Method and system for configuring strategy parameters by decision engine
By conducting correlation analysis and gradient enhancement algorithm training on the business history data of the decision engine, a decision tree model is built, which solves the subjectivity problem of policy parameter configuration of decision engine and improves the accuracy and robustness of decisions.
Patent Information
- Application Number
- CN202411205250.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-12-24
AI Technical Summary
The existing decision engine policy parameter configuration is greatly affected by subjectivity, making it difficult to ensure the objectivity and accuracy of decisions.
By collecting user's business history decision data, conducting correlation analysis, building a decision model, using gradient enhancement algorithm to train a decision tree model, and optimizing decision parameter configuration.
It improves the configuration speed and accuracy of decision engine policy parameters, and enhances the objectivity and robustness of decisions.
Smart Images
Figure CN119180513A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent decision-making technology, and more specifically to a method and system for configuring strategy parameters of a decision engine. Background Art
[0002] With the digital transformation of banks and the development of Internet finance, intelligent risk control and marketing businesses such as financial risk control and content risk control are increasing. The decision-making engine, as the core system supporting the above business scenarios, has also received more attention.
[0003] Generally speaking, any scenario involving the use of rules or models to make business decisions can be implemented using a decision engine. The decision engine can be used in various fields including financial risk control, content risk control, recommendation marketing, IoT monitoring, etc. For example, when making decisions on a user's loan limit, credit card limit, transfer limit, etc., the decision engine can be used to make corresponding limit decisions based on the collected decision data. However, most existing decision engines are expert systems, and the engine's rules and policy parameters are formulated by experts or expert groups. The configuration of the decision engine's policy parameters is subject to a large degree of subjective influence, and the objectivity and accuracy of the decisions made by the decision engine are easily affected.
[0004] Therefore, how to accurately configure the decision engine policy parameters is an urgent problem that those skilled in the art need to solve. Summary of the invention
[0005] In view of this, the present invention provides a method and system for configuring policy parameters of a decision engine, thereby improving the configuration speed and accuracy of the policy parameters in the policy engine.
[0006] In order to achieve the above object, the present invention provides the following technical solutions:
[0007] The present invention discloses a method for configuring policy parameters of a decision engine, such as Figure 1 As shown, the specific steps are as follows:
[0008] Collect the historical business decision data of each user and get the sample data set after sorting;
[0009] Performing correlation analysis on the decision parameters of each sample in the sample data set to obtain a correlation matrix;
[0010] Determine direct strategy parameters of each business according to existing business decision templates, and determine associated strategy parameters of the direct strategy parameters according to the correlation matrix;
[0011] A decision model is constructed with the direct policy parameters and the associated policy parameters, and training and verification are performed through sample data of the corresponding business in the sample data set to obtain a decision engine and policy parameters for the corresponding business.
[0012] Furthermore, the sample data includes: business type, decision parameter, decision value, and result attribute;
[0013] The arrangement includes: deleting abnormal samples in the historical decision data of each business, normalizing parameters of numerical type, encoding parameters of character type, and obtaining Boolean type result attributes according to the decision execution result.
[0014] Furthermore, the correlation analysis is specifically as follows:
[0015] Clustering the numerical type parameters in the decision parameters and classifying the character type parameters to obtain classification decision parameters, and constructing a classification correlation matrix according to the classification decision parameters;
[0016] According to the weights of each category of the classification decision parameters, the classification correlation matrix is converted into the correlation matrix representing the correlation of each decision parameter.
[0017] Furthermore, the decision parameter is expressed as c i , the classification decision parameter is represented by c ik , where i represents the i-th item among all n decision parameters, and k represents the k-th category among the i-th decision parameters; the classification correlation matrix is expressed as in represents the classification decision parameter c ik With c jk′ The similarity value of is the total number of classification decision parameters, m i is the classification number of the i-th decision parameter; the similarity value calculation formula is:
[0018]
[0019] in, are the classification decision parameters c contained in the sample data set ik 、c jk′ The subset of ,|| represents the number of samples in the subset.
[0020] Furthermore, the correlation matrix is represented as X=(x ij ) n×n , where x ij represents the decision parameter c i With c j The similarity value of x ij The formula is:
[0021]
[0022] in, represents the classification decision parameter c ik With cjk′ The weight coefficient of the similarity value is The sample data sets contain decision parameters c i 、c j A subset of .
[0023] Furthermore, the decision model is a decision tree model, and the number G of decision subtrees in the decision tree is determined according to the total number of the direct strategy parameters and the associated strategy parameters, and the decision tree model is obtained by merging the G decision subtrees; each of the decision subtrees is a weak classifier, and the decision tree model is trained by a gradient boosting algorithm to determine the weights of each decision subtree.
[0024] Furthermore, the training specifically includes:
[0025] Selecting sample data whose business type corresponds to the current decision tree model in the sample data set, and selecting positive samples according to the result attributes of the sample data; supplementing the positive samples according to the direct strategy parameters and the associated strategy parameters, and the user database to obtain training samples;
[0026] The number of outer loops h is determined according to the number of decision subtrees G, and the number of classifications m of each decision parameter in the training sample is determined according to the number of decision subtrees G. i Determine the number of inner loops H; the training steps are:
[0027] S1, initialize the number of cycles, set h = 0, H = 0;
[0028] S2, h+1, inner loop training for the h-th decision subtree;
[0029] S3, H+1, train and update the weak classifier in the gradient direction with reduced error until H=m h ;
[0030] S4, perform Logistic transformation on the prediction function F(C) and use the gradient boosting algorithm to update the weights of the weak classifier;
[0031] S5, when h=G, H=m G When , the loop ends and the classification tree model is obtained, otherwise it returns to S2;
[0032] Among them, m h Represents the number of categories of the h-th decision parameter.
[0033] Furthermore, the weak classifier is represented by T h (C;θ h ), where C is the input parameter value, θ h is the loss function; the loss function formula is Among them, r represents the rth sample in the total R samples of the current inner loop training, F h-1 (C r ) is the predicted value of the prediction function for the rth sample before updating, T h (C r θ h ) represents the predicted value of the h-th weak classifier for the r-th sample, P r Represents the true value of the rth sample; Logistic transformation of the prediction function is expressed as Where W h is the weight of the h-th weak classifier.
[0034] The present invention also discloses a system for configuring policy parameters of a decision engine, comprising:
[0035] Data acquisition module: collects the business historical decision data of each user and obtains the sample data set after sorting;
[0036] Data analysis module: performing correlation analysis on the decision parameters of each sample in the sample data set to obtain a correlation matrix;
[0037] Parameter selection module: determining direct strategy parameters of each business according to existing business decision templates, and determining associated strategy parameters of the direct strategy parameters according to the correlation matrix;
[0038] Parameter configuration module: construct a decision model with the direct policy parameters and the associated policy parameters, and perform training and verification with the sample data of the corresponding business in the sample data set to obtain the decision engine and policy parameters of the corresponding business.
[0039] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a method and system for configuring strategy parameters of a decision engine. By acquiring historical decision data and analyzing and mining the relationship between decision parameters, it is possible not only to select direct decision parameters according to existing decision strategies, but also to select associated parameters according to the correlation between the parameters, so that the decision parameters of the decision engine are more comprehensive and the model is constructed more accurately to improve the accuracy of the decision; and the decision tree model is trained using the gradient boosting algorithm using the supplemented and corrected training data, so that the trained decision tree model has better prediction ability and robustness, and the decision parameters are configured using the prediction function of the decision tree model to more accurately configure the strategy parameters of the decision engine. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0041] Figure 1 It is a schematic diagram of the overall process of an embodiment of the present invention. DETAILED DESCRIPTION
[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0043] The embodiment of the present invention discloses a method for configuring policy parameters of a decision engine, such as Figure 1 As shown, the specific steps are as follows:
[0044] Collect the historical business decision data of each user and get the sample data set after sorting;
[0045] Perform correlation analysis on the decision parameters of each sample in the sample data set to obtain a correlation matrix;
[0046] Determine the direct strategy parameters of each business according to the existing business decision templates, and determine the associated strategy parameters of the direct strategy parameters according to the correlation matrix;
[0047] A decision model is constructed with direct policy parameters and associated policy parameters, and is trained and verified through sample data of the corresponding business in the sample data set to obtain the decision engine and policy parameters of the corresponding business.
[0048] In a specific embodiment, the sample data includes: business type, decision parameter, decision value, result attribute;
[0049] The sorting includes: deleting abnormal samples in the historical decision data of each business, normalizing parameters of numerical type, encoding parameters of character type, and obtaining Boolean type result attributes according to the decision execution results.
[0050] Specifically, business types include personal loans, transfers, clearing services, financial product trading, foreign exchange transactions, credit cards and other businesses. The transaction limits and loan amounts of these businesses can be given decision results through the decision engine, thereby reducing the financial risks of financial institutions or individuals. Decision parameters include user personal parameters, such as the user's historical credit rating, age, fixed assets, current assets, income, labor consumption behavior, work and company type, as well as macroeconomic parameters such as bank interest rates, foreign exchange rates, bond yields, deposit reserve ratios, GDP growth rates, inflation rates, etc. Decision values can be loan limits, loan time limits, transfer limits, credit card limits, foreign exchange transaction limits, etc. The result attribute is based on whether financial risks occur after the business decision is executed, such as loan defaults, credit card overdue, foreign exchange transaction losses, suspicious transactions, etc. The result attribute of the decision sample data that has financial risks is set to "0", and the result attribute of the decision sample data that has not financial risks is set to "1".
[0051] In a specific embodiment, the correlation analysis is specifically:
[0052] Cluster the numerical parameters in the decision parameters, classify the character parameters, obtain the classified decision parameters, and construct the classification correlation matrix according to the classified decision parameters. Among them, the numerical parameters, such as fixed asset value and income value, are clustered by K-means method, and the classification results of the numerical parameters are obtained; the character parameters, such as the user's credit rating and job type, already have independent classification attributes, and only need to encode the parameters of different character types and classify the parameters according to the encoding. There may be correlations between two different parameters, but these correlations are often not linearly related, or can be obtained by simply calculating the distance between the parameters; and for character parameters, it is often difficult to directly calculate the correlation between them and other parameters. Therefore, by classifying different decision parameters, according to the correlation between the classified decision parameters of different types of decision parameters, the correlation between the two decision parameters is comprehensively evaluated, so as to explore the intrinsic correlation between the parameters, better select the decision parameters of the decision task, and improve the accuracy of the decision.
[0053] According to the weights of each category of the classification decision parameters, the classification correlation matrix is converted into a correlation matrix representing the correlation of each decision parameter.
[0054] In a specific embodiment, the decision parameter is represented by c i , the classification decision parameter is represented by c ik , where i represents the i-th item among all n decision parameters, and k represents the k-th category among the i-th decision parameters; the classification correlation matrix is expressed as in represents the classification decision parameter c ik With cjk′ The similarity value of is the total number of classification decision parameters, m i is the number of classifications of the i-th decision parameter; the similarity value calculation formula is:
[0055]
[0056] in, are the classification decision parameters c contained in the sample data set ik 、c jk′ The subset of ,|| represents the number of samples in the subset.
[0057] In a specific embodiment, the correlation matrix is represented as X=(x ij ) n×n , where x ij represents the decision parameter c i With c j The similarity value of x ij The formula is:
[0058]
[0059] in, represents the classification decision parameter c ik With c jk′ The weight coefficient of the similarity value is The sample data sets contain decision parameters c i 、c j A subset of .
[0060] In a specific embodiment, the decision model is a decision tree model. The number G of decision subtrees in the decision tree is determined according to the total number of direct strategy parameters and associated strategy parameters. The decision tree model is obtained by merging G decision subtrees. Each decision subtree is a weak classifier. The decision tree model is trained by a gradient boosting algorithm to determine the weights of each decision subtree.
[0061] In a specific embodiment, the training specifically includes:
[0062] Select sample data whose business type corresponds to the current decision tree model in the sample data set, and select positive samples according to the result attributes of the sample data, that is, select samples with the result attribute of "1"; supplement the positive samples according to the direct strategy parameters and the associated strategy parameters, as well as the user database, that is, supplement the associated strategy parameters missing in the original samples according to the historical data of the financial institution on the user and the historical data of macroeconomic parameters, so as to obtain training samples, and further obtain the training sample set and the test sample set.
[0063] The number of outer loops h is determined according to the number of decision subtrees G, and the number of classifications m of each decision parameter in the training sample is determined according to the number of decision subtrees G. i Determine the number of inner loops H; the training steps are:
[0064] S1, initialize the number of cycles, set h = 0, H = 0;
[0065] S2, h+1, inner loop training for the h-th decision subtree;
[0066] S3, H+1, train and update the weak classifier in the gradient direction with reduced error until H=m h ;
[0067] S4, perform Logistic transformation on the prediction function F(C) and use the gradient boosting algorithm to update the weights of the weak classifier;
[0068] S5, when h=G, H=m G When , the loop ends and the classification tree model is obtained, otherwise it returns to S2;
[0069] Among them, m h Indicates the number of classifications of the h-th decision parameter. In the inner loop training, the weak classifier of the h-th decision subtree is trained by selecting each classification decision parameter in the h-th decision parameter one by one, and the relationship between the h-th decision parameter and the decision value is fitted. When an inner loop training is completed, the residual between the current model prediction value and the actual value after updating the weak classifier is calculated through the gradient boosting algorithm, so as to update the weights in the prediction function.
[0070] In a specific embodiment, the weak classifier is represented by T h (C;θ h ), where C is the input parameter value, θ h is the loss function; the loss function formula is Among them, r represents the rth sample in the total R samples of the current inner loop training, F h-1 (C r ) is the predicted value of the prediction function for the rth sample before updating, T h (C r θ h ) represents the predicted value of the h-th weak classifier for the r-th sample, P r Represents the true value of the rth sample; the weight parameters of the weak classifier are set by minimizing the loss value, and the prediction function is transformed into Logistic form: Where W h is the weight of the h-th weak classifier.
[0071] In a specific embodiment, the training step also includes: S6, using the test sample set to test the trained classification tree model, when the average value and standard deviation of the deviation value between the predicted accurate value and the true value are within the set threshold range, the final decision model is obtained, otherwise the weights of each weak classifier in the classification tree model trained by S5 are used as initial values, and steps S2 to S5 are re-executed to train the classification tree model again until the average value and standard deviation of the deviation value between the predicted accurate value and the true value are within the set threshold range. Since the decision tree model is obtained by merging multiple sub-decision trees in series, the weight value obtained by a single training may have a local convergence problem. By re-training and updating the weights of the decision tree model that has been trained to determine the weights, the weight values of each sub-decision tree can be better re-optimized to improve the accuracy of the decision model.
[0072] The embodiment of the present invention also discloses a system for configuring policy parameters of a decision engine, including:
[0073] Data acquisition module: collects the business historical decision data of each user and obtains the sample data set after sorting;
[0074] Data analysis module: performs correlation analysis on the decision parameters of each sample in the sample data set to obtain a correlation matrix;
[0075] Parameter selection module: Determine the direct strategy parameters of each business according to the existing business decision templates, and determine the associated strategy parameters of the direct strategy parameters according to the correlation matrix;
[0076] Parameter configuration module: Build a decision model with direct policy parameters and associated policy parameters, and train and verify it through sample data of the corresponding business in the sample data set to obtain the decision engine and policy parameters of the corresponding business.
[0077] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0078] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for configuring policy parameters of a decision engine, characterized in that: The specific steps are as follows: Collect the historical business decision data of each user and get the sample data set after sorting; Performing correlation analysis on the decision parameters of each sample in the sample data set to obtain a correlation matrix; Determine direct policy parameters of each service according to existing service decision templates, and determine associated policy parameters of the direct policy parameters according to the correlation matrix; A decision model is constructed with the direct policy parameters and the associated policy parameters, and training and verification are performed through sample data of the corresponding business in the sample data set to obtain a decision engine and policy parameters for the corresponding business.
2. A method for configuring policy parameters of a decision engine according to claim 1, characterized in that: The sample data includes: business type, decision parameters, decision value, and result attributes; The arrangement includes: deleting abnormal samples in the historical decision data of each business, normalizing parameters of numerical type, encoding parameters of character type, and obtaining Boolean type result attributes according to the decision execution result.
3. The method for configuring policy parameters of a decision engine according to claim 1, characterized in that: The correlation analysis is specifically as follows: Clustering the numerical type parameters in the decision parameters and classifying the character type parameters to obtain classification decision parameters, and constructing a classification correlation matrix according to the classification decision parameters; According to the weights of each category of the classification decision parameters, the classification correlation matrix is converted into the correlation matrix representing the correlation of each decision parameter.
4. A method for configuring policy parameters of a decision engine according to claim 3, characterized in that: The decision parameter is denoted as c i , the classification decision parameter is represented by c ik , where i represents the i-th item among all n decision parameters, and k represents the k-th category among the i-th decision parameters; the classification correlation matrix is expressed as in represents the classification decision parameter c ik With c jk′ The similarity value of is the total number of classification decision parameters, m i is the classification number of the i-th decision parameter; the similarity value calculation formula is: in, are the classification decision parameters c contained in the sample data set ik 、c jk′ The subset of ,|| represents the number of samples in the subset.
5. A method for configuring policy parameters of a decision engine according to claim 4, characterized in that: The correlation matrix is represented as X=(x ij ) n×n , where x ij represents the decision parameter c i With c j The similarity value of x ij The formula is: in, represents the classification decision parameter c ik With c jk′ The weight coefficient of the similarity value is The sample data sets contain decision parameters c i 、c j A subset of .
6. A method for configuring policy parameters of a decision engine according to claim 1, characterized in that: The decision model is a decision tree model. The number G of decision subtrees in the decision tree is determined according to the total number of the direct strategy parameters and the associated strategy parameters. The decision tree model is obtained by merging G decision subtrees. Each decision subtree is a weak classifier. The decision tree model is trained by a gradient boosting algorithm to determine the weight of each decision subtree.
7. A method for configuring policy parameters of a decision engine according to claim 6, characterized in that: The training specifically includes: Selecting sample data whose business type corresponds to the current decision tree model in the sample data set, and selecting positive samples according to the result attributes of the sample data; supplementing the positive samples according to the direct strategy parameters and the associated strategy parameters, and the user database to obtain training samples; The number of outer loops h is determined according to the number of decision subtrees G, and the number of classifications m of each decision parameter in the training sample is determined according to the number of decision subtrees G. i Determine the number of inner loops H; the training steps are: S1, initialize the number of cycles, set h = 0, H = 0; S2, h+1, inner loop training for the h-th decision subtree; S3, H+1, train and update the weak classifier in the gradient direction with reduced error until H=m h ; S4, perform Logistic transformation on the prediction function F(C) and use the gradient boosting algorithm to update the weights of the weak classifier; S5, when h=G, H=m G When , the loop ends and the classification tree model is obtained, otherwise it returns to S2; Among them, m h Represents the number of categories of the h-th decision parameter.
8. A method for configuring policy parameters of a decision engine according to claim 7, characterized in that: The weak classifier is denoted as T h (C;θ h ), where C is the input parameter value, θ h is the loss function; the loss function formula is Among them, r represents the rth sample in the total R samples of the current inner loop training, F h-1 (C r ) is the predicted value of the prediction function for the rth sample before updating, T h (C r θ h ) represents the predicted value of the h-th weak classifier for the r-th sample, P r Represents the true value of the rth sample; Logistic transformation of the prediction function is expressed as Where W h is the weight of the h-th weak classifier.
9. A system for configuring policy parameters of a decision engine, characterized in that: include: Data acquisition module: collects the business historical decision data of each user and obtains the sample data set after sorting; Data analysis module: performing correlation analysis on the decision parameters of each sample in the sample data set to obtain a correlation matrix; Parameter selection module: determining direct strategy parameters of each business according to existing business decision templates, and determining associated strategy parameters of the direct strategy parameters according to the correlation matrix; Parameter configuration module: construct a decision model with the direct policy parameters and the associated policy parameters, and perform training and verification with the sample data of the corresponding business in the sample data set to obtain the decision engine and policy parameters of the corresponding business.