A method and apparatus for recommending financial products

By constructing a probability grammar model, the problem of weak expansion capabilities in the construction of user portraits is solved, the accuracy and expansion of financial product recommendations are achieved, and the sales rate and user experience of bank sales personnel are improved.

CN118469679BActive Publication Date: 2024-11-08DMAI (GUANGZHOU) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410931710.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-11-08
Estimated Expiration
2044-07-12

AI Technical Summary

Technical Problem

In the prior art, the user portrait construction process has weak ability to accurately portray user portraits, resulting in inappropriate recommendations of financial products, affecting user experience and sales rates, and it is impossible to combine the account information of each branch to make reasonable financial products recommendations.

Method used

By constructing a probability grammar model, using product transaction records to uniform unit and quantitatively classify the node relationship in the probability knowledge graph, count node probability, perform group clustering, calculate customer group similarity, and recommend financial products based on new customer nodes.

Benefits of technology

It realizes accurate financial product recommendations, improves the sales rate of bank sales personnel, understands the characteristics of financial customer groups, adapts to the addition of new nodes and attributes, does not require artificial definition of new rules, and has strong expansion capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118469679B_ABST
    Figure CN118469679B_ABST
Patent Text Reader

Abstract

The embodiment of the present application proposes a financial product recommendation method and device, which is applied to the field of smart financial technology. First, the acquired product transaction records are unified and quantitatively graded to obtain product transaction levels, and then the product transaction levels are used to characterize the node relationships in the probabilistic knowledge graph. The node probabilities in the probabilistic knowledge graph are statistically obtained to obtain a probabilistic grammar model. The probabilistic grammar model is then used to group and cluster the node probabilities to obtain cluster center feature vectors. The customer group similarity is calculated based on the cluster center feature vectors between different probabilistic grammar models. Finally, based on the new customer node, the corresponding customer node is found according to the customer group similarity and financial product recommendation information is sent. By constructing a probabilistic grammar model to count the user's behavior in selecting financial products, estimate the customer's choice, accurately recommend financial products to customers, and improve the sales rate of financial products of bank sales personnel.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of smart financial technology, and more specifically, to a method and device for recommending financial products. Background Art

[0002] Before recommending financial products, user portraits are generally drawn, and then the financial products that meet the requirements are screened and recommended to users based on the user's product preferences, target intentions and user affordability. User portraits are an effective tool for outlining target users and linking user demands with design directions, and they have been widely used in various fields. User portraits help financial product operation departments improve and improve products and promote business development. They are mainly divided into: First, accurate market positioning. The customer portraits constructed by this method can more accurately locate the characteristics and needs of different customer groups. Banks can launch financial products that meet their needs in a targeted manner based on the portrait information of different customer groups, thereby improving the accuracy and effectiveness of product promotion. Second, personalized product recommendations. Based on the constructed customer portraits, banks can provide a basis for personalized product recommendations. Banks can recommend financial products that meet their needs and preferences based on their portrait information, thereby improving customer satisfaction and product usage. Third, reducing marketing costs: Accurate customer portraits can help banks reduce marketing costs. Through in-depth analysis of customer characteristics and behaviors, banks can carry out marketing activities in a targeted manner, reduce resource waste, and improve marketing efficiency. Fourth, increase cross-selling opportunities. By using the customer profile constructed by this method, banks can discover the various financial needs that customers may have and conduct targeted cross-selling. This helps to increase the bank's sales and customer satisfaction. Fifth, continuously improve products and services. By analyzing customer profile data, banks can timely understand customer feedback and behavior, as well as market trends, so as to adjust and improve existing financial products and services and maintain competitive advantages.

[0003] However, in the existing technology, various problems often occur in the construction of user portraits. For example, when new attributes are added to the data, the portrait construction process in the existing technology basically needs to be rebuilt, and the expansion capability is weak. If the user portrait cannot be accurately portrayed, the recommendation of financial products will not be appropriate, resulting in a poor user experience and failure to find accurate customers. In addition, each bank has multiple branches, but when making financial recommendations, it is often impossible to integrate the account information of each branch to make reasonable financial product recommendations.

[0004] Therefore, how to accurately portray user portraits and recommend financial products based on account information from different sources is one of the problems that technicians in this field need to solve. Summary of the invention

[0005] In view of this, the purpose of this application is to provide a financial product recommendation method and device, which constructs a probabilistic grammar model to statistically analyze the user's behavior in selecting financial products, estimate the customer's choice, accurately recommend financial products to customers, and improve the sales rate of financial products of bank sales personnel.

[0006] In order to achieve the above purpose, the technical solution adopted in the embodiment of the present application is as follows:

[0007] In a first aspect, an embodiment of the present application provides a method for recommending financial products, the method comprising:

[0008] The acquired product transaction records are unified and quantified to obtain product transaction grades;

[0009] The node relationships in the probabilistic knowledge graph are characterized by using the product transaction level, and the node probabilities in the probabilistic knowledge graph are statistically analyzed to obtain a probabilistic grammar model.

[0010] The probability grammar model is used to group and cluster the node probabilities to obtain the cluster center feature vector;

[0011] Calculate customer group similarity based on cluster center feature vectors between different probabilistic grammar models;

[0012] Based on the new customer node, find the corresponding customer node according to the similarity of customer groups and send financial product recommendation information.

[0013] In a possible implementation, the steps of characterizing the node relationships in the probabilistic knowledge graph using the product transaction level and obtaining the probabilistic grammar model by counting the node probabilities in the probabilistic knowledge graph include:

[0014] The node probabilities of financial product nodes and customer nodes in the product transaction level statistical probability knowledge graph are used to obtain a discrete probability distribution;

[0015] The parameterized probability model is used to fit the discrete probability distribution to obtain the parameterized description of the node;

[0016] The node parameterized descriptions of all financial product nodes and customer nodes are statistically analyzed to obtain a probabilistic grammar model.

[0017] In a possible implementation, the node probability includes a node connection probability, a priori probability, a state jump probability, a node state change probability, a node state continuity probability, and a port usage probability.

[0018] In a possible implementation, the step of grouping and clustering the node probabilities using the probabilistic grammar model to obtain the cluster center feature vector includes:

[0019] The node probabilities are arranged using the probabilistic grammar model to obtain a structured feature vector;

[0020] According to the similarity of node feature vectors, customer nodes and product nodes in the probabilistic grammar model are grouped and clustered to obtain product node cluster centers and customer node cluster centers;

[0021] The cluster center feature vector is obtained according to the product node cluster center and the customer node cluster center.

[0022] In a possible implementation, the calculation formula for customer group similarity is:

[0023] ;

[0024] in and are the first probability grammar model and the second probability grammar model respectively, and are adjustable weights, represents the similarity of the cluster center feature vectors between two probabilistic grammar models, represents the cluster center feature vector of the first probabilistic grammar model, represents the cluster center feature vector of the second probabilistic grammar model, represents the number of cluster center feature vectors that meet the threshold requirements, Represents the number of cluster center feature vectors in the second probabilistic grammar model.

[0025] In a possible implementation, based on the new customer node, the steps of finding the corresponding customer node according to the customer group similarity and sending the financial product recommendation information are as follows:

[0026] Constructing a feature vector for a new customer node to obtain a feature vector of the new customer node;

[0027] According to the similarity of customer groups, the distance between the feature vector of the new customer node and the feature vector of the cluster center is calculated to obtain the corresponding customer node;

[0028] Send financial product recommendation information to new customer nodes based on the relationship between customer nodes and financial nodes.

[0029] In one possible implementation, the product transaction record includes customer information, types of products sold, and user financial product usage. The customer information includes customer number, asset amount, daily consumption, credit rating, invisible asset amount, occupation, income amount and credit limit. The types of products sold include product name, branch, product type, interest rate and number of installments. The user financial product usage includes user repayment status, repayment amount, overdue status, overdue amount, installment repayment status and installment amount.

[0030] In a possible implementation, the product transaction level includes asset level, daily consumption level, invisible asset level, career stability level, and income level.

[0031] In a first aspect, an embodiment of the present application provides a financial product recommendation device, the device comprising:

[0032] A preprocessing module is used to unify the units of the acquired product transaction records and perform quantitative grading processing to obtain the product transaction level;

[0033] The probability statistics module is used to characterize the node relationship in the probability knowledge graph by using the product transaction level, and to obtain the probability grammar model by counting the node probabilities in the probability knowledge graph;

[0034] The grouping and clustering module is used to group and cluster the node probabilities using the probabilistic grammar model to obtain the cluster center feature vector;

[0035] A similarity calculation module is used to calculate the customer group similarity between customer nodes and cluster centers between different probability grammar models;

[0036] The recommendation information sending module is used to find the corresponding customer node based on the new customer node and the similarity of the customer groups and send the financial product recommendation information.

[0037] In a possible implementation, the probability statistics module is further used for:

[0038] The node probabilities of financial product nodes and customer nodes in the product transaction level statistical probability knowledge graph are used to obtain a discrete probability distribution;

[0039] The parameterized probability model is used to fit the discrete probability distribution to obtain the parameterized description of the node;

[0040] The node parameterized descriptions of all financial product nodes and customer nodes are statistically analyzed to obtain a probabilistic grammar model.

[0041] The above main scheme of the present application and its further options can be freely combined to form multiple schemes, all of which are schemes that can be adopted and claimed for protection in the present application; and in the present application, (non-conflicting options) options and other options can also be freely combined. After understanding the scheme of the present application, those skilled in the art can understand that there are many combinations based on the existing technology and common knowledge, all of which are technical schemes to be protected by the present application, and they are not exhaustively listed here.

[0042] The embodiment of the present application proposes a financial product recommendation method and device, which is applied to the field of smart financial technology. First, the acquired product transaction records are unified and quantitatively graded to obtain product transaction levels, and then the product transaction levels are used to characterize the node relationships in the probabilistic knowledge graph. The node probabilities in the probabilistic knowledge graph are statistically obtained to obtain a probabilistic grammar model. The probabilistic grammar model is then used to group and cluster the node probabilities to obtain cluster center feature vectors. The customer group similarity is calculated based on the cluster center feature vectors between different probabilistic grammar models. Finally, based on the new customer node, the corresponding customer node is found according to the customer group similarity and financial product recommendation information is sent. By constructing a probabilistic grammar model to count the user's behavior in selecting financial products, estimate the customer's choice, accurately recommend financial products to customers, and improve the sales rate of financial products of bank sales personnel.

[0043] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 A flow chart of a financial product recommendation method proposed in an embodiment of the present application is shown.

[0046] Figure 2 A schematic diagram showing the connection relationship between the customer node and the financial node proposed in an embodiment of the present application is shown.

[0047] Figure 3 Another preprocessing data diagram proposed in an embodiment of the present application is shown.

[0048] Figure 4 Another preprocessing data diagram proposed in an embodiment of the present application is shown.

[0049] Figure 5 A schematic diagram of the structure of a financial product recommendation device provided in an embodiment of the present application is shown.

[0050] Icons: 100 - financial product recommendation device; 110 - preprocessing module; 120 - probability statistics module; 130 - grouping and clustering module; 140 - similarity calculation module; 150 - recommendation information sending module. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0052] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0053] In the existing technology, various problems often occur in the construction of user portraits. For example, when new attributes are added to the data, the portrait construction process in the existing technology basically needs to be rebuilt, and the expansion capability is weak. If the user portrait cannot be accurately portrayed, the recommendation of financial products will be inappropriate, resulting in a poor user experience and failure to find accurate customers.

[0054] Therefore, the embodiments of the present application propose a financial product recommendation method and device, which can not only be used to provide customers with suggestions on selling financial products, but also understand the group characteristics of financial customers, find out the common points between customers in different regions, and provide a basis for the subsequent exploration and development of new financial products.

[0055] Please refer to the following Figure 1 , Figure 1 The flowchart of a financial product recommendation method proposed in an embodiment of the present application is shown, including the following steps S100 to S500:

[0056] Step S100: unify the units of the acquired product transaction records and perform quantitative grading processing to obtain product transaction grades.

[0057] The acquired data is first preprocessed. The data acquired in the embodiment of the present application is a product transaction record, which includes customer information, the type of product sold, and the user's financial product usage.

[0058] Customer information includes customer number, asset amount, daily consumption, credit rating, amount of hidden assets, occupation, income amount and credit limit.

[0059] The types of products sold include product name, branch, product type, interest rate and number of installments.

[0060] User financial product usage includes user repayment status, repayment amount, overdue status, overdue amount, installment repayment status and installment amount.

[0061] Product transaction levels include asset level, daily consumption level, invisible asset level, career stability level, and income level.

[0062] Since there are various channels for obtaining product transaction records, in order to better conduct data analysis later, the units of different sources and data records are uniformly divided, and the asset amount, daily consumption, invisible asset amount, occupation, and income amount are quantified and graded and converted into product transaction levels.

[0063] It is worth noting that the transaction records selected in the embodiment of the present application are required to be data for at least one consecutive year to ensure the accuracy of model construction.

[0064] Step S200: Use product transaction levels to characterize the node relationships in the probabilistic knowledge graph, and obtain a probabilistic grammar model by statistically analyzing the node probabilities in the probabilistic knowledge graph.

[0065] Please refer to Figure 2 , Figure 2 The schematic diagram of the connection relationship between the customer node and the financial node proposed in the embodiment of the present application is shown, with the upper layer being five financial product nodes D1, D2, D3, D4, and D5, and the lower layer being five customer nodes S1, S2, S3, S4, and S5. The connection relationship between the financial product node and the customer node is fully connected, and the edge between the financial product node and the customer node represents the situation of the customer using the financial product, including normal repayment, installment repayment, overdue repayment, etc.

[0066] Node probability includes node connection probability, prior probability, state jump probability, node state change probability, node state continuity probability and port usage probability, etc., and the embodiments of the present application do not impose any limitations on this.

[0067] The node connection probability is the probability P0(Dj|Si) that the customer node Si is connected to the financial product node Dj. The total number of edges between the customer node Si and the financial product node Dj is equal to the total number of valid training data. For example, P0(Dj|Si) =<Si, …, Dj, …, …> The number of times the sample appears / the total number of samples in the training set.

[0068] The prior probability is the probability of the node state occurring, for example, the probability of the state occurring at time t is P1(s_t), where s_t represents the state at time t, which can be s={customer asset level, consumption level, credit level, asset level, income level} or s={financial product type, interest rate, number of installments}.

[0069] The state jump probability is the probability P12(s_t2|s_t1) of a node changing from time t1 to time t2.

[0070] The node state change probability is the probability of the node state change, for example P2(△s_t), where △s = {customer asset level, consumption level, credit level, asset level, income level} or △s = {financial product type, interest rate, number of installments} or △s = { …,-200, -100, -50, 0, 50, 100, 200, …}, where a negative number represents a decrease in the customer level or financial product attribute value, a positive number represents an increase in the customer level or financial product attribute value, and 0 represents that the customer level or financial product attribute value remains unchanged.

[0071] The node state persistence probability is the probability that the node state persists, for example, P3(△s=0 for T days), where T is the statistical interval and takes values ​​[1, Tmax], where Tmax is the maximum idle time.

[0072] The port usage probability is the probability of node port usage, P4(s|Si) = the number of occurrences of node attribute s / the number of all behaviors of node Si, where s = {customer asset level, consumption level, credit level, asset level, income level}.

[0073] Step S200 includes:

[0074] The node probabilities of financial product nodes and customer nodes in the product transaction level statistical probability knowledge graph are used to obtain a discrete probability distribution;

[0075] The parameterized probability model is used to fit the discrete probability distribution to obtain the parameterized description of the node;

[0076] The node parameterized descriptions of all financial product nodes and customer nodes are statistically analyzed to obtain a probabilistic grammar model.

[0077] First, the variable values ​​are quantified and standardized, and then the probability parameters of various nodes are statistically analyzed to obtain the probability distribution (discrete type). The parametric probability model is selected to fit the probability distribution to obtain the parametric description of each node ( , , ,……,in represents the probability parameter), and parameterized modeling is performed to obtain the probabilistic grammar model.

[0078] It is worth noting that the above-mentioned parameterized probability model can be a normal distribution model, a multinomial distribution model, a Gaussian mixture model, etc.

[0079] Step S300: using a probabilistic grammar model to group and cluster the node probabilities to obtain a cluster center feature vector.

[0080] After flattening the features using the node probabilities obtained by the probability grammar model, clustering is performed according to different nodes to obtain multiple cluster centers, and multiple corresponding feature vectors are obtained in turn to complete feature clustering.

[0081] Step S300 includes:

[0082] The node probabilities are arranged using the probabilistic grammar model to obtain a structured feature vector;

[0083] According to the similarity of node feature vectors, customer nodes and product nodes in the probabilistic grammar model are grouped and clustered to obtain product node cluster centers and customer node cluster centers;

[0084] The cluster center feature vector is obtained according to the product node cluster center and the customer node cluster center.

[0085] The probabilistic grammar model is used to arrange and combine the obtained node probabilities to obtain a structured feature vector (TZ={ }), in the structured feature vector, according to the similarity between the feature vectors of different nodes, group and cluster the customer nodes and product nodes respectively, and obtain Nd product node cluster centers { } and Ns customer node cluster centers { }, the above structured feature vectors and customer node cluster centers are converted into Ns+Nd cluster center feature vectors, each of which represents a type of customer or a type of product. For example, the feature vector of the customer node cluster center represents the characteristics of this type of customer, and the feature vector of the product node cluster center represents the characteristics of this type of product.

[0086] After obtaining the probabilistic grammar model, it is necessary to analyze and compare customer groups within the same bank or between different banks. When new customer data is available, in order to find customer nodes similar to the customer, and then find financial product nodes related to the customer nodes to recommend financial products, perform the following steps.

[0087] Step S400: Calculate customer group similarity based on cluster center feature vectors between different probabilistic grammar models.

[0088] The source of product transaction records can be different banks, several branches of a bank, or several parts of data from the same bank. After receiving multiple product transaction records, the above steps are processed to obtain corresponding probability grammar models, such as the first probability grammar model, the second probability grammar model, ..., the Nth probability grammar model. The first probability grammar model has multiple customer nodes, and the similarity of each customer node with Ns cluster centers of other probability grammar models (such as the second probability grammar model) other than the first probability grammar model is calculated. If the similarity of the cluster center of the second probability grammar model is greater than a preset threshold, it is considered that the customer node of the first probability grammar model belongs to the class of the second probability grammar model, indicating the similarity between the bank customer groups.

[0089] It is worth mentioning that in addition to judging the class of a customer node by similarity, it can also be judged by distance. For example, the distance between each customer node and other probability grammar models (such as the second probability grammar model) except the first probability grammar model is calculated. When the distance is less than the threshold, it indicates the similarity between the bank's customer groups.

[0090] The calculation formula for the similarity between the customer node and the customer group of the cluster center is:

[0091] ;

[0092] in and are the first probability grammar model and the second probability grammar model respectively, and are adjustable weights, represents the similarity of the cluster center feature vectors between two probabilistic grammar models, represents the cluster center feature vector of the first probabilistic grammar model, represents the cluster center feature vector of the second probabilistic grammar model, represents the number of cluster center feature vectors that meet the threshold requirements, Represents the number of cluster center feature vectors in the second probabilistic grammar model.

[0093] When calculating the similarity between each two customer groups, set the first probability grammar model For the unknown, second probability grammar model As is known, each probability grammar model contains multiple customer nodes and multiple product nodes. For example, the first probability grammar model P1 is set to contain 20 customer nodes {C1, C2, ..., C20}, and the customer nodes are assumed to be clustered into 5 categories through the clustering algorithm. Each category has a customer node cluster center, represented as {center1, center2, ..., center5}, and each customer node cluster center corresponds to a cluster center feature vector, and the dimension of the vector is the same as that of the customer node. After the clustering operation, the customer nodes of the second probability grammar model P2 have 3 cluster centers { center1 , center2 , center3 }. Assume that the class center closest to center1 and center2 in P2 is center1 , center1 and center1 The distance between center2 and center1 is less than the threshold. The distance is greater than the threshold. The class center closest to center3 in PKG2 is center2 , center3 and center2 The distance is less than the threshold. The class center closest to center4 and center5 in P2 is center3 , center4 and center3 The distance between center5 and center3 is less than the threshold. The distance is greater than the threshold.

[0094] It represents the distance between any class center in P1 and its nearest class center in P2. In addition, the distance between the class centers must also satisfy the class threshold constraint in P2. Specifically, it includes Sim(center1, center1 ),Sim(center3,center2 )、Sim(center4,center3 ), represents the number of cluster center feature vectors that meet the threshold requirements, Indicates the number of cluster center feature vectors in the second probability grammar model, combined with the above example is 3, is 5.

[0095] exist middle, Represents the similarity between node A and cluster center B. Possible calculation methods include Gaussian similarity measurement , cosine similarity measure etc., among which and Respectively represent the cluster center feature vectors corresponding to different probability grammar models, In this way, we can calculate the similarity between any two customer groups of banks or different customer groups of the same bank.

[0096] Step S500: Based on the new customer node, find the corresponding customer node according to the customer group similarity and send financial product recommendation information.

[0097] Step S500 includes:

[0098] Constructing a feature vector for a new customer node to obtain a feature vector of the new customer node;

[0099] According to the similarity of customer groups, the distance between the feature vector of the new customer node and the feature vector of the cluster center is calculated to obtain the corresponding customer node;

[0100] Send financial product recommendation information to new customer nodes based on the relationship between customer nodes and financial nodes.

[0101] When facing a new customer node, the node features are used to construct the new customer node feature vector Snew. Then, the distance between Snew and multiple cluster center feature vectors in the existing probabilistic grammar model is calculated. The distance is calculated in the same way as the customer group similarity. The closest customer node is found, and financial product recommendation information is sent to the new customer node based on the relationship between the customer node and the financial node.

[0102] If any customer group class is found in the probabilistic grammar model, it indicates that the model needs to be updated, and the data set will be regenerated into a probabilistic grammar model according to the construction of the probabilistic grammar model and the node clustering method.

[0103] The embodiment of the present application can recommend to the new customer the product with the lowest probability of overdue payment among similar customers, thereby maximizing the bank's profit.

[0104] In a possible embodiment, in order to better explain the above-mentioned financial product recommendation method, the embodiment of the present application selects the annual customer financial transaction activity data of two branches of a bank as a data set, totaling 97,000 financial transaction records, including: customer number, customer deposit amount, customer consumption amount, customer fixed asset status, customer occupation, customer income, customer credit line, name of the financial product traded, financial product interest rate, number of financial product installments, financial product processing amount, whether the customer repays normally, normal repayment amount, whether the customer chooses installment payment, installment repayment amount, whether the customer is overdue and overdue amount, totaling 17 items.

[0105] In order to facilitate analysis, some data items are graded and quantified. Customer deposits are divided into three asset levels: less than 10,000 yuan, 10,000 to 100,000 yuan, and more than 100,000 yuan. Customer consumption is divided into three levels: less than 3,000 yuan, 3,000 to 10,000 yuan, and more than 10,000 yuan. Customer fixed assets or other assets such as stocks and options are unified as invisible assets and divided into three levels: less than 2 million, 2 million to 8 million, and more than 8 million. For customer occupations, state-owned enterprises and institutions are classified into one category, and other occupations are classified into another category. Therefore, customer occupation data can be divided into two levels according to the degree of stability. Customer income is divided into three levels: less than 8,000 yuan / month, 8,000 yuan / month to 20,000 yuan / month, and more than 20,000 yuan / month. The names of financial products are coded using positive integers 1, 2, ... according to their names. The three different states of customer installment repayment, normal repayment, and overdue repayment are recorded as 1, 2, and 3 respectively. Other attributes are used according to numerical values. For example, Table 1 shows the preprocessed data:

[0106] Table 1

[0107]

[0108] According to the different branches, the data is divided into two data sets, denoted as Data1 and Data2. The probabilistic knowledge graph contains multiple customer nodes Si and multiple financial product nodes Di. For the customer node Si, the common probabilities are selected:

[0109] The probability P0(Dj|Si) that a customer node Si is connected to a financial product node Dj indicates that normal repayment, installment repayment, and overdue repayment appear in the customer's transaction record.

[0110] Usage distribution P0(Di) of different products of customer node Si.

[0111] The distribution of different products used by customer node Si in different situations is P4(s|Si).

[0112] The results of the Data1 dataset are shown. In order to calculate the probability P0(Dj|Si), we first extract all the data packet records of a certain customer node Si from the dataset, and count the connection methods of the node to different financial product nodes Di, that is, count the number of normal repayments and overdue repayments when different customers choose different products. The probability distribution of the number of normal repayments, installment repayments, and overdue repayments in all products can be obtained Figure 3 and Figure 4 , Figure 3 Another pre-processing data diagram proposed in the embodiment of the present application is shown. Figure 4Another preprocessing data diagram proposed in an embodiment of the present application is shown.

[0113] For the unselected products, the probabilities of normal repayment and overdue repayment are both 0. There are 5 products in the data set, and each product contains two probabilities: normal and overdue. Therefore, the probability distribution of the statistical results is vectorized to obtain a 5 2=10-dimensional vector v1.

[0114] For the usage distribution P0(Di) of different products of customer node Si, the probability of customer node Si using different products is counted. Since the data set only contains 5 products, the result of P0(Di) is a 5-dimensional vector v2.

[0115] P4(s|Si) represents the situation where customers use different products under different conditions. The probability distribution here is a two-dimensional probability distribution of state-product. The customer's status here includes indicators such as asset level, consumption level, credit level, invisible asset level, career stability level, income level, and credit limit. According to the previous data preprocessing part, asset level, consumption level, credit level, invisible asset level, and income level can all be divided into 3 different levels. The career stability level contains 2 levels. For the convenience of processing, the credit limit can also be graded and quantified, that is, divided into 3 levels: less than 10,000 yuan, 10,000 to 50,000 yuan, and more than 50,000 yuan. Therefore, the customer status can exist at most There are 1458 states in total. The state of a product includes three attributes: product type, interest rate, and number of installments. There are 5 types of products in the dataset. The interest rate of the product can also be classified into 3 levels. The number of installments of the product includes 3 installments, 6 installments, and 12 installments. Therefore, the state of the product exists There are 45 product states in total. Therefore, P4(s|Si) is a The two-dimensional probability distribution of . In order to further compress the data features, a four-component two-dimensional Gaussian mixture model is used for fitting. The parameterized model contains 8 means, 8 covariances, 4 component weights, and a total of 20 parameters. Finally, it is converted into a 20-dimensional vector v3.

[0116] Flattening the three types of features together (v1, v2, v3) can obtain a 35-dimensional customer node feature vector. Data1 contains 695 customer nodes Si. Clustering the 695 customer nodes Si according to the 35-dimensional vector in the model construction can obtain 12 35-dimensional class centers. For the data set Data2, the discrete probability distribution of customer node features is fitted in the parameterized model, and the dimension is 35 dimensions. However, in the clustering process, the number of class centers is different, which is 48. Table 2 shows the probability grammar model corresponding to the data set:

[0117] Table 2

[0118]

[0119] Consider the similarity of the customer groups of PKG1 and PKG2. The similarities and differences between the customer groups of several banks can be analyzed. For each customer node cluster center in PKG1, calculate its similarity (or distance) with the Ns cluster centers of PKG2. If it is similar to a certain center, If the similarity is greater than the threshold (or the distance is less than the threshold), the node is considered to belong to PKG2. The category they belong to represents the similarities between the two bank customer groups.

[0120] The similarity calculation uses the formula:

[0121] ;

[0122] in the formula and The values ​​are 0.6 and 0.4 respectively. The similarity is measured by cosine similarity. The calculation results show that the first customer group in Data1 and the fifth customer group in Data2 have a high similarity of 80%. The eighth customer group in Data1 and the 36th customer group in Data2 have a high similarity of 75%.

[0123] Get a new customer data from the same bank as Data1, and construct the customer's feature vector according to the above method. Then use this feature vector to calculate the distance with the 12 class centers in Data1. It can be found that the distance between this customer and the first class center of Data1 is the closest. It can be seen that recommending the product with the lowest probability of overdue in the first class customer group to this customer can ensure the maximum profit of the bank.

[0124] In summary, the financial product recommendation method provided in the embodiment of the present application utilizes a probabilistic grammar model to capture the statistical laws of different customers' behaviors in selecting bank financial products, thereby estimating the customers' selection preferences, inferring the financial products most needed by the customers, and recommending them to the customers, thereby improving the sales rate of financial products by bank sales staff.

[0125] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0126] First, it can not only be used to provide customers with advice on selling financial products, but also to understand the group characteristics of financial customers and the commonalities between customers in different regions, providing a basis for the subsequent exploration and development of new financial products. It has the advantage of strong functionality.

[0127] Second, it provides an intuitive way to understand and explain the process of recommending financial products to customers, which facilitates the improvement and optimization of subsequent financial product sales and strategy adjustments, and has the advantage of strong explanatory power.

[0128] Third, it can adapt to the addition of new nodes and new attributes. It does not require manual definition of new rules and can be expanded automatically, which has the advantage of strong expansion capability.

[0129] A possible implementation of the financial product recommendation device 100 is given below, which is used to execute the various steps and corresponding technical effects of the financial product recommendation method shown in the above embodiments and possible implementations. Figure 5 , Figure 5 A schematic diagram of the structure of a financial product recommendation device 100 provided in an embodiment of the present application is shown, and the device includes:

[0130] A preprocessing module 110 is used to unify the units of the acquired product transaction records and perform quantitative grading processing to obtain product transaction grades;

[0131] The probability statistics module 120 is used to characterize the node relationship in the probability knowledge graph by using the product transaction level, and to obtain a probability grammar model by counting the node probabilities in the probability knowledge graph;

[0132] A grouping and clustering module 130 is used to group and cluster the node probabilities using a probabilistic grammar model to obtain a cluster center feature vector;

[0133] A similarity calculation module 140, used to calculate the customer group similarity based on the cluster center feature vectors between different probability grammar models;

[0134] The recommendation information sending module 150 is used to find the corresponding customer node based on the new customer node and according to the customer group similarity and send the financial product recommendation information.

[0135] In a possible implementation, the probability statistics module 120 is further configured to:

[0136] The node probabilities of financial product nodes and customer nodes in the product transaction level statistical probability knowledge graph are used to obtain a discrete probability distribution;

[0137] The parameterized probability model is used to fit the discrete probability distribution to obtain the parameterized description of the node;

[0138] The node parameterized descriptions of all financial product nodes and customer nodes are statistically analyzed to obtain a probabilistic grammar model.

[0139] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A financial product recommendation method, characterized in that: The method comprises: The acquired product transaction records are unified and quantified to obtain product transaction grades; The node probabilities of financial product nodes and customer nodes in the product transaction level statistical probability knowledge graph are used to obtain a discrete probability distribution; The parameterized probability model is used to fit the discrete probability distribution to obtain the parameterized description of the node; The node parameterized descriptions of all financial product nodes and customer nodes are counted to obtain a probabilistic grammar model; The probability grammar model is used to group and cluster the node probabilities to obtain the cluster center feature vector; The customer group similarity is calculated based on the customer group cluster center feature vectors between different probability grammar models. The calculation formula for customer group similarity is: ; in and are the first probability grammar model and the second probability grammar model respectively, and are adjustable weights, represents the similarity of the cluster center feature vectors between two probabilistic grammar models, represents the cluster center feature vector of the first probabilistic grammar model, represents the cluster center feature vector of the second probabilistic grammar model, represents the number of cluster center feature vectors that meet the threshold requirements, represents the number of cluster center feature vectors in the second probabilistic grammar model; Constructing a feature vector for a new customer node to obtain a feature vector of the new customer node; According to the similarity of customer groups, the distance between the feature vector of the new customer node and the feature vector of the cluster center is calculated to obtain the corresponding customer node; Send financial product recommendation information to new customer nodes based on the relationship between customer nodes and financial nodes.

2. The financial product recommendation method according to claim 1, characterized in that: The node probability includes node connection probability, prior probability, state jump probability, node state change probability, node state continuity probability and port usage probability.

3. The financial product recommendation method according to claim 1, characterized in that: The steps of grouping and clustering the node probabilities using the probabilistic grammar model to obtain the cluster center feature vector include: The node probabilities are arranged using the probabilistic grammar model to obtain a structured feature vector; According to the similarity of node feature vectors, customer nodes and product nodes in the probabilistic grammar model are grouped and clustered to obtain product node cluster centers and customer node cluster centers; The cluster center feature vector is obtained according to the product node cluster center and the customer node cluster center.

4. The financial product recommendation method according to claim 1, characterized in that: The product transaction record includes customer information, types of products sold, and user financial product usage. The customer information includes customer number, asset amount, daily consumption, credit rating, invisible asset amount, occupation, income amount and credit limit. The types of products sold include product name, branch, product type, interest rate and number of installments. The user financial product usage includes user repayment status, repayment amount, overdue status, overdue amount, installment repayment status and installment amount.

5. The financial product recommendation method according to claim 1, characterized in that: The product transaction levels include asset level, daily consumption level, invisible asset level, career stability level, and income level.

6. A financial product recommendation device, characterized in that: The device comprises: A preprocessing module is used to unify the units of the acquired product transaction records and perform quantitative grading processing to obtain the product transaction level; The probability statistics module is used to obtain a discrete probability distribution by using the node probabilities of financial product nodes and customer nodes in the product transaction level statistical probability knowledge graph; The parameterized probability model is used to fit the discrete probability distribution to obtain the parameterized description of the node; The node parameterized descriptions of all financial product nodes and customer nodes are counted to obtain a probabilistic grammar model; The grouping and clustering module is used to group and cluster the node probabilities using the probabilistic grammar model to obtain the cluster center feature vector; The similarity calculation module is used to calculate the customer group similarity based on the customer group cluster center feature vectors between different probability grammar models. The calculation formula for customer group similarity is: ; in and are the first probability grammar model and the second probability grammar model respectively, and are adjustable weights, represents the similarity of the cluster center feature vectors between two probabilistic grammar models, represents the cluster center feature vector of the first probabilistic grammar model, represents the cluster center feature vector of the second probabilistic grammar model, represents the number of cluster center feature vectors that meet the threshold requirements, represents the number of cluster center feature vectors in the second probabilistic grammar model; The recommendation information sending module is used to construct a feature vector of the new customer node to obtain the feature vector of the new customer node; According to the similarity of customer groups, the distance between the feature vector of the new customer node and the feature vector of the cluster center is calculated to obtain the corresponding customer node; Send financial product recommendation information to new customer nodes based on the relationship between customer nodes and financial nodes.