KR102428405B1 - System and method for recommendation of customized financial products - Google Patents

Abstract

A customized financial product recommendation system according to one technical aspect of the present invention includes a pre-processing unit that performs pre-processing by reducing the dimension of customer data by performing principal component analysis on customer data, and multiple customer data pre-processed by the pre-processing unit. A label model unit for generating a label model using the label classification model, and an imbalance data processing unit for resolving imbalance by performing random sampling on the label model generated by the label model unit.

Description

Customized financial product recommendation system and method {SYSTEM AND METHOD FOR RECOMMENDATION OF CUSTOMIZED FINANCIAL PRODUCTS}

The present invention relates to a customized financial product recommendation system and method.

As financial products with various and complex structures appear one after another, various technologies are being developed to provide financial services by judging the investment propensity of financial products by using information such as customer's financial product transaction details.

As an example of such a technology, a customized financial product recommendation technology that predicts and provides a recommendation for a personalized financial product using a user's transaction data is being developed.

In order to recommend customized financial products, statistical analysis is performed using customer's financial data. In this statistical analysis, conventionally, a single-label classification model is generally used due to the complexity of the system and the like. However, since this single-label classification model handles each financial product independently, it has a limitation in not reflecting the case where customers purchase multiple financial products at the same time.

The prior art has a limitation in that it is difficult to compare various financial products at the same time and reflect the actual needs of financial consumers who construct a portfolio with a plurality of financial products.

That is, there is a demand for a customized financial product recommendation technology that can consider various financial products at the same time, but the conventional technology has a problem in that it cannot satisfy this demand.

Korean Patent Publication No. 10-2017-0125631

One technical aspect of the present invention is to solve the problems of the prior art, and to increase the accuracy of customer's financial product recommendation by using a multi-label classification model that is an extension of the single-label classification model. It is to provide a customized financial product recommendation system and recommendation method.

In addition, a technical aspect of the present invention provides a customized financial product recommendation system and recommendation method that can efficiently configure system resources by performing principal component analysis on customer data to reduce the dimension of customer data and perform preprocessing will do

In addition, one technical aspect of the present invention is to provide a customized financial product recommendation system and a method for providing the same, which can increase prediction accuracy and efficiently configure system resources by performing unbalanced data processing on a label model.

The above objects and various advantages of the present invention will become more apparent from preferred embodiments of the present invention by those skilled in the art.

One technical aspect of the present invention proposes a customized financial product recommendation system. The customized financial product recommendation system includes a preprocessor that performs principal component analysis on customer data to reduce the dimension of customer data to perform preprocessing, and labels using a multi-label classification model for customer data preprocessed by the preprocessor It may include a label model unit for generating a model and an imbalance data processing unit for resolving imbalance by performing random sampling on the label model generated by the label model unit.

In one embodiment, the customer data includes at least one of financial product transaction amount information, financial product transaction number information, deposit and withdrawal amount information, financial product average balance information, customer profile information, customer investment propensity information, and long-term dormancy information can do.

In one embodiment, the preprocessor sets a new coordinate system such that, when the data of the customer data is mapped to one axis, the axis with the largest variance is the first main component and the second axis is the second main component, and this can be used to linearly transform customer data.

In an embodiment, the label model unit may use at least one of a random forest model, a logistic regression model, an Adaboost model, and an artificial neural network model as a single label classification model.

In an embodiment, the label model unit may generate a multi-label model using a binary association technique that converts the preprocessed customer data into a plurality of single label data.

In one embodiment, the label model unit generates a classification result for a previous label according to a binary association technique with respect to the preprocessed customer data, and uses the classification result for the previous label as an explanatory variable when predicting the next label. A classifier chain can be created by reflecting the correlations that exist between labels.

In an embodiment, the label model unit may generate a label model using a probabilistic classifier chain that sequentially calculates conditional probabilities of labels and calculates a joint probability distribution therefor.

In an embodiment, the imbalance data processing unit may process the imbalance data by performing a first sampling process for performing fixed sampling and a second sampling process for arbitrarily performing sampling using an explanatory variable.

In one embodiment, the imbalance data processing unit, by using the long-term dormancy information included in the customer data, if there is no transaction for a specific financial product for a preset period, using information such as a profile irrelevant to the financial product transaction to determine the purchase probability can be calculated

Another technical aspect of the present invention proposes a customized financial product recommendation method. The customized financial product recommendation method includes: performing principal component analysis on customer data to reduce the dimension of customer data to perform preprocessing; generating a label model using a multi-label classification model for preprocessed customer data; and resolving imbalance by performing random sampling on the label model.

In one embodiment, the performing of the pre-processing may include setting an axis having the largest variance as a first principal component when the data of the customer data is mapped to one axis, and setting the axis having the second largest variance as a second principal component It may include the steps of setting , setting a new coordinate system using the first principal component and the second principal component, and linearly transforming the customer data using the new coordinate system.

In one embodiment, the generating of the label model includes generating a multi-label model using a binary association technique that converts the preprocessed customer data into a plurality of single label data; Generating a classification result for the previous label according to the association technique, and using the classification result for the previous label as an explanatory variable when predicting the next label to reflect the correlation existing between the labels to generate a classifier chain and a label The method may include at least one of generating a label model using a probabilistic classifier chain that sequentially calculates conditional probabilities of and calculates a joint probability distribution for them.

In an embodiment, the linear transformation of the customer data may include a first sampling processing step of performing fixed sampling using an explanatory variable, and a second sampling processing step of optionally performing sampling.

In one embodiment, the step of linearly transforming the customer data includes excluding from the training data about customers who have not traded for more than a preset period using long-term dormancy information included in the customer data, and a specific financial product for a preset period. If there is no transaction for , the method may include calculating a purchase probability by using information such as a profile that is not related to financial product transactions.

The means for solving the above-described problems do not enumerate all the features of the present invention. Various means for solving the problems of the present invention may be understood in more detail with reference to specific embodiments of the detailed description below.

According to an embodiment of the present invention, there is an effect of increasing the accuracy of a customer's financial product recommendation using a multi-label classification model that is an extension of the single-label classification model.

In addition, according to an embodiment of the present invention, there is an effect of reducing the dimension of data by performing principal component analysis on customer data, thereby alleviating the multicollinearity problem, and efficiently configuring system resources.

In addition, according to an embodiment of the present invention, there is an effect of increasing the accuracy of prediction by performing unbalanced data processing on the label model and efficiently configuring system resources.

1 is a view for explaining an application example of a customized financial product recommendation system according to an embodiment of the present invention.
2 is a block diagram illustrating a customized financial product recommendation system according to an embodiment of the present invention.
FIG. 3 is a view for explaining a classifier chain performed by the label model unit shown in FIG. 2 .
4 is a diagram illustrating a simplified comparison between a binary association technique and a classifier chain.
FIG. 5 is a diagram illustrating a result of applying the ensemble classifier chain in the label model unit shown in FIG. 2 .
6 is a graph illustrating an example of a customer's purchase probability distribution for three types of financial products.
FIG. 7 is a graph illustrating a case in which a classifier chain is applied with respect to FIG. 6 .
8 is a flowchart illustrating a method for recommending a customized financial product according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

However, the embodiments of the present invention may be modified in various other forms, and the scope of the present invention is not limited to the embodiments described below. In addition, the embodiments of the present invention are provided in order to more completely explain the present invention to those of ordinary skill in the art.

That is, the above-described objects, features and advantages will be described below in detail with reference to the accompanying drawings, and accordingly, a person of ordinary skill in the art to which the present invention pertains will be able to easily implement the technical idea of the present invention. In describing the present invention, if it is determined that a detailed description of a known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description will be omitted. Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings. In the drawings, the same reference numerals are used to refer to the same or similar components.

Also, as used herein, the singular expression includes the plural expression unless the context clearly dictates otherwise. In the present application, terms such as "consisting of" or "comprising" should not be construed as necessarily including all of the various components or various steps described in the specification, and some components or some steps are included. It should be construed that it may not, or may further include additional components or steps.

In addition, various components and sub-components thereof are described below in order to describe the system according to the present invention. These components and sub-components thereof may be implemented in various forms, such as hardware, software, or a combination thereof. For example, each element may be implemented as an electronic configuration for performing a corresponding function, or may be software itself operable in an electronic system or implemented as a functional element of such software. Alternatively, it may be implemented with an electronic configuration and corresponding driving software.

Although various flowcharts are disclosed to describe the embodiments of the present invention, this is for convenience of description of each step, and each step is not necessarily performed according to the order of the flowchart. That is, each step in the flowchart may be performed simultaneously with each other, performed in an order according to the flowchart, or may be performed in an order opposite to the order in the flowchart.

Hereinafter, a customized financial product recommendation system and a method for providing the same according to some embodiments of the present invention will be described.

1 is a view for explaining an application example of a customized financial product recommendation system according to an embodiment of the present invention.

Referring to FIG. 1 , the financial product recommendation system 100 may receive customer data from the financial product service server 200 and provide individual financial product recommendation services to each customer using the customer data.

The financial product service server 200 may be a single integrated service server, or may be an aggregate of a plurality of financial service servers that are each separated.

The financial product recommendation system 100 obtains customer data related to the customer's financial product from the financial product service server 200, performs pre-processing on the customer data, and generates a label model to provide financial product recommendation data for each user. can create

Hereinafter, the financial product recommendation system 100 for a complete sales guarantee will be described in more detail with reference to FIGS. 2 to 7 .

Referring to FIG. 2 , the financial product recommendation system 100 may include a preprocessor 110 , a label model unit 120 , an imbalance data processing unit 130 , and a customer data DB 140 .

The customer data DB 140 may store customer data provided from the financial product service server 200 .

In an embodiment, the customer data may include at least one of financial product transaction amount information, financial product transaction count information, deposit and withdrawal amount information, financial product average balance information, customer profile information, customer investment propensity information, and long-term dormancy information. have.

The financial product transaction amount information is data on the amount of purchase and sale of a financial product by a customer, and may be aggregated for a certain period, for example, every month. For example, financial product transaction amount information is stored in at least one of CMAs, domestic stocks, loans, wraps, trusts, deposits, bonds, derivative-linked securities, derivatives, funds, foreign stocks and other financial products for a certain period of time. It may include purchase and sale amounts for

The financial product transaction count information is data on the number of times a customer purchases and sells a financial product, and may be aggregated for a certain period, for example, every month. For example, information on the number of transactions in financial products is stored in at least one of CMA, domestic stocks, loans, wraps, trusts, deposits, bonds, derivative-linked securities, derivatives, funds, overseas stocks and other financial products for a certain period of time. It may include the number of purchases and sales for

The deposit and withdrawal amount information is information on the customer's deposit, warehousing, withdrawal, and shipment amount, and may be aggregated for a certain period, for example, every month.

The financial product average balance information is average balance information on financial products owned by the customer, and may be aggregated for a certain period, for example, every month. For example, the average balance information of financial products includes at least one of CMA, domestic stocks, loans, wraps, trusts, deposits, bonds, derivative-linked securities, derivatives, funds, foreign stocks and other financial products held by each customer for a certain period of time. may include a balance for

The customer profile information is profile information that distinguishes an individual customer from other customers. For example, the customer profile information may include at least one of the customer's age, gender, and classification code.

Customer investment propensity information is investment propensity information for each individual customer, and is calculated based on the questionnaire questions in the customer's investor information confirmation letter. Through this, the customer's propensity score is calculated, and the propensity score or propensity classification code can be used. Based on the customer's questionnaire information, the propensity score can be calculated, and information such as the degree of risk tolerance and expected return can be obtained. may be included. Based on this, the customer's investment propensity (eg, risk preference, etc.) can be classified.

The long-term dormancy information is information set for a customer who has not made a transaction for more than a preset period. The long-term dormancy information may be set for a customer or may be set for a specific product of a certain customer. For example, if a customer does not make a transaction for more than a preset period, the customer may be set to dormancy for a long time. As another example, if a customer does not make a transaction for a fund for a predetermined period or more, the customer may be set as a long-term dormant for a fund product.

Long-term dormancy information can be used to remove disproportionate data by excluding the dormant customer's data from training or by resetting the purchase probability.

The preprocessor 110 performs preprocessing on customer data using principal component analysis. The preprocessor 110 may perform the preprocessing by reducing the dimension of the customer data by performing principal component analysis on the customer data.

Here, principal component analysis (PCA) is to reduce high-dimensional data to low-dimensional data.

For example, the preprocessor 110 may use orthogonal transformation to transform samples in a high-dimensional space that are likely to be related to each other into samples in a low-dimensional space in which there is no linear correlation. In this case, the number of dimensions of the principal component is less than or equal to the number of dimensions of the original sample.

The preprocessor 110 may use an orthogonal transformation by setting an axis according to the size of the variance. For example, the preprocessor 110 sets a new coordinate system so that when the data of customer data is mapped to one axis, the axis with the largest variance is the first main component and the axis with the second largest as the second main component, and using this The data can be linearly transformed.

In this way, various applications are possible by decomposing the sample difference into the components that best represent the difference. This transformation can be defined in such a way that the first principal component has the largest variance, and the subsequent principal components have the largest variance under the constraint that they are orthogonal to the previous principal components. The principal components are orthogonal because they are eigenvectors of the covariance matrix.

As an example, principal component analysis may be performed as follows. First, suppose we have an nХp dimension data X. Let each row of data X be x _i . Assume that the mean of the data is 0, and if the mean is not 0, subtract the mean of the data. The principal component w ₁ of data X is defined by the following equation.

[Equation 1]

라고 하면 번째 주성분 W_k는 다음의 수학식 2와 같이 정의된다.Now w ₁ , w ₂ , … Let w _k-1 be 1, 2, … Let it be the k-1th principal component,

, the th principal component W _k is defined as in Equation 2 below.

[Equation 2]

w ₁ , w ₂ , … w _p to 1, 2, ?? Assuming the p-th principal component, W becomes a matrix consisting of each column of w _i , and the total principal component decomposition of the data X is as shown in Equation 3 below.

[Equation 3]

In addition, the singular value decomposition (SVD) of the data X can be expressed as in Equation 4 below.

[Equation 4]

where D is a diagonal matrix of positive d _i called singular values of X and U and W are orthogonal matrices. Here, W coincides with W used in principal component analysis.

Here, instead of using the entire p, if an appropriate k is selected to generate W _k and T _k is generated through the above transformation, T _k is an nХk-dimensional matrix and has a reduced dimension. Here, the method of selecting k generally selects the minimum k having a certain ratio (for example, 90%) of the _variance of the data X. This method is based on the expression It is the same method as finding the minimum k that satisfies the condition.

[Equation 5]

The label model unit 120 may generate a label model using a multi-label classification model for customer data preprocessed by the preprocessor.

The label model unit 120 may generate a label model based on the single label classification model using a multi-label classification model described below.

Here, the single label classification model is a model used in a one-dimensional classification problem, for example, a random forest model, a logistic regression model, an AdaBoost model, and a neural network. ) model may be used, and other types of single label models may also be applied.

In the case of the single-label model, since different labels are mutually excluded, a suitable estimation is possible when a single product is purchased, but it is difficult to estimate when multiple products are purchased at the same time.

Therefore, the label model unit 120 can provide a label model that can be applied even when purchasing several products at the same time using the multi-label classification model.

In an embodiment, the label modeler 120 may generate a label model by using a binary relevance (BR) method as a multi-label model. That is, the label model unit 120 may convert the preprocessed customer data into a plurality of single label data.

For example, suppose that financial product purchase prediction data as shown in Table 1 below is given.

client domestic stocks bond fund overseas stocks One O X X O 2 X X O O 3 O X X X 4 X O O X

For the binary association technique, the independence between each label is established.

Specifically, assuming that the set of all labels is L, for each different label l in the set L, the label model unit 120 constructs a binary classification model such as H _l :X → {1, -1} to learn. can That is, the label model unit 120 converts the original data into |L| data sets. The label model unit 120 may configure each data set in such a way that if the original sample includes the label l, it is displayed as l, otherwise it is expressed as -l.

The financial product purchase prediction data in Table 1 may be converted into Tables 2 to 5 below when the binary association technique is applied.

client stock trading stock not traded One O X 2 X O 3 O X 4 X O

client fund transaction Fund non-trading One X O 2 O X 3 X O 4 O X

client bond trading non-trading of bonds One X O 2 X O 3 X O 4 O X

client Overseas stock trading Non-trading of foreign stocks One O X 2 O X 3 X O 4 X O

The new sample x by this binary association technique is classified as a label corresponding to the union of labels output from the |L| analysis models (classifier) as shown in Equation 6 below.

[Equation 6]

In an embodiment, the label model unit 120 may generate a label model using classifier chains (CC) as a multi-label model. Classifier chains are advantageous in that the correlation between labels is reflected in the model.

The label model unit 120 first generates a classification result for the previous label according to the binary association technique with respect to the preprocessed customer data, and uses the classification result for the previous label as an explanatory variable when predicting the next label to exist between the labels. A classifier chain can be created by reflecting the correlation.

FIG. 3 is a view for explaining a classifier chain performed by the label model unit shown in FIG. 2 . For example, assuming that there are a total of five multi-labels, the label model unit 120 performs training and prediction shown in FIG. 3 , respectively.

As can be seen from FIG. 3 , the label model unit 120 may apply the classifier chain algorithm by using a portion of the input variable x and the output variable y given in the training process as input variables of the binary classification algorithm. Also, in the prediction process, both the input variable x given in the past and the y predicted in the previous process can be used for prediction of the binary classification algorithm.

4 shows a simplified comparison of the binary association technique and the classifier chain. As shown, the classifier chain is advantageous over the binary association technique in that it can represent the correlation between labels.

Also, the label model unit 120 may generate a label model by using Ensemble Classifier Chains (ECC) as a multi-label model. The ensemble classifier chain applies the classifier chain methodology multiple times by randomly determining the order in which it predicts the labels. And all the results for several generated chains are used in an ensemble.

5 is a diagram illustrating the result of applying the ensemble classifier chain in the label model unit 120. Referring to this, it can be seen that the classifier chain is applied 5 times to data with all 6 labels and the result is ensembled and predicted. have. Here, a simple average was used for the ensemble, and 0 and 1 were classified based on 0.5 when predicting the label.

In an embodiment, the label modeler 120 may generate a label model using probabilistic classifier chains (PCC) as a multi-label model. The probabilistic classifier chain does not predict the labels sequentially, but generates correlation by estimating the joint probability distribution of all labels, so more accurate estimation is possible.

6 is a graph illustrating an example of a customer's purchase probability distribution for three financial products, and shows, for example, a probability distribution for which customer's stock, fund, and bond to purchase. In the graph shown, individual nodes mean the purchase probability of a combination of (stocks), (stocks, funds), and (stocks, funds, bonds) in order from the second level to the fourth level, and the line connecting the nodes means the conditional probability indicates Looking at the probabilities of all financial product combinations at level 4, the probability of not buying anything is the highest at 0.324, so this customer is most likely not buying anything.

FIG. 7 is a graph illustrating a case in which a classifier chain is applied with respect to FIG. 6 . Since the classifier chain simply follows the direction in which the conditional probability is high in the corresponding node, a prediction that all three products will be purchased is derived as shown in FIG. 7 when the classifier chain is applied. Since the classifier chains are compared only sequentially, the actual prediction probability may not operate correctly in the case of FIG. 6 .

On the other hand, unlike classifier chains that predict labels sequentially, probabilistic classifier chains still follow the philosophy of binary association techniques, but estimate the joint distribution of all labels. That is, the label model unit 120 may calculate the joint probability distribution by using Bayes theorem after sequentially calculating the conditional probabilities of the labels. The label model unit 120 satisfies Equation 7 below.

[Equation 7]

Also, the label model unit 120 may apply Ensemble Probabilistic Classifier Chains (EPCC). When estimating the conditional probability of a label, the ensemble probabilistic classifier chain fits several probabilistic classifier chains that randomly determine the order of label prediction. Finally, by ensemble of these results, the accuracy of the joint probability distribution of the label can be further increased.

On the other hand, since most customers trade only a few financial products, the transaction data of each financial product contains an imbalance. Accordingly, the imbalance data processing unit 130 may solve the imbalance by performing random sampling on the label model generated by the label model unit.

In an embodiment, the imbalance data processing unit 130 may reduce the imbalance by undersampling multiple labels. The imbalance data processing unit 130 may arbitrarily select a sample having a plurality of labels so that the number of labels included in the label model is equal to the number of decimal labels, and perform undersampling thereon. Since the multi-label model is actually a method of applying the single-label model several times, the data imbalance can be resolved by applying the undersampling method to every label.

Alternatively, the imbalance data processing unit 130 may oversampling the label model. To this end, the imbalance data processing unit 130 first randomly samples a sample from a group having a fractional label, and selects a sample having k number of fractional labels in the vicinity of the corresponding sample. And sampling can be performed by adding a random sample to the data between the line connecting the corresponding sample and k adjacent samples.

In an embodiment, the imbalance data processing unit 130 may perform fixed sampling using an explanatory variable. Some of the customer data may be set as explanatory variables. For example, the imbalance data processing unit 130 may perform fixed sampling using long-term dormancy information among explanatory variables. That is, long-term dormant customers can be excluded from the training data. This is to use as a fixed sampling the phenomenon that customers who trade certain financial products continue to trade while customers who do not trade do not trade consistently.

For example, the imbalance data processing unit 130 excludes data about customers who have not traded for more than a preset period (eg, 6 months) from training, and when predicting whether to purchase a specific product, If there is no transaction for financial products, the purchase probability can be calculated using information such as profiles that are not related to financial product transactions.

This is sufficiently effective because the loss of predictive power caused by deleting the long-term dormant customer data is insignificant compared to the accuracy improvement obtained by deleting the long-term dormant customer data.

In an embodiment, the imbalance data processing unit 130 may perform two-step sampling. The unbalanced data processing unit 130 may process the unbalanced data by performing a first sampling process of performing fixed sampling and a second sampling process of arbitrarily performing sampling using an explanatory variable.

In the above, a customized financial product recommendation system according to an embodiment of the present invention has been described with reference to FIGS. 1 to 7 .

Hereinafter, a customized financial product recommendation method according to an embodiment of the present invention will be described with reference to FIG. 8 .

Since the customized financial product recommendation method to be described below is performed based on the previously described customized financial product recommendation system, it can be more easily understood with reference to the above description based on FIGS. 1 to 8 .

8 is a flowchart illustrating a method for recommending a customized financial product according to an embodiment of the present invention.

Referring to FIG. 8 , the financial product recommendation system 100 may perform principal component analysis on customer data to reduce the dimension of customer data to perform preprocessing ( S810 ).

The financial product recommendation system 100 may generate a label model using a multi-label classification model for the preprocessed customer data (S820).

The financial product recommendation system 100 may eliminate the imbalance by performing random sampling on the label model (S830).

In one embodiment for step S810, the financial product recommendation system 100 sets the axis with the largest variance as the first principal component when the data of customer data is mapped to one axis, and the axis with the second largest variance is set as two axes. It can be set as the second principal component. The financial product recommendation system 100 may set a new coordinate system using the first principal component and the second principal component, and linearly transform the customer data using the new coordinate system.

In one embodiment for step S820, the financial product recommendation system 100 generates a multi-label model using a binary association technique that converts the preprocessed customer data into a plurality of single label data; generating a classification result for the previous label according to the binary association technique for A label model may be generated using at least one of the steps of generating a label model using a probabilistic classifier chain that sequentially calculates the conditional probability of a label and calculates a joint probability distribution for it.

In one embodiment for step S830, the financial product recommendation system 100 performs a first sampling process of performing fixed sampling using an explanatory variable, and performs a second sampling process of arbitrarily sampling to perform unbalanced sex can be relieved.

In one embodiment for step S830, the financial product recommendation system 100 excludes from the training the data for the customer who has not made a transaction for more than a preset period using the long-term dormancy information included in the customer data from the training, and If there is no product transaction, the imbalance can be resolved by calculating the purchase probability using information such as profiles that are not related to financial product transactions.

The present invention described above is not limited by the above-described embodiments and the accompanying drawings, but is limited by the claims described below, and the configuration of the present invention may vary within the scope without departing from the technical spirit of the present invention. Those of ordinary skill in the art to which the present invention pertains can easily recognize that it can be changed and modified.

10, 11, 12: user terminal
100: Financial product recommendation system
110: preprocessor
120: label model part
130: unbalanced data processing unit
140: customer data DB
200: financial product service server

Claims (14)

a preprocessing unit for performing preprocessing by reducing the dimension of customer data by performing principal component analysis on customer data;
a label model unit for generating a label model using a multi-label classification model for customer data preprocessed by the preprocessor; and
An imbalance data processing unit for resolving imbalance by performing random sampling on the label model generated by the label model unit;
The label model part
generating a label model using a probabilistic classifier chain that calculates the conditional probability of a label and calculates a joint probability distribution for the conditional probability,
When estimating the conditional probability of the label, a plurality of probabilistic classifier chains for estimating the conditional probability of the label in an arbitrary order are applied, and then the results obtained from each probabilistic classifier chain are ensembled using the result of the label. A customized financial product recommendation system that calculates a combined probability distribution.
The method of claim 1, wherein the customer data is
A customized financial product recommendation system including at least one of financial product transaction amount information, financial product transaction number information, deposit and withdrawal amount information, average financial product balance information, customer profile information, customer investment propensity information, and long-term dormancy information.
According to claim 1, wherein the pre-processing unit
When the data of the customer data is mapped to one axis, a new coordinate system is set so that the axis with the largest variance is the first principal component and the axis with the second largest as the second principal component, and using this, the customer data is linearly transformed. Financial product recommendation system.
According to claim 1, wherein the label model part
A customized financial product recommendation system using at least one of a random forest model, a logistic regression model, an Adaboost model, and an artificial neural network model as a single label classification model.
According to claim 1, wherein the label model part
A customized financial product recommendation system for generating a multi-label model using a binary association technique that converts the pre-processed customer data into a plurality of single-label data.
According to claim 1, wherein the label model part
For the preprocessed customer data, a classification result for the previous label is generated according to the binary association technique, and the classification result for the previous label is used as an explanatory variable when predicting the next label to reflect the correlation existing between the labels. A custom financial product recommendation system that creates a chain of classifiers.
delete The method of claim 1, wherein the imbalance data processing unit
A customized financial product recommendation system that processes imbalanced data by performing a first sampling process that performs fixed sampling using an explanatory variable and a second sampling process that performs random sampling.
The method of claim 1, wherein the imbalance data processing unit
Using the long-term dormancy information included in the customer data, data on customers who have not traded for more than a preset period are excluded from training, and if there is no transaction for a specific financial product for a preset period, information such as profiles unrelated to financial product transactions are stored. A customized financial product recommendation system that calculates the purchase probability using
performing the preprocessing by reducing the dimension of the customer data by the preprocessor performing principal component analysis on the customer data;
generating a label model using a multi-label classification model for the preprocessed customer data by the label model unit; and
resolving the imbalance by performing random sampling on the label model by the imbalance data processing unit; including,
The label model part
generating a label model using a probabilistic classifier chain that calculates the conditional probability of a label and calculates a joint probability distribution for the conditional probability,
When estimating the conditional probability of the label, a plurality of probabilistic classifier chains for estimating the conditional probability of the label in an arbitrary order are applied, and then the results obtained from each probabilistic classifier chain are ensembled using the result of the label. A customized financial product recommendation method that calculates a joint probability distribution.
The method of claim 10, wherein the pre-processing unit performing the pre-processing comprises:
setting, as a first principal component, an axis having the largest variance when the preprocessor maps the customer data data to one axis;
setting, by the pre-processing unit, an axis on which the dispersion is secondly increased as a second principal component;
setting, by the preprocessor, a new coordinate system using the first principal component and the second principal component; and
linearly transforming the customer data by the preprocessor using the new coordinate system; A method of recommending customized financial products, including
The method of claim 10, wherein the step of generating the label model by the label model unit,
generating a multi-label model using a binary association technique in which the label model unit converts the preprocessed customer data into a plurality of single label data;
The label model unit generates a classification result for the previous label according to the binary association technique with respect to the preprocessed customer data, and uses the classification result for the previous label as an explanatory variable when predicting the next label. Correlation existing between labels generating a classifier chain by reflecting the relationship; A customized financial product recommendation method including at least one of.
The method of claim 11, wherein the pre-processing unit linearly transforms the customer data
a first sampling processing step in which the preprocessor performs fixed sampling using an explanatory variable; and
a second sampling processing step in which the pre-processing unit arbitrarily performs sampling; A method of recommending customized financial products, including
The method of claim 11, wherein the pre-processing unit linearly transforms the customer data
excluding from the training by the pre-processing unit data about a customer who has not made a transaction for more than a preset period using the long-term dormancy information included in the customer data; and
and setting, by the pre-processing unit, a probability of purchasing a corresponding financial product to 0 if there is no transaction for a specific financial product for a preset period.

Images