Structured nonlinear ordered regression method for user credit rating
By introducing an adaptive weight matrix and a structured nonlinear ordered regression method that minimizes the constraints of misclassified samples, the problem of insufficient model robustness in traditional user credit evaluation is solved, and more accurate user credit rating classification and risk control are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN VOCATIONAL & TECHNICAL COLLEGE (DALIAN OPEN UNIVERSITY)
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional user credit rating methods ignore global semantic information, cannot guarantee the order of categories, and have difficulty handling unbalanced sample distributions and inter-class noise interference, resulting in insufficient model robustness and inability to achieve accurate credit rating classification.
A structured nonlinear ordered regression method is adopted, and the KDLOR algorithm is improved by introducing an adaptive weight matrix and minimizing the constraint of misclassified samples. The Lagrange multiplier method is used to solve the problem, and a training model is built to identify blacklisted and whitelisted users.
It achieves more accurate and robust user credit rating classification, improves risk control and precision marketing decision support for Internet applications, and performs particularly well in unbalanced sample scenarios.
Smart Images
Figure CN122241302A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of data mining and machine learning technology, and in particular to a method for classifying user credit ratings based on structured nonlinear ordered regression. Background Technology
[0002] With the rapid development of e-commerce, mobile devices, the Internet of Things, and 5G technology, the number of mobile applications has surged, and massive amounts of users' basic and behavioral information constitute core big data resources. Mining the potential value of this data to achieve accurate user credit evaluation has become crucial for enhancing application competitiveness and optimizing user stickiness—it can not only support the identification of fraudulent users and the screening of high-quality users, but also enable differentiated service recommendations.
[0003] User credit rating is essentially an ordered regression problem. Traditional user credit rating methods typically score users through credit surveys or predict user tags using classification models. However, traditional methods have significant drawbacks. The original credit scoring model ignores global semantic information, cannot guarantee the orderliness of categories, and struggles to achieve differentiated service recommendations and risk management for users with ordered credit categories. Ordered regression algorithms achieve ordered classification of category tags by learning a suitable embedding plane, but they do not consider the unbalanced distribution of user samples in real-world applications (the sample size of blacklisted and whitelisted users is far smaller than that of regular users). This easily leads to underfitting of a few categories and overfitting of most categories, and it does not effectively handle the interference of inter-class noise samples on the model, making it difficult to meet the practical requirement of high robustness in credit rating classification. Summary of the Invention
[0004] Traditional user credit rating methods suffer from several problems, including ignoring global semantic information, failing to guarantee category ordering, and difficulty handling imbalanced sample distributions and inter-class noise interference. These issues lead to insufficient model robustness and an inability to achieve accurate credit rating classification. This invention proposes a structured nonlinear ordered regression method for user credit rating classification. Specifically, it improves the KDLOR algorithm by utilizing an adaptive full-time matrix and minimizing misclassified sample constraints, and solves the problem using the Lagrange multiplier method. This results in more accurate and robust identification of key blacklisted and whitelisted users, providing strong decision support for risk control and precision marketing in internet applications.
[0005] The technical means employed in this invention are as follows:
[0006] A structured nonlinear ordered regression method for classifying user credit ratings includes the following steps: Collect basic user registration information and user behavior data from mobile devices, define user credit rating tags based on business rules, and construct a dataset based on user basic registration information, user behavior data, and user credit rating tags; The dataset is preprocessed by dividing it into a training set, a validation set, and a test set. A structured nonlinear ordered regression model is constructed based on an improved KDLOR algorithm. The improvement is that an adaptive weight matrix is introduced to balance the sample distribution and a constraint is introduced to minimize misclassified samples. The Lagrange multiplier method is used to solve the problem, and the structured nonlinear ordered regression model is trained using the training set and the validation set to obtain the trained structured nonlinear ordered regression model. The test set is input into the trained structured nonlinear ordered regression model to obtain the classification results of user credit rating.
[0007] Furthermore, the basic user registration information includes a basic user registration information table, and the user behavior data includes an order transaction table, a recommendation timetable, and a recommendation relationship table.
[0008] Furthermore, the definition of user credit rating tags based on business rules includes: Based on user registration information and user behavior data, user credit ratings are divided into three ordered categories: whitelist users, regular users, and blacklist users. The criteria for determining whitelisted users include: normal registration information, having a recommendation volume and a recommendation success rate exceeding the first threshold, and having an order number lower than the second threshold or a withdrawal rate lower than the third threshold. The criteria for determining blacklisted users include: garbled registration information, abnormal user or child status information, order volume exceeding the fourth threshold or withdrawal rate exceeding the fifth threshold. Users who meet any of the following conditions are classified as blacklisted users. The first to fifth thresholds are preset based on the actual data distribution and business needs of the application. The regular users are other users who are not marked as whitelisted or blacklisted users.
[0009] Furthermore, the mathematical model of the improved KDLOR algorithm is as follows:
[0010] in, Let be the projection matrix. It is a constant. This is the weighted intraclass distance matrix. The first regularization parameter is used. This is the second regularization parameter. This is the sum of the distances between misclassified samples in adjacent categories. For loss function, For category Distance from misclassified samples to the separating plane For category The distance between misclassified samples and the samples projected to the class mean. Let be the mean vector of the k-th class of samples. Let be the mean vector of the (k+1)th class of samples. For category indexing, This represents the total number of category indexes.
[0011] Furthermore, the formula for calculating the weighted intraclass distance matrix is as follows:
[0012] in, As weight, Given the intra-class scatter matrix, the formula for calculating each element of the weights is:
[0013] in, Let i be the element in the i-th row and i-th column of the weight. For category k Sample size This represents the total sample size.
[0014] Furthermore, the formula for calculating the sum of distances between misclassified samples in adjacent categories is as follows:
[0015] in, Let k be the sample size of category k. Let k be the sample size for category k+1. Let be the distance from the a-th misclassified element in category k to the separating plane. Let be the distance from the b-th misclassified sample in category k+1 to the separating plane, where a is the count index of the misclassified sample in category k, and b is the count index of the misclassified sample in category k+1.
[0016] Furthermore, the mathematical model obtained by solving using the Lagrange multiplier method is as follows:
[0017] in, Let be the projection matrix. It is a constant. This is the weighted intraclass distance matrix. The first regularization parameter is used. This is the second regularization parameter. It is the sum of the distances between misclassified samples in adjacent categories. For loss function, For category Distance from misclassified samples to the separating plane For category The distance between misclassified samples and the samples projected to the class mean. Let be the mean vector of the k-th class of samples. Let be the mean vector of the (k+1)th class of samples. For category indexing, The total number of category indexes, As the first Lagrange multiplier, For the second Lagrange multiplier, It is the third Lagrange multiplier. It is the k-th first Lagrange multiplier.
[0018] Compared with the prior art, the present invention has the following advantages: This invention proposes a user credit rating classification method based on structured nonlinear ordered regression. Designed to address the challenges of ordered, imbalanced, and noisy user credit data in the context of explosive growth in internet application data, this method introduces an adaptive weight matrix and minimizes the constraint of misclassified samples. It can more accurately and robustly identify key blacklisted and whitelisted users, providing strong decision support for risk control and precision marketing in internet applications. Compared to traditional support vector machine models, ordered binary classification models, and other ordered regression threshold models, this method achieves superior AUC values and ROC curve performance on multiple real-world datasets, especially showing significant advantages in scenarios with high sample imbalance, demonstrating good application value and promising prospects for wider adoption.
[0019] Based on the above reasons, this invention can be widely applied in fields such as data mining and machine learning. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a schematic diagram comparing the KDLOR projection with the correct projection direction under the sample distribution in an embodiment of the present invention.
[0022] Figure 2 This is a schematic diagram of inter-class noise samples that are difficult to handle during the KDLOR projection process of this invention.
[0023] Figure 3 This is a flowchart of a structured nonlinear ordered regression method for classifying user credit ratings according to the present invention.
[0024] Figure 4(a) is a comparison chart of ROC curves for blacklisted users in an embodiment of the present invention.
[0025] Figure 4(b) is a comparison chart of ROC curves for regular users in the embodiments of the present invention.
[0026] Figure 4(c) is a comparison chart of ROC curves for whitelisted users in the embodiments of the present invention. Detailed Implementation
[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0029] like Figure 3 As shown, this invention provides a method for classifying user credit ratings using structured nonlinear ordered regression, the specific steps of which are as follows: S1. Collect basic user registration information and user behavior data from mobile devices, define user credit rating tags based on business rules, and construct a dataset based on user basic registration information, user behavior data, and user credit rating tags.
[0030] User basic registration information includes a user basic registration information table, and user behavior data includes an order transaction table, a recommendation time table, and a recommendation relationship table.
[0031] Furthermore, user credit rating labels are defined based on business rules, including: The first step is to divide users’ credit ratings into three ordered categories based on user registration information and user behavior data: whitelist users (high-frequency or high-value users), regular users (ordinary users), and blacklist users (who may commit fraud or attacks). The sample size of each category is unbalanced.
[0032] The second step involves determining whitelisted users based on the following criteria: normal registration information, a high number of recommendations with a success rate exceeding the first threshold, and a number of orders below the second threshold or a withdrawal rate below the third threshold.
[0033] The third step is to determine the criteria for blacklisted users, including: garbled registration information, abnormal user or child status information, order volume exceeding the fourth threshold or withdrawal rate exceeding the fifth threshold. Users who meet any of these criteria are blacklisted. The first to fifth thresholds are preset based on the actual data distribution and business needs of the application.
[0034] Step 4: Regular users are other users who are not marked as whitelisted or blacklisted users.
[0035] Specifically, a large number of discrete, independent scores cannot fully represent the potential common characteristics of users. To address this issue, this paper proposes to divide user credit into different naturally ordered categories according to human cognition, similar to student grade evaluation (bad, good, very good, excellent). This paper defines user credit level as three ordered categories: (1) blacklisted users; (2) regular users; (3) whitelisted users.
[0036] Based on user information and behavioral data, and combined with the background of Internet applications, the user credit rating category standards are defined as follows: (1) Whitelist users. Includes 3 conditions: ① Normal registration information, such as normal user and email names without garbled characters; ② Has a recommendation volume, the recommendation success rate exceeds the first threshold (set to 0.8 in this embodiment) and the child information is normal; ③ The number of orders is lower than the second threshold (set to 2 in this embodiment) or the withdrawal rate is lower than the third threshold (set to 0.2 in this embodiment). (2) Blacklist users only need to meet any of the following conditions: ① The user or email name is abnormal garbled characters; ② The child status is abnormal information; ③ The number of orders exceeds the fourth threshold (set to 5 in this embodiment) or the withdrawal rate exceeds the fifth threshold (set to 0.5 in this embodiment). (3) The remaining unmarked users are all regular users. The above data categories are all manually marked according to the actual situation of the application. At the same time, the marking principle can help to classify user credit ratings in Internet applications of the same category.
[0037] In this embodiment, the ratio of blacklisted users, regular users, and whitelisted users is 5:408:11. Since the ratio varies across different datasets, this does not limit the specific ratio used in this invention.
[0038] S2. Preprocess the dataset by dividing the preprocessed dataset into training, validation, and test sets.
[0039] Specifically, the preprocessing steps are as follows: The raw data is cleaned, correlated, and numerically transformed. By associating and integrating user IDs, core features related to credit evaluation are selected, such as device type, application version, order volume, withdrawal ratio, and recommendation success rate. These features are then converted into numerical values to build a standardized data foundation for model input.
[0040] Specifically, the cleaning and association steps are as follows: 1. Integration via user ID association 2. Missing value handling (removing features with a high missing value ratio and filling features with a low missing value ratio) 3. Outlier Identification and Handling Then, numerical conversion is performed.
[0041] The principles for numerical transformation of features are as follows: (1) Category-type features (such as registered device type, application version number) are converted into category integers; (2) Numerical features (such as order volume, recommendation success rate) should retain their original integer or decimal format; (3) Boolean features (such as whether the recommender is normal) are converted into 0 / 1 codes.
[0042] S3. Construct a structured nonlinear ordered regression model. The structured nonlinear ordered regression model is based on the improved KDLOR algorithm. The improvements are: introducing an adaptive weight matrix to balance the sample distribution and introducing a constraint to minimize misclassified samples.
[0043] This invention addresses the problem of insufficient generalization ability of traditional ordered regression algorithms such as KDLOR on imbalanced data by enhancing the basic model.
[0044] Kernel Discriminant Ordinal Regression (KDLOR) is an improvement on Linear Discriminant Analysis (LDA). The goal of LDA is to find optimal projection directions that maximize the distance between the two classes of samples in the projection directions while minimizing the intra-class distance. The LDA model is as follows:
[0045] in, For the LDA algorithm model, The scatter matrix is the intra-class scatter matrix. The inter-class scatter matrix, This is the projection matrix.
[0046] The KDLOR algorithm, based on the LDA algorithm, ensures that the classes are ordered after sample projection, and that the difference between the mean vectors of adjacent classes after projection is greater than or equal to zero. .in, Let be the mean vector of the k-th class of samples, a constant. Therefore, the KDLOR model can be rewritten from the LDA model as follows:
[0047] in, This is the first regularization parameter, used to expand the difference in the projection mean between two adjacent classes. Let be the mean vector of the k-th class of samples. Let be the mean vector of the (k+1)th class of samples. For category indexing, This represents the total number of category indexes.
[0048] To adapt to various data distributions, the model was further extended to a nonlinear method. Subsequent steps, based on the kernel concept and utilizing mapping functions... This method maps samples to a high-dimensional feature space, enabling the classification of nonlinear data. Simultaneously, solving the projection space requires satisfying the following assumptions: ,in, The projection matrix after "kerneling" the sample data. Given a dataset containing N samples (x i y i ), i=1,…,N. The transformed KDLOR algorithm model is:
[0049] in, H This is the intra-class distance matrix after applying the kernel function. Let be the mean of the kernel inner product of all samples in class k+1. The mean of the kernel inner product of all samples in the k-th class is calculated as follows:
[0050]
[0051] in, Let be the mean of the kernel product of sample j and all samples of class k. Let X be the sample size of the k-th class. k For the set of samples of class k, The j-th sample belongs to the k-th class. All element values are matrix, for The nonlinear kernel matrix.
[0052] The calculation method is as follows:
[0053] in, Let x represent the i-th sample of class k in the nonlinear kernel matrix. i and the j-th sample x j Inner product in a high-dimensional feature space.
[0054] The improvement in this embodiment is as follows: (1) Introduce an adaptive weight matrix to balance the sample distribution. To address the shortcomings of the KDLOR algorithm in handling imbalanced data, a weight matrix is introduced to balance the sample distribution. This matrix is adaptively generated based on the sample size of each ordered class, assigning greater weight to classes with smaller sample sizes (such as blacklist / whitelist users) and less weight to classes with larger sample sizes (regular users). The formula for calculating the weighted intra-class distance matrix is as follows:
[0055] in, As weight, It is the intra-class scatter matrix.
[0056] The formula for calculating each element of the weight is:
[0057] in, Let i be the element in the i-th row and i-th column of the weight. For category k Sample size Let this be the total sample size. This satisfies... .
[0058] At this point, the model's formula is updated to:
[0059] This method obtains an optimized projection direction that is more beneficial to minority class samples by balancing the sample distribution, thereby alleviating the underfitting and overfitting problems caused by unbalanced distribution.
[0060] (2) Introduce a constraint to minimize misclassified samples to improve robustness.
[0061] During classification, noisy samples between classes can affect the optimal projection direction. By introducing a constraint to minimize misclassified samples, the loss caused by potentially misclassified samples between adjacent ordered classes or by noisy samples between classes is minimized. The updated model formula (i.e., the mathematical model of the improved KDLOR algorithm) is as follows:
[0062] in, This is the second regularization parameter, used to balance the difference in projected mean between two adjacent classes. It is the sum of the distances between misclassified samples in adjacent categories. For loss function, For category Distance from misclassified samples to the separating plane For category The distance between misclassified samples and the samples projected to the class mean. Let be the projection matrix. It is a constant. This is the weighted intra-class distance matrix. Misclassified samples include noisy samples.
[0063] Figure 2 This study revealed inter-class noise samples that are difficult to handle during the KDLOR projection process. This patent aims to minimize the loss from misclassified samples, thereby improving classification accuracy and obtaining more optimized, ordered projection results.
[0064] The formula for calculating the sum of distances between misclassified samples in adjacent categories is:
[0065] in, Let k be the sample size of category k. Let k be the sample size for category k+1. Let be the distance from the a-th misclassified element in category k to the separating plane. Let be the distance from the b-th misclassified sample in category k+1 to the separating plane, where a is the count index of the misclassified sample in category k, and b is the count index of the misclassified sample in category k+1.
[0066] This constraint forces the model to consider not only intra-class compactness and inter-class order when searching for projection directions, but also to actively increase the distance between misclassified samples and class boundaries, thereby obtaining projection directions that are less sensitive to noise and have clearer classification boundaries, thus improving the robustness of the model.
[0067] S4. Solve using the Lagrange multiplier method, and then train the structured nonlinear ordered regression model using the training set and validation set to obtain the trained structured nonlinear ordered regression model.
[0068] Specifically, the mathematical model obtained by solving using the Lagrange multiplier method is as follows:
[0069] in, Let be the projection matrix. It is a constant. This is the weighted intraclass distance matrix. The first regularization parameter is used. This is the second regularization parameter. It is the sum of the distances between misclassified samples in adjacent categories. For loss function, For category Distance from misclassified samples to the separating plane For category The distance between misclassified samples and the samples projected to the class mean. Let be the mean vector of the k-th class of samples. Let be the mean vector of the (k+1)th class of samples. For category indexing, The total number of category indexes, As the first Lagrange multiplier, For the second Lagrange multiplier, It is the third Lagrange multiplier. It is the k-th first Lagrange multiplier.
[0070] S5. Input the test set into the trained structured nonlinear ordered regression model to obtain the classification results of user credit rating.
[0071] right L The partial derivatives of the variables to be solved are calculated separately, and numerical optimization methods such as the interior point method, the effective set method, and the conjugate gradient method are used to solve for the optimal projection. This allows for the determination of the user's credit rating.
[0072] Figure 3 In this context, the original data space refers to the original data distribution, and the ordered subspace refers to the data distribution after mapping obtained by this invention.
[0073] Figure 1 In this algorithm, KDLOR ensures that the projected sample classes are ordered by maximizing the constraint term. However, when the data distribution is... Figure 1 In case (a), considering only minimizing the sum of intra-class distances and maximizing the mean difference between ordered classes after projection, misclassification of class samples after projection may occur. Noise samples existing between ordered classes will also affect the discrimination model and the optimal projection direction. A more optimized projection plane should be as follows: Figure 1 As shown in (b), the samples on the projection plane are basically correctly classified and ordered between classes.
[0074] This invention uses ROC curves and AUC values to measure the classification performance of the model.
[0075] This invention compares the performance of three algorithms through experiments to verify its own algorithm: traditional models (SVC1V1, SVC1VA), ordered binary classification models (SVMOP, ELMOP), and threshold models (POM, REDSVM, KDLOR). Simultaneously, to analyze the effectiveness of the proposed algorithm, the results of RDLOR_Weight (an improvement in this embodiment: introducing an adaptive weight matrix, abbreviated as RDLOR_W) and RDLOR_Parameters (an improvement in this embodiment: introducing a constraint to minimize misclassified samples, abbreviated as RDLOR_P) are statistically analyzed. Figure 4 shows a comparison of the ROC curves of each algorithm on different categories of samples on the dataset. Because other methods do not fully consider the sample distribution, they fail to adequately learn the blacklist and whitelist users with smaller sample sizes during classification, resulting in less satisfactory classification results. The experimental results curves verify that the classification accuracy of this algorithm is slightly better than other algorithms.
[0076] Table 1 Comparison of AUC values of each algorithm on the dataset
[0077] Table 1 shows the AUC values of each algorithm on the dataset. The closer the AUC value is to 1, the better the classification performance of the algorithm. The algorithm presented in this paper achieved better results than other traditional algorithms on the dataset.
[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A structured nonlinear ordered regression method for classifying user credit ratings, characterized in that, Includes the following steps: Collect basic user registration information and user behavior data from mobile devices, define user credit rating tags based on business rules, and construct a dataset based on user basic registration information, user behavior data, and user credit rating tags; The dataset is preprocessed by dividing it into a training set, a validation set, and a test set. A structured nonlinear ordered regression model is constructed based on an improved KDLOR algorithm. The improvement is that an adaptive weight matrix is introduced to balance the sample distribution and a constraint is introduced to minimize misclassified samples. The Lagrange multiplier method is used to solve the problem, and the structured nonlinear ordered regression model is trained using the training set and the validation set to obtain the trained structured nonlinear ordered regression model. The test set is input into the trained structured nonlinear ordered regression model to obtain the classification results of user credit rating.
2. The user credit rating classification method based on structured nonlinear ordered regression according to claim 1, characterized in that, The basic user registration information includes a basic user registration information table, and the user behavior data includes an order transaction table, a recommendation timetable, and a recommendation relationship table.
3. The user credit rating classification method based on structured nonlinear ordered regression according to claim 1, characterized in that, The definition of user credit rating tags based on business rules includes: Based on user registration information and user behavior data, user credit ratings are divided into three ordered categories: whitelist users, regular users, and blacklist users. The criteria for determining whitelisted users include: normal registration information, having a recommendation volume and a recommendation success rate exceeding the first threshold, and having an order number lower than the second threshold or a withdrawal rate lower than the third threshold. The criteria for determining blacklisted users include: garbled registration information, abnormal user or child status information, order volume exceeding the fourth threshold or withdrawal rate exceeding the fifth threshold. Users who meet any of the following conditions are classified as blacklisted users. The first to fifth thresholds are preset based on the actual data distribution and business needs of the application. The regular users are other users who are not marked as whitelisted or blacklisted users.
4. The user credit rating classification method based on structured nonlinear ordered regression according to claim 1, characterized in that, The mathematical model of the improved KDLOR algorithm is as follows: in, For the projection matrix, It is a constant. This is the weighted intraclass distance matrix. The first regularization parameter is used. This is the second regularization parameter. This is the sum of the distances between misclassified samples in adjacent categories. For loss function, For category Distance from misclassified samples to the separating plane For category The distance between misclassified samples and the samples projected to the class mean. Let be the mean vector of the k-th class of samples. Let be the mean vector of the (k+1)th class of samples. For category indexing, This represents the total number of category indexes.
5. The user credit rating classification method based on structured nonlinear ordered regression according to claim 4, characterized in that, The formula for calculating the weighted intraclass distance matrix is as follows: in, As weight, Given the intra-class scatter matrix, the formula for calculating each element of the weights is: in, Let i be the element in the i-th row and i-th column of the weight. For category k Sample size This represents the total sample size.
6. The user credit rating classification method based on structured nonlinear ordered regression according to claim 4, characterized in that, The formula for calculating the sum of distances between misclassified samples in adjacent categories is: in, Let k be the sample size of category k. Let k be the sample size for category k+1. Let be the distance from the a-th misclassified element in category k to the separating plane. Let be the distance from the b-th misclassified sample in category k+1 to the separating plane, where a is the count index of the misclassified sample in category k, and b is the count index of the misclassified sample in category k+1.
7. The user credit rating classification method based on structured nonlinear ordered regression according to claim 1, characterized in that, The mathematical model obtained by solving using the Lagrange multiplier method is as follows: in, For the projection matrix, It is a constant. This is the weighted intraclass distance matrix. The first regularization parameter is used. This is the second regularization parameter. It is the sum of the distances between misclassified samples in adjacent categories. For loss function, For category Distance from misclassified samples to the separating plane For category The distance between misclassified samples and the samples projected to the class mean. Let be the mean vector of the k-th class of samples. Let be the mean vector of the (k+1)th class of samples. For category indexing, The total number of category indexes, As the first Lagrange multiplier, For the second Lagrange multiplier, It is the third Lagrange multiplier. It is the k-th first Lagrange multiplier.