CN118798360A - Inference model training and credit attribute inference method, device and terminal - Google Patents

Abstract

The embodiment of the specification discloses an inference model training and credit attribute inference method, device and terminal. Determining a plurality of sample user features of a plurality of sample users and priori credit attributes corresponding to each sample user feature respectively, and determining loss weights of each sample user feature, wherein the loss weights of each sample user feature are positively correlated with the confidence of the priori credit attributes; inputting all sample user features into an initial inference model, controlling the initial inference model to output predicted credit attributes of each sample user feature, and adjusting parameters of the initial inference model according to the predicted credit attributes, priori credit attributes and loss weights of all sample user features; updating prior credit attributes and loss weights of target sample user characteristics meeting preset updating conditions in all sample user characteristics based on the adjusted initial inference model, and using the prior credit attributes and loss weights for next training of the initial inference model until the initial inference model converges to obtain a credit attribute inference model.

Description

Inference model training and credit attribute inference method, device and terminal

Technical Field

The embodiments of this specification relate to the field of computer technology, and in particular, to an inference model training and credit attribute inference method, device and terminal.

Background Art

In the context of financial credit, effectively managing the user's credit risk is undoubtedly a core and vital task. Generally speaking, the user's personal information and behavioral data on credit products can explain the user's credit habits and preferences. Therefore, when assessing the user's credit risk, financial institutions will abstract and quantify the user's personal information and behavioral data on credit products to obtain the user's credit attributes, and describe the user's credit characteristics through credit attributes. Therefore, in the huge system of risk management, understanding the user's credit attributes is particularly critical.

Summary of the invention

The embodiments of this specification provide an inference model training and credit attribute inference method, device and terminal, which can solve the technical problem that the credit attributes of users cannot be accurately inferred in the related art.

In a first aspect, an embodiment of the present specification provides an inference model training method, the method comprising:

Determine multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and determine the loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute;

Input all sample user characteristics into the initial inference model, control the initial inference model to output the predicted credit attributes of each sample user characteristic, and adjust the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user characteristics;

Determine the target sample user features that meet the preset update conditions among all sample user features, update the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and use all sample user features for the next training of the initial inference model after the update until the initial inference model converges to obtain a credit attribute inference model.

In a second aspect, an embodiment of this specification provides a credit attribute inference method, the method comprising:

Obtain user characteristics of target users;

Inputting the user characteristics into a credit attribute inference model to obtain the credit attributes corresponding to the target user output by the credit attribute inference model;

Providing services to the above target users based on the above credit attributes;

Among them, the above-mentioned credit attribute inference model is the credit attribute inference model obtained in the inference model training method provided in the above-mentioned first aspect.

In a third aspect, an embodiment of the present specification provides an inference model training device, the device comprising:

A sample acquisition module, used to determine multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and to determine the loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute;

A model training module, used to input all sample user features into an initial inference model, control the initial inference model to output the predicted credit attributes of each sample user feature, and adjust the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user features;

The sample update module is used to determine the target sample user features that meet the preset update conditions among all sample user features, update the prior credit attributes and loss weights of the above target sample user features based on the adjusted above initial inference model, and use all sample user features for the next training of the above initial inference model after the update until the above initial inference model converges to obtain a credit attribute inference model.

In a fourth aspect, an embodiment of the present specification provides a credit attribute inference device, the device comprising:

A user characterization module is used to obtain user characteristics of target users;

A model output module, used to input the user characteristics into a credit attribute inference model to obtain the credit attributes corresponding to the target user output by the credit attribute inference model;

A service provision module, used for providing services to the target users based on the credit attributes;

Among them, the above-mentioned credit attribute inference model is the credit attribute inference model obtained in the inference model training method provided in the above-mentioned first aspect.

In a fifth aspect, an embodiment of the present specification provides a computer program product comprising instructions, which, when the computer program product is run on a computer or a processor, enables the computer or the processor to execute the steps of the above method.

In a sixth aspect, an embodiment of the present specification provides a computer storage medium, wherein the computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the steps of the above method.

In a seventh aspect, an embodiment of the present specification provides a terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is suitable for being loaded by the processor and executing the steps of the method described above.

The beneficial effects brought by the technical solutions provided by some embodiments of this specification include at least:

The embodiments of the present specification provide an inference model training method, which determines multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and determines the loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute; inputs all sample user features into an initial inference model, controls the initial inference model to output the predicted credit attributes of each sample user feature, and adjusts the parameters of the initial inference model according to the predicted credit attributes, the prior credit attributes and the loss weight of all sample user features; determines the target sample user features that meet the preset update conditions among all sample user features, updates the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and uses all sample user features for the next training of the initial inference model after the update, until the initial inference model converges to obtain a credit attribute inference model. When obtaining samples, the prior credit attributes of sample users are used as their sample labels, and sample users will receive different loss weights based on the confidence of their prior credit attributes. The higher the confidence level of the sample, the higher the loss weight. Therefore, the contribution of the sample to the model loss value is adjusted according to the confidence level of the sample, so that the model training will rely more on sample optimization with high label accuracy to reduce the error caused by inaccurate labels. After each model parameter adjustment, the prior credit attributes and loss weights of the samples with low confidence will be updated, so that the prior credit attributes of the low-confidence samples can be corrected, thereby helping the model to rely on more and more high-confidence samples for fitting in the future, and ultimately obtaining better model inference effects.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

FIG1 is an exemplary system architecture diagram of an inference model training method provided in an embodiment of this specification;

FIG2 is a flow chart of an inference model training method provided in an embodiment of this specification;

FIG3 is a flow chart of an inference model training method provided in an embodiment of this specification;

FIG4 is a flow chart of a credit attribute inference method provided in an embodiment of this specification;

FIG5 is a structural block diagram of an inference model training device provided in an embodiment of this specification;

FIG6 is a structural block diagram of a credit attribute inference device provided in an embodiment of this specification;

FIG. 7 is a schematic diagram of the structure of a terminal provided in an embodiment of this specification.

DETAILED DESCRIPTION

In order to make the features and advantages of the embodiments of this specification more obvious and easy to understand, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the embodiments of this specification.

When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this specification. Instead, they are only examples of devices and methods consistent with some aspects of the embodiments of this specification as detailed in the attached claims.

In the context of financial credit, effectively managing and controlling the credit risk of users is undoubtedly a core and vital task. This is not only related to the sound operation of financial institutions, but also a manifestation of responsibility to every credit user. Generally speaking, a user's personal information and behavioral data on credit products can illustrate the user's credit habits and preferences. Therefore, when assessing a user's credit risk, the user's personal information and behavioral data on credit products can be abstracted and quantified to obtain the user's credit attributes, and the user's credit characteristics can be described by the specific values of the credit attributes. Therefore, in the huge system of risk management, it is particularly important to understand the user's credit attributes.

The credit attributes of users reflect rich information and deep meanings with simple numerical values. They are not only a list of basic information about users, but also an accurate portrayal of the comprehensive credit status of users. When generating credit attributes for each user, it is usually based on the user's personal information and information on the use of credit products. Among them, the user's personal information generally includes the user's occupational information, such as the industry and position he is engaged in, which reflects the user's career development prospects; social attributes, such as the user's family situation, which reveals the user's social role and responsibilities; and behavioral preferences, such as the user's consumption habits and borrowing frequency, which show the user's activity and risk preference in the credit field. The user's use information of credit products comes from a large amount of data traces left by the user in the process of using the credit product. These data include but are not limited to the user's borrowing behavior data, such as the loan amount, repayment record, etc., which can intuitively reflect the user's repayment ability and credit status; occupational information, such as work unit, job change, etc., can provide a reference for the user's career stability; and asset information, such as real estate, vehicles, etc., provide important material guarantees for evaluating the user's repayment ability.

The most common credit attribute calculation methods currently use in-depth mining and analysis of existing user data, and use logical rules in related transaction scenarios to infer the credit attributes of the users corresponding to these data. These users are then used as samples and the corresponding credit attributes are used as sample labels to input into the model for training. A model for inferring credit attributes is obtained, and then the converged model is used to infer the credit attributes of other users without credit attributes.

However, due to missing data, inconsistent data for each user, and untimely updates, the credit attributes inferred by logical rules are not completely accurate, resulting in some inaccurate and wrong sample labels in the modeled samples, resulting in low sample credibility. Training a model on erroneous samples will make the model's evaluation performance inaccurate, and may even amplify the error, resulting in low credibility of the final model's output results.

Therefore, the embodiments of this specification provide a method for inference model training and credit attribute inference to solve the above-mentioned technical problem of not being able to accurately infer the credit attributes of users.

Please refer to Figure 1, which is an exemplary system architecture diagram of an inference model training method provided in an embodiment of this specification.

As shown in FIG1 , the system architecture may include a terminal 101, a network 102, and a server 103. The network 102 is used to provide a medium for a communication link between the terminal 101 and the server 103. The network 102 may include various types of wired communication links or wireless communication links, for example, the wired communication link includes an optical fiber, a twisted pair, or a coaxial cable, and the wireless communication link includes a Bluetooth communication link, a Wireless-Fidelity (Wi-Fi) communication link, or a microwave communication link.

The terminal 101 can interact with the server 103 through the network 102 to receive messages from the server 103 or send messages to the server 103, or the terminal 101 can interact with the server 103 through the network 102 to receive messages or data sent by other users to the server 103. The terminal 101 can be hardware or software. When the terminal 101 is hardware, it can be various electronic devices, including but not limited to smart watches, smart phones, tablet computers, laptop portable computers and desktop computers. When the terminal 101 is software, it can be installed in the electronic devices listed above, which can be implemented as multiple software or software modules (for example: used to provide distributed services), or it can be implemented as a single software or software module, which is not specifically limited here.

In an embodiment of the present specification, first, the terminal 101 determines multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and determines the loss weight of each sample user feature, and the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute; further, the terminal 101 inputs all sample user features into the initial inference model, controls the initial inference model to output the predicted credit attributes of each sample user feature, and adjusts the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user features; after the model is adjusted once, the terminal 101 determines the target sample user features that meet the preset update conditions among all sample user features, updates the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and uses all sample user features for the next training of the initial inference model after the update, until the initial inference model converges to obtain a credit attribute inference model.

The server 103 may be a business server that provides various services. It should be noted that the server 103 may be hardware or software. When the server 103 is hardware, it may be implemented as a distributed server cluster consisting of multiple servers, or it may be implemented as a single server. When the server 103 is software, it may be implemented as multiple software or software modules (for example, for providing distributed services), or it may be implemented as a single software or software module, which is not specifically limited here.

Alternatively, the system architecture may not include the server 103. In other words, the server 103 may be an optional device in the embodiments of this specification, that is, the method provided in the embodiments of this specification may be applied to a system structure that only includes the terminal 101, and the embodiments of this specification do not limit this.

It should be understood that the number of terminals, networks and servers in FIG1 is only illustrative, and any number of terminals, networks and servers may be used according to implementation requirements. Similarly, the exemplary system architecture diagram of a credit attribute inference method provided in the embodiments of this specification is also similar to the exemplary system architecture diagram of the inference model training method in FIG1, and will not be repeated here.

Please refer to Figure 2, which is a flowchart of an inference model training method provided in an embodiment of this specification. The execution subject of the embodiment of this specification can be a terminal that executes the inference model training method, or a processor in the terminal that executes the inference model training method, or an inference model training service in the terminal that executes the inference model training method. For the convenience of description, the specific execution process of the inference model training method is introduced below by taking the execution subject as the processor in the terminal as an example.

As shown in FIG2 , the inference model training method may at least include:

S202, determining multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and determining the loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute.

Optionally, samples are used as the basis for model learning. Before training the model, it is necessary to first determine multiple sample user features of multiple sample users. The sample user features are generally embodied in the form of embedded features that can be input into the model. When using sample user features to train the model, the sample user features also need to have a priori standard labels. The standard labels are used to describe the samples realistically and accurately. After the model outputs the predicted labels for the sample user features, the predicted labels are compared with the standard labels of the sample user features to obtain the difference between the two values, that is, the loss value. The model further adjusts its parameters according to the loss value so that its prediction results fit in the direction of the standard labels. When the standard labels are correct, the model gradually obtains better prediction performance until convergence.

In the embodiments of this specification, it is necessary to determine the prior credit attributes corresponding to each sample user feature for multiple sample user features. The prior credit attributes serve as standard labels for the sample user features and are subsequently used to guide the fitting direction of the model.

However, in actual application scenarios, each user may have inconsistent data, missing data, or untimely updates, so the credit attributes inferred by logical rules are not completely accurate, resulting in inaccurate and erroneous sample labels for the credit attributes of these users after they are modeled as samples, leading to low sample credibility. Training a model on erroneous samples will make the model's evaluation performance inaccurate, and may even amplify the error, resulting in low credibility of the final model's output results.

Optionally, in order to reduce the misleading effect of samples with low confidence in credit attributes on the model and increase the attention of samples with high confidence to the model, it is necessary to consider the loss weight of each sample user feature, so that the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute, that is, high-confidence samples can obtain high loss weights, while low-confidence samples have low corresponding loss weights. In this way, the loss weights are used to influence the contribution of samples with different confidence levels to the loss when calculating the loss value, thereby avoiding misleading the model fitting due to inaccurate information, making the model training more dependent on sample optimization with high label accuracy, and reducing the errors caused by inaccurate labels.

S204. Input all sample user features into the initial inference model, control the initial inference model to output the predicted credit attributes of each sample user feature, and adjust the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user features.

Optionally, after determining the sample user characteristics and determining the prior credit attributes and loss weights corresponding to each sample user characteristic, all sample user characteristics can be input into the initial inference model, and the initial inference model can be controlled to output the predicted credit attributes of each sample user characteristic. Then, the loss value is calculated based on the predicted credit attributes, prior credit attributes and loss weights of all sample user characteristics, and the parameters of the initial inference model are determined based on the loss value.

Optionally, the basic model based on which the initial inference model in the embodiment of this specification can be a tree model, and the model algorithm used can be a LightGBM (Light Gradient Boosting Machine) algorithm, which is an efficient and scalable machine learning algorithm based on a gradient boosted decision tree (Gradient Boosted Decision Trees, GBDT). As one of the algorithms of the GBDT framework, its advantage is that it can solve the problem of low computational efficiency of the GBDT algorithm framework when processing massive data, and improve the computational efficiency of GBDT by nearly 20 times at the cost of sacrificing extremely small computational accuracy. LightGBM Tonggu uses a technology called "histogram-based learning" to efficiently process data, which makes LightGBM very fast in training on large-scale data sets, so it is particularly useful when processing large amounts of data. In addition to the tree model, the initial inference model can also be constructed based on some common logistic regression models, and can also be constructed based on a combination of multiple models. Different basic models differ only in model principles and parameter adjustment methods, so the initial inference model does not limit the scope of implementation of the embodiment of this specification of the basic model.

S206. Determine the target sample user features that meet the preset update conditions among all sample user features, update the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and use all sample user features for the next training of the initial inference model after the update until the initial inference model converges to obtain a credit attribute inference model.

Optionally, the model often needs to go through multiple rounds of training and parameter adjustment to achieve the target performance requirements. When constructing samples, in addition to some high-confidence samples obtained due to complete and accurate data, there are also a large number of low-confidence samples with inaccurate prior credit attributes. The loss weight of these low-confidence samples is low, so the model will rely more on high-confidence samples, but the number of high-confidence samples is only a small part relative to all samples. With a small number of samples, the model either fails to converge during training and underfits, or learns the characteristics of the sample itself and overfits. Either case will result in poor model performance. Therefore, it is necessary to ensure that the model is learned based on a sufficient number of high-confidence samples to ensure the final inference effect of the model.

Optionally, during the model training process, the prior credit attributes and loss weights of some low-confidence sample user features can also be updated so that these low-confidence sample user features can be corrected. Then, as the confidence of the low-confidence sample user features increases and their weights change, all sample user features will continue to be used for the next training of the initial inference model after the update. As the model undergoes rounds of parameter adjustments, the low-confidence sample user features will gradually tend towards high-confidence iterations, and the sample user features with higher confidence will be used for model training with the newly obtained weights, so that the model relies on more and more high-confidence samples for fitting in subsequent stages, and ultimately the model achieves convergence to the target performance based on a sufficient number of reliable samples.

Specifically, among all sample user features, not all low-confidence sample user features may need to be updated. For example, if the credit attribute predicted by the model for the low-confidence sample user feature is not much different from its prior credit attribute, it means that the predicted credit attribute is not accurate enough and the confidence is not high. In this case, the update is unnecessary. If the credit attribute predicted by the model is different from its prior credit attribute, it is possible that the predicted credit attribute is more reliable than the prior credit attribute. Therefore, the update conditions for screening sample user features should be set in advance. Every time the model is adjusted once, the target sample user features that meet the preset update conditions are determined among all sample user features. Then, the model at this time is used to update the prior credit attribute and loss weight of the target sample user features. After the update, all sample user features are used for the next training of the initial inference model until the initial inference model converges to obtain the credit attribute inference model. In this way, by repeatedly iteratively correcting the weights and credit attribute labels of low-confidence samples, the model can learn from more samples with high accuracy, thereby improving the effect and generalization ability of the model.

In an embodiment of the present specification, an inference model training method is provided, which determines multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and determines the loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute; inputs all sample user features into an initial inference model, controls the initial inference model to output the predicted credit attributes of each sample user feature, and adjusts the parameters of the initial inference model according to the predicted credit attributes, the prior credit attributes and the loss weights of all sample user features; determines the target sample user features that meet the preset update conditions among all sample user features, updates the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and uses all sample user features for the next training of the initial inference model after the update, until the initial inference model converges to obtain a credit attribute inference model. When obtaining samples, the prior credit attributes of sample users are used as their sample labels, and sample users will receive different loss weights based on the confidence of their prior credit attributes. The higher the confidence level of the sample, the higher the loss weight. Therefore, the contribution of the sample to the model loss value is adjusted according to the confidence level of the sample, so that the model training will rely more on sample optimization with high label accuracy to reduce the error caused by inaccurate labels. After each model parameter adjustment, the prior credit attributes and loss weights of the samples with low confidence will be updated, so that the prior credit attributes of the low-confidence samples can be corrected, thereby helping the model to rely on more and more high-confidence samples for fitting in the future, and ultimately obtaining better model inference effects.

Please refer to FIG. 3 , which is a flow chart of an inference model training method provided in an embodiment of this specification.

As shown in FIG3 , the inference model training method may at least include:

S302: If the initial inference model has not been trained at this time, multiple sample user features of the multiple sample users and the prior credit attributes corresponding to each sample user feature are determined based on the original user data of the multiple sample users and the credit attribute logic rules.

Optionally, when initially constructing the sample, it is necessary to embed feature representations of each sample user based on the original user data of each sample user to obtain sample user features of each sample user. At the same time, based on the original user data and credit attribute logic rules, the initial prior credit attributes of each sample user feature can be inferred.

S304. Determine the loss weight of each sample user feature based on the prior credit attribute of each sample user feature.

Optionally, since the original user data may not be complete and accurate, the prior credit attributes of each sample user feature may not be accurate. In order to reduce the impact of low-confidence samples on the model, it is necessary to assign different loss weights to sample user features with different confidence levels, where the loss weight of each sample user feature is positively correlated with the confidence level of the prior credit attribute. Specifically, it is necessary to determine the loss weight of each sample user feature based on the prior credit attribute of each sample user feature.

In a feasible implementation, considering that in financial scenarios, the evaluation dimension of user credit attributes is usually not single, but is evaluated from multiple dimensions such as the user's borrowing assets, industry, geographic location, interest preferences, etc. According to experience, in each dimension, the distribution of users' credit attributes is close to Gaussian distribution, that is, in each dimension, the number of users with credit attributes in the middle is the largest, while the number of users with extreme values is small. For this reason, the distribution of credit attributes of all users can be regarded as the superposition of multiple Gaussian distributions, that is, Gaussian mixed distribution. This means that if a user's credit attributes clearly belong to the Gaussian distribution of a certain dimension, then the certainty that it belongs to this dimension is relatively high, and then the credit attributes of this user are more reliable and have a high confidence level.

Optionally, when determining the loss weight of each sample user feature through the prior credit attribute of each sample user feature, the Gaussian mixture distribution corresponding to the credit attribute scenario can be preset first, and the multiple Gaussian distributions included in the preset Gaussian mixture distribution are all obtained based on multiple dimensions of the credit attribute. Specifically, a clustering model can be used to analyze the credit attribute scenario to obtain several obvious Gaussian distributions in the credit attribute scenario.

Furthermore, the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution in the preset Gaussian mixture distribution is determined, and then the loss weight of each sample user feature is calculated according to the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution. The loss weight of each sample user feature is positively correlated with the highest distribution probability among the probabilities that the prior credit attribute belongs to each Gaussian distribution. If the probability that the prior credit attribute of the sample user feature belongs to a certain distribution is significantly higher than that of other distributions, for example, there are a total of 3 Gaussian distributions in the preset Gaussian mixture distribution, and the probability that the prior credit attribute of the sample user feature belongs to the first Gaussian distribution is 70%, and the probabilities of belonging to the other two Gaussian distributions are 22% and 8% respectively, then it means that the certainty that the prior credit attribute of the sample user feature belongs to the dimension of the first Gaussian distribution is relatively high, and then the confidence of this credit attribute is high, and accordingly, this sample user feature should also receive a higher loss weight. However, if the prior credit attributes of the sample user characteristics do not obviously belong to a certain Gaussian distribution, that is, the probabilities in multiple Gaussian distributions are similar, for example, the probabilities that the prior credit attributes of the sample user characteristics belong to three Gaussian distributions are 40%, 30% and 30% respectively, then it means that the confidence level of the prior credit attributes of the sample user characteristics is not high.

S306. Input all sample user features into an initial inference model, control the initial inference model to output the predicted credit attributes of each sample user feature, and adjust the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user features.

Regarding step S306, please refer to the detailed description in step S204, which will not be repeated here.

S308: Determine at least one low-confidence sample user feature whose loss weight is less than a preset weight threshold among all sample user features.

Optionally, when determining the target sample user features that meet the preset update conditions among all sample user features, it is necessary to first select the sample user features with low confidence for updating, and the sample user features that originally have high confidence do not need to be updated and iterated. Then, since the confidence of the sample user features is also reflected in their loss weights, the confidence of the sample user features with high loss weights is relatively high, while the confidence of the sample user features with low loss weights is relatively low. Based on this, at least one low-confidence sample user feature whose loss weight is less than the preset weight threshold can be determined among all sample user features.

Optionally, the preset weight threshold can be set to 0.7, that is, the loss weight less than 0.7 is determined as a low-confidence sample user feature, and the loss weight greater than or equal to 0.7 is determined as a high-confidence sample user feature. However, it should be noted that the preset weight threshold can be set according to the needs of the actual scenario, and the embodiments of this specification do not limit its specific data.

S310, inputting each low-confidence sample user feature into the adjusted initial inference model, and controlling the adjusted initial inference model to output the modified credit attributes of each low-confidence sample user feature.

S312: Determine the current existing prior credit attributes of each low-confidence sample user feature, and determine the low-confidence sample user feature whose difference between the revised credit attribute and the existing prior credit attribute is greater than a preset difference threshold as the target sample user feature.

Optionally, after selecting the low-confidence sample user features, it is necessary to further select the sample user features that can be updated this time. If the credit attributes predicted by the model for the low-confidence sample user features are not much different from their prior credit attributes, it means that the predicted credit attributes are not accurate enough and the confidence is not high. In this case, the update is unnecessary. If the model predicts a certain difference between its predicted credit attributes and its prior credit attributes, it is possible that the predicted credit attributes are more reliable than the prior credit attributes. Then input each low-confidence sample user feature into the adjusted initial inference model, and control the adjusted initial inference model to output the corrected credit attributes of each low-confidence sample user feature. The next step is to determine the sample user features that need to be updated based on the corrected credit attributes.

Specifically, the current existing prior credit attributes of each low-confidence sample user feature are determined, and the difference between the revised credit attribute and the existing prior credit attribute is calculated. If the difference is greater than the preset difference threshold, the sample user feature is determined as the target sample user feature; conversely, if the difference is less than or equal to the preset difference threshold, it means that the result output by the model at this time is not much different from the low-confidence prior credit attribute, and the sample user feature will not be updated temporarily.

It should be noted that, in order to facilitate the training and calculation of the model, the credit attributes can be expressed in the form of numerical values with a value range of (0,1), and the above-mentioned preset difference threshold can be set to 0.1, that is, when the difference between the modified credit attributes of the sample user characteristics and the existing prior credit attributes reaches more than 0.1, it is considered that the sample user characteristics can be modified once.

S314. Calculate the prior credit attributes of the target sample user features and increase the loss weight of the target sample user features based on the modified credit attributes of the target sample user features and the existing prior credit attributes of the target sample user features; and use all sample user features for the next training of the initial inference model after the update.

Optionally, for the target sample user feature, the prior credit attribute of the target sample user feature is recalculated and the loss weight of the target sample user feature is increased according to the revised credit attribute of the target sample user feature and the existing prior credit attribute of the target sample user feature. Specifically, the average of the revised credit attribute and the existing prior credit attribute can be taken as the new prior credit attribute of the target sample user feature. For example, if the revised credit attribute of the target sample user feature is 0.4 and its existing prior credit attribute is 0.2, then the new prior credit attribute of the target sample user feature is the average value of 0.3 after adding 0.4 and 0.2, so as to ensure smooth updating of the prior credit attribute. As the confidence of the prior credit attributes of the target sample user characteristics increases, its loss weight should also increase to improve the fitting effect of this part of the samples during model training. However, it should be noted that because the target sample user characteristics are originally low-confidence samples, their weights are not increased to the same as those of high-confidence samples, but are between their initial weights and the weights of high-confidence samples. The greater the deviation between the revised credit attributes of the target sample user characteristics and their existing prior credit attributes, the smaller the increased weight, and vice versa.

S316. Determine the number of low-confidence sample user features whose difference between the revised credit attribute and the existing prior credit attribute is less than or equal to a preset difference threshold; if the number meets the preset correction condition, determine that the adjusted initial inference model has converged.

Optionally, as the model is trained and low-confidence samples are updated, if the difference between the revised credit attribute and the existing prior credit attribute of most low-confidence sample user features is less than a preset difference threshold, it means that the model has converged and training can be stopped. That is, the number of low-confidence sample user features whose difference between the revised credit attribute and the existing prior credit attribute is less than or equal to the preset difference threshold is determined. If the number meets the preset correction condition, for example, the number reaches 80% of the number of all low-confidence sample user features, it can be determined that the adjusted initial inference model has converged.

In an embodiment of the present specification, an inference model training method is provided to determine the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution in a preset Gaussian mixture distribution, and then calculate the loss weight of each sample user feature based on the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution. In this way, by adding prior knowledge to the sample using the Gaussian mixture distribution method, a method for evaluating the accuracy of the sample is provided, and the sample accuracy is incorporated into the model loss function in the form of the loss weight of the sample, thereby reducing the model error caused by inaccurate samples. The prior credit attributes and loss weights of the target sample user features are updated based on the adjusted initial inference model, and all sample user features are used for the next training of the initial inference model after the update, thereby increasing the high-accuracy samples during training and improving the effect and generalization ability of the model. In the case where the sample accuracy is unknown, the inference effect of the model can still be improved, and inaccurate samples can also be corrected at the same time.

Please refer to FIG4, which is a flowchart of a credit attribute inference method provided in an embodiment of this specification. The execution subject of the embodiment of this specification can be a terminal that executes credit attribute inference, a processor in the terminal that executes the credit attribute inference method, or a credit attribute inference service in the terminal that executes the credit attribute inference method. For the convenience of description, the specific execution process of the credit attribute inference method is introduced below by taking the execution subject as a processor in a terminal as an example.

As shown in FIG4 , the credit attribute inference method may at least include:

S404: Obtain user characteristics of the target user.

Optionally, the credit attribute inference model obtained in the inference model training method provided in the above embodiments of this specification can be deployed in an actual scenario of credit attribute inference. In the application scenario, if the user has complete and accurate original user data, then you can choose to use credit attribute logic rules to infer the user's credit attributes; and if the user's original user data has problems such as insufficient data, missing key data, or inaccurate data, then you can use the credit attribute inference model deployed in the scenario to infer the user's credit attributes. When using the credit attribute inference model to infer credit attributes, you first need to obtain the user characteristics of the target user.

S404: Input the user characteristics into the credit attribute inference model to obtain the credit attributes corresponding to the target user output by the credit attribute inference model.

S406: Provide services to target users based on credit attributes.

Optionally, the user characteristics are input into the credit attribute inference model, which analyzes the user characteristics and can directly output the credit attributes corresponding to the target user. Furthermore, more appropriate financial credit services can be provided to the target user directly based on the credit attributes, thereby improving the user experience.

In an embodiment of the present specification, a credit attribute inference method is provided, which obtains the user characteristics of a target user; inputs the user characteristics into a credit attribute inference model to obtain the credit attributes corresponding to the target user output by the credit attribute inference model; and provides services to the target user based on the credit attributes; wherein the credit attribute inference model is a credit attribute inference model obtained in the inference model training method provided in the above embodiment of the present specification. Since the credit attribute inference model is trained with samples of unknown accuracy, and inaccurate samples are corrected during the training process, the credit attribute inference model has efficient and accurate credit attribute inference performance, and the credit attribute inference model can be used to accurately infer the credit attributes of the target user, thereby providing the target user with more appropriate financial credit services and improving the user experience.

Please refer to Figure 5, which is a structural block diagram of an inference model training device provided in an embodiment of this specification. As shown in Figure 5, the inference model training device 500 includes:

The sample acquisition module 510 is used to determine a plurality of sample user features of a plurality of sample users and a priori credit attribute corresponding to each sample user feature, and determine a loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the priori credit attribute;

Model training module 520, used to input all sample user features into the initial inference model, control the initial inference model to output the predicted credit attribute of each sample user feature, and adjust the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user features;

The sample update module 530 is used to determine the target sample user features that meet the preset update conditions among all sample user features, update the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and use all sample user features for the next training of the initial inference model after the update until the initial inference model converges to obtain a credit attribute inference model.

Optionally, the sample acquisition module 510 is also used to determine multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature based on the original user data of multiple sample users and credit attribute logic rules if the initial inference model has not been trained at this time; and determine the loss weight of each sample user feature through the prior credit attributes of each sample user feature.

Optionally, the sample acquisition module 510 is also used to determine the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution in a preset Gaussian mixture distribution, wherein the preset Gaussian mixture distribution includes multiple Gaussian distributions obtained based on multiple dimensions of credit attributes; according to the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution, the loss weight of each sample user feature is calculated, and the loss weight of each sample user feature is positively correlated with the highest distribution probability among the probabilities that the prior credit attribute belongs to each Gaussian distribution.

Optionally, the sample update module 530 is also used to determine at least one low-confidence sample user feature among all sample user features whose loss weight is less than a preset weight threshold; input each low-confidence sample user feature into the adjusted initial inference model, and control the adjusted initial inference model to output the corrected credit attributes of each low-confidence sample user feature; determine the current existing prior credit attributes of each low-confidence sample user feature, and determine the low-confidence sample user feature whose difference between the corrected credit attribute and the existing prior credit attribute is greater than a preset difference threshold as the target sample user feature.

Optionally, the sample updating module 530 is further used to calculate the prior credit attributes of the target sample user characteristics and to increase the loss weight of the target sample user characteristics according to the revised credit attributes of the target sample user characteristics and the existing prior credit attributes of the target sample user characteristics.

Optionally, the inference model training device 500 also includes: a model convergence judgment module, used to determine the number of low-confidence sample user features whose difference between the revised credit attribute and the existing prior credit attribute is less than or equal to a preset difference threshold; if the number meets the preset correction condition, it is determined that the adjusted initial inference model has converged.

Optionally, the model training module 520 is further used to calculate the loss value according to the predicted credit attributes, prior credit attributes and loss weights of all sample user characteristics, and initially infer the parameters of the model based on the loss value.

Please refer to FIG6 , which is a structural block diagram of a credit attribute inference device provided in an embodiment of this specification. As shown in FIG6 , the credit attribute inference device 600 includes:

A user characterization module 610 is used to obtain user characteristics of a target user;

Model output module 620, used to input user characteristics into the credit attribute inference model to obtain the credit attribute corresponding to the target user output by the credit attribute inference model;

A service providing module 630, for providing services to target users based on credit attributes;

The credit attribute inference model is the credit attribute inference model obtained by the inference model training method provided in the above embodiments of this specification.

The embodiments of this specification provide a computer program product including instructions. When the computer program product is executed on a computer or a processor, the computer or the processor executes the steps of any one of the methods in the above embodiments.

The embodiments of this specification also provide a computer storage medium, which can store multiple instructions, and the instructions are suitable for being loaded by a processor and executing the steps of any method in the above embodiments.

Please refer to Figure 7, which is a schematic diagram of the structure of a terminal provided in an embodiment of this specification. As shown in Figure 7, the terminal 700 may include: at least one terminal processor 701, at least one network interface 704, a user interface 703, a memory 705, and at least one communication bus 702.

The communication bus 702 is used to realize the connection and communication between these components.

The user interface 703 may include a display screen (Display) and a camera (Camera), and the optional user interface 703 may also include a standard wired interface and a wireless interface.

The network interface 704 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

Among them, the terminal processor 701 may include one or more processing cores. The terminal processor 701 uses various interfaces and lines to connect various parts in the entire terminal 700, and executes various functions and processes data of the terminal 700 by running or executing instructions, programs, code sets or instruction sets stored in the memory 705, and calling data stored in the memory 705. Optionally, the terminal processor 701 can be implemented in at least one hardware form of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field-Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA). The terminal processor 701 can integrate one or a combination of a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU) and a modem. Among them, the CPU mainly processes the operating system, user interface and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the terminal processor 701, and it can be implemented separately through a chip.

The memory 705 may include a random access memory (RAM) or a read-only memory (ROM). Optionally, the memory 705 includes a non-transitory computer-readable storage medium. The memory 705 may be used to store instructions, programs, codes, code sets or instruction sets. The memory 705 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 705 may also be at least one storage device located away from the aforementioned terminal processor 701. As shown in FIG. 7 , the memory 705 as a computer storage medium may include an operating system, a network communication module, a user interface module, and an inference model training program or a credit attribute inference program.

In a feasible implementation manner, in the terminal 700 shown in FIG. 7 , the user interface 703 is mainly used to provide an input interface for the user and obtain data input by the user; and the terminal processor 701 can be used to call the inference model training program stored in the memory 705, and specifically perform the following operations:

Determine multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and determine the loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute;

Input all sample user features into the initial inference model, control the initial inference model to output the predicted credit attributes of each sample user feature, and adjust the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user features;

Determine the target sample user features that meet the preset update conditions among all sample user features, update the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and use all sample user features for the next training of the initial inference model after the update until the initial inference model converges to obtain a credit attribute inference model.

In some embodiments, when the terminal processor 701 determines multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and determines the loss weight of each sample user feature, it specifically performs the following steps: if the initial inference model has not been trained at this time, then based on the original user data of the multiple sample users and the credit attribute logic rules, determine multiple sample user features of the multiple sample users and the prior credit attributes corresponding to each sample user feature; determine the loss weight of each sample user feature through the prior credit attributes of each sample user feature.

In some embodiments, when the terminal processor 701 determines the loss weight of each sample user feature through the prior credit attribute of each sample user feature, it specifically performs the following steps: determining the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution in a preset Gaussian mixture distribution, wherein the preset Gaussian mixture distribution includes multiple Gaussian distributions obtained based on multiple dimensions of credit attributes; calculating the loss weight of each sample user feature according to the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution, and the loss weight of each sample user feature is positively correlated with the highest distribution probability among the probabilities that the prior credit attribute belongs to each Gaussian distribution.

In some embodiments, when the terminal processor 701 determines the target sample user features that meet the preset update conditions among all sample user features, it specifically performs the following steps: determining at least one low-confidence sample user feature among all sample user features whose loss weight is less than a preset weight threshold; inputting each low-confidence sample user feature into the adjusted initial inference model, and controlling the adjusted initial inference model to output the corrected credit attributes of each low-confidence sample user feature; determining the current existing prior credit attributes of each low-confidence sample user feature, and determining the low-confidence sample user feature whose difference between the corrected credit attribute and the existing prior credit attribute is greater than the preset difference threshold as the target sample user feature.

In some embodiments, when the terminal processor 701 updates the prior credit attributes and loss weights of the target sample user characteristics based on the adjusted initial inference model, it specifically performs the following steps: based on the revised credit attributes of the target sample user characteristics and the existing prior credit attributes of the target sample user characteristics, calculate the prior credit attributes of the target sample user characteristics and increase the loss weights of the target sample user characteristics.

In some embodiments, before determining the low-confidence sample user features whose difference between the revised credit attributes and the existing prior credit attributes is greater than a preset difference threshold as the target sample user features, the terminal processor 701 further specifically performs the following steps: determining the number of low-confidence sample user features whose difference between the revised credit attributes and the existing prior credit attributes is less than or equal to the preset difference threshold; if the number meets the preset correction condition, determining that the adjusted initial inference model has converged.

In some embodiments, when the terminal processor 701 adjusts the parameters of the initial inference model based on the predicted credit attributes, prior credit attributes and loss weights of all sample user characteristics, it specifically performs the following steps: calculate the loss value based on the predicted credit attributes, prior credit attributes and loss weights of all sample user characteristics, and initially infer the parameters of the model based on the loss value.

In another feasible implementation, in the terminal 700 shown in FIG. 7 , the user interface 703 is mainly used to provide an input interface for the user and obtain the data input by the user; and the terminal processor 701 can also be used to call the credit attribute inference program stored in the memory 705 and specifically perform the following operations:

Obtain user characteristics of target users;

Inputting user characteristics into the credit attribute inference model to obtain the credit attributes corresponding to the target user output by the credit attribute inference model;

Providing services to target users based on credit attributes;

The credit attribute inference model is the credit attribute inference model obtained by the inference model training method provided in the above embodiments of this specification.

As mentioned above, the devices, computer-readable storage media, computer program products or chips provided in the embodiments of the present application are all used to execute the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

In the several embodiments provided in this specification, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The above computer program product includes one or more computer instructions. When the above computer program instructions are loaded and executed on a computer, the above process or function according to the embodiment of this specification is generated in whole or in part. The above computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The above computer instructions can be stored in a computer-readable storage medium or transmitted by the above computer-readable storage medium. The above computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (Digital Subscriber Line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode. The above computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server, data center, etc. that contains one or more available media integrated. The above-mentioned available media can be magnetic media (for example, floppy disks, hard disks, tapes), optical media (for example, digital versatile discs (DVD)), or semiconductor media (for example, solid state drives (SSD)), etc.

It should be noted that, for the convenience of description, the aforementioned method embodiments are all described as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

In addition, it should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions. For example, the original user data involved in this specification is obtained with full authorization.

The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

The above is a description of an inference model training and credit attribute inference method, device and terminal provided in the embodiments of this specification. For technical personnel in this field, according to the ideas of the embodiments of this specification, there may be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the embodiments of this specification.

Claims (13)

1. A method for training an inference model, the method comprising: Determine multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and determine the loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute; Input all sample user features into an initial inference model, control the initial inference model to output the predicted credit attributes of each sample user feature, and adjust the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user features; Determine the target sample user features that meet the preset update conditions among all sample user features, update the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and use all sample user features for the next training of the initial inference model after the update until the initial inference model converges to obtain a credit attribute inference model. 2. According to the method of claim 1, the step of determining a plurality of sample user features of a plurality of sample users and a priori credit attributes corresponding to each sample user feature, and determining a loss weight of each sample user feature comprises: If the initial inference model has not been trained at this time, then based on the original user data of the multiple sample users and the credit attribute logic rules, determine multiple sample user features of the multiple sample users and the prior credit attributes corresponding to each sample user feature; The loss weight of each sample user feature is determined by the prior credit attribute of each sample user feature. 3. The method according to claim 2, wherein determining the loss weight of each sample user feature by using the prior credit attribute of each sample user feature comprises: Determine the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution in a preset Gaussian mixture distribution, wherein the preset Gaussian mixture distribution includes multiple Gaussian distributions obtained based on multiple dimensions of the credit attribute; According to the probability that the prior credit attribute of each sample user feature belongs to each Gaussian distribution, the loss weight of each sample user feature is calculated. The loss weight of each sample user feature is positively correlated with the highest distribution probability among the probabilities that the prior credit attribute belongs to each Gaussian distribution. 4. The method according to claim 1, wherein determining the target sample user features that meet the preset update condition among all sample user features comprises: Determine at least one low-confidence sample user feature whose loss weight is less than a preset weight threshold among all sample user features; Inputting each low-confidence sample user feature into the adjusted initial inference model, and controlling the adjusted initial inference model to output the modified credit attribute of each low-confidence sample user feature; The current existing prior credit attributes of each low-confidence sample user feature are determined, and the low-confidence sample user feature whose difference between the revised credit attribute and the existing prior credit attribute is greater than a preset difference threshold is determined as the target sample user feature. 5. The method according to claim 4, wherein updating the prior credit attributes and loss weights of the target sample user characteristics based on the adjusted initial inference model comprises: According to the modified credit attribute of the target sample user feature and the existing prior credit attribute of the target sample user feature, the prior credit attribute of the target sample user feature is calculated and the loss weight of the target sample user feature is increased. 6. The method according to claim 4, before determining the low-confidence sample user feature whose difference between the revised credit attribute and the existing prior credit attribute is greater than a preset difference threshold as the target sample user feature, further comprising: Determining the number of low-confidence sample user features whose difference between the revised credit attribute and the existing prior credit attribute is less than or equal to the preset difference threshold; If the quantity satisfies a preset correction condition, it is determined that the adjusted initial inference model has converged. 7. The method according to claim 1, wherein adjusting the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user characteristics comprises: A loss value is calculated based on the predicted credit attributes, prior credit attributes and loss weights of all sample user characteristics, and parameters of the initial inference model are determined based on the loss value. 8. A credit attribute inference method, the method comprising: Obtain user characteristics of target users; Inputting the user characteristics into a credit attribute inference model to obtain the credit attribute corresponding to the target user output by the credit attribute inference model; Providing services to the target user based on the credit attribute; The credit attribute inference model is a credit attribute inference model obtained in the inference model training method described in claims 1 to 7. 9. An inference model training device, the device comprising: A sample acquisition module, used to determine multiple sample user features of multiple sample users and the prior credit attributes corresponding to each sample user feature, and to determine the loss weight of each sample user feature, wherein the loss weight of each sample user feature is positively correlated with the confidence of the prior credit attribute; A model training module, used to input all sample user features into an initial inference model, control the initial inference model to output the predicted credit attribute of each sample user feature, and adjust the parameters of the initial inference model according to the predicted credit attributes, prior credit attributes and loss weights of all sample user features; The sample update module is used to determine the target sample user features that meet the preset update conditions among all sample user features, update the prior credit attributes and loss weights of the target sample user features based on the adjusted initial inference model, and use all sample user features for the next training of the initial inference model after the update until the initial inference model converges to obtain a credit attribute inference model. 10. A credit attribute inference device, comprising: A user characterization module is used to obtain user characteristics of target users; A model output module, used for inputting the user characteristics into a credit attribute inference model to obtain the credit attribute corresponding to the target user output by the credit attribute inference model; A service providing module, used for providing services to the target user based on the credit attribute; The credit attribute inference model is a credit attribute inference model obtained in the inference model training method described in claims 1 to 7. 11. A computer program product comprising instructions, which, when executed on a computer or a processor, causes the computer or the processor to execute the steps of the method according to any one of claims 1 to 7 or 8. 12. A computer storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the steps of the method according to any one of claims 1 to 7 or 8. 13. A terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 7 or 8 when executing the computer program.