Method and device for recommending financial products, storage medium and electronic device

Abstract

The application discloses a financial product recommendation method and device, a storage medium and an electronic device, and relates to the field of artificial intelligence, wherein the method comprises the following steps: obtaining target data information of a target financial product; determining a feature vector set based on the target data information; fusing a plurality of feature vectors in the feature vector set by using a preset vector fusion algorithm to obtain a combined feature vector; inputting the combined feature vector and the plurality of feature vectors into a preset cross-mode model to determine a behavior mode vector of a customer; inputting the behavior mode vector of the customer into a preset prediction model to determine a preference degree of the customer to the target financial product, and determining a strategy for recommending the target financial product to the customer according to the preference degree of the customer to the target financial product. The application solves the technical problem that simple data statistics are performed on customer behavior data information in the related art, and the accuracy of product recommendation for customers is low.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a financial product recommendation method and device, a storage medium and an electronic device. BACKGROUND

[0002] For the financial industry, how to develop a mining strategy for potential financial customers in the financial industry wisdom marketing, instead of the traditional marketing mode, has become an important task in the banking industry.

[0003] Due to the diversification of existing data types, containing text, image and sequence information, etc., the text has both system generated part (such as product introduction, marketing activity information, etc.), and customer generated part (such as consultation, notes, etc.), and the wisdom marketing method in the related art often ignores the information carried by the text, and cannot fully integrate the text information, accurately analyze the behavior of the customer, so that the wisdom marketing system cannot fully understand the personalized characteristics of the customer and the product, and thus cannot generate accurate recommendation results.

[0004] Secondly, the current financial industry realizes wisdom marketing mainly by collecting customer behavior information in self-owned channels and marketing channels, and constructing a customer portrait based on the customer behavior information to predict customer demand, and then recommending financial products to customers to realize bank wisdom marketing delivery. Compared with the traditional customer acquisition and customer activation method relying on outlets, this processing method has great improvement in marketing cost control and customer stickiness maintenance. However, the existing method only performs simple mathematical statistical analysis on the collected customer behavior information, ignores the importance of feature engineering, and the prediction accuracy is not high, resulting in unsatisfactory wisdom marketing effect.

[0005] In view of the above problems, no effective solution has been proposed so far. SUMMARY

[0006] The embodiments of the present application provide a financial product recommendation method and device, a storage medium and an electronic device, to at least solve the technical problem that the related art performs simple data statistics on customer behavior data information, resulting in low accuracy of recommending products to customers.

[0007] According to an aspect of some embodiments of the present application, a method for recommending a financial product is provided. The method includes: obtaining target data information of a target financial product, wherein the target data information includes at least product data information of the target financial product, customer portrait information of the target financial product, and behavior data information of a customer of the target financial product; determining a feature vector set based on the target data information, wherein the feature vector set includes a plurality of feature vectors; fusing the plurality of feature vectors in the feature vector set by a preset vector fusion algorithm to obtain a merged feature vector; inputting the merged feature vector and the plurality of feature vectors into a preset cross-modal model to determine a behavior mode vector of the customer, wherein the preset cross-modal model is configured to fuse the merged feature vector and the plurality of feature vectors and learn an association between the merged feature vector and the plurality of feature vectors; inputting the behavior mode vector of the customer into a preset prediction model to determine a preference degree of the customer for the target financial product, and determining a strategy for recommending the target financial product to the customer based on the preference degree of the customer for the target financial product.

[0008] Further, the step of determining the feature vector set based on the target data information includes: inputting the target data information into a language representation model to output a first feature vector of the target data information, wherein the language representation model is configured to perform word vector representation on the target data information; and determining the feature vector set based on the first feature vector and the target data information.

[0009] Further, the step of determining the feature vector set based on the first feature vector and the target data information includes: pre-processing the target data information to obtain initial feature data, and determining derived feature data based on the initial feature data, wherein the derived feature data is dynamic behavior data of the customer of the target financial product; determining a second feature vector of the initial feature data based on the initial feature data; determining a third feature vector of the derived feature data based on the derived feature data; and determining the feature vector set based on the first feature vector, the second feature vector, and the third feature vector.

[0010] Further, the step of inputting the merged feature vector and the plurality of feature vectors into a preset cross-modal model to determine the behavior pattern vector of the customer comprises: inputting the merged feature vector and the plurality of feature vectors into a preset cross-modal model, determining a weight vector set for processing the merged feature vector and the plurality of feature vectors through a self-attention learning module of the preset cross-modal model, wherein the weight vector set comprises a plurality of weight vectors; and processing the merged feature vector and the plurality of feature vectors based on the weight vector set to obtain the behavior pattern vector of the customer.

[0011] Further, the step of determining derived feature data based on the initial feature data comprises: performing data exploration on the initial feature data, dividing the explored initial feature data according to time points, and determining feature data of a plurality of time periods; and performing classification statistics on the feature data of each time period to determine the derived feature data.

[0012] Further, the step of obtaining target data information of a target financial product comprises: extracting product data information of the target financial product and behavior data information of the customer from a target database; and extracting customer portrait information of the customer from a preset customer portrait system.

[0013] Further, the step of preprocessing the target data information to obtain initial feature data comprises: integrating the product data information, the behavior data information and the customer portrait information through a third-party database to determine integrated data information; and performing data cleaning and data filling on the integrated data information to determine the initial feature data.

[0014] According to another aspect of the embodiments of the present application, a financial product recommendation device is further provided, comprising: an acquisition unit configured to acquire target data information of a target financial product, wherein the target data information at least comprises product data information of the target financial product, customer portrait information of the target financial product, and behavior data information of a customer of the target financial product; a first determination unit configured to determine a feature vector set based on the target data information, wherein the feature vector set comprises a plurality of feature vectors; a fusion unit configured to fuse the plurality of feature vectors in the feature vector set by a preset vector fusion algorithm to obtain a merged feature vector; a second determination unit configured to input the merged feature vector and the plurality of feature vectors into a preset cross-modal model to determine a behavior mode vector of the customer, wherein the preset cross-modal model is configured to fuse the merged feature vector and the plurality of feature vectors and learn the correlation between the merged feature vector and the plurality of feature vectors; and a third determination unit configured to input the behavior mode vector of the customer into a preset prediction model to determine a preference degree of the customer for the target financial product, and determine a strategy of recommending the target financial product to the customer according to the preference degree of the customer for the target financial product.

[0015] Further, the first determination unit comprises: an output subunit configured to input the target data information into a language representation model to output a first feature vector of the target data information, wherein the language representation model is configured to perform word vector representation on the target data information; and a first determination subunit configured to determine the feature vector set based on the first feature vector and the target data information.

[0016] Further, the first determination subunit comprises: a processing module configured to pre-process the target data information to obtain initial feature data and determine derived feature data based on the initial feature data, wherein the derived feature data is dynamic behavior data of the customer of the target financial product; a first determination module configured to determine a second feature vector of the initial feature data based on the initial feature data; a second determination module configured to determine a third feature vector of the derived feature data based on the derived feature data; and a third determination module configured to determine the feature vector set based on the first feature vector, the second feature vector, and the third feature vector.

[0017] Further, the second determining unit comprises: a second determining sub-unit, configured to input the merged feature vector and the plurality of feature vectors into a preset cross-modal model, determine a weight vector set for processing the merged feature vector and the plurality of feature vectors through a self-attention learning module of the preset cross-modal model, wherein the weight vector set comprises a plurality of weight vectors; and a first processing sub-unit, configured to process the merged feature vector and the feature vector set based on the weight vector set to obtain the behavior pattern vector of the customer.

[0018] Further, the processing module comprises: a processing sub-module, configured to perform data exploration on the initial feature data, divide the explored initial feature data according to time points, and determine feature data of a plurality of time periods; and a first determining sub-module, configured to classify and count the feature data of each time period, and determine the derived feature data.

[0019] Further, the acquisition unit comprises: a first extracting unit, configured to extract product data information of the target financial product and behavior data information of the customer from a target database; and a second extracting unit, configured to extract customer portrait information of the customer from a preset customer portrait system.

[0020] Further, the processing module further comprises: a second determining sub-module, configured to perform data integration on the product data information, the behavior data information and the customer portrait information through a third-party database to determine integrated data information; and a third determining sub-module, configured to perform data cleaning and data filling on the integrated data information to determine the initial feature data.

[0021] In this invention, target data information of a target financial product is obtained, wherein the target data information includes at least: product data information of the target financial product, customer profile information of the target financial product, and customer behavior data information of the target financial product; based on the target data information, a feature vector set is determined, wherein the feature vector set includes: multiple feature vectors; multiple feature vectors in the feature vector set are fused using a preset vector fusion algorithm to obtain a merged feature vector; the merged feature vector and the multiple feature vectors are input into a preset cross-model model to determine the customer's behavior pattern vector, wherein the preset cross-model model is used to fuse the merged feature vector and the multiple feature vectors and learn the correlation between the merged feature vector and the multiple feature vectors; the customer's behavior pattern vector is input into a preset prediction model to determine the customer's preference for the target financial product, and based on the customer's preference for the target financial product, a strategy for recommending the target financial product to the customer is determined. This solves the technical problem in related technologies where simple data statistics on customer behavior data lead to low accuracy in recommending products to customers. In this invention, multiple feature vectors are obtained through target data information. The relationships between these feature vectors are then fused to obtain a customer behavior pattern vector. Based on this customer behavior pattern vector, a strategy for recommending target financial products is derived, thereby improving the accuracy of product recommendations to customers. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0023] Figure 1 This is a flowchart of an optional financial product recommendation method according to an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of an optional financial product recommendation device according to an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of an optional electronic device according to an embodiment of the present invention. Detailed Implementation

[0026] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort should belong to the scope of the present application.

[0027] It should be noted that the terms "first", "second" and the like in the description and claims of the present application and the above drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in other than the order illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a list of steps or units need not be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products or devices.

[0028] It should be noted that the recommendation method of financial products and the device thereof in the present disclosure can be used in the field of artificial intelligence in the case of financial product recommendation strategy, and can also be used in any field other than the field of artificial intelligence in the case of marketing recommendation of financial products. The application field of the recommendation method of financial products and the device thereof in the present disclosure is not limited.

[0029] It should be noted that the related information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present disclosure are all information and data authorized by the user or authorized by all parties. For example, an interface is provided between the system and the related user or institution. Before obtaining the related information, the interface needs to send a request for obtaining to the aforementioned user or institution, and after receiving the consent information fed back by the aforementioned user or institution, the related information is obtained.

[0030] The present application can be applied to various software products, control systems, client marketing systems (including but not limited to mobile clients, PC machines, etc.) of various financial institutions. Taking the client as an example, through marketing recommendation scheme analysis of the client, the business content of the financial institution (including but not limited to transfer, financial management, fund, payment, account checking, advertising, recommendation, etc.) can be realized.

[0031] The present application will be further described in conjunction with the embodiments.

[0032] Embodiment one

[0033] According to an embodiment of the present application, an optional financial product recommendation method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0034] Figure 1 is a flowchart of an optional financial product recommendation method according to an embodiment of the present application, as shown in the figure, the method comprises the following steps: Figure 1

[0035] Step S101, obtaining target data information of a target financial product, wherein the target data information at least includes product data information of the target financial product, customer portrait information of the target financial product, and behavior data information of the customer of the target financial product.

[0036] The target data information described above can include product data information of the recommended target financial product. The product data information can be basic data information of the product (such as basic information such as product introduction, product parameters, and product functions), and can also include customer portrait information of the target financial product (such as personal attributes, life-related attributes, and occupation-related attributes of the customer), and can also include behavior data information of the customer (such as consumption data and attention preference data of the customer).

[0037] Step S102, determining a feature vector set based on the target data information, wherein the feature vector set includes a plurality of feature vectors.

[0038] According to the target data information, a feature vector set composed of a plurality of feature vectors can be determined. The plurality of feature vectors can include feature vectors directly generated from the target data information, and can also include feature vectors derived from dynamic behavior data of the customer (such as calculating the dynamic behavior data of the customer as the annual daily deposit, the annual daily deposit of the current account, and the annual daily deposit of the fixed deposit of the customer) from the target data information. That is, the dynamic behavior data can be behavior data counted by stage or period.

[0039] Step S103, fusing the plurality of feature vectors in the feature vector set through a preset vector fusion algorithm to obtain a merged feature vector.

[0040] In order to avoid the problem of high dimension in the fusion process of the plurality of feature vectors, the preset vector fusion algorithm described above can adopt a low-quality decomposition factor algorithm. Through the preset vector fusion algorithm, the plurality of feature vectors can be fused to obtain the merged feature vector. ​

[0041] The following takes the first feature vector P, the second feature vector D, and the third feature vector Y as an example to illustrate the way to obtain the merged feature vector.

[0042] The first feature vector P, the second feature vector D, and the third feature vector Y of the target financial product are characterized by using a low-rank matrix factorization (LMF) algorithm. The low-rank matrix factorization algorithm is used as the preset vector fusion method to avoid the high-dimensional problem caused by directly connecting and merging the feature vectors. The specific principle is that the input vector (corresponding to the feature vector set) is processed by a linear layer to generate a vector representation: where W is the weight, R is a real number, is the dimension of the vector b is the offset. At the same time, based on the decomposition of W, according to , the following formula can be derived:

[0043] ;

[0044] The Z feature vector dimension of m modalities (m feature vectors) is The low-rank matrix factorization algorithm is used instead of the original direct connection and merging method of the feature vectors to obtain a low-dimensional multi-modal merged feature vector h.

[0045] Step S104, input the merged feature vector and the plurality of feature vectors into a preset cross-modal model to determine the behavior pattern vector of the customer, wherein the preset cross-modal model is used to fuse the merged feature vector and the plurality of feature vectors, and learn the association relationship between the merged feature vector and the plurality of feature vectors.

[0046] The above-mentioned preset cross-modal model can adopt a cross-modal Transformer model to realize the fusion of information between different feature vectors and learn the association relationship between different feature vector information, and further obtain a rich fusion feature vector customer behavior pattern vector.

[0047] Step S105, input the behavior pattern vector of the customer into a preset prediction model to determine the preference degree of the customer to the target financial product, and determine the strategy of recommending the target financial product to the customer according to the preference degree of the customer to the target financial product.

[0048] The preset prediction module can adopt a prediction model constructed by combining a Bi-LSTM (Bidirectional Long Short-Term Memory) algorithm and a softmax method, to predict the preference degree of the customer to the target financial product, determine the recommendation probability, and then determine the strategy for recommending the target financial product to the customer. The recommendation strategy can refer to the recommendation frequency and the number of the target financial product recommended to the customer.

[0049] The customer behavior mode vector M is input into a Bi-LSTM (Bidirectional Long Short-Term Memory) model, and a prediction model is constructed by combining a softmax method, to realize the prediction of the preference degree of the customer to the financial asset pool product, and further realize the provision of personalized intelligent marketing strategies for the customer. Secondly, the Bi-LSTM model is composed of a forward long short-term memory network and a backward long short-term memory network, and has a wide application in sequence data processing and has achieved good results. Specifically, it is represented by formulas (3.1), (3.2) and (3.3):

[0050] (3.1)

[0051] (3.2)

[0052] (3.3)

[0053] Among them, represents a forward variable of the long short-term memory network, represents a backward variable of the long short-term memory network, represents a long short-term memory network variable represents a forward long short-term memory network, represents a backward long short-term memory network, represents a forward hidden layer state at time t, is a backward hidden layer state at time t, represents a splicing result of the forward hidden layer state and the backward hidden layer state at time t, that is, an output result of the model .

[0054] Then, the output vector H obtained by processing the Bi-LSTM model is taken as an input value of a normalization method, to obtain a prediction result of the financial asset pool intelligent recommendation model, which is specifically defined as formula:

[0055]

[0056] Among them, H is an input vector, For model training parameters, y is the model output category. Meanwhile, the embodiment can use cross-entropy as the optimization objective function, which is defined by the formula where N represents the number of classification categories, is the real label, is the predicted label.

[0057] Through the above steps, the technical problem that the related art simply statistically processes customer behavior data information, resulting in low accuracy of product recommendation for customers, is solved. In the embodiment, a plurality of feature vectors are obtained through target data information, the relationship of the plurality of feature vectors is fused to obtain a customer behavior pattern vector, and then a strategy of recommending a target financial product is obtained according to the customer behavior pattern vector, thereby achieving the technical effect of improving the accuracy of product recommendation for customers.

[0058] In order to avoid the case that the data of the direct target data information is sparse, in the embodiment, the step of determining the feature vector set based on the target data information is further limited: inputting the target data information into a language representation model to output a first feature vector of the target data information, wherein the language representation model is used for word vector representation of the target data information; determining the feature vector set based on the first feature vector and the target data information.

[0059] The language representation model described above can be a pre-trained BERT model. The target data information can be represented by a word vector through the language representation model. Based on the first feature vector and other plurality of feature vectors determined through the target data information, the feature vector set described above can be determined.

[0060] The following takes a financial asset pool as an example of a target financial product. Through a pre-trained BERT model (Bidirectional Encoder Representations from Transformers, BERT for short, language representation model), the text information of the financial asset pool product (corresponding to the target data information described above) is input into the Self-Attention layer of the BERT model, and the pre-trained BERT model is used to represent the input text information by a word vector to obtain the financial asset pool product feature vector P.

[0061] By representing the target data information by a word vector to obtain the first feature vector, and determining the feature vector set through the first feature and the target data information, the technical effect of reducing the sparsity of the target data information and facilitating the improvement of data analysis efficiency can be achieved.

[0062] In order to avoid accurately obtaining each feature vector in the feature vector set, in the embodiment, the step of determining the feature vector set based on the first feature vector and the target data information is further limited: the target data information is preprocessed to obtain initial feature data, and derived feature data is determined based on the initial feature data, wherein the derived feature data is the dynamic behavior data of the customer of the target financial product; based on the initial feature data, a second feature vector of the initial feature data is determined; based on the derived feature data, a third feature vector of the derived feature data is determined; and based on the first feature vector, the second feature vector and the third feature vector, the feature vector set is determined.

[0063] In the embodiment, since the obtained target data information involves many categories and has a large amount of data, there may be deviation in selecting features only according to the data science method. Therefore, feature derivation is performed according to the business logic of the target financial product to obtain features closer to the business meanings of the target financial product, further widening the precision and breadth of the original target data information, and the target data information can be preprocessed to obtain initial feature data.

[0064] Then the initial feature data is vectorized to determine the second feature vector D of the initial feature data, the derived feature data obtained based on the initial feature data is vectorized to obtain the derived feature vector Y, and the feature vector set Z is determined based on the first feature vector P, the second feature vector D and the third feature vector Y, so as to achieve the technical effect of improving the precision of the data represented by each feature vector in the feature vector set.

[0065] In order to further improve the accuracy of the recommended customer behavior information by learning and fusing the correlation of different features, the step of inputting the merged feature vector and the plurality of feature vectors into the preset cross-modal model to determine the behavior pattern vector of the customer includes: inputting the merged feature vector and the plurality of feature vectors into the preset cross-modal model, determining a weight vector set for processing the merged feature vector and the plurality of feature vectors through a self-attention learning module of the preset cross-modal model, wherein the weight vector set includes a plurality of weight vectors; and processing the merged feature vector and the plurality of feature vectors based on the weight vector set to obtain the behavior pattern vector of the customer.

[0066] For example: based on the cross-modal Transformer model, the multi-modal merged feature vector h and each modal representation vector Z (corresponding to the feature vector set) are inputted as inputs to realize fusion and learning of the information between different modalities and the correlation between different modal information, and further obtain rich multi-modal fusion feature vectors. The cross-modal Transformer model is composed of D-layer cross-modal self-attention modules, specifically:

[0067] (1) Three weight vectors are obtained by training based on the input feature vectors, namely the query weight vector Q, the key weight vector K, and the value weight vector V.

[0068] (2) Perform a dot product calculation based on vector Q and vector K to obtain the degree of influence of the input value on other input sequence parts, and perform normalization to ensure gradient stability;

[0069] (3) Based on the output vector obtained in step (2), perform a dot product with vector V and add them together to obtain the weighted input vector. This is specifically represented by the following formula:

[0070]

[0071] in, It is the dimension of the query weight vector Q and the key weight vector K. Based on the weighted input vector, the input feature vector can be weighted to obtain the customer behavior pattern representation vector M that is related and complementary among multiple modal information, thereby achieving more accurate customer demand prediction for target financial products.

[0072] To avoid the initial feature data not fully representing customer behavior and thus leading to inaccurate recommendations for target financial products, this embodiment further defines the steps for determining derived feature data based on the initial feature data: data exploration is performed on the initial feature data, and the explored initial feature data is divided according to time points to determine feature data for multiple time periods; the feature data for each time period is classified and statistically analyzed to determine derived feature data.

[0073] For example, data exploration can be conducted on aspects such as feature correlation, data distribution, and data offset distribution. Secondly, based on the initial feature data and the business implications involved, financial asset pools, bills, and pledges can be divided into stages according to time points. Derivative feature data can be calculated based on these stages, such as: number of days held, average daily value of wealth management products, total amount and frequency of pledges, total amount and number of bills held in finance companies, total amount and quantity of commercial bills held, total amount and number of discounts, total amount and frequency of bill receipts and acceptances, average number of days held, total amount and frequency of endorsements and acceptances, average number of days held, total amount and quantity of bank drafts held, total amount of financing in target financial institutions, total credit line, credit rating, asset transaction amount, average daily value of structured deposits, average daily value of demand deposits, and average daily value of time deposits, etc., to characterize the dynamic behavioral trajectory of customers in the asset pool. This achieves the technical effect of improving the accuracy of customer behavior data.

[0074] To avoid the possibility of unreliable sources of target data information, this implementation further limits the steps for obtaining target data information of target financial products: extracting product data information and customer behavior data information of target financial products from the target database; and extracting customer profile information of customers from the preset customer profile system.

[0075] For example, data related to the target financial product (corresponding to the aforementioned target data information) can be extracted from the financial institution's data lake / database and corporate customer profiling system. This includes product data information and customer behavior data information. Customer profile information can be extracted from a pre-set customer profiling system. The specific data involved may include financial asset pools, bills, pledges, acceptance amounts, and customer profiles, thereby achieving the technical effect of improving the reliability of the target data information.

[0076] To avoid situations where the target data is large in volume and suffers from severe data redundancy, this embodiment further limits the steps of preprocessing the target data to obtain initial feature data: integrating product data, behavioral data, and customer profile information through a third-party database to determine integrated data; and cleaning and filling the integrated data to determine the initial feature data.

[0077] For example, third-party libraries in Python (a programming language) such as Pandas (a tool created to solve data analysis tasks) and NumPy (a matrix-based mathematical computing module) can be used to perform some basic data preprocessing, including data integration (e.g., merging various business data), data cleaning (e.g., deleting records with empty customer behavior information), and data imputation (e.g., interpolating missing fields based on the characteristics of the values). This achieves the goal of obtaining accurate initial feature data, which facilitates the improvement of the accuracy of the analysis results using the initial feature data.

[0078] The invention will be further illustrated below with reference to another optional embodiment.

[0079] Example 2

[0080] This embodiment provides another optional method for recommending financial products, addressing the problem of neglecting textual information and failing to consider the correlation and complementarity between multimodal information in customer financial demand analysis and smart marketing based on customer business behavior data of financial institutions. It uses a low-rank factorization algorithm and a cross-modal Transformer fusion method to obtain customer behavior pattern feature representation vectors, and then uses the Bi-LSTM method combined with a normalization method to construct a prediction model to achieve more accurate customer demand prediction.

[0081] According to the customer basic information data, the customer business behavior data (corresponding to the behavior information data in embodiment one), the financial asset pool product information (corresponding to the target financial product in embodiment one), the original data is preprocessed by using the third-party library Pandas, Numpy and the like of Python, and the financial asset pool, bills, pledge and the like are divided into stages according to the time point according to the business meaning, and the derived features (corresponding to the derived feature data in embodiment one) closer to the business meaning are derived; then the financial asset pool product text information is processed by using the Bert algorithm to obtain the financial asset pool product feature vector, and the feature vector is further enriched; then the low-rank decomposition factor algorithm is combined with the cross-modal Transformer fusion method to obtain the customer behavior mode feature representation vector related and complementary among multiple modal information; finally, based on the customer behavior mode feature representation vector, a prediction model is constructed by using the Bi-LSTM (bidirectional long short-term memory network) algorithm combined with the softmax method, and then the preference degree of the customer to the financial asset pool product is predicted to generate a target financial customer recommendation list. Considering the information carried by the text, it is beneficial to comprehensively understand the personalized characteristics of the customer and the product, has a practical reference value, and is also an indispensable factor in predicting customer demand; considering the multi-modal information fusion, the data sparseness and cold start problem can be effectively alleviated, which is beneficial to accurately analyzing and describing the customer behavior mode, and can improve the prediction accuracy of customer demand.

[0082] The application will be further described below in combination with another alternative embodiment.

[0083] Embodiment three

[0084] The embodiment provides an alternative financial product recommendation device, and each implementation unit included in the recommendation device corresponds to each implementation step in the above-mentioned embodiment one.

[0085] Figure 2 It is a schematic diagram of an alternative financial product recommendation device according to the embodiment of the application, as shown in Figure 2 The recommendation device comprises an acquisition unit 21, a first determination unit 22, a fusion unit 23, a second determination unit 24 and a third determination unit 25.

[0086] The acquisition unit 21 is used for acquiring target data information of a target financial product, wherein the target data information at least includes product data information of the target financial product, customer portrait information of the target financial product and behavior data information of a customer of the target financial product.

[0087] The first determination unit 22 is used for determining a feature vector set based on the target data information, wherein the feature vector set comprises a plurality of feature vectors.

[0088] The fusion unit 23 is configured to fuse a plurality of feature vectors in the feature vector set by a preset vector fusion algorithm to obtain a merged feature vector.

[0089] The second determination unit 24 is configured to input the merged feature vector and the plurality of feature vectors into a preset cross-modal model to determine a behavior pattern vector of the customer, wherein the preset cross-modal model is configured to fuse the merged feature vector and the plurality of feature vectors and learn the correlation between the merged feature vector and the plurality of feature vectors.

[0090] The third determination unit 25 is configured to input the behavior pattern vector of the customer into a preset prediction model to determine the preference degree of the customer to the target financial product, and determine the strategy of recommending the target financial product to the customer according to the preference degree of the customer to the target financial product.

[0091] The above-mentioned financial product recommendation device first acquires target data information of a target financial product by the acquisition unit 21, wherein the target data information at least includes product data information of the target financial product, customer portrait information of the target financial product, and behavior data information of the customer of the target financial product, then determines a feature vector set based on the target data information by the first determination unit 22, wherein the feature vector set includes a plurality of feature vectors, then fuses the plurality of feature vectors in the feature vector set by the fusion unit 23 by a preset vector fusion algorithm to obtain a merged feature vector, inputs the merged feature vector and the plurality of feature vectors into a preset cross-modal model by the second determination unit 24 to determine a behavior pattern vector of the customer, wherein the preset cross-modal model is configured to fuse the merged feature vector and the plurality of feature vectors and learn the correlation between the merged feature vector and the plurality of feature vectors, and finally inputs the behavior pattern vector of the customer into a preset prediction model by the third determination unit 25 to determine the preference degree of the customer to the target financial product, and determines the strategy of recommending the target financial product to the customer according to the preference degree of the customer to the target financial product. The technical problem of low accuracy of product recommendation to the customer caused by simple data statistics on the customer behavior data information in the related art is solved. In the present application, the plurality of feature vectors are obtained from the target data information, the relationship of the plurality of feature vectors is fused to obtain the behavior pattern vector of the customer, and then the strategy of recommending the target financial product is obtained according to the behavior pattern vector of the customer, thereby achieving the technical effect of improving the accuracy of product recommendation to the customer.

[0092] Optionally, in the financial product recommendation device of the third embodiment of the present application, the first determination unit 22 includes an output subunit configured to input the target data information into a language representation model to output a first feature vector of the target data information, wherein the language representation model is configured to perform word vector representation on the target data information; and a first determination subunit configured to determine the feature vector set based on the first feature vector and the target data information.

[0093] Optionally, in the financial product recommendation apparatus of the third embodiment of the present application, the first determining subunit comprises: a processing module, configured to pre-process the target data information to obtain initial feature data, and determine derivative feature data based on the initial feature data, wherein the derivative feature data is dynamic behavior data of a customer of the target financial product; a first determining module, configured to determine a second feature vector of the initial feature data based on the initial feature data; a second determining module, configured to determine a third feature vector of the derivative feature data based on the derivative feature data; and a third determining module, configured to determine a feature vector set based on the first feature vector, the second feature vector, and the third feature vector.

[0094] Optionally, in the financial product recommendation apparatus of the third embodiment of the present application, the second determining unit 24 comprises: a second determining subunit, configured to input the merged feature vector and the plurality of feature vectors into a preset cross-modal model, and determine a weight vector set for processing the merged feature vector and the plurality of feature vectors through a self-attention learning module of the preset cross-modal model, wherein the weight vector set comprises a plurality of weight vectors; and a first processing subunit, configured to process the merged feature vector and the feature vector set based on the weight vector set to obtain a behavior pattern vector of the customer.

[0095] Optionally, in the financial product recommendation apparatus of the third embodiment of the present application, the processing module comprises: a processing sub-module, configured to perform data exploration on the initial feature data, and divide the explored initial feature data according to time points to determine feature data of a plurality of time periods; and a first determining sub-module, configured to classify and count the feature data of each time period to determine the derivative feature data.

[0096] Optionally, in the financial product recommendation apparatus of the third embodiment of the present application, the acquisition unit 21 comprises: a first extraction unit, configured to extract product data information of the target financial product and behavior data information of the customer from a target database; and a second extraction unit, configured to extract customer portrait information of the customer from a preset customer portrait system.

[0097] Optionally, in the financial product recommendation apparatus of the third embodiment of the present application, the processing module further comprises: a second determining sub-module, configured to perform data integration on the product data information, the behavior data information, and the customer portrait information through a third-party database to determine integrated data information; and a third determining sub-module, configured to perform data cleaning and data filling on the integrated data information to determine the initial feature data.

[0098] The recommendation device of the financial product can further include a processor and a memory, and the acquisition unit 21, the first determination unit 22, the fusion unit 23, the second determination unit 24, the third determination unit 25, and the like are stored in the memory as program units, and the corresponding functions are realized by the processor executing the program units stored in the memory.

[0099] The processor includes a core, and the core calls the corresponding program units in the memory. The core can be one or more, and the target data information is obtained by adjusting the core parameters to obtain a plurality of feature vectors. The relationship of the plurality of feature vectors is fused to obtain a customer behavior mode vector, and then the strategy of recommending the target financial product is obtained according to the customer behavior mode vector, so that the technical effect of improving the accuracy of recommending the product to the customer is achieved.

[0100] The memory can include a non-permanent memory in a computer readable medium, a random access memory (RAM), and / or a non-volatile memory such as a read-only memory (ROM) or a flash memory (flash RAM), and the memory includes at least one memory chip.

[0101] According to another aspect of the embodiment of the present application, an electronic device is further provided, including a processor and a memory for storing executable instructions of the processor, wherein the processor is configured to execute the financial product recommendation method of any one of the above by executing the executable instructions.

[0102] According to another aspect of the embodiment of the present application, a computer readable storage medium is further provided, including a stored computer program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute the financial product recommendation method of any one of the above when the computer program runs.

[0103] Figure 3 is a schematic diagram of an optional electronic device according to the embodiment of the present application, as Figure 3 shown, the embodiment of the present application provides an electronic device 30, the device includes a processor, a memory and a program stored in the memory and executable on the processor.

[0104] The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0105] In the above-mentioned embodiments of the present application, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0106] In several embodiments provided in the present application, it should be understood that the disclosed technology can be implemented by other ways. Among them, the above-described device embodiments are only schematic, for example, the division of the units can be a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, units or modules, and can be electrical or other forms.

[0107] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed to multiple units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0108] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0109] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0110] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method for recommending financial products, characterized in that, include: Obtain target data information of the target financial product, wherein the target data information includes at least: product data information of the target financial product, customer profile information of the target financial product, and behavioral data information of the customer of the target financial product; Based on the target data information, a set of feature vectors is determined, wherein the set of feature vectors includes multiple feature vectors; the method further includes performing feature derivation processing on the target data information according to the business logic of the target financial product. Multiple feature vectors in the feature vector set are fused using a preset vector fusion algorithm to obtain a merged feature vector. The preset vector fusion algorithm is a low-quality factorization algorithm. The merged feature vector and the plurality of feature vectors are input into a preset cross-model to determine the customer's behavior pattern vector. The preset cross-model is used to fuse the merged feature vector and the plurality of feature vectors and learn the correlation between the merged feature vector and the plurality of feature vectors. The customer's behavioral pattern vector is input into a preset prediction model to determine the customer's preference for the target financial product, and a strategy for recommending the target financial product to the customer is determined based on the customer's preference for the target financial product. The step of determining the feature vector set based on the target data information includes: The target data information is input into a language representation model, and the first feature vector of the target data information is output. The language representation model is used to represent the target data information by word vectors. Based on the first feature vector and the target data information, the feature vector set is determined; The step of determining the feature vector set based on the first feature vector and the target data information includes: The target data information is preprocessed to obtain initial feature data, and derived feature data is determined based on the initial feature data, wherein the derived feature data is the dynamic behavior data of the customers of the target financial product; Based on the initial feature data, determine the second feature vector of the initial feature data; Based on the derived feature data, determine the third feature vector of the derived feature data; The feature vector set is determined based on the first feature vector, the second feature, and the third feature vector.

2. The recommended method according to claim 1, characterized in that, The step of inputting the merged feature vector and the multiple feature vectors into a preset cross-model to determine the customer's behavior pattern vector includes: The merged feature vector and the plurality of feature vectors are input into a preset cross-modal model. The self-attention learning module of the preset cross-modal model determines a set of weight vectors for processing the merged feature vector and the plurality of feature vectors, wherein the set of weight vectors includes a plurality of weight vectors. Based on the set of weight vectors, the merged feature vector and the set of feature vectors are processed to obtain the customer's behavior pattern vector.

3. The recommended method according to claim 1, characterized in that, The step of determining derived feature data based on the initial feature data includes: The initial feature data is explored, and the explored initial feature data is divided according to time points to determine feature data for multiple time periods; The characteristic data for each time period are classified and statistically analyzed to determine the derived characteristic data.

4. The recommended method according to claim 1, characterized in that, The steps to obtain target data information for a target financial product include: Extract product data information of the target financial product and customer behavior data information from the target database; Extract the customer profile information of the customer from the preset customer profile system.

5. The recommended method according to claim 1, characterized in that, The step of preprocessing the target data information to obtain initial feature data includes: By integrating the product data, behavioral data, and customer profile information through a third-party database, the integrated data information is determined. Integrate data information to perform data cleaning and data filling, and determine the initial feature data.

6. A device for recommending financial products, characterized in that, include: The acquisition unit is used to acquire target data information of the target financial product, wherein the target data information includes at least: product data information of the target financial product, customer profile information of the target financial product, and behavioral data information of the customer of the target financial product; The first determining unit is configured to determine a set of feature vectors based on the target data information, wherein the set of feature vectors includes multiple feature vectors; the device is further configured to perform feature derivation processing on the target data information according to the business logic of the target financial product. The fusion unit is used to fuse multiple feature vectors in the feature vector set using a preset vector fusion algorithm to obtain a merged feature vector, wherein the preset vector fusion algorithm is a low-quality factorization algorithm. The second determining unit is used to input the merged feature vector and the plurality of feature vectors into a preset cross-mode model to determine the customer's behavior pattern vector, wherein the preset cross-mode model is used to fuse the merged feature vector and the plurality of feature vectors, and learn the correlation between the merged feature vector and the plurality of feature vectors; The third determining unit is used to input the customer's behavior pattern vector into a preset prediction model, determine the customer's preference for the target financial product, and determine a strategy for recommending the target financial product to the customer based on the customer's preference for the target financial product. The first determining unit includes: an output subunit, used to input target data information into a language representation model and output a first feature vector of the target data information, wherein the language representation model is used to represent the target data information using word vectors; and a first determining subunit, used to determine a set of feature vectors based on the first feature vector and the target data information. The first determining subunit includes: a processing module, used to preprocess the target data information to obtain initial feature data, and determine derived feature data based on the initial feature data, wherein the derived feature data is the dynamic behavior data of the target financial product's customers; a first determining module, used to determine a second feature vector of the initial feature data based on the initial feature data; a second determining module, used to determine a third feature vector of the derived feature data based on the derived feature data; and a third determining module, used to determine a feature vector set based on the first feature vector, the second feature vector, and the third feature vector.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform a method for recommending the financial product according to any one of claims 1 to 5.

8. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of recommending the financial product according to any one of claims 1 to 5.

Images