set is input into its respective local prediction model. The output of the local prediction model is compared with the risk label of each data point to obtain the respective loss value and the common aligned loss value, wherein the risk label represents the user's credit risk level. The loss value of one credit institution is sent to the other credit institution, and combined with the aligned loss value, the respective reconstruction loss value and the respective comprehensive gradient are obtained; Based on the comprehensive gradient, the parameters of the local prediction model are updated until the reconstruction loss value converges to a preset accuracy. Then, the training of the local prediction model is stopped, and the trained local prediction model is used as the credit risk control prediction model corresponding to the credit institution.

2. The training method for the credit risk control prediction model according to claim 1, characterized in that, The step of inputting the data from each of the completed training sets into their respective local prediction models, and comparing the output of the local prediction models with the risk labels of each data point to obtain their respective loss values ​​and a common alignment loss value includes: Each of the supplemented data is input into its respective local prediction model. Based on the output of the local prediction model and the risk label corresponding to the supplemented data, the loss value of each credit institution is obtained. The data containing the hidden features of both parties is input into the local prediction model. Based on the output of the local prediction model and the risk label corresponding to the data containing the hidden features of both parties, a common alignment loss value is obtained.

3. The credit risk control prediction model training method according to claim 2, characterized in that, The step of sending the loss value of one credit institution to the other credit institution and combining it with the aligned loss value to obtain their respective reconstructed loss values ​​and their respective comprehensive gradients includes: The loss value of one credit institution is sent to the other credit institution, and the loss values ​​of both parties and the alignment loss value are added together to obtain the reconstruction loss value; Based on the reconstruction loss value, determine the aligned gradient and the independent gradient; The individual gradient is added to the aligned gradient to obtain the combined gradient.

4. The credit risk control prediction model training method according to claim 2, characterized in that, The step of sending the loss value from one credit institution to the other credit institution, and adding the loss values ​​from both parties and the aligned loss value to obtain the reconstructed loss value includes: The loss value of one credit institution is sent to the other credit institution. The loss values ​​of both parties are added together, and then the product of the aligned loss value and the preset weight is added to obtain the reconstructed loss value.

5. The credit risk control prediction model training method according to claim 3, characterized in that, The step of updating the parameters of the local prediction model according to the comprehensive gradient until the reconstructed loss value converges to a preset accuracy, stopping the training of the local prediction model, and using the trained local prediction model as the credit risk control prediction model corresponding to the credit institution includes: Based on the comprehensive gradient, the parameters of the local prediction model are updated until their respective loss values, the common alignment loss value, and their respective reconstruction loss values ​​all converge to the corresponding preset accuracy, at which point the training of the local prediction model is stopped.

6. The training method for the credit risk control prediction model according to claim 1, characterized in that, Before inputting the data from the respective training sets of both credit institutions into their respective local learning models to obtain the hidden features of the data in their respective training sets, the method further includes: Preprocess user data, add the risk label to the credit risk of each user, and construct sample data; The sample data is differentially processed to obtain the training set.

7. The credit risk control prediction model training method according to claim 6, characterized in that, The preprocessing of user data, including adding risk tags to the credit risk of each user and constructing sample data, includes: Extract the data features from the user data, and add the risk label to the credit risk of each user based on the credit records in the user data to construct the sample data. The data features include user identity information, income level, assets and liabilities, tax payment status, business abnormalities and historical credit. The risk label of each user corresponds to the user data of each user. Examine the data type in the sample data, wherein the data type includes string type and numeric type; When the data in the sample data is of the string type, the string type data is converted into a floating-point number type, and any missing data after conversion is filled in; When the data in the sample data is of the numerical type, the data of the numerical type is standardized.

8. The training method for the credit risk control prediction model according to claim 7, characterized in that, The step of performing differential processing on the sample data to obtain the training set includes: When the data column corresponding to the data feature in the sample data is a continuous data column, independent random noise following a Gaussian mechanism is added to each data item in the data column, wherein the data corresponding to each data feature in the sample data constitutes a data column; When the data column corresponding to the data feature in the sample data is a discrete data column, an exponential mechanism is used to perform difference processing on each data item in the data column. The data in the processed data column is used to replace the original data in the sample data to obtain the training set.

9. A credit risk control prediction method, characterized in that, include: The user data of users who apply for loans is input into a local learning model to obtain the hidden features of the user data; Based on the user data, the hidden features, the local learning model, and the credit risk control prediction model training method as described in any one of claims 1-8, the predicted risk label corresponding to the user data is obtained. Assess the user's credit risk based on the predicted risk label.

10. A credit risk control prediction system, characterized in that, include: The learning module is used to input user data of users applying for loans into a local learning model to obtain the hidden features of the user data; The prediction module is used to obtain the predicted risk label corresponding to the user data based on the user data, the hidden features, the local learning model, and the credit risk prediction model training method as described in any one of claims 1-8, and to assess the credit risk of the user based on the predicted risk label.

Images