Model training method, evaluation method, device, apparatus, medium and program product

Abstract

This embodiment discloses a model training method, evaluation method, apparatus, device, medium, and program product. The model training method includes: determining a first community to which a first user belongs; determining a first user tag for the first user based on the first user's first social relationship data; updating the community to which the first user belongs and determining a second community to which the first user belongs when the difference between the first user tag and the first community tag of the first community is greater than or equal to a first threshold; determining a first evaluation tag for the first user based on the second community tag of the second community; constructing a training dataset using the first evaluation tag of the first user; and training an evaluation model based on the training dataset so that the trained evaluation model can evaluate the user's credit.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and specifically relates to a model training method, evaluation method, device, equipment, medium and program product. Background Technology

[0002] With the advancement of mobile internet, online loan application processes are typically designed to be quite simple to facilitate individual loan applications. Users are usually only required to fill in basic personal information, but this is not conducive to building a user's credit profile. This simplified application process presents significant challenges to credit assessment and delinquency risk detection, thus necessitating the effective identification of potential user risks and the implementation of differentiated risk control strategies. Summary of the Invention

[0003] This application provides a model training method, evaluation method, apparatus, device, medium, and program product.

[0004] This application provides a model training method, the method comprising:

[0005] Determine the first community to which the first user belongs;

[0006] The first user tag of the first user is determined based on the first user's first social relationship data;

[0007] If the difference between the first user tag and the first community tag of the first community is greater than or equal to a first threshold, update the community to which the first user belongs and determine the second community to which the first user belongs; based on the second community tag of the second community, determine the first evaluation tag of the first user;

[0008] A training dataset is constructed using the first user's first evaluation label. The evaluation model is then trained based on the training dataset, enabling the trained evaluation model to assess the user's credit.

[0009] In some embodiments, determining the first community to which the first user belongs includes: directly determining the first community to which the first user belongs based on the first user's first social relationship data; updating the community to which the first user belongs and determining the second community to which the first user belongs includes: updating the first community to which the first user belongs based on the first social relationship data; if the difference between the first community tag of the updated first community and the first user tag is less than the first threshold, determining the updated first community as the second community to which the first user belongs.

[0010] It can be seen that by comparing the first community tag with the more accurate first user tag to identify the differences, and updating the first community when the differences are large, the tag of the second community to which the first user belongs can be made closer to the first user tag, thus improving the accuracy of community division for the first user.

[0011] In some embodiments, determining the first community to which the first user belongs includes: determining the first community to which the first user belongs through a first model; determining the first user tag of the first user based on the first social relationship data of the first user includes: processing the first social relationship data based on a second model to obtain the first user tag of the first user; before updating the community to which the first user belongs, the method further includes: obtaining a first loss function of the first model and a second loss function of the second model; determining a third loss function based on the difference between the first user tag and the first community tag; constructing a joint loss function based on the first loss function, the second loss function, and the third loss function; updating the community to which the first user belongs includes: updating the first community to which the first user belongs with the objective of minimizing the value of the joint loss function.

[0012] It can be seen that by aiming to minimize the joint loss function, it is beneficial to obtain accurate community segmentation results through the first model, accurate first user labels through the second model, first community labels that are closest to the first user labels, and second community labels that are accurate to the first user, thus achieving accurate segmentation of the community to which the first user belongs.

[0013] In some embodiments, the method further includes: obtaining an evaluation model loss function by reducing the weight of easily classified samples in the binary classification loss function or increasing the weight of difficult-to-classify samples in the binary classification loss function; and optimizing the model parameters of the evaluation model using the evaluation model loss function.

[0014] It can be seen that by reducing the weight of easily classified samples or increasing the weight of difficult-classified samples in the binary classification loss function, the evaluation of the model's loss function can focus more on the model's handling of difficult-classified samples during the model parameter update process. Optimizing the model parameters by evaluating the model's loss function can improve the accuracy of the evaluation model in handling difficult-classified sample data.

[0015] In some embodiments, before determining the first community to which the first user belongs, the method further includes: obtaining first basic data of the first user; the first basic data includes one or more of the first user's identity information and the first user's payment information with the operator; determining a second evaluation label of the first user based on the first basic data; the step of constructing a training dataset using the first evaluation label of the first user includes: constructing a training dataset using the first evaluation label of the first user and the second evaluation label of the first user.

[0016] It can be seen that in practical applications, since the corresponding operator data can be directly obtained from the operator network through user identity information, when users conduct credit assessments, data diversity can be achieved by combining the data from the operator network with the user's input identity information, thereby improving the training effect of the assessment model.

[0017] In some embodiments, before determining the first user tag of the first user based on the first user's first social relationship data, the method further includes: acquiring the first user's first communication behavior data, and determining the first social relationship data based on the first communication behavior data.

[0018] It can be seen that determining the first social relationship data through the first user's first communication behavior data is beneficial when users build credit profiles and apply for loans through mobile devices. By directly using the communication behavior data obtained from the operator, the user's first user tag can be determined. This further facilitates obtaining an accurate second community based on the first user tag and obtaining accurate training data.

[0019] This application also provides an evaluation method, the method comprising:

[0020] Determine the third community to which the second user belongs;

[0021] The second user tag of the second user is determined based on the second user's second social relationship data;

[0022] If the difference between the second user's tag and the third community tag of the third community is greater than or equal to a first threshold, update the community to which the second user belongs and determine the fourth community to which the second user belongs; based on the fourth community tag of the fourth community, determine the third evaluation tag of the second user;

[0023] The credit assessment result of the second user is obtained by processing the third assessment label through the assessment model; wherein the assessment model is trained based on the above model training method.

[0024] In some embodiments, before determining the third community to which the second user belongs, the method further includes: obtaining second basic data of the second user; the second basic data includes one or more of the second user's identity information and the second user's payment information with the operator; determining a fourth evaluation tag of the second user based on the second basic data; and processing the third evaluation tag through an evaluation model, including: processing the third evaluation tag and the fourth evaluation tag through an evaluation model.

[0025] As can be seen, when a user conducts credit assessment based on a mobile device, the method provided in this embodiment can obtain the fourth assessment tag of the second user by acquiring data from the operator's network, thereby enriching the characteristics of the second user and improving the accuracy of credit assessment.

[0026] In some embodiments, before determining the second user's second user tag based on the second user's second social relationship data, the method further includes: obtaining the second user's second communication behavior data, and determining the second social relationship data based on the second communication behavior data.

[0027] As can be seen, the method given in this embodiment can obtain the third evaluation label of the second user by acquiring communication behavior data in the operator's network, enriching the characteristics of the second user and facilitating accurate assessment of user credit through mobile terminals.

[0028] This application embodiment also provides a model training apparatus, the apparatus comprising:

[0029] The first determining module is used to determine the first community to which the first user belongs; and to determine the first user tag of the first user based on the first user's first social relationship data.

[0030] The first processing module is configured to update the community to which the first user belongs and determine the second community to which the first user belongs when the difference between the first user tag and the first community tag of the first community is greater than or equal to a first threshold; and determine the first evaluation tag of the first user based on the second community tag of the second community.

[0031] The training module is used to construct a training dataset using the first evaluation label of the first user, and to train the evaluation model based on the training dataset so that the trained evaluation model can evaluate the user's credit.

[0032] This application embodiment also provides an evaluation apparatus, the apparatus comprising:

[0033] The second determining module is used to determine the third community to which the second user belongs; and to determine the second user's second user tag based on the second user's second social relationship data;

[0034] The second processing module is used to update the community to which the second user belongs and determine the fourth community to which the second user belongs when the difference between the second user tag and the third community tag of the third community is greater than or equal to a first threshold; determine the third evaluation tag of the second user based on the fourth community tag of the fourth community; and process the third evaluation tag through an evaluation model to obtain the credit evaluation result of the second user; wherein the evaluation model is trained based on the above model training method.

[0035] This application provides an electronic device, which includes a processor and a memory for storing computer programs capable of running on the processor; wherein,

[0036] The processor is used to run the computer program to perform any of the above-described model training or evaluation methods.

[0037] This application provides a computer storage medium storing a computer program that, when executed by a processor, implements any of the above-described model training or evaluation methods.

[0038] This application provides a computer program product, including a computer program that, when executed by a processor, implements any of the above-described model training methods or evaluation methods.

[0039] This application provides a model training method, evaluation method, apparatus, device, medium, and program product. Based on the model training and evaluation methods provided in this application, the accuracy of user community segmentation is improved by correcting the results based on user tags. The evaluation tags of a first user are determined by the community to which the first user belongs, and the evaluation model is trained based on these tags. The trained evaluation model is then used to process the evaluation tags reflecting social relationships of a second user to obtain the credit evaluation result for the second user, effectively improving the accuracy of user credit evaluation. Attached Figure Description

[0040] Figure 1 A flowchart of a model training method provided in an embodiment of this application;

[0041] Figure 2 A flowchart of an evaluation model training process provided in this application embodiment;

[0042] Figure 3A flowchart of an evaluation method provided in an embodiment of this application;

[0043] Figure 4 This application provides a schematic diagram of the structure of a model training device according to an embodiment of the present application;

[0044] Figure 5 This application provides a schematic diagram of the structure of an evaluation device according to an embodiment of the present application;

[0045] Figure 6 This is a schematic diagram of the composition structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0046] With the advancement of mobile internet, mobile applications (APPs) have become an important platform for financial institutions to provide online services. Credit lending is a major source of revenue for financial institutions' personal business, and credit risk decisions are a key determinant of their success. Failure to effectively identify delinquency risks will not only lead to a significant increase in bad debt rates but also easily generate systemic risks, greatly increasing the management costs of financial institutions.

[0047] To facilitate loan applications for individual users, online loan application processes are typically designed to be simple, meaning users are not required to fill in detailed application information. This hinders accurate credit assessment. Therefore, the simplification of application information presents a significant challenge to credit evaluation; lower accuracy leads to higher delinquency rates and greater financial losses for banks. From a bank's security perspective, credit assessment has become a crucial and widely studied task in the financial industry. The key to credit assessment lies in distinguishing between applicants with good and bad repayment ability and willingness (creditworthiness).

[0048] Consumer loans constitute a large proportion of mobile financial apps, but the application process is typically simple and quick. To improve user experience and enhance the competitiveness of app-based loan services, users are usually not required to fill in excessive personal information. This reduces the reference information available to the risk control system for credit assessment, hindering the control of delinquency rates. Therefore, maximizing the predictability of delinquency risk without increasing the complexity of the online application process is a significant challenge for many financial institutions during the credit assessment process.

[0049] Traditional credit assessment and loan risk assessment methods typically rely on credit data modeling, primarily employing time-window-based statistical strategies to extract users' historical statistical characteristics, such as the number of borrowings, repayments, and overdue payments in the past week or three months, as well as the amount of overdue payments. This method, using only user credit data, struggles to improve the model's predictive accuracy. This application's research revealed that low-credit individuals often exhibit clustering characteristics, meaning there are often some kind of correlation between low-credit users. Therefore, it is necessary to process information related to applicants and their associates, build community relationships, identify potential risks from applicants, and implement differentiated risk control strategies to effectively reduce losses for financial institutions.

[0050] Currently, commonly used methods in community partitioning include the Label Propagation Algorithm (LPA), Louvain's algorithm, and Graph Neural Networks (GNNs). These methods perform well in small- to medium-sized relational networks, but they suffer from limitations in computational performance and insufficient model effectiveness in large- and ultra-large relational networks. These limitations include high computational complexity, slow computation, high demand for Graphics Processing Unit (GPU) resources, unstable community partitioning results, and lack of support for complex overlapping communities. In practical applications, these methods are not conducive to the implementation of credit assessment in complex network environments.

[0051] The model training and evaluation methods provided in this application can reduce computational complexity, computational cost, and resource dependence while supporting the identification of complex overlapping communities and ensuring the stability of community partitioning results. During the training of the evaluation model, to avoid introducing noisy data, the model parameters are optimized by reducing the impact of severe imbalance in the positive and negative proportions of credit samples, without using oversampling methods. This improves the accuracy of the evaluation model in assessing user credit.

[0052] The embodiments of this application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the embodiments provided herein are merely illustrative of the embodiments of this application and are not intended to limit the embodiments of this application. Furthermore, the embodiments provided below are some embodiments for implementing this application, and not all embodiments for implementing this application. Unless otherwise specified, the technical solutions described in the embodiments of this application can be implemented in any combination.

[0053] It should be noted that, in the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a method or apparatus that includes a list of elements includes not only the elements expressly described, but also other elements not expressly listed, or elements inherent to implementing the method or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other related elements in the method or apparatus that includes that element (e.g., steps in the method or units in the apparatus; for example, a unit in the apparatus may be a portion of circuitry, a portion of a processor, a portion of a program or software, etc.).

[0054] The model training method and evaluation method provided in this application include a series of steps. However, the model training method and evaluation method provided in this application are not limited to the steps described. Similarly, the model training device and evaluation device provided in this application include a series of modules. However, the device provided in this application is not limited to the modules explicitly described. It may also include modules that need to be set up to obtain relevant information or to process information.

[0055] This application provides a model training method, such as... Figure 1 As shown, the model training method provided in this application includes:

[0056] Step 101: Determine the first community to which the first user belongs.

[0057] Determining the first community to which a first user belongs can be achieved in various ways. Specifically, users can be segmented based on their identity information to obtain the first community to which any one of the multiple users belongs. Alternatively, users can be segmented based on their relationships within the social network and their interactions with each other. By segmenting users, users with some shared characteristics or tags can be grouped into the same community. As generally understood, this application, in determining the first community to which a first user belongs, includes segmenting multiple users using the methods described above to obtain one or more communities, thereby obtaining the first community to which any one of the multiple users belongs.

[0058] In the process of dividing users into communities as described above, algorithms such as LPA, Louvain's algorithm, and GNN can be used to divide users and obtain the first community to which the first user belongs. Taking the LPA algorithm as an example, LPA is a local community partitioning algorithm based on label propagation. The logic of the LPA algorithm is relatively simple, and its running speed is very fast compared to other optimization modularity algorithms. However, the partitioning results of the LPA algorithm may have problems such as instability and strong randomness.

[0059] Step 102: Determine the first user's first user tag based on the first user's first social relationship data.

[0060] In this step, the first user's first user tag is determined through the first user's first social relationship data. In practice, this data can be obtained from the first user's data on social media platforms, or by utilizing social media analytics tools. When the first user is a mobile user, their call information, SMS messages, etc., can also be obtained based on carrier data from the carrier network to acquire their first social relationship data.

[0061] After obtaining the initial social relationship data, features can be extracted from this data, and methods such as cluster analysis and relationship mining can be used to identify the behavioral patterns of the first user. By performing behavioral analysis based on these patterns, a first user tag can be obtained. Since the first user tag is derived from the first user's initial social relationship data, it reflects the characteristics of the first user's social relationships.

[0062] In the specific implementation process, steps 101 and 102 can be performed simultaneously, or the first user tag can be determined first, and then the first community to which the first user belongs can be determined.

[0063] Step 103: If the difference between the first user tag and the first community tag of the first community is greater than or equal to the first threshold, update the community to which the first user belongs and determine the second community to which the first user belongs; based on the second community tag of the second community, determine the first evaluation tag of the first user.

[0064] It can be seen that since the first community tag can reflect the characteristics of all users in the first community to a certain extent, and the first user tag can reflect the user's social relationship characteristics, the first community tag and the first user tag should have a certain degree of similarity. In particular, when users are divided into communities based on their social relationships, the first user tag and the first community tag should have a high degree of similarity.

[0065] By comparing the difference between the first user tag and the first community tag, if the difference is greater than or equal to a first threshold, it is considered that there is an error in the first community to which the first user belongs, as determined in step 101. The first community to which the first user belongs needs to be re-determined, and the first user tag is compared with the re-determined first community tag until the difference between the first user tag and the re-determined first community tag is less than the first threshold. The final community to which the first user belongs, i.e., the second community, is then obtained. If the difference between the first user tag and the first community tag is less than the first threshold, the first community to which the first user belongs is considered relatively accurate, and the first community can be directly used as the second community to which the first user belongs. Here, the first threshold can be determined based on empirical values.

[0066] Taking the LPA algorithm in the above steps as an example, since the partitioning result of the LPA algorithm is unstable, the method given in this embodiment can ensure that the accuracy of the LPA algorithm for community partitioning is improved on the basis of achieving efficient community partitioning, and quickly obtain the second community to which the first user belongs.

[0067] Step 104: Construct a training dataset using the first user's first evaluation label, and train the evaluation model based on the training dataset so that the trained evaluation model can evaluate the user's credit.

[0068] A training dataset is constructed by using a first evaluation label that reflects the social relationships of the first user. The evaluation model is then trained using this dataset, enabling the trained evaluation model to obtain the user's credit evaluation result based on the user's social relationships, thereby improving the accuracy of credit evaluation.

[0069] This application provides a model training method that corrects the user's community division results by using user tags that reflect the user's social relationships, obtains the accurate community to which the user belongs, determines the user's corresponding evaluation tag based on the user's final community tag, and trains the evaluation model based on the evaluation tag, which can obtain accurate credit evaluation results through the trained evaluation model.

[0070] In practical applications, steps 101 to 104 can be implemented based on a processor, which can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Model Trainer, Micro Model Trainer, and Microprocessor.

[0071] In some embodiments, determining the first community to which the first user belongs includes: directly determining the first community to which the first user belongs based on the first user's first social relationship data; updating the community to which the first user belongs and determining the second community to which the first user belongs includes: updating the first community to which the first user belongs based on the first social relationship data; and determining the updated first community as the second community to which the first user belongs if the difference between the first community tag of the updated first community and the first user tag is less than a first threshold.

[0072] Since the first community is derived based on the first user's first social relationship data, and the first user's tag is also derived based on the first social relationship data, the first community tag and the first user tag should have a high degree of similarity. By determining whether the difference between the first community tag and the first user tag is greater than or equal to a first threshold, we can judge whether the community division of the first user is accurate.

[0073] When the difference between the first community tag and the first user tag is greater than or equal to a first threshold, the community classification of the first user is considered inaccurate. In this case, the first community to which the first user belongs needs to be updated, i.e., the first community to which the first user belongs needs to be redefined. During the process of updating the first community, it should be ensured that the method for classifying the first user is consistent with the previous method. Specifically, the first community to which the first user belongs can be redefined based on the first user's first social relationship data; alternatively, the first social relationship data can be supplemented or corrected, and the first community to which the first user belongs can be redefined based on the supplemented or corrected first social relationship data.

[0074] Taking the determination of the first community to which the first user belongs using the LPA algorithm as an example, when it is necessary to update the first community to which the first user belongs, based on the characteristics of the LPA algorithm, since it is an iterative algorithm, the community partitioning result of the LPA algorithm for the first user can be affected by setting different maximum iterations, thus obtaining the new first community to which the first user belongs. Alternatively, because in the LPA algorithm, when multiple neighbor labels appear with the same frequency, a label is randomly selected as the new label, and the user's partitioning result is determined by the new label, this randomness may lead to different partitioning results from the LPA algorithm. Therefore, by controlling the randomness of the LPA algorithm, different randomness settings can be tried to find a better community partitioning and achieve the update of the first community.

[0075] During the updating of the first community, after each update, the updated first community tag is compared with the first user tag to obtain the difference between the updated first community tag and the first user tag. This process continues until the difference is less than a first threshold, at which point the corresponding updated first community is designated as the second community to which the first user belongs. It can be seen that updating the first community based on the difference between the first community tag and the first user tag can yield an accurate evaluation tag that reflects the user's social relationships.

[0076] In some embodiments, determining the first community to which the first user belongs includes: determining the first community to which the first user belongs through a first model; determining the first user tag of the first user based on the first social relationship data of the first user includes: processing the first social relationship data based on a second model to obtain the first user tag of the first user; before updating the community to which the first user belongs, the method further includes: obtaining a first loss function of the first model and a second loss function of the second model; determining a third loss function based on the difference between the first user tag and the first community tag; constructing a joint loss function based on the first loss function, the second loss function, and the third loss function; updating the community to which the first user belongs includes: updating the first community to which the first user belongs with the objective of minimizing the value of the joint loss function.

[0077] This embodiment further provides a method for updating the first community to which the first user belongs and determining the second community. A first model can be constructed based on the LPA algorithm, Louvain algorithm, GNN, etc., given in the above embodiments. Based on the first model, multiple users are then divided into communities to obtain the first community to which the first user belongs. Taking the construction of the first model based on the LPA algorithm as an example, each user is treated as a node in the LPA algorithm. According to the community division method of the LPA algorithm, a small portion of the nodes are pre-labeled nodes with annotation results. The annotation results of the nodes can be overdue, fraudulent, or blacklisted, etc. Among the multiple nodes to be divided, there are also a large number of unlabeled nodes. The first model divides the remaining unlabeled nodes into communities based on the labeled nodes, obtaining the community division result for each node and the annotation result for each node. Here, the node annotation result can be obtained based on the community tag of the user's community.

[0078] When the first model constructed using the LPA algorithm divides users into communities, it first initializes the label of each node. This initial label uses the pre-labeled results of the node, or, for nodes without pre-labeled results, the label can be set to 0. Based on the social relationship data of each node, the relationships between nodes are determined, and based on these relationships, the weights of the edges between nodes are determined.

[0079] Based on the LPA algorithm, for the first node formed by the first user, the label of the first node is determined by voting on the label of all its neighboring nodes and the weights of the corresponding edges. For example, in a risk control scenario, the density of defaulting users in the community will significantly affect the user's probability of normal repayment. If the default probability of a user's first-degree contacts and second-degree contacts is high and their credit assessment is low, the user himself will also have a high probability of default.

[0080] If the label corresponding to the voting result of the first node is different from the label of the first node itself, then the label of the first node is updated. During the community partitioning process, if the labels of all nodes remain unchanged (i.e., the label of a node after voting is the same as the node's own label), then the community partitioning is stable. Specifically, the labels of all adjacent nodes corresponding to the first node are obtained. Each adjacent node votes for the first node; the greater the weight of the edge between nodes, the greater its influence in the vote. If two adjacent nodes have the same vote, the labels corresponding to the votes are accumulated. The label with the largest final vote count is calculated to obtain the first community label f ​​corresponding to the first node. i .

[0081] The first loss function of the first model is determined by comparing the community segmentation results obtained from the first model with the corresponding user pre-labeled results.

[0082] Simultaneously, a second model can be constructed to process social relationship data and obtain user tags. For example, a neural network binary classifier can be built to learn mini-batch gradient descent on all user tags, and then the user's social relationship data X can be calculated. 1：n The predicted user tag g is obtained. i and pre-labeled real user tags y i By making comparisons, the predicted user tag g is continuously minimized. i With real user tags y i The differences between them are used to update the parameters of the binary classifier, i.e., the model parameters of the second model. This can also be achieved by using a randomized regularization coefficient λ. L To prevent overfitting, based on predicted user label g. i With real user tags y i Based on the differences between them, a second loss function corresponding to the second model is constructed.

[0083] For the unlabeled nodes (such as the first user), the first user's first social relationship data is processed using the trained second model to obtain the predicted first user label g. i By using the first user tag g i The first community label f ​​obtained from the first model i A comparison is then made to construct a third loss function. In constructing the third loss function, randomized regularization coefficients λ can be used. U Noise reduction is performed. If the first user tag g i The first community label f ​​obtained from the first model i If the difference is less than the first threshold, then the first community label f ​​obtained by the first model is considered to be... i That's correct, the first community tag f i The second community tag y, belonging to the first user's second community. i ′ Otherwise, the first community needs to be updated, and its tags need to be redefined. i .

[0084] A joint loss function is constructed by using a first loss function, a second loss function, and a third loss function, such that the value of the joint loss function is positively correlated with the values ​​of the first loss function, the second loss function, and the third loss function. For example, the joint loss function L can be obtained based on formula (1):

[0085]

[0086] Among them, L LP Denotes the first loss function. This represents the second loss function. This represents the third loss function. KL stands for divergence, or relative entropy, which measures the degree of difference between two values.

[0087] This means that for a pre-labeled node, the features of that node are calculated using the second model to obtain a predicted user label g. i KL algorithm is used to calculate and predict user tag g. i With real user tags y i The discreteness, and through λ L Noise reduction is performed. This means that for users who are not pre-labeled, the second model is used to calculate the user's features to obtain a predicted user label g. i KL algorithm is used to calculate and predict user tag g. i Compared with the community label f ​​obtained through the first model i The discreteness, and through λ U Noise reduction is performed.

[0088] The first model is used to divide unlabeled users into communities and determine their corresponding second community label y. i ′ As predicted user labels, users are divided into communities, and a joint loss function is calculated. With the goal of minimizing the joint loss function, the community division of all users is completed, the second community to which each user belongs is obtained, the evaluation label of each user is obtained, and the first evaluation label of the first user is obtained.

[0089] As can be seen, the joint loss function given in this embodiment is used to determine the community partitioning result for users. Using the method given in this embodiment, when the first model is constructed using the LPA algorithm for initial community partitioning, and the processing result of the second model is used to update and correct the initial community partitioning of the first model, the overall computational cost of community partitioning is close to that of the LPA algorithm. In other words, the computational cost of the community partitioning method given in this embodiment is relatively low. Compared to algorithms such as Louvain and GNN, it has higher computational efficiency and requires far less CPU and GPU resources than existing algorithms like Louvain and GNN, effectively reducing computational costs while improving the accuracy of community partitioning. Furthermore, when constructing the first model based on the LPA algorithm, compared to other label propagation algorithms, it has the advantages of faster convergence speed, the ability to support complex overlapping communities, and no reduction in community partitioning accuracy. In practice, less than 10 rounds of training are needed. After multiple iterations, the model accuracy improves and becomes generally stable through the graph features obtained by the LPA algorithm.

[0090] In some embodiments, the method further includes: obtaining an evaluation model loss function by reducing the weight of easily classified samples in the binary classification loss function or increasing the weight of difficult-to-classify samples in the binary classification loss function; and optimizing the model parameters of the evaluation model by evaluating the model loss function.

[0091] In the above embodiments, the user data used to construct the training dataset typically consists of users who have already taken out loans and have demonstrated their behavior after taking out the loans (on-time repayment or delinquency). Therefore, user behavior after taking out loans usually has two labeled results: one is a defaulting user (delinquent user), and the other is a user who has made regular repayments. Thus, the evaluation model can be a binary classification model. In practical applications, the evaluation model can be built based on the LightGBM algorithm. LightGBM is a tree-based gradient boosting algorithm that is relatively fast in both training and prediction.

[0092] In real-world lending scenarios, due to stringent risk control and review processes, the statistical distribution of user numbers based on the two labeling results is often extremely unbalanced. Generally, the number of defaulting users is far less than the number of users making regular repayments; that is, defaulting users are the minority class, while regular repayment customers are the majority class. In other words, in the constructed training dataset, samples labeled as defaulting users are minority class samples, while samples labeled as regular repayment users are majority class samples.

[0093] For imbalanced labeling results, common methods include undersampling the majority class or synthesizing minority class data, such as the SMOTE and ADASYN algorithms, to bring the distribution of different labeled samples in the training dataset closer to equilibrium. However, both the SMOTE and ADASYN algorithms have drawbacks. For example, in practical applications, the ADASYN algorithm is often used in conjunction with undersampling techniques, randomly discarding a portion of the majority class samples. This approach may lead to the loss of important information about the majority class samples during model training.

[0094] The SMOTE algorithm is an improvement on the random oversampling algorithm. Its basic idea is to analyze minority class samples and artificially synthesize new samples based on these minority class samples, adding them to the dataset. Instead of simply repeating minority class samples, SMOTE generates new minority class samples in local regions using the K-nearest neighbor algorithm. However, the SMOTE algorithm has several drawbacks. First, the neighbor selection process is somewhat arbitrary. During the data synthesis algorithm, the value of K—how many nearest neighbor samples to select for synthesis—needs to be determined by the modelers. The optimal value of K is usually uncertain and can only be determined through trial and error, requiring significant experimental time. Second, SMOTE cannot address the data distribution problem in imbalanced datasets, easily leading to marginalization. Since the data synthesis algorithm synthesizes new data by selecting features from samples near the minority class samples, if a minority class sample happens to be on the edge of the minority class distribution, the newly synthesized sample generated from this minority class sample and its neighbors will also be in this marginal region, blurring the boundary between the minority and majority classes. While the SMOTE algorithm can improve the imbalanced distribution of datasets, this boundary ambiguity increases the difficulty of classification algorithms, affecting the evaluation of model training performance. Compared to the method combining the ADASYN algorithm with undersampling techniques, the method of synthesizing data using the SMOTE algorithm retains more information from the majority class samples, but it also introduces noisy data, potentially reducing model performance.

[0095] To improve the accuracy of the evaluation model and reduce the impact of imbalanced samples, this embodiment provides a method for determining the loss function of the evaluation model. The model parameters are optimized using the loss function provided in this embodiment. For example, when constructing an evaluation model based on the LightGBM algorithm, the original loss function of the LightGBM algorithm can be replaced using the loss function provided in this embodiment.

[0096] When evaluating user credit using an assessment model, a large portion of user data consists of easily classifiable samples. This means the assessment model can directly derive the assessment result from the majority of user data, resulting in typically low loss for the sample data based on user data. Due to the highly imbalanced data distribution, the relatively large number of easily classifiable samples may dominate the loss function of the assessment model. To make the assessment model focus more on difficult-to-classify samples, the method presented in this embodiment can construct a new loss function based on the original binary classification loss function. This new loss function can be achieved by either reducing the weight of easily classifiable samples or increasing the weight of difficult-to-classify samples in the original binary classification loss function.

[0097] For example, based on the cross-entropy loss function and the method given in this embodiment, the loss function LOSS shown in formula (2) below can be obtained as the loss function for evaluating the model:

[0098]

[0099] Where y represents the preset real annotation results in the training dataset, which can be overdue or not overdue. 1 represents overdue, and 0 represents not overdue. ′ y represents the label predicted by the evaluation model. ′ This can be represented as a probability value between 0 and 1. γ is a hyperparameter greater than 0, and its value can be determined by continuously adjusting the parameters to obtain the optimal value. Empirically, the initial value of γ can be set to 2. α is a hyperparameter between 0 and 1. In the credit scenario involved in this application, it can be set according to the sample proportion. The initial value of α can be determined based on the ratio of positive to negative samples.

[0100] In formula (2), (1-y ′ y is a modulation factor that is applied when a sample is misclassified, such as when an overdue sample is classified as a non-overdue sample by the evaluation model. ′ The value is very small, at which point the modulation factor (1-y) ′ When y is close to 1, it has little impact on the original loss value. ′ When it approaches 1, the modulation factor (1-y) ′ When the weighted value approaches 0, the loss value of easily classified samples is reduced. In actual business, a large number of negative samples in the evaluation model are easily classified samples. Easily classified samples (high-confidence samples) have very little effect on improving the processing capability of the evaluation model. The evaluation model should focus more on difficult-to-classify samples and reduce the loss of high-confidence samples. This can be achieved by combining parameters γ and α to effectively reduce the impact of high-confidence samples on the overall loss of the evaluation model.

[0101] As can be seen, the evaluation model loss function given in this embodiment can reduce the impact of imbalanced training data on model training during the evaluation model training process, and effectively improve the evaluation performance of the evaluation model. This embodiment optimizes the problem of imbalanced positive and negative sample distribution in credit scenarios. By using the custom evaluation model loss function in this embodiment, it avoids the data noise introduced when using the SMOTE or ADASYN algorithm to synthesize data, and significantly improves the performance of the evaluation model in handling imbalanced datasets.

[0102] In some embodiments, before determining the first community to which the first user belongs, the method further includes: obtaining the first basic data of the first user; the first basic data includes one or more of the first user's identity information and the first user's payment information with the operator; determining the second evaluation label of the first user based on the first basic data; the above-mentioned construction of a training dataset using the first evaluation label of the first user includes: constructing a training dataset using the first evaluation label of the first user and the second evaluation label of the first user.

[0103] This embodiment further provides a method for constructing a training dataset. The identity information of the first user includes, but is not limited to, the user's age and gender. When the user conducts credit assessment or applies for a loan via a mobile device, the user's payment information with the operator can also be obtained. This payment information includes, but is not limited to, one or more of the user's network access date, payment status, and historical service outage duration. A second assessment label for the first user is determined using this first basic data. When constructing the training dataset for the assessment model, the labels corresponding to the first user in the training dataset can be constructed based on the aforementioned first assessment label and the second assessment label in this embodiment, and the assessment model can be trained accordingly.

[0104] As can be seen, when a user conducts credit assessment or applies for a loan via a mobile device, the method presented in this embodiment can construct an assessment tag for the first user based on their relevant information in the operator's network, and then train the model. This eliminates the need for the user to input multiple pieces of information, improving the efficiency of obtaining the user's assessment tag and enhancing the diversity of user tags. Based on the method presented in this embodiment, multiple user data, such as basic information, outstanding payments, and whether the phone is suspended or switched off, can be combined to form tabular data features, enabling the assessment of customer credit from multiple dimensions.

[0105] In some embodiments, before determining the first user tag of the first user based on the first user's first social relationship data, the method further includes: obtaining the first user's first communication behavior data, and determining the first social relationship data based on the first communication behavior data.

[0106] When a user conducts a credit assessment or applies for a loan via a mobile device, the method provided in this embodiment can directly determine the first user's first social relationship by obtaining the first user's first communication behavior data in the operator's network. The first communication behavior data includes, but is not limited to, one or more of the first user's call behavior data, SMS behavior data, Internet communication information data, call details with financial institutions, and the number of contacts.

[0107] This application provides a model training method that uses modified, accurate community partitioning results as user evaluation labels to train the evaluation model, effectively improving the model's accuracy. Furthermore, by using tabular data and community partitioning results as the training dataset, the variety of sample data is greatly enriched, resulting in a better processing performance for the trained evaluation model.

[0108] In practical applications, when constructing the second model using a neural network, g can be selected as a two-layer neural network, with the first layer having 128 hidden units and the second layer having 64 hidden units. When constructing the first model based on the LPA algorithm and the second model based on the neural network, community partitioning is performed using both models. Compared to the existing LPA algorithm, this improves processing performance by 5% to 30% and effectively reduces noise in the graph features from the LPA algorithm. Furthermore, during community partitioning, the method described in this application exhibits higher convergence speed while maintaining accuracy, with computation time increasing by only about 20% to 30% compared to the LPA algorithm.

[0109] Based on the credit assessment method combining the first and second models and the assessment model given in this embodiment, taking the construction of the assessment model based on the LightGBM algorithm, the construction of the first model based on the LPA algorithm, and the construction of the second model based on the neural network as an example, the final model assessment results are obtained by evaluating the values ​​of the Area Under the ROC Curve (AUC) and the Kolmogorov-Smirnov statistic (KS), as shown in the table below:

[0110] Table 1

[0111]

[0112] The model training method provided in this application improves model performance without requiring more computation time and has a lower resource consumption rate.

[0113] Based on the model training method given in the above embodiments, Figure 2 A flowchart for evaluating model training is shown, including:

[0114] Step 201: Data extraction.

[0115] Specifically, user identity information, payment and arrears information, shutdown and shutdown behavior, communication behavior data, and APP behavior data can be extracted from the operator's entire network database.

[0116] Step 202: Data cleaning.

[0117] In the acquired operator data, abnormal data is cleaned and missing data is filled in.

[0118] Step 203: Determine user tags.

[0119] The processed data obtained in step 202 is divided into tabular data and social relationship data. Tabular data may include user identity information, payment and arrears information, shutdown and shutdown behavior, etc., while social relationship data may include communication behavior data, APP behavior data, etc.

[0120] The method for determining user tags in step 203 specifically includes:

[0121] Step 2031: Determine Tabular data.

[0122] In the processed data of step 202, Tabular data is determined, wherein Tabular data includes, but is not limited to, one or more of the user's identity information, payment and arrears information, and shutdown / shutdown behavior.

[0123] Step 2032: Determine the second evaluation label based on Tabular data.

[0124] The tabular data is processed to extract its features, resulting in the user's second evaluation label.

[0125] Step 2033: Determine social relationship data.

[0126] In the processed data of step 202, social relationship data is determined. This social relationship data includes, but is not limited to, one or more of the user's communication behavior data and app behavior data.

[0127] Step 2034: Determine the first evaluation label based on social relationship data.

[0128] The social relationship data is processed using the methods described in the above embodiments through the first model and the second model to obtain the first evaluation label.

[0129] The methods for obtaining the second evaluation label given in steps 2031 and 2032 can be performed simultaneously with the methods for obtaining the first evaluation label given in steps 2033 and 2034.

[0130] Step 204: Train the evaluation model.

[0131] Based on the first evaluation label and the second evaluation label obtained in step 203, the evaluation model is trained, and the model parameters of the evaluation model are optimized based on the evaluation model loss function given in the above embodiment.

[0132] Step 205: Model Deployment.

[0133] Specifically, the method described in this application embodiment can utilize the Spark framework for distributed computing within a Yarn cluster, improving computational efficiency. The Spark Structured Query Language (SparkSQL) module under the Spark framework can be used to process tabular data to obtain a second evaluation label, the platograph computation process under the Spark framework can generate a first evaluation label, and the PySpark module can be used to call the evaluation model.

[0134] Based on the model training method proposed in the foregoing embodiments, this application also provides an evaluation method, such as... Figure 3 As shown, the evaluation method proposed in this application includes:

[0135] Step 301: Determine the third community to which the second user belongs.

[0136] When a second user applies for a loan or credit assessment, the community to which the second user belongs can be divided based on the method given in step 101 above. Specifically, the community division can be based on algorithms such as LPA, Louvain's algorithm, and GNN. When the first model is built based on the LPA algorithm, the third community to which the second user belongs can be determined based on the first model.

[0137] Step 302: Determine the second user's second user tag based on the second user's second social relationship data.

[0138] Corresponding to the method given in step 102 above, the second user tag can be determined based on the second user's second social relationship data. Alternatively, based on the method given in the above embodiments, the second social relationship data can also be processed by a second model to obtain the second user tag.

[0139] Step 303: If the difference between the second user's tag and the third community tag of the third community is greater than or equal to the first threshold, update the community to which the second user belongs and determine the fourth community to which the second user belongs; based on the fourth community tag of the fourth community, determine the third evaluation tag of the second user.

[0140] Corresponding to the method given in step 103 above, since the third community tag can reflect the characteristics of all users in the third community to a certain extent, and the second user tag can reflect the social relationship characteristics of users, the third community tag and the second user tag should have a certain similarity.

[0141] If the difference between the second user's tag and the third community tag is greater than or equal to the first threshold, it is considered that there is an error in the community classification of the second user. In this case, the third community to which the second user belongs is re-determined until the difference between the third community tag and the second user's tag is less than the first threshold. The corresponding third community is then determined as the community to which the second user finally belongs, i.e., the fourth community. Here, the first threshold is the same as the first threshold in step 103 above.

[0142] Specifically, when updating the third community, the method given in the above embodiments can be used as a basis. If the third community to which the second user belongs is determined by the first model and the second user label of the second user is determined by the second model, the third community can be updated based on the joint loss function constructed in the above embodiments with the goal of minimizing the joint loss function, and the fourth community can be determined to obtain the third evaluation label of the second user.

[0143] Step 304: Process the third evaluation label using the evaluation model to obtain the credit evaluation result of the second user.

[0144] The evaluation model is trained using the model training method given in the above embodiments.

[0145] After obtaining the user's third evaluation label, the evaluation model obtained based on the model training method given in the above embodiment processes the third evaluation label to obtain the credit evaluation result of the second user.

[0146] As can be seen, the evaluation method provided in this application can fully consider the user's social relationships during the credit evaluation process, further evaluate the user's credit, and help obtain accurate credit evaluation results.

[0147] In some embodiments, before determining the third community to which the second user belongs, the method further includes: obtaining second basic data of the second user; the second basic data includes one or more of the second user's identity information and the second user's payment information with the operator; determining the second user's fourth evaluation label based on the second basic data; the above-mentioned processing of the third evaluation label through the evaluation model includes: processing the third evaluation label and the fourth evaluation label through the evaluation model.

[0148] Corresponding to the above-described model training method implementation, when a second user applies for a loan or conducts a credit assessment on a mobile device, the second user's second basic data can be directly obtained from the operator's network, and a fourth assessment label can be obtained based on the second basic data. The third and fourth assessment labels are then processed by the assessment model to obtain the second user's credit assessment result, which helps to further improve the accuracy of credit assessment.

[0149] In some embodiments, before determining the second user's second user tag based on the second user's second social relationship data, the method further includes: obtaining the second user's second communication behavior data, and determining the second social relationship data based on the second communication behavior data.

[0150] Corresponding to the above-described model training method embodiment, when the second user conducts credit assessment or applies for a loan via a mobile device, the method provided in this embodiment can directly determine the second user's second social relationship data by obtaining the second user's second communication behavior data from the operator's network. The second communication behavior data includes, but is not limited to, the second user's call behavior data, SMS behavior data, and Internet communication information data.

[0151] This application provides an assessment method that can perform credit assessments based on a user's social relationships, further improving the accuracy of the assessment results. It can be seen that when a user conducts a credit assessment or applies for a loan via a mobile device, the method provided in this application can obtain the user's social relationship data through the operator's network based on the user's identity information. Therefore, when a user actually applies for a loan or undergoes a credit assessment, they only need to enter their name, mobile phone number, and identity information, without needing to fill in excessive personal information. The method can obtain the user's relevant social relationship data through the operator's network based on the user's name, mobile phone number, and identity information. This simplifies the user's input while improving the accuracy of the credit assessment results and predicting the user's probability of delinquency in the future.

[0152] Based on the model training method proposed in the foregoing embodiments, this application also provides a model training apparatus, such as... Figure 4 As shown, Figure 4 The model training apparatus shown includes:

[0153] The first determining module 401 is used to determine the first community to which the first user belongs; and to determine the first user tag of the first user based on the first user's first social relationship data.

[0154] The first processing module 402 is used to update the community to which the first user belongs and determine the second community to which the first user belongs when the difference between the first user tag and the first community tag of the first community is greater than or equal to a first threshold; and determine the first evaluation tag of the first user based on the second community tag of the second community.

[0155] The training module 403 is used to construct a training dataset using the first user's first evaluation label, and to train the evaluation model based on the training dataset, so that the trained evaluation model can evaluate the user's credit.

[0156] In practical applications, the first determining module 401, the first processing module 402, and the training module 403 can be implemented based on a processor and a communication device.

[0157] In some embodiments, the first determining module 401 is specifically used to directly determine the first community to which the first user belongs based on the first social relationship data of the first user; the first processing module 402 is specifically used to update the first community to which the first user belongs based on the first social relationship data; and if the difference between the first community tag of the updated first community and the first user tag is less than a first threshold, the updated first community is determined as the second community to which the first user belongs.

[0158] In some embodiments, the first determining module 401 is specifically used to determine the first community to which the first user belongs through the first model; process the first social relationship data based on the second model to obtain the first user tag of the first user; before updating the community to which the first user belongs, the first processing module 402 is further used to obtain the first loss function of the first model and the second loss function of the second model; determine the third loss function based on the difference between the first user tag and the first community tag; construct a joint loss function based on the first loss function, the second loss function and the third loss function; and update the first community to which the first user belongs with the goal of minimizing the value of the joint loss function.

[0159] In some embodiments, the first processing module 402 is further configured to obtain an evaluation model loss function by reducing the weight of easily classified samples in the binary classification loss function or increasing the weight of difficult-to-classify samples in the binary classification loss function; the training module 403 is further configured to optimize the model parameters of the evaluation model by using the evaluation model loss function.

[0160] In some embodiments, the model training apparatus further includes a first acquisition module, which is used to acquire first basic data of the first user before determining the first community to which the first user belongs; the first basic data includes one or more of the first user's identity information and the first user's payment information with the operator; the first processing module 402 is further used to determine the second evaluation label of the first user based on the first basic data; the training module 403 is further used to construct a training dataset using the first evaluation label of the first user and the second evaluation label of the first user.

[0161] In some embodiments, before determining the first user tag of the first user based on the first user's first social relationship data, the first acquisition module is further configured to acquire the first user's first communication behavior data; the first processing module 402 is further configured to determine the first social relationship data based on the first communication behavior data.

[0162] Based on the evaluation method proposed in the foregoing embodiments, this application also provides an evaluation apparatus, such as... Figure 5 As shown, Figure 5 The model training apparatus shown includes:

[0163] The second determining module 501 is used to determine the third community to which the second user belongs; and to determine the second user's second user tag based on the second user's second social relationship data.

[0164] The second processing module 502 is used to update the community to which the second user belongs and determine the fourth community to which the second user belongs when the difference between the second user tag and the third community tag of the third community is greater than or equal to a first threshold; determine the third evaluation tag of the second user based on the fourth community tag of the fourth community; and process the third evaluation tag through an evaluation model to obtain the credit evaluation result of the second user; wherein the evaluation model is trained based on the model training method given in the above embodiment.

[0165] In practical applications, the second determining module 501 and the second processing module 502 can be implemented based on a processor and a communication device.

[0166] In some embodiments, the evaluation model further includes a second acquisition module, which is used to acquire second basic data of the second user before determining the third community to which the second user belongs; the second basic data includes one or more of the second user's identity information and the second user's payment information with the operator; a fourth evaluation label of the second user is determined based on the second basic data; and a second processing module 502 is specifically used to process the third evaluation label and the fourth evaluation label through the evaluation model.

[0167] In some embodiments, before determining the second user's second user tag based on the second user's second social relationship data, the second acquisition module is further configured to acquire the second user's second communication behavior data; the second processing module 502 is further configured to determine the second social relationship data based on the second communication behavior data.

[0168] It should be noted that the description of the above device embodiments is similar to the description of the above model training method embodiments, and has similar beneficial effects as the same method embodiments. For technical details not disclosed in the device embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0169] It should be noted that, in the embodiments of this application, if the above methods are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a terminal, server, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0170] This application also provides an electronic device. Figure 6 This is a schematic diagram of the composition structure of an electronic device provided in an embodiment of this application, as shown below. Figure 6 As shown, the electronic device 60 may include:

[0171] Memory 601 is used to store executable instructions.

[0172] The processor 602 is used to execute executable instructions stored in the memory 601 to implement any of the above-mentioned model training methods and evaluation methods.

[0173] The processor 602 mentioned above can be at least one of ASIC, DSP, DSPD, PLD, FPGA, CPU, model trainer, micro-model trainer, and microprocessor.

[0174] The aforementioned computer-readable storage medium or memory 601 may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM), etc.; it may also be various terminals that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0175] This application embodiment further provides a computer storage medium storing computer-executable instructions, which are used to implement any of the model training methods and evaluation methods provided in the above embodiments.

[0176] Correspondingly, this application embodiment further provides a computer program product, which includes computer-executable instructions for implementing any of the model training methods and evaluation methods provided in the above embodiments.

[0177] In some embodiments, the functions or modules of the apparatus, model trainer, and NTN base station provided in this application can be used to execute the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0178] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0179] The methods disclosed in the various method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0180] The features disclosed in the various product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0181] The features disclosed in the various method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0182] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0183] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these forms are within the protection scope of this application.

Claims

1. A model training method, characterized in that, The method includes: Determine the first community to which the first user belongs; The first user tag of the first user is determined based on the first user's first social relationship data; If the difference between the first user tag and the first community tag of the first community is greater than or equal to a first threshold, update the community to which the first user belongs and determine the second community to which the first user belongs; based on the second community tag of the second community, determine the first evaluation tag of the first user; A training dataset is constructed using the first user's first evaluation label. The evaluation model is then trained based on the training dataset, enabling the trained evaluation model to assess the user's credit.

2. The method according to claim 1, characterized in that, The process of determining the first community to which the first user belongs includes: Based on the first user's first social relationship data, the first community to which the first user belongs is directly determined; The step of updating the community to which the first user belongs and determining the second community to which the first user belongs includes: Based on the first social relationship data, update the first community to which the first user belongs; If the difference between the updated first community label and the first user label is less than the first threshold, the updated first community is determined to be the second community to which the first user belongs.

3. The method according to claim 1, characterized in that, The step of determining the first community to which the first user belongs includes: determining the first community to which the first user belongs through a first model; the step of determining the first user tag of the first user based on the first social relationship data of the first user includes: processing the first social relationship data based on a second model to obtain the first user tag of the first user; Before updating the community to which the first user belongs, the method further includes: Obtain the first loss function of the first model and the second loss function of the second model; Based on the difference between the first user tag and the first community tag, a third loss function is determined; A joint loss function is constructed based on the first loss function, the second loss function, and the third loss function; The update of the community to which the first user belongs includes: With the goal of minimizing the value of the joint loss function, the first community to which the first user belongs is updated.

4. The method according to claim 1, characterized in that, The method further includes: The evaluation model loss function is obtained by reducing the weight of easily classified samples in the binary classification loss function or increasing the weight of difficult-to-classify samples in the binary classification loss function. The model parameters of the evaluation model are optimized using the loss function of the evaluation model.

5. The method according to claim 1, characterized in that, Before determining the first community to which the first user belongs, the method further includes: Obtain the first basic data of the first user; the first basic data includes one or more of the first user's identity information and the first user's payment information with the operator; determine the second evaluation tag of the first user based on the first basic data; The step of constructing a training dataset using the first user's first evaluation label includes: A training dataset is constructed using the first user's first evaluation label and the first user's second evaluation label.

6. The method according to claim 1, characterized in that, Before determining the first user tag of the first user based on the first user's first social relationship data, the method further includes: Obtain the first user's first communication behavior data, and determine the first social relationship data based on the first communication behavior data.

7. An evaluation method, characterized in that, The method includes: Determine the third community to which the second user belongs; The second user tag of the second user is determined based on the second user's second social relationship data; If the difference between the second user's tag and the third community tag of the third community is greater than or equal to a first threshold, update the community to which the second user belongs and determine the fourth community to which the second user belongs; based on the fourth community tag of the fourth community, determine the third evaluation tag of the second user; The credit assessment result of the second user is obtained by processing the third assessment label through the assessment model; wherein the assessment model is trained based on the method described in any one of claims 1 to 6.

8. The method according to claim 7, characterized in that, Before determining the third community to which the second user belongs, the method further includes: Obtain the second basic data of the second user; the second basic data includes one or more of the second user's identity information and the second user's payment information with the operator; determine the fourth evaluation label of the second user based on the second basic data; The process of processing the third evaluation label using the evaluation model includes: The third and fourth evaluation labels are processed using an evaluation model.

9. The method according to claim 7, characterized in that, Before determining the second user's second user tag based on the second user's second social relationship data, the method further includes: Obtain the second user's second communication behavior data, and determine the second social relationship data based on the second communication behavior data.

10. A model training device, characterized in that, The device includes: The first determining module is used to determine the first community to which the first user belongs; and to determine the first user tag of the first user based on the first user's first social relationship data. The first processing module is configured to update the community to which the first user belongs and determine the second community to which the first user belongs when the difference between the first user tag and the first community tag of the first community is greater than or equal to a first threshold; and determine the first evaluation tag of the first user based on the second community tag of the second community. The training module is used to construct a training dataset using the first evaluation label of the first user, and to train the evaluation model based on the training dataset so that the trained evaluation model can evaluate the user's credit.

11. An evaluation device, characterized in that, The device includes: The second determining module is used to determine the third community to which the second user belongs; and to determine the second user's second user tag based on the second user's second social relationship data; The second processing module is configured to update the community to which the second user belongs and determine the fourth community to which the second user belongs when the difference between the second user tag and the third community tag of the third community is greater than or equal to a first threshold; determine the third evaluation tag of the second user based on the fourth community tag of the fourth community; and process the third evaluation tag through an evaluation model to obtain the credit evaluation result of the second user; wherein the evaluation model is trained based on the method described in any one of claims 1 to 6.

12. An electronic device, characterized in that, The electronic device includes a processor and a memory for storing computer programs capable of running on the processor; wherein, The processor is used to run the computer program to perform the method according to any one of claims 1 to 6, or any one of claims 7 to 9.

13. A computer storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method described in any one of claims 1 to 6, or any one of claims 7 to 9.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 6, or any one of claims 7 to 9.

Images