Financial product recommendation method and device, equipment, medium and product

Abstract

The embodiment of the invention provides a financial product recommendation method and device, equipment, a medium and a product, and relates to the field of financial science and technology or the technical field of artificial intelligence. According to the method, a historical transaction record of a financial product purchased by a user, a search index of each candidate financial product in a plurality of candidate financial products and an emotion index of each candidate financial product are jointly input into a pre-trained recommendation model, so that candidate financial recommended products output by the recommendation model can be obtained. According to the method, through the intelligent recommendation process, the technical problem that it is difficult for recommendation personnel to comprehensively and deeply understand and compare all financial products is solved, meanwhile, the problem that the recommendation result is inaccurate and unfair due to incomplete information mastering or subjective prejudice is avoided, then the accuracy and fairness of the recommendation result are improved, and the recommendation efficiency is improved. Therefore, the purchasing success rate of the financial product by the user is improved.

Description

Technical Field

[0001] The present application relates to the field of financial technology or artificial intelligence technology, and in particular to a method, device, equipment, medium and product for recommending financial products. Background Art

[0002] As the economy continues to grow and prosper, the financial sector continues to innovate and launch more and more financial products to meet the needs of different consumers and investors.

[0003] In the existing technology, the recommendation of financial products usually relies on human intervention. This method requires the recommender to not only have in-depth knowledge of financial products, but also to have a deep understanding of customer needs in order to provide customers with financial product recommendations that best suit their needs.

[0004] However, since it is difficult for recommenders to have a comprehensive and in-depth understanding of all financial products and the specific needs of customers, and they may also be affected by incomplete information or subjective bias, these factors often lead to inaccurate and unfair recommendations, which in turn damages customer satisfaction. Summary of the invention

[0005] The present application provides a method, device, equipment, medium and product for recommending financial products, which are used to solve the technical problem that it is difficult for recommenders to have a comprehensive and in-depth understanding and comparison of all financial products, and they may be affected by incomplete information or subjective bias, resulting in inaccurate and unfair recommendation results, thereby damaging customer satisfaction and reducing the success rate of financial product purchases.

[0006] In a first aspect, the present application provides a method for recommending financial products, comprising:

[0007] Acquire historical transaction records of users purchasing financial products, search indexes of each of the candidate financial products among multiple candidate financial products, and sentiment indexes of each of the candidate financial products, wherein the sentiment index of the candidate financial product is used to characterize the sentiment direction of the evaluation of the candidate financial product;

[0008] The historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products are input into a pre-trained recommendation model to obtain a recommended financial product output by the recommendation model, wherein the recommended financial product is part of the multiple candidate financial products.

[0009] In a second aspect, the present application provides a training method for a recommendation model for financial products, including:

[0010] Acquire multiple training samples, each of which includes a historical transaction record of a user purchasing a financial product, a search index of each of multiple candidate financial products, a sentiment index of each of the candidate financial products, and a real selected financial product, wherein the sentiment index of the candidate financial product is used to characterize the sentiment direction of the evaluation of the candidate financial product;

[0011] Inputting the historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products into a generator of a recommendation model to obtain a recommended financial product output by the generator, wherein the recommended financial product is a part of the multiple candidate financial products;

[0012] Inputting the recommended financial product and the actually selected financial product into the discriminator of the recommendation model to obtain a discriminant result output by the discriminator;

[0013] A loss of the recommendation model is determined based on the recommended financial product and the discrimination result, and a parameter of the recommendation model is adjusted based on the loss.

[0014] In a third aspect, the present application provides a financial product recommendation device, comprising:

[0015] An acquisition module, used to acquire historical transaction records of users purchasing financial products, a search index of each of a plurality of candidate financial products, and a sentiment index of each of the candidate financial products, wherein the sentiment index of the candidate financial product is used to represent the sentiment direction of the evaluation of the candidate financial product;

[0016] An input module is used to input the historical transaction records, the search index of each candidate financial product, and the sentiment index of each candidate financial product into a pre-trained recommendation model to obtain a recommended financial product output by the recommendation model, wherein the recommended financial product is part of the multiple candidate financial products.

[0017] In a fourth aspect, the present application provides a financial product recommendation device, comprising:

[0018] an acquisition module, configured to acquire a plurality of training samples, each of which includes a historical transaction record of a user purchasing a financial product, a search index of each of a plurality of candidate financial products, a sentiment index of each of the candidate financial products, and a financial product actually selected, wherein the sentiment index of the candidate financial product is used to characterize the sentiment direction of the evaluation of the candidate financial product;

[0019] An input module, used to input the historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products into a generator of a recommendation model, to obtain a recommended financial product output by the generator, wherein the recommended financial product is a part of the multiple candidate financial products;

[0020] The input module is further used to input the recommended financial product and the actually selected financial product into the discriminator of the recommendation model to obtain a discrimination result output by the discriminator;

[0021] A determination module, configured to determine a loss of the recommendation model based on the recommended financial product and the discrimination result;

[0022] An adjustment module is used to adjust the parameters of the recommendation model based on the loss.

[0023] In a fifth aspect, an embodiment of the present application provides a financial product recommendation device, including: a memory, a processor;

[0024] The memory stores computer-executable instructions;

[0025] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementations of the first aspect.

[0026] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-readable storage medium is stored computer execution instructions, and when the computer execution instructions are executed by a processor, they are used to implement the first aspect above and / or various possible implementation methods of the first aspect.

[0027] In a seventh aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above first aspect and / or various possible implementation methods of the first aspect.

[0028] The financial product recommendation method, device, equipment, medium and product provided in the present application can obtain the candidate financial recommendation products output by the recommendation model by inputting the historical transaction records of users' purchase of financial products, the search index of each candidate financial product among multiple candidate financial products, and the sentiment index of each candidate financial product into a pre-trained recommendation model. This method solves the technical problem that it is difficult for recommenders to have a comprehensive and in-depth understanding and comparison of all financial products through an intelligent recommendation process, while avoiding the problem of inaccurate and unfair recommendation results caused by incomplete information or subjective bias, thereby improving the accuracy and fairness of the recommendation results, thereby increasing the user's success rate in purchasing financial products. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0030] Figure 1 Schematic diagram of the process of recommending financial products provided for this application Figure 1 ;

[0031] Figure 2 Model structure diagram of the recommended model provided for this application;

[0032] Figure 3 Schematic diagram of the process of recommending financial products provided for this application Figure 2 ;

[0033] Figure 4 The convolutional neural network provided for this application uses a parallel structure to link the hole convolution structure diagram;

[0034] Figure 5 Schematic diagram of the process of recommending financial products provided for this application Figure 3 ;

[0035] Figure 6 A diagram of the basic atomic structure of the grid network provided for this application;

[0036] Figure 7 A unit sequence structure diagram of the grid network provided for this application;

[0037] Figure 8 The basic framework structure diagram of the discriminator provided in this application;

[0038] Fig. 9 The overall process structure diagram of the recommendation method of the financial products provided for this application;

[0039] Fig.10 A flowchart for determining the sentiment index of a candidate financial product provided for this application;

[0040] Fig.11 A schematic diagram of the structure of the financial product recommendation device provided in this application;

[0041] Fig.12 A schematic diagram of the structure of the training device for the financial product provided in this application;

[0042] Fig.13 A schematic diagram of the structure of a configuration device for a network device provided in this application.

[0043] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0044] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0045] It should be noted that the financial product recommendation method, device, equipment, medium and product provided in this application can be used in the field of financial technology or artificial intelligence technology, and can also be used in any field outside the field of financial technology or artificial intelligence technology. This application does not limit the application field of the financial product recommendation method, device, equipment, medium and product.

[0046] With the continuous growth and prosperity of the global economy, various financial institutions have launched a variety of financial products to meet the increasingly diverse wealth management and investment needs of consumers and investors. These financial products cover a wide range of areas, from traditional savings and loans to complex funds, stocks, insurance, etc., injecting new vitality and opportunities into the market.

[0047] Under current technical conditions, the recommendation of financial products mainly relies on manual methods. Financial institutions usually arrange professional financial advisors or recommenders to communicate with customers face-to-face or online to understand their needs and risk tolerance, and then recommend the most suitable financial products to customers based on their own financial product knowledge base. This method requires recommenders to have not only deep knowledge of financial products, but also high customer insight and communication skills to ensure the accuracy and effectiveness of recommendations.

[0048] However, the existing financial product recommendation methods also have some obvious flaws. Due to the wide variety of financial products and their continuous updates, it is often difficult for recommenders to have a comprehensive and in-depth understanding and comparison of all products. At the same time, due to incomplete information or subjective bias, the recommendation results may be inaccurate and unfair. This will not only damage customer satisfaction and trust, but may also affect the reputation and market competitiveness of financial institutions.

[0049] In response to the above problems, the present application provides a method for recommending financial products. This method can obtain candidate financial recommendation products output by the recommendation model by inputting the historical transaction records of users purchasing financial products, the search index of each candidate financial product among multiple candidate financial products, and the sentiment index of each candidate financial product into a pre-trained recommendation model. This method solves the technical problem that it is difficult for recommenders to have a comprehensive and in-depth understanding and comparison of all financial products through an intelligent recommendation process, while avoiding the problem of inaccurate and unfair recommendation results caused by incomplete information or subjective bias, thereby improving the accuracy and fairness of the recommendation results, thereby increasing the user's success rate in purchasing financial products.

[0050] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0051] Figure 1 Schematic diagram of the process of recommending financial products provided for this application Figure 1 .like Figure 1 As shown, the method includes:

[0052] S101. Obtain historical transaction records of users purchasing financial products, search indexes of each candidate financial product among multiple candidate financial products, and sentiment indexes of each candidate financial product, wherein the sentiment index of the candidate financial product is used to represent the sentiment direction of the evaluation of the candidate financial product.

[0053] Among them, historical transaction records include but are not limited to: the number of transactions, transaction amount, transaction product type, highest product price, and lowest product price of multiple historical financial products.

[0054] Candidate financial products include, but are not limited to: stock funds, bond funds or money market funds.

[0055] The search index of a candidate financial product is used to characterize the frequency of searches for the candidate financial product by the user groups in the market. The higher the search index of a candidate financial product, the higher the frequency of searches for the candidate financial product by the user groups in the market. On the contrary, the lower the search index of a candidate financial product, the lower the frequency of searches for the candidate financial product by the user groups in the market.

[0056] The purpose of this step of obtaining the historical transaction records of the financial products purchased by the user is to analyze the user's purchasing behavior, preferences and investment trends based on the user's historical transaction records.

[0057] The purpose of obtaining the search index of each candidate financial product among multiple candidate financial products and the sentiment index of each candidate financial product in this step is to determine the popularity and sentiment satisfaction of the candidate financial product in the market.

[0058] Optionally, after obtaining the historical transaction records of the user purchasing the financial product, the present application will clean the historical transaction records of the financial product according to preset rules to obtain the historical transaction records after the cleaning process is completed.

[0059] Among them, the preset rules include but are not limited to: the holding period of historical financial products is short, and the transaction amount of historical financial products is small.

[0060] Understandably, the short holding period of historical financial products means that users sell financial products soon after purchasing them, resulting in a relatively short holding period of the financial products. This situation may reflect that users are not very satisfied with the financial products. For example, suppose a user purchases a fixed-term financial product with a term of 3 years, but chooses to redeem it after only holding it for half a year due to urgent need for funds. In this case, the holding period of the transaction record is relatively short, so the historical transaction records of this financial product will be cleaned up.

[0061] The low transaction amount of historical financial products refers to the transaction amount of the user's purchase of the financial product. If the transaction amount of the user is relatively small when purchasing a financial product, then such transaction records may be considered unimportant. This situation may reflect that the user is not willing to invest in the financial product. For example, suppose a user purchases a fund product, but the transaction amount is only 100 yuan, which is far below the minimum investment amount of the fund. Therefore, the historical transaction records of this financial product will be cleaned up.

[0062] Optionally, after obtaining the historical transaction records of users purchasing financial products and completing the cleaning process of the historical transaction records, the present application also needs to perform feature extraction on the historical transaction records of the financial products to obtain the historical transaction records after extraction.

[0063] The historical transaction records after feature extraction include but are not limited to: the number of transactions of historical financial products, the transaction amount, the type of transaction products, the highest price of the products, and the lowest price of the products.

[0064] Optionally, after the feature extraction of the transaction product type of historical financial products is completed, the present application also uses feature coding technology to feature encode the transaction product type in the historical transaction records, with the aim of converting the transaction product type into a numerical form, thereby facilitating the processing of the transaction product type.

[0065] The feature encoding technology includes but is not limited to: label encoding technology. For example, converting stock funds into 011, bond funds into 012, and money market funds into 013.

[0066] Optionally, the present application provides a method for obtaining a search index for each candidate financial product, including:

[0067] The first step is to determine multiple product types for each candidate financial product.

[0068] Among them, the product types of candidate financial products include but are not limited to: savings products, funds, bonds or insurance products.

[0069] The second step is to determine the keywords of the candidate financial products based on each type of candidate financial products.

[0070] Among them, different types of candidate financial products correspond to different keywords.

[0071] For example, if the type of the candidate financial product is a savings product, the keywords of the savings product include time deposit, demand deposit, notice deposit or lump sum deposit.

[0072] If the type of the candidate financial product is a fund, the keywords of the fund include stock fund, bond fund, hybrid fund, money fund or index fund.

[0073] If the type of the candidate financial product is a bond, the keywords for the bond include government bonds, corporate bonds, financial bonds or local bonds.

[0074] If the type of the candidate financial product is an insurance product, the keywords of the insurance product include life insurance, health insurance, accident insurance, automobile insurance or property insurance.

[0075] The third step is to determine the data collection tool and set the time range and geographical range on the data search tool.

[0076] Among them, the data search tool is used to provide a data source related to the search index of candidate financial products.

[0077] Time range refers to specifying a specific time period to collect data when using a data search tool for data collection. For example, if you need to collect the search results of user groups on stock funds in the first quarter of 2023 (January 1 to March 31), you can set this time range in the data collection tool to collect the search index of stock funds during this time period.

[0078] Geographic scope refers to specifying one or more specific geographic areas to collect data when using data search tools for data collection. For example, if you need to collect the search results of user groups in the Shandong province market for stock funds, you can set the geographic scope in the data collection tool to collect the search index of stock funds in that region.

[0079] The fourth step is to execute the third step according to the keywords of the candidate financial products using data crawling tools.

[0080] Among them, data crawling tools are used to assist data search tools in automatically extracting and collecting data.

[0081] For example, suppose the current task is to collect the search index of time deposits in savings products within a specific time period and geographical scope. Therefore, the system will first set the time range from January 1, 2023 to March 31, 2023 in the data search tool to capture the user group's search for time deposits in that quarter; at the same time, the geographical scope is set to Beijing to focus on analyzing the user's attention trend to time deposits in the Beijing market. Next, the data crawler will assist the data search tool to extract and collect search index information related to time deposits from existing data sources. In the end, it was found that the search index for time deposits was 20,000 within the specified time period and region.

[0082] S102, inputting historical transaction records, search indexes of each candidate financial product, and sentiment index of each candidate financial product into a pre-trained recommendation model to obtain a recommended financial product output by the recommendation model, where the recommended financial product is part of the multiple candidate financial products.

[0083] Among them, the model structure diagram of the recommended model is as follows: Figure 2 Actual data It is the historical transaction records, the search index of each candidate financial product, and the sentiment index of each candidate financial product By using real data Input Generator In the generator Output of recommended financial products ; Then the real data and generator Output of recommended financial products Input to the discriminator To judge the truth or falsity of these two types of data; finally, the Loss value output by the recommendation model is obtained, which is the recommended financial product output by the recommendation model.

[0084] The purpose of this step is to use the pre-trained recommendation model, combined with the user's historical transaction records, the search index and sentiment index of the candidate financial products, to screen out some recommended financial products from multiple candidate financial products.

[0085] Understandably, historical transaction records record in detail key information such as the number of transactions, transaction amount, transaction product type, product maximum price and product minimum price of each financial product, which helps the system to have a more comprehensive understanding of the characteristics and patterns of users in selecting financial products.

[0086] The search index reflects the frequency of searches for candidate financial products by user groups in the market, and is a key indicator for measuring the popularity and attention of the product market. The sentiment index represents the emotional direction of users' evaluation of candidate financial products, which helps the system understand the emotional tendencies of user groups in the market, such as satisfaction and trust in the product.

[0087] Therefore, by taking the user's historical transaction records, the search index and sentiment index of each candidate financial product as input and inputting them into the pre-trained recommendation model, the model can fully and deeply understand the user's purchasing behavior, preferences and investment trends, while also accurately grasping the market popularity of financial products on the market and user sentiment tendencies, and thus can output financial products that not only meet the user's personalized needs but also keep up with market hotspots and user sentiment tendencies, thereby helping the system provide users with more accurate product recommendations.

[0088] Optionally, after the recommendation model outputs the recommended financial product, the present application also needs to display the recommended financial product output by the model on a terminal device used by the recommender to help the recommender provide more accurate financial product recommendations to the user.

[0089] The financial product recommendation method provided in the embodiment of the present application first obtains the historical transaction records of users purchasing financial products, the search index of multiple candidate financial products, and the sentiment index of each candidate financial product (used to reflect the sentiment tendency of the evaluation of the candidate financial products). Then, this information is input into the pre-trained recommendation model, and the model will screen out some of the candidate financial products as recommended financial products based on these comprehensive data. This method effectively solves the problem that it is difficult for recommenders to have a comprehensive and in-depth understanding and comparison of all financial products, and that the recommendation results may be inaccurate and unfair due to incomplete information or subjective bias, thereby avoiding the risk of damaging customer satisfaction and reducing the success rate of financial product purchases. At the same time, the method also improves the objectivity and accuracy of the recommendation results, enhances the user experience, and thus brings significant commercial benefits and technological innovations to the financial industry.

[0090] Figure 3 Schematic diagram of the process of recommending financial products provided for this application Figure 2 ,like Figure 3 As shown, in this embodiment Figure 1 Based on the embodiment, a method for recommending financial products is described in detail, and the method includes:

[0091] S201. Obtain historical transaction records of users purchasing financial products, search indexes of each candidate financial product among multiple candidate financial products, and sentiment indexes of each candidate financial product, wherein the sentiment index of the candidate financial product is used to represent the sentiment direction of the evaluation of the candidate financial product.

[0092] The explanation of step S201 is similar to that of the above step S101 and will not be repeated here.

[0093] S202: For any candidate financial product, obtain multiple evaluation texts for the candidate financial product.

[0094] The purpose of this step is to collect multiple evaluation texts for any candidate financial product.

[0095] It is understandable that evaluation texts usually come from user feedback, market reviews or professional reviews, etc. They contain users' direct evaluations of the financial product's usage experience, functional effects, cost-effectiveness, etc. Therefore, by obtaining multiple evaluation texts of candidate financial products, we can have a more comprehensive understanding of the market reputation and user satisfaction of the candidate financial products.

[0096] Optionally, this application provides a possible implementation method, which specifically includes:

[0097] The first step is to obtain multiple initial evaluation texts for candidate financial products.

[0098] Among them, the initial evaluation text refers to the preliminary evaluation or comments written by users, market reviewers, professional evaluation agencies, etc. for a candidate financial product, reflecting its usage experience, functional effects, cost-effectiveness, etc.

[0099] The method of obtaining multiple initial review texts for the candidate financial products may be, for example, obtaining them through online forums, professional review websites, or social media. This application does not impose any special restrictions on this.

[0100] For example, if the candidate financial product is a money fund, the system will obtain multiple initial evaluation texts related to the money fund from online forums and professional evaluation websites.

[0101] In the second step, if the first initial evaluation text among the multiple initial evaluation texts contains the preset keywords, the first initial evaluation text is deleted, and if there is a repeated second initial evaluation text among the multiple initial evaluation texts, the second initial evaluation text is deleted to obtain multiple evaluation texts.

[0102] Among them, the preset keywords include but are not limited to: stock trading tutorial, adding WeChat or joining a group.

[0103] The purpose of this step is to screen multiple initial evaluation texts of candidate financial products to eliminate those meaningless initial evaluation texts.

[0104] For example, suppose that for money funds, the system currently obtains 50 initial evaluation texts about money funds. Next, the system will first filter according to the preset keywords. The system will check these 50 evaluation texts one by one to determine whether they contain these keywords. If these keywords appear in an evaluation text, the system will regard it as a meaningless evaluation text and remove it; conversely, if these keywords do not appear in an evaluation text, the system will retain it.

[0105] At the same time, the system will also perform deduplication processing on these 50 evaluation texts, that is, check whether there are duplicate evaluation texts. Once the same evaluation text is found, the system will directly delete it to avoid redundant information from interfering with subsequent analysis. In the end, the system obtained 40 evaluation texts that do not contain preset keywords and are not repeated.

[0106] S203: extract the word vector matrix of each evaluation text.

[0107] The purpose of this step is to convert each evaluation text into numerical form to facilitate subsequent sub-processing and analysis.

[0108] It can be understood that in the process of converting each evaluation text into a word vector, each word in the evaluation text will be mapped into a high-dimensional space to form a word vector. These word vectors are combined to form the word vector matrix of the evaluation text.

[0109] Optionally, this application provides a possible method for extracting a word vector matrix of each evaluation text, including:

[0110] The first step is to segment the evaluation text and delete the stop words in the segmented words to obtain the segmentation results.

[0111] The word segmentation result is to divide the evaluation text into independent and meaningful word units. The number of words included in the word segmentation result can be one or more.

[0112] The word segmentation method may be, for example, the Jieba word segmentation algorithm, or other word segmentation algorithms, which are not specifically limited in the present application.

[0113] The method of deleting the stop words in the words obtained by word segmentation may be, for example, to delete them using a stop word list or to delete them according to stop word rules.

[0114] It is understandable that when processing the evaluation text, the system first performs word segmentation on the evaluation text. Next, in order to further improve the quality of text processing, the system also needs to remove stop words from the word segmentation results. Finally, the system obtains the word segmentation results that do not include stop words.

[0115] For example, suppose the evaluation text for money funds includes "I bought a financial product - Tianlibao from a certain bank, and the recent returns are very stable. Based on the above information, the system will first use the Jieba word segmentation algorithm to segment the evaluation text and obtain multiple words, including "I", "at", "a certain bank", "purchased", "an item", "financial product", "Tianlibao", "recently", "returns", "very", "stable", etc.; then, the system will also remove stop words from these words according to the stop word list, removing the words "I", "at", "an item", and finally, the system obtains the segmentation results of the evaluation text, including "a certain bank", "purchase", "Tianlibao", "returns", and "stable".

[0116] In the second step, the word segmentation results are input into the word embedding model to obtain the initial word vector matrix of the evaluation text.

[0117] Among them, the word embedding model (word2vec) is a model for efficiently training word vectors. It can quickly and effectively express a word into a word vector representation through an optimized training model to obtain the semantic similarity between words. These word vectors form a word vector matrix , when using word vectors to represent text, use Indicates the text words, among which ,but It can be expressed as:

[0118]

[0119] in, Represents the number of word vectors, Represents the dimension of word vector. The word vector is trained based on the CBOW training model based on negative sampling, mainly through the word Context To determine the current word The sentiment is "negative", "positive" or "neutral".

[0120] First, the word segmentation results are used as input to the word embedding model; then, the word embedding model uses its internal processing algorithm to convert each word in the word segmentation results into a vector represented by a numerical value. These vectors together form an initial word vector matrix; finally, the word embedding model outputs the initial word vector matrix.

[0121] For example, suppose the evaluation text of a money fund includes "I bought a financial product - Tianlibao from a certain bank, and the recent returns are very stable. It is known that the word segmentation results of this evaluation text include "a certain bank", "purchase", "Tianlibao", "return" and "stable". Based on the above information, the system first takes these word segmentation results as input and inputs them into the word embedding model. Subsequently, the word embedding model converts "a certain bank" into a vector [0.1, 0.2, 0.3, 0.1, 0.2, 0.3], "purchase" into a vector [0.4, 0.5, 0.6, 0.7, 0.8, 0.9], and "Tianlibao" into a vector [0.3, 0.4, 0.5, 0.1, 0.2, 0 .3], “benefit” is converted into a vector [0.6, 0.7, 0.8, 0.1, 0.2, 0.3], and “stability” is converted into a vector [0.2, 0.3, 0.4, 0.1, 0.3, 0.4]. Finally, the word embedding model outputs the initial vector matrix corresponding to this review text {[0.1, 0.2, 0.3, 0.1, 0.2, 0.3], [0.4, 0.5, 0.6, 0.7, 0.8, 0.9], [0.3, 0.4, 0.5, 0.1, 0.2, 0.3], [0.6, 0.7, 0.8, 0.1, 0.2, 0.3], [0.2, 0.3, 0.4, 0.1, 0.3, 0.4]}.

[0122] The third step is to determine the word vector matrix of the evaluation text based on the initial word vector matrix and the preset matrix dimension.

[0123] The preset matrix dimension is used to indicate the target number of rows of the initial word vector matrix. For example, if the preset matrix dimension is 6, it indicates that the target number of rows of the initial word vector matrix is ​​6.

[0124] The purpose of this step is to convert the initial word vector matrices with different numbers of rows into word vector matrices with the same number of rows.

[0125] It is understandable that different evaluation texts will form different initial word vector matrices, and these matrices may have different numbers of rows. Therefore, in order to unify the dimensions of these matrices, the preset matrix dimensions can be used as the adjustment target, and the initial word vector matrices formed by different evaluation texts can be adjusted so that the number of rows meets the requirements of the preset matrix dimensions.

[0126] Optionally, the present application proposes a method for determining various situations that may exist in the word vector matrix of the evaluation text according to the initial word vector matrix and the preset matrix dimension, and provides corresponding implementation methods for these situations, including:

[0127] In the first case, when the dimension of the initial word vector matrix is ​​a preset matrix dimension, the initial word vector matrix is ​​determined as the word vector matrix of the evaluation text.

[0128] The purpose of this step is to directly determine the initial word vector matrix as the word vector matrix of the evaluation text while ensuring that the dimension of the initial word vector matrix matches the preset matrix dimension.

[0129] For example, assume that the preset matrix dimension is 6. It is known that the initial vector matrix of a certain evaluation text of a money fund is {[0.1, 0.2, 0.3, 0.1, 0.2, 0.3], [0.4, 0.5, 0.6, 0.7, 0.8, 0.9], [0.3, 0.4, 0.5, 0.1, 0.2, 0.3], [0.6, 0.7, 0.8, 0.1, 0.2, 0.3], [0.2, 0.3, 0.4, 0.1, 0.3, 0.4]}. Based on the above information, the system will directly use the initial vector matrix as the word vector matrix of the evaluation text.

[0130] In the second case, when the dimension of the initial word vector matrix is ​​lower than the preset matrix dimension, the keywords in the word segmentation result are determined, and the word vectors corresponding to the keywords are added to the initial word vector matrix as filling vectors to obtain the word vector matrix of the evaluation text.

[0131] It can be understood that when the dimension of the initial word vector matrix is ​​lower than the preset matrix dimension, it means that the number of rows in the initial word vector matrix is ​​less than the target number of rows in the initial word vector matrix. Therefore, the system will identify keywords from the word segmentation results of the evaluation text and find the word vectors corresponding to these keywords. Then, these word vectors are added as filling vectors to the initial word vector matrix until the number of rows in the initial word vector matrix reaches the target number of rows.

[0132] Optionally, the present application provides a possible implementation method for the second situation, including:

[0133] In the first step, the word frequency-inverse text frequency algorithm is used to determine the weight of each word in the word segmentation result.

[0134] The basic idea of ​​the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm is that the more frequently a word appears in a document, the more important it is to the document. If a word appears in many documents, it may be a common word and contribute less to distinguishing a document. It solves the problem that the weight of a word is determined only by its frequency, and can be used to evaluate the importance of a word in a document.

[0135] Among them, term frequency (TF) indicates the frequency of a word appearing in a document. The calculation formula of term frequency is:

[0136]

[0137] Among them, the molecule Expressing words In the documentation The number of times it appears in the denominator Representation Document The sum of the number of all words in the document. The higher the frequency of a word in a document, the more important it is to the document, and the larger the corresponding TF value. For example, in a review text containing 200 words, it is known that the word segmentation result of the review text includes "income", and it is known that "income" appears 10 times in this initial review text. Based on the above information, the word frequency of "income" can be calculated to be 10 / 200=0.05.

[0138] Inverse document frequency (IDF) is mainly used to correct the importance of words. The calculation formula of inverse document frequency is:

[0139]

[0140] Among them, the molecule Represents the total number of documents, the denominator Indicates that a word has appeared If the word appears in a lot of documents, it may be a common word, which contributes less to distinguishing a document, and the corresponding IDF value is smaller. For example, it is known that the number of evaluation texts obtained for money funds is 10,000, and it is known that "income" appears in 100 documents. Based on the above information, the system will first determine that the total number of documents is 10,000, and then calculate the inverse document frequency of "income" log(10000 / 100)=log(100)=2.

[0141] Term frequency-inverse document frequency (TF-IDF) is the product of term frequency and inverse document frequency. The calculation formula for term frequency-inverse document frequency is:

[0142]

[0143] The higher the term frequency-inverse document frequency, the more important the term is to the document. To prevent the denominator from being 0, an extra 1 is added to the IDF denominator. For example, the term frequency of "income" is 0.05 and the inverse document frequency is 2. Based on the above information, we can determine that the term frequency-inverse document frequency of "income" is 0.05*2=0.1.

[0144] Term frequency-inverse document frequency is obtained by calculating the term frequency (TF) and inverse document frequency (IDF) of a word in a document, which reflects the relative importance of a word in a document. However, due to differences in document length and vocabulary, the original term frequency-inverse document frequency may have different dimensions and ranges, which makes it difficult to directly compare different words in the same document. Therefore, in order to eliminate the impact of this dimension and range and make the word importance index more intuitive and comparable, it is necessary to use the following calculation formula to normalize the term frequency-inverse document frequency. The purpose of normalization is to normalize each word. The term frequency-inverse document frequency is scaled to a uniform range, thereby obtaining the weight of the term frequency-inverse document frequency of each word, which is the normalized term frequency-inverse document frequency.

[0145]

[0146] Among them, the molecule Expressing words In the documentation Term frequency - inverse document frequency, denominator Representation Document The sum of the term frequency-inverse document frequency of all words in , For Documentation The total number of different words contained in .

[0147] For example, it is known that in a 200-word evaluation text, the word segmentation results include "a certain line", "purchase", "Tianlibao", "income" and "stable". At the same time, it is known that the word frequency inverse document frequency of each of these five words is 0.1 for "a certain line", 0.3 for "purchase", 0.5 for "Tianlibao", 0.1 for "income" and 0.3 for "stable". Based on the above information, the system can first calculate the sum of the word frequency-inverse document frequency of all words as 0.1+0.3+0.5+0.1+0.3=1.3. Subsequently, the system continues to normalize these five words and obtains the weight of each word, which is 0.077 for "a certain line", 0.231 for "purchase", 0.3846 for "Tianlibao", 0.077 for "income" and 0.231 for "stable".

[0148] The second step is to determine the keywords from the word segmentation results based on the weight of each word in the word segmentation results.

[0149] Among them, the weight is used to represent the importance of the word in the evaluation text. The higher the weight of the word, the more important the word is in the evaluation text. The lower the weight of the word, the less important the word is in the evaluation text.

[0150] The purpose of this step is to extract important words from the word segmentation results and use the extracted words as keywords.

[0151] For example, assume that the preset matrix dimension is 6. It is known that the initial vector matrix of a certain evaluation text of a money fund is {[0.1, 0.2, 0.3, 0.1, 0.2, 0.3], [0.4, 0.5, 0.6, 0.7, 0.8, 0.9], [0.3, 0.4, 0.5, 0.1, 0.2, 0.3], [0.6, 0.7, 0.8, 0.1, 0.2, 0.3], [0.2, 0.3, 0.4, 0.1, 0.3, 0.4]}. And it is known that the word segmentation results of this evaluation text include "a certain line", "buy", "Tianlibao", "income" and "stable". Based on the above information, the system first calculates the weights corresponding to these word segmentation results, which are "a certain line" for 0.077, "purchase" for 0.231, "Tianlibao" for 0.3846, "income" for 0.077 and "stable" for 0.231; then, the system will extract "Tianlibao" according to the weights corresponding to these word segmentation results and use it as a keyword.

[0152] The third step is to use the word vectors corresponding to the keywords as filling vectors to fill in the initial word vector matrix to obtain the word vector matrix of the evaluation text.

[0153] The purpose of this step is to integrate the word vectors corresponding to the keywords into the initial word vector matrix, so as to construct a complete evaluation text word vector matrix.

[0154] For example, assume that the preset matrix dimension is 6. It is known that the initial vector matrix of a certain evaluation text of a money fund is {[0.1, 0.2, 0.3, 0.1, 0.2, 0.3], [0.4, 0.5, 0.6, 0.7, 0.8, 0.9], [0.3, 0.4, 0.5, 0.1, 0.2, 0.3], [0.6, 0.7, 0.8, 0.1, 0.2, 0.3], [0.2, 0.3, 0.4, 0.1, 0.3, 0.4]}. And it is known that "Tianlibao" is a keyword. Based on the above information, the system directly uses the word vector [0.3, 0.4, 0.5, 0.1, 0.2, 0.3] corresponding to "Tianlibao" as the filling vector and fills it into the initial word vector matrix, thereby obtaining the word vector matrix of the evaluation text {[0.1, 0.2, 0.3, 0.1, 0.2, 0.3], [0.4, 0.5, 0.6, 0.7, 0.8, 0.9], [0.3, 0.4, 0.5, 0.1, 0.2, 0.3], [0.6, 0.7, 0.8, 0.1, 0.2, 0.3], [0.2, 0.3, 0.4, 0.1, 0.3, 0.4], [0.3, 0.4, 0.5, 0.1, 0.2, 0.3]}.

[0155] In the third case, when the dimension of the initial word vector matrix is ​​higher than the preset matrix dimension, the initial word vector matrix is ​​truncated according to the preset matrix dimension to obtain the word vector matrix of the evaluation text.

[0156] The truncation rule may be, for example, to truncate rows one by one from the beginning of the initial word vector matrix according to a preset matrix dimension until the number of rows corresponding to the preset matrix dimension is met. This application does not impose any special restrictions on this.

[0157] The purpose of this step is to truncate the initial word vector matrix when the dimension of the initial word vector matrix is ​​higher than the preset matrix dimension, so as to ensure that the obtained evaluation text word vector matrix meets the expected dimension requirements.

[0158] It can be understood that when the dimension of the initial word vector matrix is ​​lower than the preset matrix dimension, it means that the number of rows of the initial word vector matrix is ​​higher than the target number of rows of the initial word vector matrix. Therefore, the system will intercept the initial word vector matrix according to the preset matrix dimension to obtain a word vector matrix that meets the system expectations.

[0159] For example, assume that the preset matrix dimension is 5. It is known that the initial vector matrix of a certain evaluation text of a stock fund is {[0.2, 0.23, 0.13, 0.11, 0.2, 0.3], [0.14, 0.54, 0.63, 0.17, 0.81, 0.19], [0.13, 0.14, 0.15, 0.41, 0.32, 0.33], [0.16, 0.71, 0.18, 0.61, 0.12, 0.13], [0.12, 0.13, 0.14, 0.11, 0.13, 0.14], [0.54, 0.44, 0.33, 0.17, 0.81, 0.19]}. Based on the above information, the system can directly extract the first 5 rows as the word vector matrix of the evaluation text {[0.2, 0.23, 0.13, 0.11, 0.2, 0.3], [0.14, 0.54, 0.63, 0.17, 0.81, 0.19], [0.13, 0.14, 0.15, 0.41, 0.32, 0.33], [0.16, 0.71, 0.18, 0.61, 0.12, 0.13], [0.12, 0.13, 0.14, 0.11, 0.13, 0.14]}.

[0160] S204: Determine the sentiment index of the candidate financial product according to the word vector matrices of the multiple evaluation texts.

[0161] The purpose of this step is to comprehensively calculate the sentiment index of the candidate financial products based on the word vector matrices converted from multiple evaluation texts.

[0162] It is understandable that since each candidate financial product is accompanied by multiple evaluation texts, these texts represent the sentiment inclinations of different users towards the financial product. In order to fully grasp the overall sentiment inclination of the user groups in the market towards the candidate financial products, each evaluation text can be converted into a corresponding word vector matrix, and these matrices are used as the target of overall analysis to obtain the overall sentiment inclination of the user groups towards the candidate financial products.

[0163] Optionally, the present application provides a possible method for determining the sentiment index of a candidate financial product based on the word vector matrices of each of a plurality of evaluation texts, including:

[0164] In the first step, the word vector matrices of multiple evaluation texts are input into the long short-term memory network to obtain the contextual features of the multiple evaluation texts.

[0165] Among them, the Long Short-Term Memory (LSTM) network is a time-recurrent neural network. The LSTM network contains multiple LSTM units, each of which is composed of three middle gate units. They are respectively introduced to forget gate , Input Gate and output gate When the LSTM unit passes information to the next unit, the forget gate of the next unit can process the received information and delete useless information. Adjust the unit to make it present a new state, with an output gate Responsible for passing the new information to the next LSTM unit.

[0166] In the long short-term memory network, the sigmoid function of the forget gate controls the degree of retention of historical information:

[0167]

[0168] in, is the weight matrix, The output of the previous time node of the hidden layer; for The output of the input layer at each moment; is the bias vector.

[0169] Will Output at any time and the eigenvectors added to At this moment, in the LSTM unit, the two states are merged to update the current state:

[0170]

[0171] in, Determine the value that needs to be updated; for The temporary state at the moment, including new candidate values; for The updated status at that moment.

[0172] The sigmoid function of the output gate determines the latest output information, and the tanh function is used to obtain a value between -1 and 1, which is then multiplied by the output value to determine the final output value:

[0173]

[0174] Through the above steps, we get the output of each step ,in, is the number of LSTM neurons; secondly, As the input of the convolutional neural network.

[0175] It is understandable that the LSTM network can make full use of the temporal information in the evaluation text, understand the relationship between words and the semantic structure of the entire sentence. Therefore, by inputting the word vector matrices of multiple evaluation texts into the LSTM network, each evaluation text can be converted into a feature representation rich in contextual information. This feature representation not only captures the key words in the evaluation text, but also incorporates the usage and meaning of the words in a specific context, thereby providing comprehensive and accurate information support for subsequent text analysis.

[0176] In the second step, the context feature information of multiple evaluation texts is input into the convolutional neural network to obtain the semantic features of the multiple evaluation texts, wherein the convolutional neural network adopts void convolution.

[0177] Among them, Convolutional Neural Networks (CNN) is a type of feedforward neural network that includes convolution calculations and has a deep structure. Convolutional neural networks include: convolutional layer, pooling layer and fully connected layer.

[0178] For the convolutional layer of the convolutional neural network, the traditional larger convolution kernel is first replaced with a smaller 3×3 convolution kernel. Then, the dilated convolution is introduced, and the convolution layer is constructed using a parallel structure. Figure 4 As shown, Figure 4 The convolutional neural network uses a parallel structure to connect the dilated convolution structure diagram. C represents ordinary convolution, and DC represents dilated convolution. When the dilation rate of the dilated convolution is 1, it degenerates into an ordinary convolutional neural network.

[0179] By adopting a smaller 3×3 convolution kernel, not only can the number of parameters and the amount of calculation be reduced, but also the prediction accuracy of the convolutional neural network can be maintained to a certain extent. At the same time, by adopting a parallel approach, the number of layers of the convolutional neural network can be gradually increased. This design enables the later hole convolution kernels to have a larger receptive field when corresponding to the first layer of the input data. The increase in receptive field means that the convolutional neural network can capture broader and more global information. Therefore, in the last few layers of the convolutional neural network, it can effectively extract high-order features of the entire text segment, which are crucial for data prediction.

[0180] The processing process of the convolutional neural network: First, the context features of multiple evaluation texts output by the long short-term memory network are input into the convolutional layer of the convolutional neural network. The convolutional layer uses the following calculation formula to perform convolution operations on the input data to extract the most important semantic features:

[0181]

[0182] in, represents the feature matrix; A matrix representing the output of the input layer; represents the weight matrix; Represents the bias vector.

[0183] Subsequently, the pooling layer uses the following calculation formula to perform maximum pooling on the data output by the convolutional layer to reduce overfitting by selecting The maximum value is used as the semantic information of the filter:

[0184]

[0185] in, Indicates the length of the sentence vector; Indicates the size of the convolution window.

[0186] Finally, in the fully connected layer, and Perform dot product operations and output the semantic features of multiple evaluation texts.

[0187] As you can understand, convolutional neural networks have powerful feature extraction capabilities and can automatically learn and extract key information from texts, such as words, phrases, and the relationships between them. Therefore, by inputting the contextual feature information of multiple evaluation texts, convolutional neural networks can learn the commonalities and differences between different evaluation texts, and then generate more representative semantic features to provide data support for subsequent sentiment tendency judgment tasks.

[0188] The third step is to perform sentiment classification based on the semantic features of multiple evaluation texts to obtain the positive sentiment index and negative sentiment index of the candidate financial products.

[0189] Among them, this application gives a calculation formula for calculating emotion classification using the softmax classification function:

[0190]

[0191] in, Indicates The output value of each node, Indicates the number of output nodes.

[0192] By using the softmax classification function, the semantic features of multiple evaluation texts can be converted into a probability distribution ranging between [0, 1] and with a sum of 1.

[0193] In the process of using the softmax classification function for sentiment classification, the softmax classification function can produce two different situations. The first situation is to obtain the positive sentiment index and the negative sentiment index, the sum of which is 1, representing the probabilities of positive sentiment and negative sentiment of the candidate financial product, respectively. The second situation is to obtain the positive sentiment index, the intermediate sentiment index, and the negative sentiment index, the sum of which is 1, representing the probabilities of positive sentiment, intermediate sentiment, and negative sentiment of the candidate financial product, respectively.

[0194] The fourth step is to determine the sentiment index of the candidate financial products based on the positive sentiment index and the negative sentiment index.

[0195] The purpose of this step is to determine the overall sentiment of the user groups in the market towards the candidate financial products.

[0196] It can be understood that the positive sentiment index and negative sentiment index represent the degree of positive and negative evaluation of the candidate financial products by the user groups in the market, respectively. Therefore, through certain rules and calculations, these two indexes can be converted into a single sentiment index, which can intuitively reflect the overall sentiment tendency of the user groups in the market towards the candidate financial products.

[0197] Optionally, the present application provides a calculation formula for determining the sentiment index of a candidate financial product based on the positive sentiment index and the negative sentiment index:

[0198]

[0199] in, represents the sentiment index of the candidate financial product, represents the positive sentiment index of the candidate financial product, Represents the negative sentiment index of the candidate financial product.

[0200] If the positive sentiment index is higher than the negative sentiment index, then the sentiment index of the candidate financial product indicates that the user group in the market has a positive attitude towards the candidate financial product; conversely, if the negative sentiment index is higher than the positive sentiment index, then the sentiment index of the candidate financial product will be a negative number or a smaller positive number, indicating that the user group in the market has a negative attitude towards the candidate financial product.

[0201] For example, suppose the positive sentiment index corresponding to a review text for a stock fund is The negative sentiment index is 0.0465. is 0.9523, then the above calculation formula can be used to determine the sentiment index of the stock fund in this evaluation text. It is -0.9085.

[0202] S205. Input the historical transaction records, the search index of each candidate financial product, and the sentiment index of each candidate financial product into the generator of the recommendation model to obtain the recommended financial product output by the generator, wherein the recommendation model includes a generator and a discriminator, and the discriminator is used to determine whether the recommended financial product output by the generator is a real selected financial product during the training stage of the recommendation model.

[0203] The purpose of this step is to uniformly input historical transaction records, search indexes of candidate financial products, and sentiment indexes into the generator of the recommendation model, so that the generator can output one or more recommended financial products based on these data and using its own execution mechanism.

[0204] In this step, for example, the historical transaction records, the search index of each candidate financial product, and the sentiment index of each candidate financial product can be first formed into seven-tuple data; then, the seven-tuple data is normalized to obtain processed seven-tuple data; then, the processed seven-tuple data is used as the input of the recommendation model and input into the generator of the recommendation model; further, the generator of the recommendation model uses the internal execution mechanism to output one or more recommended financial products; then, the discriminator of the recommendation model judges the one or more recommended financial products output by the generator to determine whether the recommended financial product is a truly selected financial product; finally, the recommendation model outputs one or more recommended financial products based on the judgment result of the discriminator.

[0205] Optionally, the present application may, for example, use the following calculation formula to normalize the seven-tuple data. The following calculation formula may be used to map each element in the seven-tuple data into a unified numerical range, thereby eliminating the dimensional differences between different data elements and improving the comparability of the data and the training efficiency of the model.

[0206]

[0207] in, Represents the normalized seven-tuple data, Represents the seven-tuple data before normalization, represents the minimum value before normalization, Indicates the maximum value before normalization.

[0208] The financial product recommendation method provided in the embodiment of the present application first obtains the historical transaction records of the user's purchase of financial products, and simultaneously obtains the search index of multiple candidate financial products. For each candidate financial product, multiple relevant evaluation texts are further collected, and the word vector matrix of each text is extracted. Based on these word vector matrices, the sentiment index of the candidate financial product is comprehensively calculated, and the index can reflect the user's emotional tendency to evaluate the candidate financial product. Subsequently, this information (historical transaction records, search index and sentiment index) is input into the generator of the recommendation model, and the generator intelligently selects some candidate financial products as recommendation results based on this information. This method effectively solves the technical problem that it is difficult for recommenders to have a comprehensive and in-depth understanding and comparison of all financial products, and avoids the problem of inaccurate and unfair recommendation results caused by incomplete information or subjective bias. At the same time, this method not only significantly improves the accuracy and fairness of the recommendation results, but also enhances customer satisfaction, thereby improving the purchase success rate of financial products.

[0209] Figure 5 Schematic diagram of the training method of the recommendation model for financial products provided in this application Figure 3 .like Figure 5 As shown, the method includes:

[0210] S301. Acquire multiple training samples, each training sample including historical transaction records of users purchasing financial products, search indexes of each candidate financial product among multiple candidate financial products, sentiment indexes of each candidate financial product, and actually selected financial products, wherein the sentiment index of the candidate financial product is used to characterize the sentiment direction of the evaluation received by the candidate financial product.

[0211] It is understandable that the training samples not only reflect the preferences and habits of individual users in selecting financial products under different situations and conditions, but also reflect the search frequency and emotional evaluation of multiple candidate financial products by a large number of user groups in the market. These data provide rich and comprehensive input information for the recommendation model.

[0212] Therefore, by obtaining multiple training samples, the recommendation model can understand user needs more accurately and grasp market dynamics, thereby generating more personalized financial product recommendations that meet user expectations.

[0213] S302: Input historical transaction records, search indexes of candidate financial products, and sentiment indexes of candidate financial products into a generator of a recommendation model to obtain recommended financial products output by the generator, where the recommended financial products are part of the multiple candidate financial products.

[0214] Among them, the role of the generator of the recommendation model is to generate candidate data that is similar to the real data.

[0215] The generator of the recommendation model includes: grid network, gradient descent algorithm and loss function. The grid network is used to build the basic framework of the generator, responsible for building the structure and hierarchy of the generator; the gradient descent algorithm is used to iteratively adjust the parameter values ​​in the grid network to minimize a certain objective function, so that the generator gradually approaches the optimal solution; the loss function is a function that measures the difference between the model's predicted results and the actual results, and is the goal of the gradient descent algorithm optimization.

[0216] Among them, (1) TrellisNet is a special temporal convolutional network. A complete TrellisNet can be composed of multiple basic atomic structures tiled in the time dimension and stacked in the space dimension, combining the given time series data with its hidden output, and each layer uses the same weights.

[0217] (2) The calculation process of the basic atomic structure of the grid network is as follows: and At time step , No. Layer, after the activation function The hidden layer output after The calculation process is as follows:

[0218]

[0219]

[0220] in, Represents at time When input, Representative Layer in time The hidden layer input is Represents at time When input, Representative Layer in time The hidden layer input at time .

[0221] (3) The basic atomic structure diagram of the grid network, such as Figure 6 shown.

[0222] (4) The unit sequence structure diagram of the grid network, such as Figure 7 shown. Figure 7 It is represented as the data of time series length 8 after the Layer and The hidden layer output after the layer.

[0223] (5) The loss function of the generator is expressed as The loss function is as follows:

[0224]

[0225] in, To measure the generated data The difference between the real data. Represents input data, Represents the data generated by the generator based on the input data.

[0226] (6) The generator uses the gradient descent algorithm of the adaptive moment estimation algorithm to implement the iteration of network parameters. The adaptive moment estimation algorithm (Adam, AdaptiveMomentEstimation) has the characteristics of adaptive learning rate and second-order moment estimation, which can accelerate the convergence of the model and improve the training effect. The update formula of the adaptive moment estimation algorithm is calculated as follows:

[0227] set up ,in is the loss function, Relative to the model parameters Gradient, for The current iteration number.

[0228] First, we calculate the first moment (mean) estimate of the gradient and the second moment (uncentered variance) estimate :

[0229]

[0230]

[0231] Then, bias correction is performed to obtain and :

[0232]

[0233]

[0234] in, and is the exponential decay rate of the control moment estimate, which is set to 0.9; Represents element-wise multiplication. express Multiply by itself, that is (Multiply by times); similarly express Self-driving Second-rate.

[0235] Finally, update the model parameters :

[0236]

[0237] in, represents the learning rate; Represents a very small number (e.g. ), which is used to prevent division by zero.

[0238] In this formula, and are estimates of the first and second moments of the gradient, respectively, which are calculated by exponential moving average. Bias correction is used to compensate for the initial and The estimation error caused by initialization from 0. As increases, the influence of bias correction gradually decreases.

[0239] By adaptively adjusting the learning rate, the convergence speed has been greatly improved, and the convergence effect is also relatively ideal, achieving the effect of shortening the training time of the recommendation model.

[0240] The purpose of this step is to use the generator of the recommendation model to comprehensively consider multi-dimensional information such as historical transaction records, search index of each candidate financial product, and sentiment index to screen out some candidate financial products for recommendation.

[0241] Understandably, the system first collects the user's historical transaction records, which reflects the user's purchasing behavior and preferences; at the same time, it also considers the search index of each candidate financial product on the market, which represents the current market attention and trend; in addition, it also incorporates the user's sentiment index of the product, which reflects the user's satisfaction and reputation with the product. Then, these multi-dimensional data are input into the generator of the recommendation model, and the generator selects some candidate financial products as the generator's recommendation results through the internal algorithm mechanism.

[0242] S303: Input the recommended financial products and the actually selected financial products into the discriminator of the recommendation model to obtain the discrimination result output by the discriminator.

[0243] Among them, the role of the discriminator is to identify whether the input data is real data or data generated by the generator, which is equivalent to a classifier. If the input data is real data, the output is 1, otherwise the output is 0.

[0244] The discriminator of the recommendation model includes: convolutional neural network, gradient descent algorithm and loss function. Convolutional neural network is used to build the basic framework of the discriminator, responsible for building the structure and hierarchy of the discriminator; the gradient descent algorithm is used to iteratively adjust the parameter values ​​in the grid network to minimize a certain objective function, so that the discriminator gradually approaches the optimal solution; the loss function is a function that measures the difference between the model's predicted results and the actual results, and is the goal of the gradient descent algorithm optimization.

[0245] Among them, (1) the basic framework of the discriminator, such as Figure 8 As shown in the figure. The basic framework of the discriminator includes: input layer, convolution layer, pooling layer, fully connected layer and output layer. The input of the discriminator includes real data and data generated by the generator. The convolution layer is composed of multiple layers of one-dimensional convolutional neural network. The activation function of the convolution layer adopts ReLU function. The output layer is a fully connected layer. The activation function is Sigmoid function. It distinguishes generated data from real data by outputting 0 and 1.

[0246] (2) The loss function of the discriminator is expressed as The difference between . The loss function is as follows:

[0247]

[0248] in, It is used to measure the difference between the generator's prediction and the actual result. Represents input data, Represents the probability that the input data belongs to the true data distribution.

[0249] (3) The discriminator and the generator use the same adaptive matrix estimation algorithm to implement the iteration of network parameters.

[0250] The purpose of this step is to use the discriminator of the recommendation model to evaluate the degree of match between the recommended financial products and the actual selected financial products.

[0251] It can be understood that the recommended financial products generated by the recommendation model and the financial products actually selected by the user are simultaneously input into the discriminator. As a classifier, the discriminator can output a discrimination result based on the input data. This result is usually a probability value, which indicates the similarity or matching degree between the recommended financial products generated by the generator and the financial products actually selected by the user.

[0252] S304: Determine the loss of the recommendation model based on the recommended financial product and the discrimination result, and adjust the parameters of the recommendation model based on the loss.

[0253] Among them, this application gives a function expression for determining the loss of the recommendation model:

[0254]

[0255] in, represents the predicted probability of the discriminator for the recommended financial products generated by the generator, Represents the recommended financial products generated by the generator.

[0256] It can be understood that by calculating the loss of the recommendation model between a given recommended financial product and the discrimination result, the accuracy of the recommendation prediction can be quantified, and the model parameters can be adjusted accordingly, thereby gradually improving the accuracy of the recommendation and user satisfaction.

[0257] Optionally, after the recommendation model training is completed, in order to evaluate the prediction performance of the recommendation model for financial products, the present application provides the following two methods to evaluate the prediction performance of the recommendation model, so as to ensure the stability and reliability of the recommendation model in practical applications.

[0258] The first method is to use the mean absolute percentage error algorithm to evaluate the predictive performance of the recommendation model.

[0259] The calculation formula for mean absolute percentage error (MAPE) is as follows:

[0260]

[0261] in, Indicates the recommended value of financial products, Represents the true value of the financial product, and the range of the mean absolute percentage error is , the closer the value is to 0, the closer the predicted value is to the true value, and the better the effect of the recommendation model.

[0262] The second method uses the root mean square error algorithm,s, to evaluate the prediction performance of the recommendation model.

[0263] The calculation formula of the root mean square error (RMSE) is as follows:

[0264]

[0265] in, Indicates the recommended value of financial products, Represents the true value of the financial product, and the range of the mean absolute percentage error is , the closer the value is to 0, the closer the predicted value is to the true value, and the better the effect of the recommendation model.

[0266] The training method of the recommendation model of financial products provided in the embodiment of the present application first obtains multiple training samples, each sample covers the historical transaction records of users purchasing financial products, the search index and sentiment index of multiple candidate financial products, and the financial products actually selected by the user. Then, this information is input into the generator of the recommendation model, and the generator selects and outputs part of the candidate financial products as recommended financial products. Then, the recommended financial products and the financial products actually selected by the user are input into the discriminator of the recommendation model, and the discriminator outputs a discrimination result. Finally, the loss of the recommendation model is determined based on the recommended financial products and the discrimination results, and the parameters of the recommendation model are adjusted accordingly to optimize its performance. This method can more comprehensively evaluate the attractiveness and user preferences of candidate financial products by comprehensively considering the historical transaction records of users, the search index and sentiment index of candidate financial products, thereby improving the accuracy and user satisfaction of the recommendation model. At the same time, through the introduction of the discriminator and the calculation of the loss function, the parameter adjustment of the recommendation model can be more effectively guided, the optimization process of the model can be accelerated, and finally more accurate financial product recommendations can be achieved.

[0267] Fig. 9 The overall process structure diagram of the financial product recommendation method provided in this application. Fig. 9 As shown in the figure, the usage process of the recommendation model is explained, including:

[0268] First, the historical transaction records of users purchasing financial products, the evaluation texts of multiple candidate financial products, and the search indexes of multiple candidate financial products are obtained; then, the sentiment indexes of the candidate financial products are determined using long-short term networks and convolutional neural networks for the evaluation texts of the candidate financial products. At the same time, the historical transaction records of users purchasing financial products are feature extracted to obtain the extracted historical transaction records, including the number of transactions, transaction amount, transaction product type, highest price, and lowest price of the historical financial products; then, the extracted historical transaction records, the sentiment indexes of multiple candidate financial products, and the search indexes of multiple candidate financial products are data fused and normalized; further, the normalized historical transaction records, the sentiment indexes of multiple candidate financial products, and the search indexes of multiple candidate financial products are input into the recommendation model together, the generator of the recommendation model outputs the recommended financial products, and the discriminator of the recommendation model discriminates the recommended financial products output by the generator and the actually selected financial products; finally, the recommended financial products output by the recommendation model are obtained.

[0269] Fig.10 The process structure diagram for determining the sentiment index of a candidate financial product provided in this application. Fig.10 The figure shows the processing flow of the sentiment index of candidate financial products, including:

[0270] Firstly, for multiple evaluation texts of candidate financial products, the word embedding model (word2vec) is used to extract the word vector matrix of each evaluation text; then, the word vector matrices of the multiple evaluation texts are input into the long short-term network (LSTM) to obtain the contextual features of the multiple evaluation texts; then, the contextual feature information of the multiple evaluation texts is input into the convolutional neural network (CNN), and the semantic features of the multiple evaluation texts are obtained through the processing of the maximum pooling layer and the fully connected layer in the convolutional neural network; then, the softmax classification is used to perform sentiment classification on the semantic features of the multiple evaluation texts, and the positive sentiment index and negative sentiment index output by the softmax classification are obtained; finally, the sentiment index of the candidate financial product is determined according to the positive sentiment index and the negative sentiment index.

[0271] Fig.11 A schematic diagram of the structure of the financial product recommendation device provided in this application, such as Fig.11 As shown, the financial product recommendation device 400 provided in this embodiment includes:

[0272] The acquisition module 401 is used to acquire the historical transaction records of the user purchasing the financial product, the search index of each of the candidate financial products among the multiple candidate financial products, and the sentiment index of each of the candidate financial products, wherein the sentiment index of the candidate financial product is used to represent the sentiment direction of the evaluation of the candidate financial product;

[0273] The input module 402 is used to input the historical transaction records, the search index of each candidate financial product, and the sentiment index of each candidate financial product into a pre-trained recommendation model to obtain a recommended financial product output by the recommendation model, where the recommended financial product is part of the multiple candidate financial products.

[0274] The financial product recommendation device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and this embodiment will not be described in detail here.

[0275] Fig.12 A schematic diagram of the structure of the training device for the financial product provided in this application, such as Fig.12 As shown, the financial product recommendation device 500 provided in this embodiment includes:

[0276] An acquisition module 501 is used to acquire a plurality of training samples, each of which includes a historical transaction record of a user purchasing a financial product, a search index of each of a plurality of candidate financial products, a sentiment index of each of the candidate financial products, and a financial product actually selected, wherein the sentiment index of the candidate financial product is used to represent the sentiment direction of the evaluation of the candidate financial product;

[0277] An input module 502 is used to input the historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products into a generator of a recommendation model to obtain a recommended financial product output by the generator, wherein the recommended financial product is a part of the multiple candidate financial products;

[0278] The input module 502 is further used to input the recommended financial product and the actually selected financial product into the discriminator of the recommendation model to obtain a discrimination result output by the discriminator;

[0279] A determination module 503, configured to determine the loss of the recommendation model based on the recommended financial product and the discrimination result;

[0280] The adjustment module 504 is used to adjust the parameters of the recommendation model based on the loss.

[0281] The financial product training device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and this embodiment will not be described in detail here.

[0282] Fig.13 This is a schematic diagram of the configuration device of the network device provided in this application. Fig.13 As shown, the electronic device 600 provided in this embodiment includes: at least one processor 601 and a memory 602. Optionally, the device 600 also includes a communication component 603. The processor 601, the memory 602 and the communication component 603 are connected via a bus 604.

[0283] In a specific implementation process, at least one processor 601 executes the computer execution instructions stored in the memory 602, so that at least one processor 601 executes the above method.

[0284] The specific implementation process of the processor 601 can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0285] In the above embodiments, it should be understood that the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the invention can be directly implemented as a hardware processor, or can be implemented by a combination of hardware and software modules in the processor.

[0286] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (NVM), such as at least one disk storage.

[0287] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of this application is not limited to only one bus or one type of bus.

[0288] The present application also provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0289] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above method is implemented.

[0290] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0291] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). Of course, the processor and the readable storage medium can also exist in the device as discrete components.

[0292] The division of units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0293] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0294] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0295] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0296] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0297] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0298] It should be further noted that, although the various steps in the flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flow chart may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0299] It should be understood that the above-mentioned device embodiments are only illustrative, and the device of the present application can also be implemented in other ways. For example, the division of units / modules in the above-mentioned embodiments is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units, modules or components can be combined, or can be integrated into another system, or some features can be ignored or not executed.

[0300] In addition, unless otherwise specified, each functional unit / module in each embodiment of the present application may be integrated into one unit / module, each unit / module may exist physically separately, or two or more units / modules may be integrated together. The above-mentioned integrated unit / module may be implemented in the form of hardware or in the form of a software program module.

[0301] If the integrated unit / module is implemented in the form of hardware, the hardware may be a digital circuit, an analog circuit, etc. The physical implementation of the hardware structure includes but is not limited to transistors, memristors, etc. Unless otherwise specified, the processor may be any appropriate hardware processor, such as a CPU, a GPU, an FPGA, a DSP, an ASIC, etc. Unless otherwise specified, the storage unit may be any appropriate magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory (RRAM), a dynamic random access memory (DRAM), a static random access memory (SRAM), an enhanced dynamic random access memory (EDRAM), a high-bandwidth memory (HBM), a hybrid memory cube (HMC), etc.

[0302] If the integrated unit / module is implemented in the form of a software program module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a memory, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned memory includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0303] In the above embodiments, the description of each embodiment has its own emphasis. For the part not described in detail in a certain embodiment, please refer to the relevant description of other embodiments. The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0304] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0305] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A method for recommending financial products, characterized in that: include: Acquire historical transaction records of users purchasing financial products, search indexes of each of the candidate financial products among multiple candidate financial products, and sentiment indexes of each of the candidate financial products, wherein the sentiment index of the candidate financial product is used to characterize the sentiment direction of the evaluation of the candidate financial product; The historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products are input into a pre-trained recommendation model to obtain a recommended financial product output by the recommendation model, wherein the recommended financial product is part of the multiple candidate financial products.

2. The method according to claim 1, characterized in that Obtaining the sentiment index of each of the candidate financial products, including: For any of the candidate financial products, obtaining a plurality of evaluation texts for the candidate financial product; Extracting a word vector matrix of each evaluation text; The sentiment index of the candidate financial product is determined according to the word vector matrices of each of the multiple evaluation texts.

3. The method according to claim 2, characterized in that The step of extracting a word vector matrix of each evaluation text includes: Segmenting the evaluation text, and deleting stop words in the words obtained by segmentation to obtain a segmentation result; Input the word segmentation result into a word embedding model to obtain an initial word vector matrix of the evaluation text; The word vector matrix of the evaluation text is determined according to the initial word vector matrix and the preset matrix dimension.

4. The method according to claim 3, characterized in that The step of determining the word vector matrix of the evaluation text according to the initial word vector matrix and the preset matrix dimension includes: When the dimension of the initial word vector matrix is ​​the preset matrix dimension, determining the initial word vector matrix as the word vector matrix of the evaluation text; In the case where the dimension of the initial word vector matrix is ​​lower than the preset matrix dimension, determining keywords in the word segmentation result, and filling the word vectors corresponding to the keywords into the initial word vector matrix as filling vectors to obtain the word vector matrix of the evaluation text; When the dimension of the initial word vector matrix is ​​higher than the preset matrix dimension, the initial word vector matrix is ​​truncated according to the preset matrix dimension to obtain the word vector matrix of the evaluation text.

5. The method according to claim 2, characterized in that: The determining the sentiment index of the candidate financial product according to the word vector matrices of the plurality of evaluation texts includes: Inputting the word vector matrices of the plurality of evaluation texts into a long short-term memory network to obtain context features of the plurality of evaluation texts; Inputting the context feature information of the multiple evaluation texts into a convolutional neural network to obtain semantic features of the multiple evaluation texts, wherein the convolutional neural network adopts a dilated convolution; Performing sentiment classification based on the semantic features of the plurality of evaluation texts to obtain a positive sentiment index and a negative sentiment index of the candidate financial product; The sentiment index of the candidate financial product is determined according to the positive sentiment index and the negative sentiment index.

6. The method according to claim 2, characterized in that The obtaining of multiple evaluation texts for the candidate financial product includes: Obtaining a plurality of initial evaluation texts for the candidate financial products; If a first initial evaluation text among the multiple initial evaluation texts contains preset keywords, the first initial evaluation text is deleted, and if a duplicate second initial evaluation text exists among the multiple initial evaluation texts, the second initial evaluation text is deleted to obtain the multiple evaluation texts.

7. The method according to any one of claims 1 to 6, characterized in that The step of inputting the historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products into a pre-trained recommendation model to obtain the recommended financial products output by the recommendation model includes: The historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products are input into the generator of the recommendation model to obtain the recommended financial products output by the generator, wherein the recommendation model includes the generator and a discriminator, and the discriminator is used to determine whether the recommended financial products output by the generator are actually selected financial products during the training stage of the recommendation model.

8. A method for training a recommendation model for financial products, characterized in that: include: Acquire multiple training samples, each of which includes a historical transaction record of a user purchasing a financial product, a search index of each of multiple candidate financial products, a sentiment index of each of the candidate financial products, and a real selected financial product, wherein the sentiment index of the candidate financial product is used to characterize the sentiment direction of the evaluation of the candidate financial product; Inputting the historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products into a generator of a recommendation model to obtain a recommended financial product output by the generator, wherein the recommended financial product is a part of the multiple candidate financial products; Inputting the recommended financial product and the actually selected financial product into the discriminator of the recommendation model to obtain a discriminant result output by the discriminator; A loss of the recommendation model is determined based on the recommended financial product and the discrimination result, and a parameter of the recommendation model is adjusted based on the loss.

9. A financial product recommendation device, characterized in that: include: An acquisition module, used to acquire historical transaction records of users purchasing financial products, a search index of each of a plurality of candidate financial products, and a sentiment index of each of the candidate financial products, wherein the sentiment index of the candidate financial product is used to represent the sentiment direction of the evaluation of the candidate financial product; An input module is used to input the historical transaction records, the search index of each candidate financial product, and the sentiment index of each candidate financial product into a pre-trained recommendation model to obtain a recommended financial product output by the recommendation model, wherein the recommended financial product is part of the multiple candidate financial products.

10. A training device for a recommendation model of a financial product, characterized in that: include: an acquisition module, configured to acquire a plurality of training samples, each of which includes a historical transaction record of a user purchasing a financial product, a search index of each of a plurality of candidate financial products, a sentiment index of each of the candidate financial products, and a financial product actually selected, wherein the sentiment index of the candidate financial product is used to characterize the sentiment direction of the evaluation of the candidate financial product; An input module, used to input the historical transaction records, the search index of each of the candidate financial products, and the sentiment index of each of the candidate financial products into a generator of a recommendation model, to obtain a recommended financial product output by the generator, wherein the recommended financial product is a part of the multiple candidate financial products; The input module is further used to input the recommended financial product and the actually selected financial product into the discriminator of the recommendation model to obtain a discrimination result output by the discriminator; A determination module, configured to determine a loss of the recommendation model based on the recommended financial product and the discrimination result; An adjustment module is used to adjust the parameters of the recommendation model based on the loss.

11. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 8 when executed by a processor.

13. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 8 when being executed by a processor.

Images