Multi-view label generation method and device for risk profiling

Abstract

The application provides a multi-view label generation method and device for risk portrait construction, which has the following characteristics: steps S1-S2, constructing a sub-feature set, calculating the importance value and correlation value of each feature, and a feature screening quantity; step S3, constructing a screening feature set; step S4, calculating and constructing a positive correlation feature set and a negative correlation feature set according to the correlation value of each feature in the screening feature set; step S5, constructing a discrete feature set and a continuous feature set according to the number of unique values corresponding to each feature in the screening feature set; step S6, obtaining a multi-view model feature set according to the positive correlation feature set, the negative correlation feature set, the discrete feature set and the continuous feature set; and step S7, calculating the woe value of each feature corresponding to each target in the multi-view model feature set. In summary, the method can generate a label with better risk expression ability.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of risk profiling, and particularly relates to a multi-view label generation method and device for risk profiling construction. BACKGROUND

[0002] Risk profiling is used in multiple fields, and constructing an accurate risk profile is crucial for identifying, assessing and predicting risks. Traditional risk profiling methods usually rely on rule-driven methods or machine learning models to extract features of subjects from historical data and generate risk labels based on these features to assess the risk status of subjects. However, with the rapid growth of data volume and data types, traditional methods have exposed some significant shortcomings in feature extraction and risk assessment:

[0003] 1) Feature interference problem: In subject-related data, there are complex correlations between different features, including positive, negative and even nonlinear correlations. Traditional machine learning models often encounter the problem of gradient cancellation when dealing with positive and negative correlations, which makes it difficult for the model to effectively focus on key risk factors during the learning process, reducing the prediction accuracy of the model. In addition, feature interference can also cause instability in model training, which adversely affects the final risk profiling construction.

[0004] 2) Learning bias of feature types: In risk profiling construction, feature types are diverse, such as discrete and continuous features coexisting, and their distributions are complex. Existing deep learning models usually learn discrete or continuous column features based on the values or distribution characteristics of features, which can easily lead to learning bias of different feature types, making it difficult to capture the relevance between discrete and continuous features. This limitation affects the effectiveness of the model in constructing a global risk profile.

[0005] 3) Lack of multi-view feature representation: Current risk profiling techniques rarely consider multi-view feature representation, i.e., lack the ability to generate a comprehensive feature space from different perspectives, such as feature correlation, distribution dispersion, etc. Existing methods usually extract features based on a single perspective, which is difficult to reflect the complex risk features of subjects. This leads to insufficient dimensions of the risk profile, making it difficult to meet the risk assessment needs in complex scenarios. SUMMARY

[0006] The present application is to solve the above problems, and aims to provide a multi-view label generation method and device for risk profiling construction.

[0007] The application provides a multi-view label generation method for risk portrait construction, which is used for generating a label result corresponding to each target and related to a risk portrait according to risk-related data, wherein the risk-related data includes values of a plurality of different features corresponding to each target and a risk judgment label corresponding to each target, and the method has the following steps: step S1, dividing all features into a plurality of categories to generate a sub-feature set corresponding to each category, wherein the sub-feature set includes at least one feature and all values corresponding to the feature; step S2, for each sub-feature set, calculating an importance value and a correlation value of each feature and a feature screening quantity according to the number of features in the sub-feature set and the values corresponding to the features; step S3, selecting a plurality of features from each sub-feature set to construct a screening feature set according to the corresponding importance value, correlation value and feature screening quantity; step S4, calculating the correlation value of each feature according to the number of features in the screening feature set and the values corresponding to the features, and constructing a positively correlated feature set and a negatively correlated feature set according to the correlation value; step S5, constructing a discrete feature set and a continuous feature set according to the number of unique values corresponding to each feature in the screening feature set; step S6, obtaining a multi-view model feature set according to the positively correlated feature set, the negatively correlated feature set, the discrete feature set and the continuous feature set in combination with a machine learning model; and step S7, performing binning on the multi-view model feature set to calculate a binning value of each feature corresponding to each target in the multi-view model feature set, and calculating a woe value of each feature corresponding to each target in the multi-view model feature set according to all binning values corresponding to the multi-view model feature set and the risk judgment label, wherein the label result of the target is each feature in the multi-view model feature set and the corresponding woe value of the target.

[0008] In the multi-view label generation method for risk portrait construction provided by the application, the calculation expression of the positively correlated feature set and the negatively correlated feature set can be as follows: In the formula, X p is the positively correlated feature set, is the correlation value of feature i in the screening feature set, X n is the negatively correlated feature set.

[0009] In the multi-view label generation method for risk portrait construction provided by the application, the calculation expression of the discrete feature set and the continuous feature set can be as follows: In the formula, X d is the discrete feature set, X c is the continuous feature set, P is the total number of features in the screening feature set.

[0010] In the multi-view label generation method for risk portrait construction provided by the application, the calculation expression of the multi-view model feature set can be: X p_d = X p ∩ X d , X p_c = X p ∩ X c , X n_d = X n ∩ X d , X n_c = X n ∩ X c , wherein X multi_view is a multi-view model feature set, || is a set concatenation symbol, is a sample prediction value generated by the machine learning model according to the input feature set X p_d , is a sample prediction value generated by the machine learning model according to the input feature set X p_c , is a sample prediction value generated by the machine learning model according to the input feature set X n_d , is a sample prediction value generated by the machine learning model according to the input feature set X n_c .

[0011] In the multi-view label generation method for risk portrait construction provided by the application, the calculation expression of the correlation value can be: In the formula, is the correlation value of feature i in feature set j, is the value of feature i corresponding to the kth target in feature set j, is the mean value of feature i in feature set j, y k is the risk judgment label corresponding to the kth target, is the mean value of the risk judgment labels of all targets, and n is the total number of targets.

[0012] In the multi-view label generation method for risk portrait construction provided by the application, in step S2, the importance value corresponding to the information gain of the feature on the node of the LightGBM model tree is calculated, and the calculation expression of the importance value is: In the formula, is the importance value of feature i in the sub-feature set j, p is the total number of features in the sub-feature set j, M is the total number of trees in the LightGBM model, N m is the set of nodes of the mth tree, is an indicator function, and ΔL m,n is the information gain of feature i on node n of the mth tree, when node n is split using feature i, is the information gain of feature i on node n of the mth tree, when node n is not split using feature i, In step S2, the calculation expression of the feature selection quantity is: In the formula, select_len j is the feature selection quantity corresponding to the sub-feature set j, p is the total number of features in the sub-feature set j, is a floor operation.

[0013] In the multi-view label generation method for risk portrait construction provided by the present application, the calculation expression of the filtered feature set in step S3 can be: X j,selected =X j,corr ∪X j,lgbm , In the formula, X j,selected is the filtered set corresponding to the sub-feature set j, X j,corr is the relevance filtered feature set corresponding to the sub-feature set j, X j,lgbm is the importance filtered feature set corresponding to the sub-feature set j, X selected is the filtered feature set, m is the total number of sub-feature sets, all features of the sub-feature set j are sorted in descending order of the relevance value, and the first feature selection quantity of features is selected from the sequence to construct the relevance filtered feature set, all features of the sub-feature set j are sorted in descending order of the importance value, and the first feature selection quantity of features is selected from the sequence to construct the importance filtered feature set.

[0014] The multi-view label generation method for risk profile construction provided by this invention may also have the following features: Step S7 includes the following sub-steps: Step S7-1, generating multiple candidate split thresholds for the multi-view model feature set, and calculating the split value corresponding to each candidate split threshold; Step S7-2, taking the candidate split threshold corresponding to the largest split value as the optimal split point; Step S7-3, repeating steps S6-1 to S6-2 until the number of optimal split points reaches a preset value; Step S7-4, calculating the binning boundary based on all optimal split points; Step S7-5, calculating the bins value and bin0 value of each feature corresponding to each target in the multi-view model feature set as binning values ​​based on the binning boundary; Step S7-6, calculating the woe value of each feature corresponding to each target in the multi-view model feature set based on all binning values.

[0015] The multi-view label generation method for risk profile construction provided by this invention may also have the following feature: wherein the calculation expression for the split value is: H(D mv )=-∑ c∈C p(c)log2p(c), In the formula IG(D mv ,θ) represents the splitting value of the candidate splitting threshold θ corresponding to the feature set of the multi-view model, and |D| represents the number of samples in D. v represents the value of feature i corresponding to target k in the feature set of the multi-view model. k For target k, y k Let C be the risk assessment label corresponding to target k, C be the total number of features in the feature set of the multi-view model, and p(c) be the probability of feature c in the feature set of the multi-view model. The expression for calculating the binning boundary is: In the formula for The set of bin boundary values, To sort all optimal split points in ascending order, the expressions for calculating the bins and bin0 values ​​are as follows: In the formula The value of feature i corresponding to target k in the feature set of the multi-view model. for The left boundary value of the u-th bin interval is determined in step S7-6. Based on the bin0 value of the feature corresponding to the target, the values ​​of all target-corresponding features are divided into multiple groups, where all target-corresponding feature values ​​in each group have the same bin0 value. The risk assessment label is a binary classification label, and the expression for calculating the WOE value is: In the formula the total number of risk judgment labels corresponding to the bad class for all values in the group in which the value of the feature i corresponding to the target k in the multi-view model feature set, N bad the total number of risk judgment labels corresponding to the bad class for all values in the group in which the value of the feature i corresponding to the target k in the multi-view model feature set, N good the total number of risk judgment labels corresponding to the good class for all values in the group in which the value of the feature i corresponding to the target k in the multi-view model feature set, N total the total number of risk judgment labels corresponding to the good class for all values in the group in which the value of the feature i corresponding to the target k in the multi-view model feature set, N

[0016] The application also provides a multi-view label generation device for risk portrait construction, which is used for generating a label result related to a risk portrait corresponding to each target according to risk-related data, the risk-related data including values of a plurality of different features corresponding to each target and a risk judgment label corresponding to each target, and has the following characteristics: a sub-set division module that divides all features into a plurality of categories, generates a sub-feature set corresponding to each category, and the sub-feature set includes at least one feature and all values corresponding to the feature; a screening module that, for each sub-feature set, calculates the importance value and the correlation value of each feature and the feature screening quantity according to the number of features in the sub-feature set and the values corresponding to the features; a screening set generation module that selects a plurality of features from each sub-feature set to construct a screening feature set according to the corresponding importance value, correlation value and feature screening quantity; a positive and negative correlation feature set generation module that calculates the correlation value of each feature according to the number of features in the screening feature set and the values corresponding to each feature, and constructs a positive correlation feature set and a negative correlation feature set according to the correlation value; a discrete and continuous feature set generation module that constructs a discrete feature set and a continuous feature set according to the number of unique values corresponding to each feature in the screening feature set; a multi-view model feature set generation module that obtains a multi-view model feature set according to the positive correlation feature set, the negative correlation feature set, the discrete feature set and the continuous feature set in combination with a machine learning model; and a woe calculation module that performs binning on the multi-view model feature set, calculates the binning value of each feature corresponding to each target in the multi-view model feature set, and calculates the woe value of each feature corresponding to each target in the multi-view model feature set according to all binning values corresponding to the multi-view model feature set and the risk judgment label, wherein the label result of the target is the woe value of each feature corresponding to the target in the multi-view model feature set.

[0017] Effects of the application

[0018] According to the multi-view label generation method and device for risk portrait construction, the correlation values and the number of unique values of the features in the screening set are calculated, a positive correlation feature set, a negative correlation feature set, a discrete feature set and a continuous feature set are constructed, a multi-view model feature set is generated, and the multi-view model feature set is subjected to binning and woe value calculation, so that the label result is obtained. Therefore, the multi-view label generation method and device for risk portrait construction can generate a label with better risk expression ability, and improve the comprehensiveness and accuracy of the risk portrait. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a block diagram of the multi-view label generation device for risk portrait construction in the embodiment of the present application;

[0020] Figure 2 is a flowchart of the woe value generated by the woe calculation module in the embodiment of the present application;

[0021] Figure 3 is a flowchart of the multi-view label generation method for risk portrait construction in the embodiment of the present application;

[0022] Figure 4 is a thundercloud diagram of the high-risk user portrait in the embodiment of the present application;

[0023] Figure 5 is a thundercloud diagram of the low-risk user portrait in the embodiment of the present application;

[0024] Figure 6 is a schematic diagram of the scorecard visualization in the embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the technical means, creative features, purposes and effects achieved by the present application easy to understand, the following embodiments will be specifically described in combination with the drawings.

[0026] The embodiment provides a multi-view label generation device for risk portrait construction, which is used for generating a label result related to a risk portrait corresponding to each target according to risk-related data. The risk-related data includes values of a plurality of different features corresponding to each target, and a risk judgment label corresponding to each target. The risk judgment label is a binary classification label.

[0027] The multi-view label generation device for risk portrait construction in the embodiment is applied to the financial field, and the risk-related data is financial transaction data D=[V X Y]. Wherein, V={v1,...,v n} represents a set of transaction accounts, each transaction account is a target, and for each vi Each ∈V has a binary label indicating whether it is illegal, i.e., a separator label y. i ∈Y, where n is the number of trading accounts, and the feature space constructed from the perspectives of statistical calculation, complex rules, expert knowledge, and intelligent algorithms is X, where X={X1,X2,…,X…} d}, X i ={χ i,k |k=1,2,…,n}, representing the feature χ i The feature set, where χ i,k Let x represent the feature corresponding to the k-th transaction account. i The value of . In other embodiments, the multi-view label generation device for risk profile construction can be applied to other fields, such as the food safety field, where the target is a specific food and the risk-related data is production data related to the specified food, such as the manufacturer.

[0028] Figure 1 This is a block diagram of a multi-view tag generation device for risk profile construction in an embodiment of the present invention.

[0029] like Figure 1 As shown, the multi-view label generation device 100 for risk profile construction includes a subset partitioning module 11, a filtering module 12, a filtering set generation module 13, a positive and negative correlation feature set generation module 14, a discrete and continuous feature set generation module 15, a multi-view model feature set generation module 16, a WOE calculation module 17, and a control module 18 that controls the operation of the above modules.

[0030] Subset partitioning module 11 divides all features into multiple categories and generates sub-feature sets corresponding to each category. Each sub-feature set includes at least one feature and all values ​​corresponding to that feature.

[0031] In this embodiment, the subset partitioning module 11 divides information into m categories, thus obtaining the sub-feature set X. 1 ,X 2 ,…,X m And satisfy

[0032] The filtering module 12 calculates the importance value, relevance value, and feature filtering quantity for each sub-feature set based on the number of features in the sub-feature set and the corresponding values ​​of the features.

[0033] The filtering module 12 includes a LightGBM model. Based on the information gain of features at the nodes of the LightGBM model's tree, the filtering module 12 calculates the corresponding importance value. The expression for calculating the importance value is as follows:

[0034]

[0035] wherein is the importance value of feature i in the sub-feature set j, p is the total number of features in the sub-feature set j, M is the total number of trees in the LightGBM model, N is the total number of samples, and y m is the set of nodes of the mth tree, is an indicator function, and m,n is the information gain of feature i on node n of the mth tree. When node n is split using feature i, When node n is not split using feature i,

[0036] The loss function corresponding to the LightGBM model in this embodiment is binary cross-entropy loss L, and the calculation expression thereof is:

[0037]

[0038] wherein N is the total number of samples, y i is the risk judgment label of the ith sample, is the predicted probability of the ith sample by the LightGBM model.

[0039] The calculation expression of the correlation value is:

[0040]

[0041] wherein is the correlation value of feature i in the feature set j, is the value of feature i corresponding to the kth target in the feature set j, is the mean value of feature i in the feature set j, y k is the risk judgment label corresponding to the kth target, is the mean value of the risk judgment labels of all targets, and n is the total number of targets.

[0042] The calculation expression of the feature screening quantity is:

[0043]

[0044] wherein select len j is the feature screening quantity corresponding to the sub-feature set j, p is the total number of features in the sub-feature set j, is a floor operation.

[0045] The screening set generation module 13 selects a plurality of features from each sub-feature set to construct a screening feature set according to the corresponding importance value, correlation value, and feature screening quantity.

[0046] Wherein, the calculation expression of the screening feature set is:

[0047] X j,selected = X j,corr ∪ X j,lgbm ,

[0048]

[0049] In the formula, X j,selected is a screening set corresponding to the sub-feature set j, X j,corr is a correlation screening feature set corresponding to the sub-feature set j, X j,lgbm is an importance screening feature set corresponding to the sub-feature set j, X selected is a screening feature set, and m is the total number of sub-feature sets.

[0050] All features of the sub-feature set j are sorted in descending order according to the correlation value, and a number of features of the feature screening quantity are selected from the beginning of the sequence to construct the correlation screening feature set. The calculation expression of the correlation screening feature set X j,corr is as follows:

[0051] X j,corr = argsort (|correlations j |) [: select_len j ],

[0052]

[0053] In the formula, argsort (|correlations j |) is a descending order arrangement of the correlation value of the sub-feature set j according to the absolute value, that is, the sorting according to the size of the absolute value of the correlation value in the embodiment, and [: select_len j ] is the feature screening quantity corresponding to the sub-feature set j.

[0054] All features of the sub-feature set j are sorted in descending order according to the importance value, and a number of features of the feature screening quantity are selected from the beginning of the sequence to construct the importance screening feature set. The calculation expression of the importance screening feature set X j,lgbm is as follows:

[0055] X j,lgbm = argsort (|lgbm_importance j |) [: select_len j ],

[0056]

[0057] where argsort(|lgbm_importance j is the importance value of the sub-feature set j in descending order by absolute value.

[0058] The positive and negative correlation feature set generation module 14 calculates the correlation value of each feature according to the number of features in the screening feature set and the value corresponding to each feature, and constructs the positive correlation feature set and the negative correlation feature set according to the correlation value.

[0059] Wherein, the calculation expression of the positive correlation feature set and the negative correlation feature set is:

[0060]

[0061] where X p is the positive correlation feature set, is the correlation value of feature i in the screening feature set, X n is the negative correlation feature set.

[0062] The discrete and continuous feature set generation module 15 constructs the discrete feature set and the continuous feature set according to the number of unique values corresponding to each feature in the screening feature set.

[0063] Wherein, the calculation expression of the discrete feature set and the continuous feature set is:

[0064]

[0065] where X d is the discrete feature set, X c is the continuous feature set, is the number of unique values of feature i in the screening feature set, and p is the total number of features in the screening feature set.

[0066] The multi-view model feature set generation module 16 obtains the multi-view model feature set according to the positive correlation feature set, the negative correlation feature set, the discrete feature set and the continuous feature set, combined with the machine learning model.

[0067] Wherein, the calculation expression of the multi-view model feature set is:

[0068]

[0069] X p_d =X p ∩X d ,

[0070] X p_c =X p ∩X c ,

[0071] Xn_d =X n ∩X d ,

[0072] X n_c =X n ∩X c ,

[0073] In the formula X multi_view For a multi-view model feature set, || is the set concatenation symbol. For machine learning models based on the input feature set X p_d The generated sample prediction values, For machine learning models based on the input feature set X p _c The generated sample prediction values, For machine learning models based on the input feature set X n_d The generated sample prediction values, For machine learning models based on the input feature set X n_c The generated sample prediction value. In this embodiment, the machine learning model is the LightGBM model, which generates the sample prediction value corresponding to a target based on the values ​​of all features corresponding to a target in the input feature set.

[0074] The WOE calculation module 17 bins the feature set of the multi-view model, calculates the binning value of each feature corresponding to each target in the feature set, and calculates the WOE value of each feature corresponding to each target in the feature set based on all binning values ​​and risk assessment labels. The label result for a target is its features in the multi-view model feature set and its corresponding WOE value.

[0075] Figure 2 This is a schematic diagram of the process by which the woe calculation module generates woe values ​​in an embodiment of the present invention.

[0076] like Figure 2 As shown, the process of generating woe values ​​by the woe calculation module 16 includes the following steps:

[0077] Step S7-1: For the feature set of the multi-view model, generate multiple candidate split thresholds and calculate the split value corresponding to each candidate split threshold. In this embodiment, the candidate split thresholds are obtained by averaging the sorted adjacent samples.

[0078] The expression for calculating the splitting value is as follows:

[0079]

[0080] H(Dmv ) = -∑ c∈C p(c) log2 p(c),

[0081]

[0082] where IG(D mv , θ) is the split value of the candidate split threshold θ corresponding to the multi-view model feature set, |D| is the number of samples in D, and a sample includes the values of all features corresponding to a target in the set D, is the value of feature i corresponding to target k in the multi-view model feature set, v k is target k, y k is the risk judgment label corresponding to target k, C is the total number of features in the multi-view model feature set, and p(c) is the probability of feature c in the multi-view model feature set.

[0083] Step S7-2, the candidate split threshold corresponding to the maximum split value is taken as the optimal split point.

[0084] Step S7-3, steps S6-1 to S6-2 are repeatedly executed until the number of optimal split points reaches a preset value. In this embodiment, two layers of iteration generate up to four optimal split points.

[0085] Step S7-4, the bin boundaries are calculated according to all optimal split points.

[0086] where the calculation expression of the bin boundaries is:

[0087]

[0088] where is the bin boundary value set of is sorted in ascending order. In this embodiment, negative infinity -∞ and positive infinity ∞ are used to ensure that the final bin boundaries cover all values.

[0089] Step S7-5, according to the bin boundaries, the bins values and bin0 values of each feature corresponding to each target in the multi-view model feature set are calculated as the binning values.

[0090] where the calculation expression of the bins values and bin0 values is:

[0091]

[0092] where is the value of feature i corresponding to target k in the multi-view model feature set, is the value of feature i corresponding to target k in the multi-view model feature set, is the left boundary value of the u-th bin interval.​

[0093] Step S7-6, according to all the bin values, the woe value of each feature corresponding to each target in the multi-view model feature set is calculated.

[0094] Wherein, according to the bin0 value of the feature corresponding to the target, the values of the features corresponding to all targets are divided into multiple groups The values of the features corresponding to all targets in the group have the same bin0 value, that is, if The feature corresponding to the target

[0095] The calculation expression of the woe value is:

[0096]

[0097] In the formula The woe value of the feature i corresponding to the target k in the multi-view model feature set, N bad The total number of risk judgment labels corresponding to the values in the group where the value of the feature i corresponding to the target k in the multi-view model feature set is located, N good The total number of risk judgment labels corresponding to the values in the group where the value of the feature i corresponding to the target k in the multi-view model feature set is located, N total The total number of values of the feature i corresponding to the target k in the multi-view model feature set. In this embodiment, the risk judgment label is 0 or 1, 0 represents the good class, and 1 represents the bad class. In other embodiments, other forms can be used to replace 0 and 1 for distinguishing the good and bad classes.

[0098] The control module 18 stores a control program for controlling the operation of each module.

[0099] The following describes the process of the multi-view label generation method for risk portrait construction by using the multi-view label generation device 100 for risk portrait construction with reference to the accompanying drawings.

[0100] Figure 3 It is the flowchart of the multi-view label generation method for risk portrait construction in the embodiment of the application.

[0101] As Figure 3 shown, the multi-view label generation method for risk portrait construction includes the following steps:

[0102] Step S1, all features are divided into multiple categories by using the sub-set division module 11, and a sub-feature set corresponding to each category is generated, the sub-feature set including at least one feature and all values corresponding to the feature.

[0103] Step S2, the screening module 12 is used to calculate the importance value and the correlation value of each feature and the feature screening quantity according to the number of features in each sub-feature set and the value corresponding to the features.

[0104] Step S3, the screening set generation module 13 is used to select a plurality of features from each sub-feature set according to the corresponding importance value, the correlation value and the feature screening quantity to construct a screening feature set.

[0105] Step S4, the positive and negative correlation feature set generation module 14 is used to calculate the correlation value of each feature according to the number of features in the screening feature set and the value corresponding to each feature, and construct a positive correlation feature set and a negative correlation feature set according to the correlation value.

[0106] Step S5, the discrete and continuous feature set generation module 15 is used to construct a discrete feature set and a continuous feature set according to the number of unique values corresponding to each feature in the screening feature set.

[0107] Step S6, the multi-view model feature set generation module 16 is used to obtain a multi-view model feature set according to the positive correlation feature set, the negative correlation feature set, the discrete feature set and the continuous feature set in combination with a machine learning model.

[0108] Step S7, the woe calculation module 17 is used to bin the multi-view model feature set to calculate the bin value of each feature corresponding to each target in the multi-view model feature set, and calculate the woe value of each feature corresponding to each target in the multi-view model feature set according to all bin values corresponding to the multi-view model feature set and the risk judgment label.

[0109] In the embodiment, the radar chart of the user portrait of the target can be constructed according to the label result of the target.

[0110] Figure 4 is a radar chart of a high-risk user portrait in the embodiment of the application.

[0111] Figure 5 is a radar chart of a low-risk user portrait in the embodiment of the application.

[0112] As shown in Figure 4 and Figure 5 , the label result includes four features, i.e., multi_view_1, multi_view_2, multi_view_3 and multi_view_4, which are multi-view labels and their corresponding values. It can be seen that the abnormal subject, i.e., the high-risk user, occupies a large area, and the normal subject, i.e., the low-risk user, occupies a small area, so that the multi-dimensional risk level of the multi-view label on the subject risk can be accurately described.

[0113] In this embodiment, the label result generated by the multi-view label generation method used for risk profiling construction is the model label X. multi_view , compared with general statistical optimization label X anno A comparative experiment was conducted on the scorecard model, and the general statistical optimization label X was performed. anno To develop a logistic regression-based scoring model (SCM) for an internal expert rule-driven labeling system within a financial institution, the model uses the label X... multi_view And general statistical optimization label X anno The training dataset was used to train the model, and the trained model was used to make binary classification predictions. The specific results are shown in the table below:

[0114] KS AUC f10 ​ f11 SCM+X anno ]]> 0.7454 0.9533 0.9657 0.5707 0.6782 SCM+X multi_view ]]> 0.7937 0.9616 0.9702 0.6244 0.7268

[0115] The first column in the table above shows the models trained on different training sets. Columns two through six represent the model's KS, AUC, f10, Recall1, and f11 metrics, respectively. For example, the cell in the second row and second column indicates that the label X was optimized using general statistical methods. anno The KS metric of the trained scoring model SCM on the constructed training set is 0.7454. This indicates that the model performs well on label X. multi_view The data representation shows that the SCM classification ability of the scoring model is further improved, and the model label X multi_view The anomaly detection performance of the scoring model SCM has been enhanced.

[0116] In this embodiment, based on the scoring model SCM, feature coefficients and intercepts are extracted, a baseline score and a score PDO for each doubling of odds are set, and factor weights and offsets are calculated to create a scorecard containing features, coefficients, sample scores, and predicted probabilities. The scores are divided into several equidistant scoring intervals, and the number and proportion of positive and negative samples in each scoring interval are calculated to visually reflect the distribution of positive and negative samples in each scoring interval.

[0117] Figure 6 This is a schematic diagram of the scorecard visualization in an embodiment of the present invention.

[0118] like Figure 6 As shown, the horizontal axis represents the sample score interval values ​​(score_bins), the vertical axis of the histogram represents the number of samples (Count) within that score interval, and the vertical axis of the line graph represents the bad rate within that score interval. Therefore, based on the model label X... multi_view The scoring card intuitively displays how each module distinguishes between positive and abnormal situations, achieving a strong correlation between the abnormal subject and the risk score.

[0119] In this embodiment, some key model labels are selected, and the number and proportion of bad samples are calculated in the group by using the bin value for grouping. The results are shown in the following table:

[0120]

[0121] The first column in the above table is different model labels x m_woe1 , x lm_woe2 and x m_woe3 , the second to sixth columns are woe value, bin range, number of bad class samples, total number of samples and proportion of bad class samples. As can be seen from the above, in the given example of key model label grouping, the proportion of bad samples reaches more than 85%, which shows that the multi-view label can accurately express the bad samples in a certain dimension.

[0122] Effects of the embodiment

[0123] According to the multi-view label generation method and device for risk portrait construction, the correlation value and the number of unique values of the features in the screening set are calculated, the positive correlation feature set, the negative correlation feature set, the discrete feature set and the continuous feature set are constructed, and then the multi-view model feature set is generated. The multi-view model feature set is binned and woe value is calculated to obtain the label result. In summary, the method can generate labels with better risk expression ability, and improve the comprehensiveness and accuracy of the risk portrait.

[0124] Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-view label generation method for risk profile construction, for generating a label result related to a risk profile corresponding to each target according to risk-related data, the risk-related data comprising values of a plurality of different features corresponding to each target, and a risk judgment label corresponding to each target, characterized in that, The method comprises the following steps: Step S1, all the features are divided into multiple categories, and a sub-feature set corresponding to each category is generated, the sub-feature set comprising at least one feature and all values corresponding to the feature; Step S2, for each sub-feature set, the importance value and the correlation value of each feature are calculated according to the number of features in the sub-feature set and the values corresponding to the features; Step S3, from each sub-feature set, a plurality of features are selected to construct a screening feature set according to the corresponding importance value, correlation value and feature screening quantity; Step S4, the correlation value of each feature is calculated according to the number of features in the screening feature set and the values corresponding to each feature, and a positive correlation feature set and a negative correlation feature set are constructed according to the correlation value; Step S5, a discrete feature set and a continuous feature set are constructed according to the number of unique values corresponding to each feature in the screening feature set; Step S6, a multi-view model feature set is obtained according to the positive correlation feature set, the negative correlation feature set, the discrete feature set and the continuous feature set in combination with a machine learning model; Step S7, the multi-view model feature set is binned, and the binning value of each feature corresponding to each target in the multi-view model feature set is calculated, and the woe value of each feature corresponding to each target in the multi-view model feature set is calculated according to all the binning values corresponding to the multi-view model feature set and the risk judgment label, Wherein, the label result of the target is the woe value of each feature corresponding to the target in the multi-view model feature set.

2. The multi-view label generation method for risk portrait construction according to claim 1, wherein: wherein, The calculation expression of the positive correlation feature set and the negative correlation feature set is: where X p is a set of positively correlated features, is a correlation value of feature i in the set of screening features, X n is a set of negatively correlated features.

3. The multi-view label generation method for risk portrait construction according to claim 1, wherein: wherein The calculation expression of the discrete feature set and the continuous feature set is: where X d is a set of discrete features, X c is a set of continuous features, is the number of unique values for feature i in the screening feature set, and p is the total number of features in the screening feature set.

4. The multi-view label generation method for risk portrait construction according to claim 1, wherein: wherein The calculation expression of the multi-view model feature set is: X p_d = X p ∩ X d , X p_c = X p ∩ X c , X n_d = X n ∩ X d , X n_c = X n ∩ X c , In the formula X multi_view For a multi-view model feature set, || is the set concatenation symbol. The machine learning model is based on the input feature set X p_d The generated sample prediction values, The machine learning model is based on the input feature set X p_c The generated sample prediction values, The machine learning model is based on the input feature set X n_d The generated sample prediction values, The machine learning model is based on the input feature set X n_c The generated sample prediction values.

5. The multi-view label generation method for risk portrait construction according to claim 1, wherein: wherein The calculation expression of the correlation value is: wherein is the relevance value of feature i in feature set j, is the value of feature i corresponding to the kth target in feature set j, is the mean value of feature i in feature set j, y k is the risk judgment label corresponding to the kth target, is the mean value of risk judgment labels of all targets, and n is the total number of the targets.

6. The multi-view label generation method for risk portrait construction according to claim 1, wherein: wherein, In step S2, the importance value is calculated according to the information gain of the feature on the node of the tree of the LightGBM model, The calculation expression of the importance value is: wherein is the importance value of feature i in the sub-feature set j, p is the total number of features in the sub-feature set j, M is the total number of trees in the LightGBM model, N m is the set of nodes of the mth tree, is the indicator function, ΔL m,n is the information gain of feature i on node n of the mth tree, when node n is split using feature i, when node n does not split using feature i, In step S2, the calculation expression of the feature screening quantity is: In the formula, select len j is the feature screening quantity corresponding to the sub-feature set j, p is the total number of features in the sub-feature set j, is a floor operation.

7. The multi-view label generation method for risk portrait construction according to claim 1, wherein: wherein In the step S3, the calculation expression of the screening feature set is: X j,selected = X j,corr U X j,lgbm , where X j,selected is the screening set corresponding to the sub-feature set j, X j,corr is the relevance screening feature set corresponding to the sub-feature set j, X j,lgbm is the importance screening feature set corresponding to the sub-feature set j, X selected is the screening feature set, and m is the total number of sub-feature sets. all the features in the sub-feature set j are sorted according to the correlation values from large to small, and the features in the sequence are selected from the beginning to obtain the number of the feature screening quantity, thereby constructing the correlation screening feature set, all the features in the sub-feature set j are sorted according to the importance values from large to small, and the features in the sequence are selected from the beginning to obtain the number of the feature screening quantity, thereby constructing the importance screening feature set.

8. The multi-view label generation method for risk profiling according to claim 1, It is characterized in that: The step S7 includes the following sub-steps: Step S7-1, for the multi-view model feature set, a plurality of candidate split thresholds are generated, and the split values corresponding to each candidate split threshold are calculated; Step S7-2, the candidate split threshold corresponding to the maximum split value is taken as the optimal split point; Step S7-3, steps S6-1 to S6-2 are repeatedly executed until the number of optimal split points reaches a preset value; Step S7-4, the bin boundaries are calculated according to all the optimal split points; Step S7-5, according to the bin boundaries, the bins value and bin0 value of each feature corresponding to each target in the multi-view model feature set are calculated as the binning value; Step S7-6, according to all the binning values, the woe value of each feature corresponding to each target in the multi-view model feature set is calculated.

9. The multi-view label generation method for risk portrait construction according to claim 8, characterized in that: wherein, The calculation expression of the split value is: H(D mv ) = -∑ c∈C p(c) log2 p(c), where IG(D mv , θ) is the split value of candidate split threshold θ corresponding to the multi-view model feature set, |D| is the number of samples in D, is the value of feature i corresponding to target k in the multi-view model feature set, v k is target k, y k is the risk judgment label corresponding to target k, C is the total number of features in the multi-view model feature set, and p(c) is the probability of feature c in the multi-view model feature set. The calculation expression of the bin boundary is: wherein is a set of bin boundary values, is an ascending order sorting of all optimal split points, The calculation expression of the bins value and the bin0 value is: In the formula is the value of feature i corresponding to target k in the multi-view model feature set, is is the left boundary value of the u-th bin interval, In the step S7-6, according to the bin0 value of the feature corresponding to the target, the values of the features corresponding to all the targets are divided into a plurality of groups, all the values of the features corresponding to the targets in the group have the same bin0 value, the risk judgment label is a binary classification label, The calculation expression of the woe value is: wherein is the woe value of the feature i corresponding to the target k in the multiple view model feature set, N bad is the total number of risk judgment labels with the label of bad class corresponding to all values in the group where the value of the feature i corresponding to the target k in the multiple view model feature set, N good is the total number of risk judgment labels with the label of good class corresponding to all values in the group where the value of the feature i corresponding to the target k in the multiple view model feature set, N total is the total number of values of the feature i corresponding to the target k in the multiple view model feature set. 10.A multi-view label generating device for risk profile construction, configured to generate a label result associated with a risk profile of each target according to risk-related data, wherein the risk-related data comprises values of a plurality of different features corresponding to each target, and a risk judgment label corresponding to each target, and the device is characterized in that, It includes: The sub-set division module divides all the features into a plurality of categories, generates sub-feature sets corresponding to each category, and each sub-feature set includes at least one feature and all the values corresponding to the feature; The screening module calculates the importance value and correlation value of each feature and the feature screening quantity according to the number of features in each sub-feature set and the values corresponding to the features; The screening set generation module selects a plurality of features from each sub-feature set according to the corresponding importance value, correlation value and feature screening quantity to construct a screening feature set; The positive and negative correlation feature set generation module calculates the correlation value of each feature according to the number of features in the screening feature set and the values corresponding to each feature, and constructs a positive correlation feature set and a negative correlation feature set according to the correlation value. The discrete continuous feature set generation module constructs a discrete feature set and a continuous feature set according to the number of unique values corresponding to each feature in the screening feature set; The multi-view model feature set generation module obtains a multi-view model feature set by combining a machine learning model according to the positive correlation feature set, the negative correlation feature set, the discrete feature set and the continuous feature set; The WOE calculation module performs binning on the multi-view model feature set, calculates the binning values of each feature corresponding to each target in the multi-view model feature set, and calculates the WOE values of each feature corresponding to each target in the multi-view model feature set according to all the binning values corresponding to the multi-view model feature set and the risk judgment label, Wherein, the label result of the target is each feature in the multi-view model feature set and the corresponding WOE value of the target.

Images