A feature combination screening method, device, and medium based on a three-stage funnel type

Abstract

The present invention relates to the technical field of data mining, and discloses a feature combination screening method, device and medium based on a three-stage funnel type, including: performing feature combination on original risk control features according to a combination operator, storing the combined features in a distributed manner, calculating the correlation coefficient value between each combined feature after distributed storage and a business objective, and performing a first funnel screening on the combined features according to the correlation coefficient value; freezing the tree structure of the constructed business tree model, and using each screened combined feature to split the frozen business tree model one by one; calculating the optimization index of the split business tree model, and performing a second funnel screening on each screened combined feature according to the optimization index; calculating the importance value of each combined feature after the second screening by using the business tree model, and performing a third funnel screening according to the importance value to obtain a target feature combination that meets the business objective. The present invention can improve the efficiency during feature screening.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a feature combination screening method, device, and medium based on a three-stage funnel type. Background Art

[0002] With the rapid development of Internet finance, financial services such as mobile payment and wealth management have become an indispensable part of people's daily lives. In financial credit operations, an intelligent risk control system is the cornerstone of the Internet finance field, and how to extract valuable information from data is particularly important. Incorporating as many feature information as possible into model training can enable the model to extract more information, which is beneficial to improving the accuracy of the final result. However, with the addition of more features, the mutual influence between features and the complexity of the model are also increasing continuously, which greatly hinders the improvement of the model effect. Therefore, it is necessary to screen a large number of features to achieve the same business indicator effect as data mining experts.

[0003] Existing feature screening directly calculates the importance of each feature, and thus realizes feature screening based on the importance. In practical applications, with the growth of business and the increase in feature collection, data mining experts usually need to face thousands of features, and there are also the feature magnitudes brought by the pairwise combination or other combination forms of these features. Just calculating the importance of each feature leads to an excessive amount of calculation, resulting in low efficiency during feature screening. Summary of the Invention

[0004] The present invention provides a feature combination screening method, device, and medium based on a three-stage funnel type, and its main purpose is to solve the problem of low efficiency during feature screening.

[0005] To achieve the above object, a feature combination screening method based on a three-stage funnel type provided by the present invention includes:

[0006] Obtain the original risk control features corresponding to the risk control business scenario, and perform feature combination on the original risk control features according to a pre-generated combination operator to obtain a combined feature set;

[0007] Perform distributed storage on the combined feature set, calculate the correlation coefficient value between each combined feature in the combined feature set after distributed storage and a preset business target, and perform a first funnel screening on the combined features according to the correlation coefficient value to obtain a first screened combined feature;

[0008] Construct a business tree model according to the original risk control features, freeze the tree structure of the business tree model, add each combined feature in the first screened combined feature to the frozen business tree model one by one, and split the frozen business tree model after adding the combined feature;

[0009] Calculate the optimization metrics of the split business tree model, and perform a second funnel screening on each combined feature in the first screened combined features according to the optimization metrics to obtain second screened combined features;

[0010] Use the business tree model to calculate the importance value of each combined feature in the second screened combined features, and perform a third funnel screening on the second screened combined features according to the importance value to obtain a target feature combination that meets the business objective.

[0011] Optionally, the feature combination of the original risk control features according to pre-generated combination operators to obtain combined features includes:

[0012] Divide the original risk control features into continuous features and discrete features;

[0013] Divide the pre-generated combination operators into continuous feature operators and discrete feature operators;

[0014] Perform feature combination on the continuous features according to the number of operators of the continuous feature operators and the continuous feature operators to obtain continuous combined features;

[0015] Perform feature combination on the discrete features according to the number of operators of the discrete feature operators and the discrete feature operators to obtain discrete combined features;

[0016] Aggregate the continuous combined features and the discrete combined features into a combined feature set.

[0017] Optionally, the distributed storage of the combined feature set includes:

[0018] Extract the single features corresponding to each combined feature in the combined feature set, and store the single features in a preset database list;

[0019] Extract the single features corresponding to the preset business requirement data;

[0020] Determine the combined features of the business requirement data according to the single features, and store the combined features in a predefined storage location.

[0021] Optionally, the calculation of the correlation coefficient value between each combined feature in the distributed storage combined feature set and a preset business objective includes:

[0022] Extract the feature values corresponding to each combined feature in the distributed storage combined feature set;

[0023] Determine the target value according to the preset business objective;

[0024] Calculate the correlation coefficient value between each of the eigenvalues and the target value one by one.

[0025] Optionally, splitting the frozen business tree model for the added combined features includes:

[0026] Divide the combined features into target numerical features and target categorical features;

[0027] Input the target numerical features into the frozen business tree model, and split the target numerical features in the frozen business tree model through a preset splitting threshold;

[0028] Merge the split target numerical features into a numerical feature split set;

[0029] Input the target categorical features into the frozen business tree model, and split the target categorical features in the frozen business tree model through a preset splitting category;

[0030] Merge the split target categorical features into a categorical feature split set.

[0031] Optionally, calculating the optimization metrics of the split business tree model includes:

[0032] Count the number of true positives and false negatives in the split business tree model, and calculate the true positive rate based on the number of true positives and the number of false negatives;

[0033] Count the number of false positives and true negatives in the split business tree model, and calculate the false positive rate based on the number of false positives and the number of true negatives;

[0034] Generate a characteristic curve based on the true positive rate and the false positive rate;

[0035] Calculate the curve integral corresponding to the characteristic curve, and determine the integral value of the curve integral as the optimization metric of the classified business tree model.

[0036] Optionally, performing a second funnel screening on each combined feature in the first screened combined features according to the optimization metric to obtain second screened combined features, includes:

[0037] Determine the target combined features in the first screened combined features according to the leaf nodes in the business tree model;

[0038] Determine the optimization metric of the business tree model as the optimization metric corresponding to the target combined features;

[0039] When the optimization metric corresponding to the target combined feature is a positive metric, save the target combined feature;

[0040] When the optimization index corresponding to the target combined feature is a negative index, delete the target combined feature;

[0041] The saved target combined features are aggregated into the second screened combined features.

[0042] Optionally, calculating the importance value of each combined feature in the second screened combined features by using the business tree model includes:

[0043] Count the number of splits of each combined feature in the second screened combined features in the business tree model, and determine the participation degree of each combined feature according to the number of splits;

[0044] Normalize the participation degree, and determine the importance value of each combined feature in the second screened combined features as the normalized participation degree.

[0045] To solve the above problems, the present invention also provides an electronic device, and the electronic device includes:

[0046] At least one processor; and,

[0047] A memory communicatively connected to the at least one processor; wherein,

[0048] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the above-mentioned feature combination screening method based on a three-stage funnel type.

[0049] To solve the above problems, the present invention also provides a computer-readable storage medium, and at least one computer program is stored in the computer-readable storage medium, and the at least one computer program is executed by a processor in an electronic device to implement the above-mentioned feature combination screening method based on a three-stage funnel type.

[0050] In the embodiment of the present invention, a set of defined standard combination operators are used to perform N-order combinations on the original risk control features to form a large number of combined features, and the large number of combined features are distributedly stored. When calculation is required, the data of single features are fetched for storage, thereby greatly reducing the storage requirements; then the three-stage funnel screening is used to screen the features effective for the business, so as to automatically improve the effect of the business model. In the same business scenario MOB6 (MOB6 is an index used in the financial industry to describe the sixth consecutive month of overdue of a customer's loan or credit card account, mainly used to evaluate the customer's credit risk level) overdue ratio model, the efficiency is relatively increased by 1900% compared with that of data mining experts, and the final business indicator AUC is relatively increased by 2%. Therefore, the feature combination screening method, device and medium based on the three-stage funnel type proposed by the present invention can solve the problem of low efficiency in feature screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 FIG. is a schematic flowchart of a feature combination screening method based on a three-stage funnel type provided by an embodiment of the present invention;

[0052] Figure 2 FIG. is a schematic diagram of a feature combination provided by an embodiment of the present invention;

[0053] Figure 3 FIG. is a schematic diagram of distributed storage provided by an embodiment of the present invention;

[0054] Figure 4 FIG. is a schematic diagram of a correlation coefficient value provided by an embodiment of the present invention;

[0055] Figure 5 FIG. is a schematic diagram of freezing model improvement provided by an embodiment of the present invention;

[0056] Figure 6 FIG. is a schematic diagram of feature importance analysis provided by an embodiment of the present invention;

[0057] Figure 7 FIG. is a schematic flowchart of feature screening provided by an embodiment of the present invention;

[0058] Figure 8 FIG. is a schematic structural diagram of an electronic device for implementing the feature combination screening method based on the three-stage funnel type provided by an embodiment of the present invention.

[0059] The implementation, functional features and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0061] An embodiment of the present application provides a feature combination screening method based on a three-stage funnel type. The execution subject of the feature combination screening method based on the three-stage funnel type includes, but is not limited to, at least one of electronic devices such as a server, a terminal, etc. that can be configured to execute the method provided by the embodiment of the present application. In other words, the feature combination screening method based on the three-stage funnel type can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to: a single server, a server cluster, a cloud server, or a cloud server cluster, etc. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms.

[0062] Referring to Figure 1 As shown, it is a schematic flowchart of a feature combination screening method based on a three-stage funnel type provided by an embodiment of the present invention. In this embodiment, the feature combination screening method based on the three-stage funnel type includes:

[0063] S1. Obtain the original risk control features corresponding to the risk control business scenario, and perform feature combination on the original risk control features according to the pre-generated combination operator to obtain a combined feature set.

[0064] In the embodiment of the present invention, the risk control business scenario refers to the situation constructed for specific risk control requirements and goals during the risk control process, such as a credit risk scenario, an anti-fraud scenario, and the original risk control features refer to the user's basic information, user transaction data, and user overdue bills involved in the risk control business scenario. Among them, the user's basic information includes, but is not limited to, name, age, and gender; the user transaction data includes, but is not limited to, transaction time, transaction amount, and transaction frequency; the user overdue bills include, but is not limited to, the number of overdue times and the number of overdue days. Among them, the original risk control features corresponding to the risk control business scenario can be obtained from a pre-stored storage area through computer statements with data scraping functions (such as Java statements, Python statements, etc.), and the storage area includes, but is not limited to, a database, a blockchain. Further, it is necessary to combine based on the existing risk control features and mine effective feature information for the business from the massive combined features. When combining the current thousands of original features, such as pairwise combination or three-way combination, the magnitude of the features will reach more than hundreds of millions. Therefore, it is necessary to design a set of standard combination operators to standardize how to combine features.

[0065] In an embodiment of the present invention, the combined feature refers to operating on and transforming the original risk control features to generate new features, so as to better capture potential patterns in the data, that is, new features generated by performing mathematical operations, logical operations, or other transformations on one or more original features.

[0066] In an embodiment of the present invention, the step of performing feature combination on the original risk control features according to pre-generated combination operators to obtain combined features includes:

[0067] Dividing the original risk control features into continuous features and discrete features;

[0068] Dividing the pre-generated combination operators into continuous feature operators and discrete feature operators;

[0069] Performing feature combination on the continuous features according to the number of operators of the continuous feature operators and the continuous feature operators to obtain continuous combined features;

[0070] Performing feature combination on the discrete features according to the number of operators of the discrete feature operators and the discrete feature operators to obtain discrete combined features;

[0071] Collecting the continuous combined features and the discrete combined features into a combined feature set.

[0072] Specifically, the original risk control features include continuous and discrete features. Among them, continuous features are numerical features such as age, income, credit score, etc.; discrete features are features with limited values, such as (male / female), region (city / village), etc. Then, the generated combination operators are divided into continuous feature operators and discrete feature operators. The continuous feature operators include single-feature operators (such as: log, sqrt) and multi-feature operators (such as: +, -), and the discrete feature operators include logical symbols (such as and, or) and functions (such as group by). Then, the original risk control features are combined according to the continuous feature operators and discrete feature operators, and the combination between features is controlled by order.

[0073] Specifically, pre-generated continuous feature operators are used to combine continuous features. For example, for features A and B, new features such as A + B, A / B can be generated, and then all the generated continuous combined features are collected; pre-generated discrete feature operators are used to combine discrete features. For example, for features C and D, new features such as C and D, C or D can be generated, and then all the generated discrete combined features are collected. Finally, all the generated continuous combined features and discrete combined features are collected into a combined feature set, such as Figure 2As shown, it is a schematic diagram of feature combinations. The combinations of features are determined according to the number of combination operators. For example, the first-order combination is itself, that is, one combination operator, representing ; The second-order combination is the combination of two features, that is, two combination operators, representing ; The third-order combination is the combination of three features, that is, three combination operators, representing ; The fourth-order combination is the combination of four features, that is, four combination operators, representing func(A, B, C).

[0074] Furthermore, a huge amount of data will be formed for the combined features. Obviously, it is difficult to store and calculate such a huge amount of data in a single-machine environment, and distributed storage needs to be used. Therefore, it is necessary to perform distributed storage on the huge amount of feature data formed by combinations to improve the efficiency of subsequent feature screening.

[0075] S2. Perform distributed storage on the combined feature set, calculate the correlation coefficient value between each combined feature in the combined feature set after distributed storage and a preset business target, and perform a first funnel screening on the combined features according to the correlation coefficient value to obtain the first screened combined features.

[0076] In the embodiment of the present invention, when performing distributed storage on the combined feature set, single-feature storage and combined-feature storage are adopted, that is, the combined data is not directly stored, but the single-feature data is retrieved for storage when calculation is required, thus greatly reducing the storage requirements.

[0077] In the embodiment of the present invention, the distributed storage of the combined feature set includes:

[0078] Extract the single features corresponding to each combined feature in the combined feature set, and store the single features in a preset database list;

[0079] Extract the single features corresponding to the preset business requirement data;

[0080] Determine the combined features of the business requirement data according to the single features, and store the combined features in a predefined storage location.

[0081] Specifically, extracting the single features corresponding to each combined feature in the combined feature set means decomposing each combined feature into multiple individual features, storing the extracted single features in the database list corresponding to the feature fields, and storing the combined features corresponding to the single features in a predefined storage location. That is, each combined feature in the combined feature set, that is, each column of features, represents digital data corresponding to a combined feature in the matrix, that is, each combined feature is stored according to the column dimension.

[0082] Exemplarily, assume that the data stored on a single machine is massive and difficult to store. Then, single-feature storage and combined-feature storage are adopted. That is, instead of directly storing the combined data, the single-feature data is retrieved for storage when calculation is needed. In data storage, the basic features of the data are stored separately. For example, in an e-commerce user dataset, the user's age, gender, purchase frequency, consumption amount, etc. are all different single features, and each single feature will be stored separately in the corresponding database field or data file column. Combined-feature storage does not pre-calculate or pre-store the combined features generated by combining multiple single features, but stores the formulas for generating these combined features. For example, in an e-commerce scenario, if you want to obtain the combined feature of user consumption intensity, it may be defined as consumption amount / purchase frequency. At this time, the consumption intensity of each user will not be calculated and stored in advance, but the formula for calculating this consumption intensity will be stored. When the business needs to use the consumption intensity data, it will obtain the data from the place where the single features (consumption amount and purchase frequency) are stored according to this formula and calculate the result, as Figure 3 shown, which is a schematic diagram of distributed storage. Each combined formula is stored separately in a single location according to the column dimension, as stored in one column, stored in one column, stored in one column, thereby improving the efficiency of subsequent feature screening.

[0083] Furthermore, since a large number of combined features will be generated during the combined-feature process, it is obviously impossible to perform calculations using the traditional tree model feature importance for screening. Therefore, it is necessary to quickly screen the features, eliminate some irrelevant features, and then perform effective screening to obtain the effective features by themselves. First, the features are eliminated based on correlation.

[0084] In the embodiments of the present invention, during the screening process, first, the correlation of a large number of derivative features is eliminated, that is, the first-layer funnel correlation elimination. That is, at this stage, a large number of features are faced, and a method with fast calculation speed is required to quickly eliminate the invalid features. Therefore, a method for selecting the correlation between the features and the business objective is used for fast calculation and elimination, and the Pearson correlation coefficient is used for fast calculation and elimination. The correlation coefficient value is a statistical index used to measure the strength and direction of the linear relationship between two variables.

[0085] In the embodiments of the present invention, calculating the correlation coefficient value between each combined feature in the combined feature set after distributed storage and a preset business objective includes:

[0086] Extracting the feature values corresponding to each combined feature in the combined feature set after distributed storage;

[0087] Determining the target value according to the preset business objective;

[0088] Calculate the correlation coefficient values between the eigenvalues and the target value one by one.

[0089] Specifically, calculate the Pearson correlation coefficient between each combined feature and the business target using the Pearson correlation coefficient, where the business target can be configured differently according to different scenarios. For example, in the anti-fraud business scenario, the corresponding business target is to determine whether it is anti-fraud, 0 indicates no anti-fraud, and 1 indicates anti-fraud. In the user overdue business scenario, the corresponding business target is to determine whether the user is overdue, 0 indicates no overdue, and 1 indicates overdue. Then the business target is a sample label constructed based on the business scenario. , and then according to the eigenvalue and the sample label calculate the correlation coefficient values between each combined feature and the business target.

[0090] Specifically, the correlation coefficient calculation formula is , where represents and 's correlation coefficient, represents and 's covariance, represents the expected value, represents 's variance (standard deviation), represents 's variance, represents 's mean, represents 's mean, represents and 's product's expected value, represents 's expected value, represents 's expected value, represents 's square's expected value, represents 's square's expected value, represents 's expected value's square, represents 's expected value's square, represents the eigenvalue corresponding to the combined feature, represents the target value corresponding to the business target, and 0.8 - 1.0 is extremely strongly correlated; 0.6 - 0.8 is strongly correlated; 0.4 - 0.6 is moderately correlated; 0.2 - 0.4 is weakly correlated; 0.0 - 0.2 is extremely weakly correlated or not correlated.

[0091] Further, perform a first funnel screening on the combined features according to the correlation coefficient values to obtain the first screened combined features, that is, remove the combined features with calculated correlation coefficient values less than 0.2, and retain the combined features with correlation coefficient values greater than 0.2, so as to gather the retained combined features into the first screened combined features. As Figure 4 shown, it is a schematic diagram of the correlation coefficient values. The combined feature f1 corresponds to the data 11231, and is calculated with the value to obtain a correlation coefficient value of 1.0 between the combined feature f1 and ; the correlation coefficient value between the combined feature f2 and is 0.37; the correlation coefficient value between the combined feature f3 and is 0.1. Then, the combined feature f3 is removed, and the combined features f1 and f2 are retained.

[0092] Furthermore, although after the first layer of removal, the actual number of features faced is still huge. Therefore, a second funnel is also needed for filtering, so the boosting method is adopted, that is, this process needs to consider both the calculation speed and the effect.

[0093] S3. Construct a business tree model based on the original risk control features, freeze the tree structure of the business tree model, add each combined feature in the first screened combined features to the frozen business tree model one by one, and split the frozen business tree model with the added combined features.

[0094] In the embodiment of the present invention, the business tree model refers to a model generated based on the business scenario for risk control, which helps decision-makers understand complex business processes and decision rules through a tree structure, and is convenient for identifying key factors and their impacts on the final decision.

[0095] Specifically, first perform data cleaning on the feature data corresponding to the original risk control features, that is, handle missing values, outliers, and duplicate data to ensure the integrity and accuracy of the data. Then define the business tree structure, clarify the goal of constructing the business tree model, such as identifying high-risk users. Furthermore, select feature data as the nodes of the tree according to business logic, and divide the data according to the feature values of each node. If the credit score > 700, it is divided into "low risk"; if the credit score ≤ 700 and the income > 5000, it is divided into "medium risk"; otherwise, it is "high risk". Then, divide the feature data corresponding to the original risk control features into a training set and a test set, use the training set to construct the business tree model, and use the test set to adjust the parameters to improve the model performance.

[0096] Further, a business tree model is trained based on the original online business data, and then the business tree model is frozen, that is, no changes are made to the tree structure subsequently. The decision conditions of all nodes, the connection relationships between nodes, and the results of leaf nodes in the business tree model are fixed. Subsequently, regardless of the input data, the business tree model processes it according to the tree structure of the frozen business tree model. Therefore, each combined feature in the first screened combined features is added to the frozen business tree model one by one. That is, for each combined feature, it is added to the business tree model, and then only this combined feature is continuously trained, so as to obtain the evaluation optimization index (such as AUC) during the training process. Other parts in the training of the business tree model are frozen, and only a single combined feature can participate in the training. Therefore, the calculation speed is fast, and because it is directly associated with business indicators, a certain effect is also taken into account.

[0097] In the embodiment of the present invention, during the process of screening combined features in the second-layer funnel, both the calculation speed and the effect need to be considered. Therefore, it is necessary to freeze the business tree model generated based on the original business data and use the boosting method to evaluate the effect brought by each combined feature.

[0098] In the embodiment of the present invention, the splitting of the frozen business tree model with added combined features includes:

[0099] Dividing the combined feature into a target numerical feature and a target categorical feature;

[0100] Inputting the target numerical feature into the frozen business tree model and splitting the target numerical feature in the frozen business tree model through a preset splitting threshold;

[0101] Combining the split target numerical features into a numerical feature split set;

[0102] Inputting the target categorical feature into the frozen business tree model and splitting the target categorical feature in the frozen business tree model through a preset splitting category;

[0103] Combining the split target categorical features into a categorical feature split set.

[0104] Specifically, each combined feature in the first screened combined features after the first layer screening is divided into a target numerical feature and a target categorical feature. That is, the target numerical feature is continuous data, and the target categorical feature is discrete data. Then, for each target numerical feature, it is input into the frozen business data model one by one, that is, the target numerical feature is input into the business tree model one by one. According to business requirements or historical data, a preset splitting threshold is set. For example, income greater than 5000 yuan and less than or equal to 5000 yuan can be used as the splitting threshold. The frozen business tree is split using the splitting threshold. This process will generate new child nodes, representing the classification results according to different feature values, that is, merged into a numerical feature splitting set, and the numerical features with the same classification result are merged in one set. For each target categorical feature, it is input into the frozen business tree model one by one for splitting, that is, the target categorical feature is input into the business tree model one by one to determine the splitting rules for different categories. For example, gender can be divided into "male" and "female", and regions can be classified according to different geographical locations. It is split according to the set categories to generate new child nodes, representing the decision results for different categories, that is, merged into a categorical feature splitting set, and the categorical features with the same classification result are merged in one set.

[0105] Specifically, the frozen business tree model can be updated according to the new feature combination, and the final result will provide a more refined classification, that is, only continue to train for each combined feature, and the optimization metrics during the training process for each combined feature, such as Figure 5 As shown, it is a schematic diagram of the improvement of the frozen model. First, a business tree model is constructed based on the original feature sample data, then the business tree model is frozen, and each combined feature is input one by one. The combined feature f1 is input into the frozen business tree model, and the obtained optimization metric is -0.67. The combined feature f2 is input into the frozen business tree model, and the obtained optimization metric is 0.75. The combined feature f3 is input into the frozen business tree model, and the obtained optimization metric is 0.75.

[0106] Furthermore, a second funnel screening is performed on the combined features based on the optimization metrics, so that the order of magnitude of the screened features is within a controllable range.

[0107] S4. Calculate the optimization metrics of the split business tree model, and perform a second funnel screening on each combined feature in the first screened combined features according to the optimization metrics to obtain second screened combined features.

[0108] In the embodiments of the present invention, the optimization metric refers to a numerical metric used to evaluate the model effect during the process of training the business tree model for each combined feature, which helps analyze the performance of the model in a specific task for comparison, adjustment, and improvement.

[0109] In the embodiments of the present invention, the optimization metrics for calculating the split service tree model include:

[0110] Count the number of true positives and false negatives in the split service tree model, and calculate the true positive rate according to the number of true positives and the number of false negatives;

[0111] Count the number of false positives and true negatives in the split service tree model, and calculate the false positive rate according to the number of false positives and the number of true negatives;

[0112] Generate a characteristic curve according to the true positive rate and the false positive rate;

[0113] Calculate the curve integral corresponding to the characteristic curve, and determine the integral value of the curve integral as the optimization metric of the classified service tree model.

[0114] Specifically, first count the instance data where each combined feature is predicted as a positive class by the service tree model during the training process, that is, true positives (TP), and the number of instances where the service tree model is predicted as a negative class but is actually a positive class, that is, false negatives (FN). Then the true positive rate is TP / (TP + FN); then count the number of instances where the model is wrongly predicted as a positive class but is actually a negative class, that is, false positives (FP), and the number of instances where the model is correctly predicted as a negative class, that is, true negatives (TN). Then the false positive rate is FP / (FP + TN). Furthermore, a characteristic curve ROC (Receiver Operating Characteristic Curve) is drawn with the false positive rate as the X-axis and the true positive rate as the Y-axis, that is, the corresponding relationship between the false positive rate and the true positive rate of each combined feature is plotted as a coordinate point. Thus, a characteristic curve is generated according to the coordinate points, and then the area under the curve (AUC, Area Under the Curve) is calculated according to the area of the region below the characteristic curve, that is, the area under the curve is calculated by a numerical integration method (such as the trapezoidal rule), that is, the curve integral of the characteristic curve is calculated, and the integral value corresponding to the curve integral is determined as the optimization metric of the classified service tree model. Thus, the classification performance of the service tree model is measured according to the calculated optimization metric (such as AUC). The value range of AUC is between 0 and 1, and the value closer to 1 indicates better model performance.

[0115] Furthermore, according to the optimization metric, the combined features after the first-layer funnel screening are subjected to a second screening to ensure that the number of remaining features is usually within a controllable range.

[0116] In the embodiments of the present invention, the second screening of combined features refers to further screening and optimization of the feature combinations selected in the first layer to improve the performance and accuracy of the model.

[0117] In an embodiment of the present invention, the second funnel screening of each combined feature in the first screened combined features according to the optimization index to obtain second screened combined features includes:

[0118] Determine the target combined feature in the first screened combined features according to the leaf nodes in the business tree model;

[0119] Determine the optimization index of the business tree model as the optimization index corresponding to the target combined feature;

[0120] When the optimization index corresponding to the target combined feature is a positive index, save the target combined feature;

[0121] When the optimization index corresponding to the target combined feature is a negative index, delete the target combined feature;

[0122] Collect the saved target combined features as the second screened combined features.

[0123] Specifically, train a business tree model based on the original online business data, and then freeze the business tree model, that is, no further changes are made to the tree structure in the future. Next, for each combined feature, add it to the tree model, and then only continue to train for this combined feature, so as to view the optimization index during the training process. That is, determine the business node of the business tree model as the added target combined feature, determine the optimization index corresponding to this combined feature according to the added target combined feature. If the optimization index is positive, it indicates that this combined feature is effective for the business, that is, retain this combined feature. If it is negative, it indicates that this combined feature is ineffective for the business, that is, eliminate this combined feature. Furthermore, collect all the retained combined features to form the second screened combined features, where the optimization index includes but is not limited to AUC (Area Under the Curve).

[0124] Exemplarily, in Figure 5 the optimization index corresponding to the combined feature f1 is -0.67, the optimization index corresponding to the combined feature f2 is 0.75, and the optimization index corresponding to the combined feature f3 is 0.75. If the optimization index corresponding to the initial business tree model is 0.7, then the optimization index corresponding to the combined feature f1 is negative, so the combined feature f1 is eliminated, while the optimization indexes corresponding to the combined features f2 and f3 are positive, so the combined features f2 and f3 are retained. The combined feature f2 has improved by 0.05 compared to the original business tree model, and the combined feature f3 has also improved by 0.05 compared to the original business tree model.

[0125] Furthermore, if the number of features of the second screened combination features obtained through the second funnel screening is within a controllable range, a business tree model related to the business objective needs to be used for feature importance analysis to obtain relatively important feature data.

[0126] S5. Calculate the importance value of each combination feature in the second screened combination features using the business tree model, and perform a third funnel screening on the second screened combination features according to the importance value to obtain a target feature combination that meets the business objective.

[0127] In an embodiment of the present invention, the importance value is an index used to measure the contribution degree of a feature in the business tree model, indicating the influence of a certain feature on model decision-making, usually presented in numerical form. The higher the numerical value, the more important the feature is in the business tree model.

[0128] In an embodiment of the present invention, the calculating the importance value of each combination feature in the second screened combination features using the business tree model includes:

[0129] Count the number of splits of each combination feature in the second screened combination features in the business tree model, and determine the participation degree of each combination feature according to the number of splits;

[0130] Normalize the participation degree, and determine the normalized participation degree as the importance value of each combination feature in the second screened combination features.

[0131] Specifically, during the process of training the model, record the number of splits of each combination feature in the business tree model, then determine the number of splits as the participation degree of each combination feature. Furthermore, in order to compare the importance of different features, it is necessary to normalize the participation degree by dividing the participation degree of each feature by the sum of the participation degrees of all features, so as to determine the normalized participation degree as the importance value of each combination feature in the second screened combination features. As Figure 6 shown, it is a schematic diagram of feature importance analysis. Determine the importance of each combination feature based on the feature sample data. For example, the importance value corresponding to the combination feature f1 is 0.8, the importance value corresponding to the combination feature f2 is 0.6, and the importance value corresponding to the combination feature f3 is 0.01.

[0132] Specifically, for the calculated importance value corresponding to each combination feature, sort all the combination features according to the importance value in descending order to obtain a feature sequence corresponding to the combination features, and select the first k features in the feature sequence as the target feature combination that meets the business objective.

[0133] Furthermore, as Figure 7As shown in the figure, it is a schematic flowchart of feature screening, including a feature combination process and a feature screening process. For example, there are 1000 original features, that is, N = 1000. Using the addition combination method, about 500,000 derivative features can be obtained. Based on the screening strategy, finally about dozens of features are obtained, so that the order of magnitude of the finally screened features is within a controllable range.

[0134] In the embodiment of the present invention, a set of defined standard combination operators are used to perform N - order combination on the original risk control features, forming a large number of combined features, and storing the large number of combined features in a distributed manner. When calculation is required, the data of single features are fetched for storage, thus greatly reducing the storage requirements; then the three - stage funnel screening is used to select the features effective for the business, so as to automatically improve the effect of the business model. In the MOB6 overdue ratio model in the same business scenario, the efficiency is relatively increased by 1900% compared with that of data mining experts, and finally the business indicator AUC is relatively increased by 2%. Therefore, the feature combination screening method, device and medium based on the three - stage funnel type proposed by the present invention can solve the problem of low efficiency in feature screening.

[0135] As Figure 8 shown, it is a schematic structural diagram of an electronic device for implementing the feature combination screening method based on the three - stage funnel type provided by an embodiment of the present invention.

[0136] The electronic device may include a processor 10, a memory 11, a communication bus 12 and a communication interface 13, and may also include a computer program stored in the memory 11 and operable on the processor 10, such as a feature combination screening program based on the three - stage funnel type.

[0137] Among them, the processor 10 may be composed of integrated circuits in some embodiments. For example, it may be composed of a single - packaged integrated circuit, or may be composed of multiple integrated circuits with the same or different functions, including a combination of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors and various control chips. The processor 10 is the control core (Control Unit) of the electronic device, connecting various components of the entire electronic device through various interfaces and lines, running or executing programs or modules stored in the memory 11 (such as executing a feature combination screening program based on the three - stage funnel type, etc.), and calling data stored in the memory 11 to perform various functions of the electronic device and process data.

[0138] The memory 11 at least includes one type of readable storage medium, which includes flash memory, mobile hard disk, multimedia card, card-type memory (such as SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, etc. The memory 11 can be an internal storage unit of the electronic device in some embodiments, such as the mobile hard disk of the electronic device. The memory 11 can also be an external storage device of the electronic device in other embodiments, such as a plug-in mobile hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc. equipped on the electronic device. Further, the memory 11 can also include both the internal storage unit and the external storage device of the electronic device. The memory 11 can be used not only to store application software installed in the electronic device and various types of data, such as the code of the feature combination screening program based on the three-stage funnel type, etc., but also to temporarily store the data that has been output or will be output.

[0139] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to implement the connection and communication between the memory 11 and at least one processor 10, etc.

[0140] The communication interface 13 is used for the communication between the above-mentioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface can include a wired interface and / or a wireless interface (such as a WI-FI interface, a Bluetooth interface, etc.), which is usually used to establish a communication connection between this electronic device and other electronic devices. The user interface can be a display (Display), an input unit (such as a keyboard (Keyboard)). Optionally, the user interface can also be a standard wired interface, a wireless interface. Optionally, in some embodiments, the display can be an LED display, a liquid crystal display, a touch liquid crystal display, and an OLED (Organic Light-Emitting Diode) toucher, etc. Among them, the display can also be appropriately referred to as a display screen or a display unit, which is used to display the information processed in the electronic device and to display the visual user interface.

[0141] Only the electronic device with components is shown in the figure. Those skilled in the art can understand that the structure shown in the figure does not constitute a limitation on the electronic device, and it may include fewer or more components than those shown, or combine certain components, or have different component arrangements.

[0142] For example, although not shown, the electronic device may further include a power source (such as a battery) for supplying power to each component. Preferably, the power source can be logically connected to the at least one processor 10 through a power management device, so as to implement functions such as charging management, discharging management, and power consumption management through the power management device. The power source may also include any components such as one or more DC or AC power sources, a recharge device, a power failure detection circuit, a power converter or inverter, a power status indicator, etc. The electronic device may also include various sensors, a Bluetooth module, a Wi-Fi module, etc., which will not be elaborated here.

[0143] It should be understood that the embodiments are for illustrative purposes only and are not limited by this structure in the scope of the patent application.

[0144] The feature combination screening program based on the three-stage funnel type stored in the memory 11 in the electronic device is a combination of multiple instructions. When running in the processor 10, it can implement:

[0145] Obtain the original risk control features corresponding to the risk control business scenario, and perform feature combination on the original risk control features according to the pre-generated combination operator to obtain a combined feature set;

[0146] Perform distributed storage on the combined feature set, calculate the correlation coefficient value between each combined feature in the distributed stored combined feature set and a preset business target, and perform a first funnel screening on the combined features according to the correlation coefficient value to obtain a first screened combined feature;

[0147] Construct a business tree model according to the original risk control features, freeze the tree structure of the business tree model, add each combined feature in the first screened combined feature to the frozen business tree model one by one, and split the frozen business tree model after adding the combined features;

[0148] Calculate the optimization index of the split business tree model, and perform a second funnel screening on each combined feature in the first screened combined feature according to the optimization index to obtain a second screened combined feature;

[0149] Use the business tree model to calculate the importance value of each combined feature in the second screened combined feature, and perform a third funnel screening on the second screened combined feature according to the importance value to obtain a target feature combination that meets the business target.

[0150] Specifically, for the specific implementation method of the above instructions by the processor 10, reference may be made to the description of the relevant steps in the corresponding embodiments of the accompanying drawings, which will not be elaborated herein.

[0151] Further, if the modules / units integrated in the electronic device are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disc, a computer memory, a read-only memory (ROM, Read-Only Memory).

[0152] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor of an electronic device, it can implement:

[0153] Obtain the original risk control features corresponding to the risk control business scenario, perform feature combination on the original risk control features according to the pre-generated combination operator to obtain a combined feature set;

[0154] Perform distributed storage on the combined feature set, calculate the correlation coefficient value between each combined feature in the combined feature set after distributed storage and a preset business target, and perform a first funnel screening on the combined features according to the correlation coefficient value to obtain a first screened combined feature;

[0155] Construct a business tree model according to the original risk control features, freeze the tree structure of the business tree model, add each combined feature in the first screened combined feature to the frozen business tree model one by one, and split the frozen business tree model after adding the combined features;

[0156] Calculate the optimization index of the split business tree model, perform a second funnel screening on each combined feature in the first screened combined feature according to the optimization index to obtain a second screened combined feature;

[0157] Use the business tree model to calculate the importance value of each combined feature in the second screened combined feature, and perform a third funnel screening on the second screened combined feature according to the importance value to obtain a target feature combination that meets the business target.

[0158] In several embodiments provided by the present invention, it should be understood that the disclosed device, medium, and method can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0159] The module described as a separation component may or may not be physically separated. The component shown as a module may or may not be a physical unit, that is, it may be located in one place or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0160] In addition, each functional module in various embodiments of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a combination of hardware and software functional modules.

[0161] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms.

[0162] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is not limited only by the above description. Therefore, it is intended to cover all changes within the meaning and scope of equivalent elements falling within the protection scope of the present invention.

[0163] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0164] In addition, obviously, the term "including" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in the system can also be implemented by one unit or device through software or hardware. The terms such as first and second are used to represent names and do not represent any specific order.

[0165] 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 preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A feature combination screening method based on a three-stage funnel type, characterized in that The method includes: Obtaining the original risk control features corresponding to the risk control business scenario, and performing feature combination on the original risk control features according to the pre-generated combination operators to obtain a combined feature set, where the risk control business scenario includes a credit risk scenario and an anti-fraud scenario, the original risk control features include the basic user information, user transaction data, and user overdue bills involved in the risk control business scenario, the user transaction data includes transaction frequency, and the user overdue bills include the number of overdue times; Storing the combined feature set distributively, calculating the correlation coefficient value between each combined feature in the distributively stored combined feature set and a preset business objective, and performing a first funnel screening on the combined features according to the correlation coefficient value to obtain a first screened combined feature; Constructing a business tree model based on the original risk control features, freezing the tree structure of the business tree model, adding each combined feature in the first screened combined feature to the frozen business tree model one by one, and splitting the frozen business tree model with the added combined feature; Calculating the optimization index of the split business tree model, and performing a second funnel screening on each combined feature in the first screened combined feature according to the optimization index to obtain a second screened combined feature; Calculating the importance value of each combined feature in the second screened combined feature by using the business tree model, and performing a third funnel screening on the second screened combined feature according to the importance value to obtain a target feature combination that meets the business objective; Among them, the storing the combined feature set distributively includes: extracting the single features corresponding to each combined feature in the combined feature set, and storing the single features in a preset database list; extracting the single features corresponding to the preset business requirement data; determining the combined features of the business requirement data according to the single features, and storing the combined features in a predefined storage location.

2. The feature combination screening method based on the three-stage funnel type according to claim 1, wherein The performing feature combination on the original risk control features according to the pre-generated combination operators to obtain a combined feature set includes: Dividing the original risk control features into continuous features and discrete features; Dividing the pre-generated combination operators into continuous feature operators and discrete feature operators; Performing feature combination on the continuous features according to the number of continuous feature operators and the continuous feature operators to obtain continuous combined features; Performing feature combination on the discrete features according to the number of discrete feature operators and the discrete feature operators to obtain discrete combined features; Collecting the continuous combined features and the discrete combined features into a combined feature set.

3. The feature combination screening method based on a three-stage funnel type according to claim 1, characterized in that The calculating the correlation coefficient value between each combined feature in the distributively stored combined feature set and a preset business objective includes: Extracting the feature values corresponding to each combined feature in the distributively stored combined feature set; Determining the target value according to the preset business objective; Calculating the correlation coefficient value between the feature value and the target value one by one.

4. The feature combination screening method based on a three-stage funnel type according to claim 1, characterized in that The splitting the frozen business tree model with the added combined feature includes: Dividing each combined feature into a target numerical feature and a target category feature; Input the target numerical feature into the frozen business tree model, and split the target numerical feature in the frozen business tree model through a preset splitting threshold; Merge the split target numerical features into a numerical feature split set; Input the target categorical feature into the frozen business tree model, and split the target categorical feature in the frozen business tree model through a preset splitting category; Merge the split target categorical features into a categorical feature split set.

5. The feature combination screening method based on a three-stage funnel type as described in claim 1, wherein The calculation of the optimization index of the split business tree model includes: Count the number of true positives and false negatives in the split business tree model, and calculate the true positive rate according to the number of true positives and the number of false negatives; Count the number of false positives and true negatives in the split business tree model, and calculate the false positive rate according to the number of false positives and the number of true negatives; Generate a characteristic curve according to the true positive rate and the false positive rate; Calculate the curve integral corresponding to the characteristic curve, and determine the integral value of the curve integral as the optimization index of the classified business tree model.

6. The feature combination screening method based on a three-stage funnel type according to claim 1, characterized in that, The second funnel screening of each combined feature in the first screened combined features according to the optimization index, to obtain second screened combined features, includes: Determine the target combined feature in the first screened combined features according to the leaf nodes in the business tree model; Determine the optimization index of the business tree model as the optimization index corresponding to the target combined feature; When the optimization index corresponding to the target combined feature is a positive index, save the target combined feature; When the optimization index corresponding to the target combined feature is a negative index, delete the target combined feature; Collect the saved target combined features as the second screened combined features.

7. The feature combination screening method based on a three-stage funnel type according to claim 1, characterized in that The calculation of the importance value of each combined feature in the second screened combined features by using the business tree model includes: Count the number of splits of each combined feature in the second screened combined features in the business tree model, and determine the participation degree of each combined feature according to the number of splits; Normalize the participation degree, and determine the normalized participation degree as the importance value of each combined feature in the second screened combined features.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the feature combination screening method based on a three-stage funnel type as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program, when executed by a processor, implements the feature combination screening method based on a three-stage funnel type as described in any one of claims 1 to 7.

Images