Financial product recommendation method and system based on collaborative filtering and neural network

Abstract

The present invention discloses a financial product recommendation method and system based on collaborative filtering and neural network, the method comprising: obtaining relevant data of target users, calculating preference similarity based on tag-related data, and constructing a similar user set based on preference similarity; obtaining the initial interest of historical users in the similar user set to the target recommended product, obtaining the interest of the target user to the target recommended product through preference similarity and initial interest and constructing a target recommended product library; dynamically updating and preprocessing the relevant data of the target recommended product in the updated target recommended product library to obtain a training sample set; obtaining a risk budget model; obtaining the optimal ratio of the target recommended product based on the risk budget model and visually displaying it. Real-time matching is possible: the interest of the target user is calculated through an improved collaborative filtering algorithm to correctly identify the interest preference of the target user, and the financial products matching it at the current moment are found to construct a personalized financial product pool.

Description

Technical Field

[0001] The present invention relates to the field of search recommendation technology, and in particular to a financial product recommendation method and system based on collaborative filtering and neural network. Background Art

[0002] In the existing technology, deep learning models, as an important branch of machine learning, have been widely used and deeply studied in the field of product recommendation. Deep learning extracts features from data layer by layer through a multi-level structure, from low-level simple features to high-level abstract features, which greatly simplifies the feature engineering steps in traditional machine learning, and can understand the data more deeply, thereby improving the expression ability and accuracy of the model.

[0003] With the development of information technology and the financial industry, many businesses are now also realized through modern statistics and computer technology. Modern statistics are basically done manually or with the help of simple technology, but as the user group increases, the amount of data will also increase, and the financial industry will also use machine learning technology to recommend related financial products. For example, through intelligent investment advisory software, users' relevant behavior data, market data, product data, etc. are analyzed, and then products or suggestions that are relatively in line with user needs can be provided to users.

[0004] However, in today's smart investment advisory software, the weight values ​​of various investment influencing factors are mainly set manually, and then the user's information and the information of the investment target recommended products are multiplied by the corresponding weight values, and the investment target recommended products with the highest scores are recommended to the user. However, the situations of different users are quite diverse, and the main factors that influence users to buy different investment target recommended products are also different. The use of artificially set fixed weight values ​​to determine the investment target recommended products can only be applied to general situations, and cannot provide different users with the most suitable investment target recommended products.

[0005] At present, the steps of smart investment advisor software are as follows: classify and score financial products according to their risk-return characteristics, screen out a high-quality financial product database; match target recommended products with investors; and input user risk preference data into the mean-variance model through parameterization to determine financial products. However, the current smart investment advisor software recommendation results are not accurate enough, resulting in many users being dissatisfied with the recommendation results. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a financial product recommendation method and system based on collaborative filtering and neural network.

[0007] In order to solve the above technical problems, the present invention is solved by the following technical solutions:

[0008] A financial product recommendation method based on collaborative filtering and neural network includes the following steps:

[0009] Obtain relevant data of the target user, wherein the user-related data at least includes tag-related data and behavior-related data, the behavior-related data includes behavior data and behavior occurrence time, and the behavior data at least includes click browsing records, collection records, and purchase records;

[0010] Based on the tag-related data, calculate the similarity between the target user and the historical users in the historical user database in terms of their preferences for the target recommended product, and build a similar user set based on the preference similarity;

[0011] Based on the behavior-related data, the initial interest of historical users in the similar user set for the target recommended product is obtained, and the final interest of the target user for the target recommended product is obtained through the preference similarity and the initial interest, and the final interest is corrected to obtain the corrected interest and a target recommended product library is constructed based on the corrected interest;

[0012] Dynamically updating the target recommended product library to obtain an updated target recommended product library, and performing data preprocessing on relevant data of the target recommended products in the updated target recommended product library to obtain a training sample set;

[0013] Construct a risk budget pre-training model, train and optimize the risk budget pre-training model through the training sample set, and obtain a risk budget model;

[0014] Based on the risk budget model, the optimal ratio of target recommended products is obtained and visualized.

[0015] As an implementable method, the method of calculating the similarity between the target user and the historical users in the historical user database based on the tag-related data includes the following steps:

[0016] Obtain the number of target recommended products that the target user is interested in and the number of target recommended products that historical users are interested in, determine the total number of target recommended products that the target user and historical users are interested in and the number of target recommended products that they are interested in, and then obtain the first preference similarity between the target user and historical users;

[0017] Get the time when the target user last acted on the target product and the time when the historical user last acted on the target product, and combine the total number of products and the number of target recommended products of common interest to obtain the similarity of the second preference of the target user and the historical user;

[0018] The first preference similarity and the second preference similarity are weighted to obtain the preference similarity between the target user and the historical users, and a similar user set is constructed based on the historical users corresponding to the top several preference similarity scores.

[0019] As an implementable method, the first preference similarity is expressed as follows:

[0020]

[0021] The second preference similarity is expressed as follows:

[0022]

[0023] The preference similarity is expressed as follows:

[0024]

[0025] in, represents the first preference similarity, represents the second preference similarity, Indicates the target user and historical users Similarity of preferences, Indicates the target user The time when the last action on the target product occurred, Indicates historical users The time when the last action on the target product occurred, Indicates Target recommended products, Indicates the target user The number of target recommended products of interest, Indicates historical users The number of target recommended products of interest, Indicates the target user and historical users The number of recommended products of common interest.

[0026] As an implementable method, the final interest of the target user in the target recommended product is obtained through the preference similarity and the initial interest, and the final interest is corrected to obtain the corrected interest and a target recommended product library is constructed based on the corrected interest, including the following steps:

[0027] Determine the target user's initial interest in the target recommended product through the target user's click browsing history, collection history, and purchase history;

[0028] Based on the first preference similarity and the initial interest, the first interest of the target user in the target recommended product is obtained, and through the second preference similarity and the initial interest, the second interest of the target user in the target product is obtained;

[0029] Based on the first interest level and the second interest level, obtaining a final interest level of the target user for the target product;

[0030] The final interest degree is corrected by the Gaussian link function to obtain the corrected interest degree, and the target recommended product library is constructed by taking the target products corresponding to the top several corrected interest degree scores.

[0031] As an implementable method, the initial interest level is expressed as follows:

[0032]

[0033] The first interest level is expressed as follows:

[0034]

[0035] The second interest level is expressed as follows:

[0036]

[0037] The final interest level is expressed as follows:

[0038]

[0039] The modified interest level is expressed as follows:

[0040]

[0041] in,

[0042]

[0043]

[0044]

[0045] in, Indicates the final interest, represents a set of similar users, represents the user similarity obtained by the maximum likelihood function constructed based on the Gaussian link function, Indicates historical users The average interest score of Indicates the target user Recommend products to target Interest ranking, Indicates the number of target recommended products. Represent the normalized target users Recommend products to target Interest ranking and historical users Recommend products to target Interest ranking, Indicates user similarity, represents the initial interest, They represent the number of click-through records, favorite records, and purchase records respectively. Respectively represent the proportion of click browsing records, collection records and purchase records, Represent the first interest, the second interest and the final interest respectively. represents a set of similar users, represents the first preference similarity, represents the second preference similarity, Indicates the target user and historical users The number of target recommended products of common interest, Indicates Target recommended products, Represents the inverse function of the standard Gaussian distribution.

[0046] As an implementable method, the following steps are also included:

[0047] If the target user is a new user, the final interest of the target user in the target product is obtained through the initial similarity and initial interest;

[0048] The initial similarity is expressed as follows:

[0049]

[0050] The final interest level is expressed as follows:

[0051]

[0052] in, represents the initial similarity, Indicates the target user For target products The final interest level, represents the initial interest, represents a set of similar users, Respectively represent the target users and historical users level of risk.

[0053] As an implementable method, dynamically updating the target recommended product library to obtain an updated target recommended product library includes the following steps:

[0054] Preset the cluster value of the target recommended product library, randomly select K target recommended products as the initial cluster centers, assign each target recommended product to the cluster to which the nearest cluster center belongs, recalculate the cluster center for each clustering, and repeat the iteration until the cluster center no longer changes;

[0055] Calculate the Euclidean distance between the current target recommended product and other target recommended products in the same cluster and normalize them;

[0056] The target recommended products with the highest final similarity scores with the current target recommended product are used as the updated target recommended product library;

[0057] The Euclidean distance is expressed as follows:

[0058]

[0059] The final similarity is expressed as follows:

[0060]

[0061] in, Indicates target recommended products The Euclidean distance between represents the final similarity, Indicates Only financial product related indicators, Indicates Only financial products indicators, Indicates Only financial products indicators, Represents a collection of similar products.

[0062] As an implementable method, the risk budget layer processes the risk budget value and obtains a convex optimization result. The specific process is as follows:

[0063] Construct the objective function and make the result of the objective function minimize the prediction error, which is expressed as follows:

[0064]

[0065] The constraints are:

[0066] Where N represents the number of target recommended products in the target recommended product portfolio. represents the combination weight matrix, represents the combined return, represents the target recommended product covariance matrix, represents the combined variance, Indicates The weight of the target recommended product, Indicates The risk budget value of the target recommended product, and ;

[0067] Combined with the preset stability condition, the objective function is modified and transformed to obtain the transformed objective function, which is expressed as follows:

[0068]

[0069] The constraints are:

[0070] Based on the solver CVXPY, the optimal ratio of each target recommended product is obtained by normalizing the results, which is expressed as follows:

[0071]

[0072] in, Indicates The weight of the target recommended product after transformation under the preset stability condition, Indicates The optimal ratio of recommended products for each target.

[0073] As an implementable method, the risk budget pre-trained model is trained and optimized through a loss function and an optimizer;

[0074] The optimizer is Adam, which calculates the first-order moment estimate and the second-order moment estimate of the gradient respectively during the back-propagation process and combines the deviation correction and the adaptive learning rate adjustment to update the parameters of the risk budget pre-training model until the optimal parameters are obtained.

[0075] A financial product recommendation system based on collaborative filtering and neural network, comprising an acquisition processing module, a first processing module, a second processing module, a data update module, a construction training module and a result display module;

[0076] The data acquisition module acquires relevant data of the target user, wherein the user-related data at least includes tag-related data and behavior-related data, the behavior-related data includes behavior data and behavior occurrence time, and the behavior data at least includes click browsing records, collection records and purchase records;

[0077] The first processing module calculates the similarity of preferences of the target user and the historical users in the historical user database for the target recommended product based on the tag-related data, and constructs a similar user set based on the preference similarity;

[0078] The second processing module obtains the initial interest of historical users in the similar user set for the target recommended product based on the behavior-related data, obtains the final interest of the target user for the target recommended product through the preference similarity and the initial interest, and corrects the final interest to obtain the corrected interest and constructs a target recommended product library based on the corrected interest;

[0079] The data updating module dynamically updates the target recommended product library to obtain an updated target recommended product library, and performs data preprocessing on the relevant data of the target recommended products in the updated target recommended product library to obtain a training sample set;

[0080] Construct a training module, construct a risk budget pre-training model, train and optimize the risk budget pre-training model through the training sample set to obtain the risk budget model. The risk budget pre-training model includes a data input layer, a main structure layer, a risk budget layer and a result conversion layer. The main structure layer includes a pooling layer and a fully connected layer. The data input layer receives the training sample set, and performs multi-level feature extraction on the training sample set under different steps through the convolution layer and the pooling layer to obtain relevant features and perform dimensionality reduction operations to obtain reduced dimensionality features. The two features of the reduced dimensionality feature set that are in different steps are spliced ​​to obtain spliced ​​features and input into the fully connected layer; the fully connected layer receives the spliced ​​features and combines them with the activation function ReLu to obtain the risk budget value set of each target recommended product in the time series, where the sum of all risk budget values ​​is 1. The risk budget layer processes the risk budget value and obtains the convex optimization result, that is, the optimal proportion of the target recommended product.

[0081] The result display module visualizes the optimal ratio of the target recommended products.

[0082] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following method is implemented:

[0083] Obtain relevant data of the target user, wherein the user-related data at least includes tag-related data and behavior-related data, the behavior-related data includes behavior data and behavior occurrence time, and the behavior data at least includes click browsing records, collection records, and purchase records;

[0084] Based on the tag-related data, calculate the similarity between the target user and the historical users in the historical user database in terms of their preferences for the target recommended product, and build a similar user set based on the preference similarity;

[0085] Based on the behavior-related data, the initial interest of historical users in the similar user set in the target recommended product is obtained, and the interest of the target user in the target recommended product is obtained through the preference similarity and the initial interest, and a target recommended product library is constructed based on the interest;

[0086] Dynamically updating the target recommended product library to obtain an updated target recommended product library, and performing data preprocessing on relevant data of the target recommended products in the updated target recommended product library to obtain a training sample set;

[0087] Construct a risk budget pre-training model, train and optimize the risk budget pre-training model through the training sample set to obtain the risk budget model. The risk budget pre-training model includes a data input layer, a main structure layer, a risk budget layer and a result conversion layer. The main structure layer includes a pooling layer and a fully connected layer. The data input layer receives the training sample set, and performs multi-level feature extraction on the training sample set under different steps through the convolution layer and the pooling layer to obtain relevant features and perform dimensionality reduction operations to obtain reduced dimensionality features. Two features of different steps in the reduced dimensionality feature set are spliced ​​to obtain spliced ​​features and input to the fully connected layer; the fully connected layer receives the spliced ​​features and combines them with the activation function ReLu to obtain a set of risk budget values ​​of each target recommended product in the time series, where the sum of all risk budget values ​​is 1. The risk budget layer processes the risk budget value and obtains a convex optimization result, that is, the optimal proportion of the target recommended product.

[0088] The optimal ratio of the target recommended products is visually displayed.

[0089] A financial product recommendation device based on collaborative filtering and neural network includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the following method when executing the computer program:

[0090] Obtain relevant data of the target user, wherein the user-related data at least includes tag-related data and behavior-related data, the behavior-related data includes behavior data and behavior occurrence time, and the behavior data at least includes click browsing records, collection records, and purchase records;

[0091] Based on the tag-related data, calculate the similarity between the target user and the historical users in the historical user database in terms of their preferences for the target recommended product, and build a similar user set based on the preference similarity;

[0092] Based on the behavior-related data, the initial interest of historical users in the similar user set in the target recommended product is obtained, and the interest of the target user in the target recommended product is obtained through the preference similarity and the initial interest, and a target recommended product library is constructed based on the interest;

[0093] Dynamically updating the target recommended product library to obtain an updated target recommended product library, and performing data preprocessing on relevant data of the target recommended products in the updated target recommended product library to obtain a training sample set;

[0094] Construct a risk budget pre-training model, train and optimize the risk budget pre-training model through the training sample set to obtain the risk budget model. The risk budget pre-training model includes a data input layer, a main structure layer, a risk budget layer and a result conversion layer. The main structure layer includes a pooling layer and a fully connected layer. The data input layer receives the training sample set, and performs multi-level feature extraction on the training sample set under different steps through the convolution layer and the pooling layer to obtain relevant features and perform dimensionality reduction operations to obtain reduced dimensionality features. Two features of different steps in the reduced dimensionality feature set are spliced ​​to obtain spliced ​​features and input to the fully connected layer; the fully connected layer receives the spliced ​​features and combines them with the activation function ReLu to obtain a set of risk budget values ​​of each target recommended product in the time series, where the sum of all risk budget values ​​is 1. The risk budget layer processes the risk budget value and obtains a convex optimization result, that is, the optimal proportion of the target recommended product.

[0095] The optimal ratio of the target recommended products is visually displayed.

[0096] The present invention has significant technical effects due to the adoption of the above technical solution:

[0097] The present invention can match in real time: by calculating the interests of investors or target users through an improved collaborative filtering algorithm, correctly identifying the interest preferences of target users, finding target recommended products that match them at the current moment, and building a personalized target recommended product pool;

[0098] Alleviate cold start: Introduce user tags to improve user similarity calculation and alleviate the cold start problem of new users; use k-means clustering method to alleviate the cold start problem by calculating Euclidean distance;

[0099] Customized convolution kernels and pooling layers based on complex neural network algorithms can achieve deep factor mining and output configuration weights based on risk budgeting strategies to avoid the problem of incorrect configuration of target recommended products due to inaccurate expected future returns of traditional machine learning and deep learning;

[0100] Evaluate the model, optimize hyperparameters, determine the optimal model and compare it, and output visualization results. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0102] Figure 1 It is a schematic flow diagram of the overall method of the present invention;

[0103] Figure 2 It is a schematic diagram of the overall structure of the system of the present invention;

[0104] Figure 3 It is a schematic diagram of the network architecture of the risk budget model of the present invention. DETAILED DESCRIPTION

[0105] The present invention is further described in detail below in conjunction with embodiments. The following embodiments are for explanation of the present invention but the present invention is not limited to the following embodiments.

[0106] Embodiment 1:

[0107] A financial product recommendation method based on collaborative filtering and neural network, such as Figure 1 As shown, the following steps are included:

[0108] S100, obtaining relevant data of the target user, wherein the user-related data at least includes tag-related data and behavior-related data, the behavior-related data includes behavior data and behavior occurrence time, and the behavior data at least includes click browsing records, collection records, and purchase records;

[0109] S200, based on the tag-related data, calculating the similarity of preferences of the target user and historical users in the historical user database for the target recommended product, and building a similar user set based on the preference similarity;

[0110] S300, based on the behavior-related data, obtaining the initial interest of historical users in the similar user set for the target recommended product, obtaining the final interest of the target user for the target recommended product through the preference similarity and the initial interest, and correcting the final interest to obtain a corrected interest, and building a target recommended product library based on the corrected interest;

[0111] S400, dynamically updating the target recommended product library to obtain an updated target recommended product library, and performing data preprocessing on relevant data of the target recommended products in the updated target recommended product library to obtain a training sample set;

[0112] S500, construct a risk budget pre-training model, train and optimize the risk budget pre-training model through the training sample set to obtain the risk budget model, the risk budget pre-training model includes a data input layer, a main structure layer, a risk budget layer and a result conversion layer, the main structure layer includes a pooling layer and a fully connected layer, the data input layer receives the training sample set, and performs multi-level feature extraction on the training sample set under different steps through the convolution layer and the pooling layer to obtain relevant features and perform dimensionality reduction operations to obtain reduced dimensionality features, and splice two features of different steps in the reduced dimensionality feature set to obtain spliced ​​features and input them into the fully connected layer; the fully connected layer receives the spliced ​​features and combines them with the activation function ReLu to obtain a set of risk budget values ​​of each target recommended product in the time series, wherein the sum of all risk budget values ​​is 1, the risk budget layer processes the risk budget value, and obtains a convex optimization result, i.e., the optimal proportion of the target recommended products;

[0113] S600: Visually display the optimal ratio of the target recommended products.

[0114] In one embodiment, the target recommended product can be a financial product or a financial-related product. The interest of the investor or the target user is calculated through an improved collaborative filtering algorithm, the interest preference of the target user is correctly identified, and the financial products that match the target user at the current moment are found to build a personalized financial product pool. In order to make the recommendation more accurate, this embodiment divides the user behavior into click browsing, collection and purchase, and calculates the similarity of interest between users in certain commodities through these behavior data, specifically: obtain the number of target recommended products that the target user is interested in and the number of target recommended products that historical users are interested in, determine the total amount of target recommended products that the target user and historical users are interested in, and the number of target recommended products that they are interested in, and then obtain the first preference similarity between the target user and the historical user, which is expressed as follows:

[0115]

[0116] When time is taken into account, the time when the target user last acted on the target product and the time when the historical user last acted on the target product are obtained. Combined with the total number of products and the number of target recommended products of common interest, the similarity of the second preference between the target user and the historical user is obtained, which is expressed as follows:

[0117]

[0118] The first preference similarity and the second preference similarity are weighted to obtain the preference similarity between the target user and the historical users, and a similar user set is constructed based on the historical users with the top preference similarity scores. The preference similarity is calculated using the following formula:

[0119]

[0120] in, represents the first preference similarity, represents the second preference similarity, Indicates the target user and historical users Similarity of preferences, Indicates the target user The time when the last action on the target product occurred, Indicates historical users The time when the last action on the target product occurred, Indicates Target recommended products, Indicates the target user The number of target recommended products of interest, Indicates historical users The number of target recommended products of interest, Indicates the target user and historical users The number of target recommended products of common interest is represented by the user set formed by .

[0121] Obtain any historical user's click browsing history, collection history, and purchase history, determine any historical user's initial interest in the target recommended product, and Defined as historical user The interest in product i is calculated as follows:

[0122]

[0123] For example, if there is a certain If the number of clicks and views accounts for 10%, the number of favorites accounts for 30%, and the number of purchases accounts for 60%, it can be expressed as follows:

[0124]

[0125] Based on the first preference similarity and the initial interest, the first interest of the target user in the target recommended product is obtained. Through the second preference similarity and the initial interest, the second interest of the target user in the target product is obtained. The first interest is expressed as follows:

[0126]

[0127] The second interest level is expressed as follows:

[0128]

[0129] Finally, the final interest of the target user for the target product is obtained through the first interest and the second interest. The final interest is expressed as follows:

[0130]

[0131] After obtaining the final interest level, the final interest level will be corrected through the Gaussian link function to obtain the corrected interest level, and the target products corresponding to the top several corrected interest level scores will be used to construct a target recommended product library;

[0132] The modified interest level is expressed as follows:

[0133]

[0134] in,

[0135]

[0136]

[0137]

[0138] in, Indicates the final interest, represents a set of similar users, represents the user similarity obtained by the maximum likelihood function constructed based on the Gaussian link function, Indicates historical users The average interest score of Indicates the target user Recommend products to target Interest ranking, Indicates the number of target recommended products. Represent the normalized target users Recommend products to target Interest ranking and historical users Recommend products to target Interest ranking, Indicates user similarity, represents the initial interest, They represent the number of click-through records, favorite records, and purchase records respectively. Respectively represent the proportion of click browsing records, collection records and purchase records, Represent the first interest, the second interest and the final interest respectively. represents a set of similar users, represents the first preference similarity, represents the second preference similarity, Indicates the target user and historical users The number of target recommended products of common interest, Indicates Target recommended products, Represents the inverse function of the standard Gaussian distribution.

[0139] In one embodiment, if the target user is a new user, when the new user registers a user account, a risk preference level is preset for the user according to the user's relevant circumstances, a user tag is set according to the risk preference level, and the similarity between the new user and the historical users in the historical user database is calculated, which can be obtained in the following way:

[0140]

[0141] Then, the final interest of the target user in the target product is obtained through the initial similarity and initial interest;

[0142]

[0143] In this way, through the above process, we can find a set of users with high similarity to the target user u in the same time period, calculate the interest of the target user u in the target recommended product, and add the N products with the highest interest to the personalized target recommended product library of the target user u. Respectively represent the target users and historical users level of risk.

[0144] The target recommendation product library will be continuously tracked and updated in a timely manner according to time and time changes. During the update process, the model will be evaluated based on the recall rate and precision rate, and the parameters will be adjusted for optimization to update the target recommendation product library.

[0145] For new product j, the present invention uses k-means clustering method to calculate its type, obtain relevant data of the product, and clean, standardize or normalize the data to obtain data in n directions .

[0146] Determine how many clusters to divide the products into, that is, the K value. For each target recommended product in the data set, calculate its Euclidean distance to each cluster center:

[0147]

[0148] K products are randomly selected as the initial cluster centers, and each product is assigned to the cluster to which the cluster center closest to it belongs. For each clustering, its cluster center is recalculated, and the iteration is repeated until the cluster center no longer changes significantly, and a new cluster center is obtained.

[0149] Calculate the Euclidean distance between the products in the same cluster as new product j and new product j, and normalize them. The l target recommended products with the highest similarity to target recommended product j are set as the product set. , the similarity between new product j and target recommended product i is calculated and expressed as follows:

[0150]

[0151] Calculate the interest of any user in the target recommended product j, and recommend the target recommended product j to the top M users in terms of interest, to alleviate the cold start problem of the new target recommended product:

[0152]

[0153] Throughout the process, Indicates target recommended products The Euclidean distance between represents the final similarity, Indicates Only financial product related indicators, Indicates Only financial products indicators, Indicates Only financial products indicators, Represents a collection of similar products. Represents a set of similar users. The indicators here can be annualized rate of return, Sharpe ratio, maximum drawdown, VaR, CVaR, etc.

[0154] In one embodiment, a risk budget pre-training model is constructed, and the risk budget pre-training model is trained and optimized through a training sample set to obtain a risk budget model, as follows:

[0155] Obtain relevant data sets of target recommended products in the personalized target recommended product library and perform data preprocessing to obtain a training sample set;

[0156] The deep learning model is trained based on the training sample set, and the parameters are optimized through forward propagation and back propagation to obtain the target recommended product configuration weights, thereby completing the model training; of course, the model will also be retrained as the data set increases.

[0157] Obtain relevant data of the target recommended products in the target recommended product library, such as historical yield, historical net capital inflow and historical scale. The yield is obtained according to time series, such as the yield in the past month, the yield in the past three months, and the yield in the past year; the net capital inflow is also obtained according to time series, such as the net capital inflow in the past week, the net capital inflow in the past month, the net capital inflow in the past three months, and the net capital inflow in the past year; the scale is also obtained according to time series, such as the scale in the past week, the scale in the past month, the scale in the past three months, the scale in the past year and other data.

[0158] In this embodiment, data preprocessing will be performed on all data in the relevant data set, such as one or more of data filling, interpolation or standardization. After preprocessing, there will inevitably be unreasonable data. Therefore, the local anomaly factor of the relevant data will be calculated through the local anomaly factor identification model, and the relevant data corresponding to the local anomaly factor will be eliminated. Finally, a training sample set suitable for training the risk budget pre-training model is obtained.

[0159] Specifically, the K distance of the relevant data in the relevant data set is obtained to obtain the reachable distance of the relevant data, and based on the reachable distance of the relevant data, the local reachable density of the relevant data is obtained; combined with the local reachable density of the relevant data, the local abnormal factor of the relevant data is obtained;

[0160] K distance is obtained by:

[0161]

[0162] The reachable distance is obtained by:

[0163]

[0164] The local reachability density is obtained as follows:

[0165]

[0166] The local anomaly factor is obtained by:

[0167]

[0168] in, represents a constant, Represents distance related data Recent data, Represents the distance between two points. represents the K distance, represents the reachable distance of the relevant data, represents the local reachable density of the relevant data, Represents the local anomaly factor of the correlated data.

[0169] In one embodiment, the risk budget pre-training model includes a data input layer, a main structure layer, a risk budget layer and a result conversion layer, and the main structure layer includes a pooling layer and a fully connected layer;

[0170] The data input layer receives a training sample set;

[0171] The custom convolutional layer is used to extract multi-level features from the training sample sets under different steps, specifically:

[0172] Custom convolution function Represents the correlation coefficient of the X and Y time series with a step of N;

[0173] Custom convolution function Indicates the standard deviation of X with a step of N days;

[0174] Custom convolution function Indicates the maximum value of X with a step of N days;

[0175] Custom convolution function Indicates the minimum value of X in N days;

[0176] Define the convolution function Indicates the mean value of X with a step of N days;

[0177] Define the convolution function Indicates the rate of change of X with a step size of N days.

[0178] Get the relevant features and perform batch normalization (Batch Norm) and maximum pooling to get the dimension reduction features. Concatenate the two features with different steps in the dimension reduction feature set to get the concatenated features and input them into the fully connected layer.

[0179] The fully connected layer receives the concatenated features and sets the activation function Leaky ReLu to obtain a set of risk budget values ​​of each target recommended product in the time series, where the sum of all risk budget values ​​is 1;

[0180] The risk budget layer processes the risk budget value and obtains a convex optimization result. The specific process is as follows:

[0181] Input the risk budget value into the risk budget layer. The combination under the risk budget ratio must meet the conditions , expressed as:

[0182]

[0183] The constraints are:

[0184] According to the KKT condition, the risk budget model is transformed into a convex optimization problem:

[0185]

[0186] The constraints are:

[0187] The convex optimization problem is solved in a neural network using the solver CVXPY, and the output is finally normalized to obtain the optimal ratio:

[0188]

[0189] Where N represents the number of target recommended products in the target recommended product portfolio. represents the combination weight matrix, represents the combined return, represents the target recommended product covariance matrix, represents the combined variance, Indicates The weight of the target recommended product, Indicates The risk budget value of the target recommended product, and , Indicates The weight of the target recommended product after transformation under the preset stability condition, Indicates The optimal ratio of recommended products for each target.

[0190] In actual operation, the actual process can be found in the attached Figure 3 As shown. Input time series data, perform feature extraction, obtain feature step N1 and step N2, respectively undergo BatchNorm and pooling operations, and then perform Concat processing on the two to obtain the corresponding fusion features. The fusion features pass through the fully connected layer and the risk budget layer in turn to finally obtain the asset weight.

[0191] In one embodiment, the risk budget pre-training model is trained and optimized by a loss function and an optimizer, wherein the loss function is a negative value of the cumulative return of the combination;

[0192] In this embodiment, the learning rate is set to 0.01, the training cycle (epoch) is set to 100, and the early stopping mechanism is adopted to avoid overfitting of the model on the training set.

[0193] The optimizer used in the present invention is Adam (Adaptive Moment Estimation), which calculates the first-order moment estimation and second-order moment estimation of the gradient during the back-propagation process, and combines the deviation correction and adaptive learning rate adjustment to update the model parameters.

[0194] Initialize the first moment estimate and the second moment estimate to 0, as well as the number of iterations;

[0195] The first moment estimate is updated by the gradient, which is expressed as follows:

[0196]

[0197] in, is the decay rate of the first moment estimate, usually set to 0.9, is the gradient.

[0198] The second moment estimate is updated by the gradient, which is expressed as follows:

[0199]

[0200] in, is the decay rate of the second moment estimate, usually set to 0.999.

[0201] Compute the bias correction for the first moment estimate, expressed as follows:

[0202]

[0203] Compute the bias correction for the second moment estimate, expressed as follows:

[0204]

[0205] Update parameters, as follows:

[0206]

[0207] in, represents the learning rate, Represents a very small number to prevent the denominator from being zero. In each iteration, the Adam optimizer updates each parameter according to the above steps.

[0208] Finally, the optimal proportion of target recommended products is obtained based on the risk budget model and displayed visually. By combining it with practical applications, the configuration results of the model can be dynamically optimized according to indicators at the return and risk levels. Specific evaluation indicators include annualized rate of return, Sharpe ratio, maximum drawdown, VaR, CVaR, etc.

[0209] The annualized rate of return is a standardized way of expressing the rate of return on investment. It converts the income within a certain period into the rate of return for one year. The calculation formula is:

[0210]

[0211] N is the number of trading days.

[0212] The Sharpe Ratio is an indicator used to evaluate the risk-adjusted return performance of a portfolio. It measures the relationship between the excess return of a portfolio and its risk. Here is how the Sharpe Ratio is calculated:

[0213]

[0214] is the annualized standard deviation, is the one-year Treasury bond yield.

[0215] Maximum drawdown measures the maximum drop from the highest point to the lowest point of an investment portfolio within a certain period of time, that is, the maximum possible loss.

[0216] VaR (Value at Risk) is a method used to measure the maximum loss that a portfolio may suffer at a given confidence level within a certain period of time:

[0217]

[0218] is the target recommended product yield on day t, is the standard deviation on day t, is the quantile corresponding to the confidence interval under the standard normal distribution.

[0219] CVaR refers to the average loss of the portfolio under the conditions where the loss exceeds VaR at a given confidence level, and is calculated as:

[0220]

[0221] Adjust parameters such as learning rate and step based on model evaluation to achieve dynamic optimization. The net value of the income obtained based on the model is displayed and visualized with other similar products.

[0222] Embodiment 2:

[0223] A financial product recommendation system based on collaborative filtering and neural networks, such as Figure 2 As shown, it includes an acquisition processing module 100, a first processing module 200, a second processing module 300, a data update module 400, a construction training module 500 and a result display module 600;

[0224] The data acquisition module 100 acquires relevant data of the target user, wherein the user-related data at least includes tag-related data and behavior-related data, the behavior-related data includes behavior data and behavior occurrence time, and the behavior data at least includes click browsing records, collection records and purchase records;

[0225] The first processing module 200 calculates the similarity of preferences of the target user and the historical users in the historical user database for the target recommended product based on the tag-related data, and constructs a similar user set based on the preference similarity;

[0226] The second processing module 300 obtains the initial interest of historical users in the similar user set for the target recommended product based on the behavior-related data, obtains the final interest of the target user for the target recommended product through the preference similarity and the initial interest, and modifies the final interest to obtain the modified interest and constructs a target recommended product library based on the modified interest;

[0227] The data updating module 400 dynamically updates the target recommended product library to obtain an updated target recommended product library, and performs data preprocessing on the relevant data of the target recommended products in the updated target recommended product library to obtain a training sample set;

[0228] Constructing a training module 500, constructing a risk budget pre-training model, training and optimizing the risk budget pre-training model through a training sample set to obtain a risk budget model, the risk budget pre-training model includes a data input layer, a main structure layer, a risk budget layer and a result conversion layer, the main structure layer includes a pooling layer and a fully connected layer, the data input layer receives the training sample set, and performs multi-level feature extraction on the training sample set under different steps through a convolution layer and a pooling layer to obtain relevant features and perform a dimensionality reduction operation to obtain a reduced dimension feature, and concatenates two features of different steps in the reduced dimension feature set to obtain a concatenated feature and input it into the fully connected layer; the fully connected layer receives the concatenated feature and combines it with an activation function ReLu to obtain a set of risk budget values ​​of each target recommended product in the time series, wherein the sum of all risk budget values ​​is 1, the risk budget layer processes the risk budget value, and obtains a convex optimization result, i.e., the optimal ratio of the target recommended product;

[0229] The result display module 600 visually displays the optimal ratio of the target recommended products.

[0230] Various changes and modifications can be made without departing from the spirit and scope of the present invention, and all equivalent technical solutions also belong to the scope of the present invention.

[0231] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0232] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0233] The present invention is described with reference to the flowcharts and / or block diagrams of the method, terminal device (system), and computer program product according to the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0234] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0235] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0236] It should be noted that:

[0237] The "one embodiment" or "embodiment" mentioned in the specification means that the specific features, structures or characteristics described in conjunction with the embodiment are included in at least one embodiment of the present invention. Therefore, the phrases "one embodiment" or "embodiment" appearing in various places throughout the specification do not necessarily refer to the same embodiment.

[0238] In addition, it should be noted that the shapes and names of the parts and components of the specific embodiments described in this specification may be different. Any equivalent or simple changes made based on the structure, features and principles described in the patent concept of the present invention are included in the protection scope of the patent of the present invention. The technicians in the technical field of the present invention can make various modifications or supplements to the specific embodiments described or replace them in a similar manner, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A financial product recommendation method based on collaborative filtering and neural network, characterized in that: The following steps are involved: Obtain relevant data of the target user, wherein the user-related data at least includes tag-related data and behavior-related data, the behavior-related data includes behavior data and behavior occurrence time, and the behavior data at least includes click browsing records, collection records, and purchase records; Based on the tag-related data, calculate the similarity between the target user and the historical users in the historical user database in terms of their preferences for the target recommended product, and build a similar user set based on the preference similarity; Obtain the number of target recommended products that the target user is interested in and the number of target recommended products that historical users are interested in, determine the total number of target recommended products that the target user and historical users are interested in and the number of target recommended products that they are interested in, and then obtain the first preference similarity between the target user and historical users; Get the time when the target user last acted on the target product and the time when the historical user last acted on the target product, and combine the total number of products and the number of target recommended products of common interest to obtain the similarity of the second preference of the target user and the historical user; Performing weighted processing on the first preference similarity and the second preference similarity to obtain the preference similarity between the target user and the historical users, and constructing a similar user set based on the historical users corresponding to the top several preference similarity scores; The first preference similarity is expressed as follows: The second preference similarity is expressed as follows: The preference similarity is expressed as follows: in, represents the first preference similarity, represents the second preference similarity, Indicates the target user and historical users Similarity of preferences, Indicates the target user The time when the last action on the target product occurred, Indicates historical users The time when the last action on the target product occurred, Indicates Target recommended products, Indicates the target user The number of target recommended products of interest, Indicates historical users The number of target recommended products of interest, Indicates the target user and historical users The number of target recommended products of common interest; Based on the behavior-related data, the initial interest of historical users in the similar user set for the target recommended product is obtained, and the final interest of the target user for the target recommended product is obtained through the preference similarity and the initial interest, and the final interest is corrected to obtain the corrected interest and a target recommended product library is constructed based on the corrected interest; Determine the target user's initial interest in the target recommended product through the target user's click browsing history, collection history, and purchase history; Based on the first preference similarity and the initial interest, the first interest of the target user in the target recommended product is obtained, and through the second preference similarity and the initial interest, the second interest of the target user in the target product is obtained; Based on the first interest level and the second interest level, obtaining a final interest level of the target user for the target product; The final interest degree is corrected by using the Gaussian link function to obtain the corrected interest degree, and the target products corresponding to the top several corrected interest degree scores are used to construct a target recommended product library; The initial interest degree is expressed as follows: The first interest level is expressed as follows: The second interest level is expressed as follows: The final interest level is expressed as follows: The modified interest level is expressed as follows: in, in, Indicates the final interest, represents a set of similar users, represents the user similarity obtained by the maximum likelihood function constructed based on the Gaussian link function, Indicates historical users The average interest score of Indicates the target user Recommend products to target Interest ranking, Indicates the number of target recommended products. Represent the normalized target users Recommend products to target Interest ranking and historical users Recommend products to target Interest ranking, Indicates user similarity, represents the initial interest, They represent the number of click-through records, favorite records, and purchase records respectively. Respectively represent the proportion of click browsing records, collection records and purchase records, Represent the first interest, the second interest and the final interest respectively. represents a set of similar users, represents the first preference similarity, represents the second preference similarity, Indicates the target user and historical users The number of target recommended products of common interest, Indicates Target recommended products, Represents the inverse function of the standard Gaussian distribution; Dynamically update the target recommended product library to obtain an updated target recommended product library, and perform data preprocessing on relevant data of the target recommended products in the updated target recommended product library, wherein the relevant data of the target recommended products include historical yields, historical net capital inflows, and historical scales, to obtain a training sample set; Construct a risk budget pre-training model, train and optimize the risk budget pre-training model through the training sample set to obtain the risk budget model. The risk budget pre-training model includes a data input layer, a main structure layer, a risk budget layer and a result conversion layer. The main structure layer includes a pooling layer and a fully connected layer. The data input layer receives the training sample set, and performs multi-level feature extraction on the training sample set under different steps through the convolution layer and the pooling layer to obtain relevant features and perform dimensionality reduction operations to obtain reduced dimensionality features. Two features of different steps in the reduced dimensionality feature set are spliced ​​to obtain spliced ​​features and input to the fully connected layer; the fully connected layer receives the spliced ​​features and combines them with the activation function ReLu to obtain a set of risk budget values ​​of each target recommended product in the time series, where the sum of all risk budget values ​​is 1. The risk budget layer processes the risk budget value and obtains a convex optimization result, that is, the optimal proportion of the target recommended product. The risk budget layer processes the risk budget value and obtains a convex optimization result. The specific process is as follows: Construct the objective function and make the result of the objective function minimize the prediction error, which is expressed as follows: The constraints are: Where N represents the number of target recommended products in the target recommended product portfolio. represents the combination weight matrix, is the target recommended product covariance matrix, represents the combined variance, Indicates The weight of the target recommended product, Indicates The risk budget value of the target recommended product, and ; Combined with the preset stability condition, the objective function is modified and transformed to obtain the transformed objective function, which is expressed as follows: The constraints are: Based on the solver, the optimal ratio of each target recommended product is obtained by normalizing the results, which is expressed as follows: in, Indicates The weight of the target recommended product after transformation under the preset stability condition, Indicates The optimal ratio of recommended products for each target; The optimal ratio of the target recommended products is visually displayed.

2. The financial product recommendation method based on collaborative filtering and neural network according to claim 1, characterized in that: The following steps are also included: If the target user is a new user, the final interest of the target user in the target product is obtained through the initial similarity and initial interest; The initial similarity is expressed as follows: The final interest level is expressed as follows: in, represents the initial similarity, Indicates the target user For target products The final interest level, represents the initial interest, represents a set of similar users, Respectively represent the target users and historical users level of risk.

3. The financial product recommendation method based on collaborative filtering and neural network according to claim 1, characterized in that: The method of dynamically updating the target recommended product library to obtain an updated target recommended product library includes the following steps: Preset the cluster value of the target recommended product library, randomly select K target recommended products as the initial cluster centers, assign each target recommended product to the cluster to which the nearest cluster center belongs, recalculate the cluster center for each clustering, and repeat the iteration until the cluster center no longer changes; Calculate the Euclidean distance between the current target recommended product and other target recommended products in the same cluster and normalize them; The target recommended products with the highest final similarity scores to the current target recommended product are used as the updated target recommended product library; The Euclidean distance is expressed as follows: The final similarity is expressed as follows: in, Indicates target recommended products The Euclidean distance between represents the final similarity, Indicates Only financial products indicators, Indicates Only financial products indicators, Represents a collection of similar products.

4. The financial product recommendation method based on collaborative filtering and neural network according to claim 1, characterized in that: Train and optimize the risk budget pre-trained model through loss function and optimizer; The optimizer is Adam, which calculates the first-order moment estimate and the second-order moment estimate of the gradient respectively during the back-propagation process and combines the deviation correction and the adaptive learning rate adjustment to update the parameters of the risk budget pre-training model until the optimal parameters are obtained.

5. A financial product recommendation system based on collaborative filtering and neural network, characterized in that: It includes an acquisition processing module, a first processing module, a second processing module, a data update module, a construction training module and a result display module; The data acquisition module acquires relevant data of the target user, wherein the user-related data at least includes tag-related data and behavior-related data, the behavior-related data includes behavior data and behavior occurrence time, and the behavior data at least includes click browsing records, collection records and purchase records; The first processing module calculates the similarity of preferences of the target user and the historical users in the historical user database for the target recommended product based on the tag-related data, and constructs a similar user set based on the preference similarity; Obtain the number of target recommended products that the target user is interested in and the number of target recommended products that historical users are interested in, determine the total number of target recommended products that the target user and historical users are interested in and the number of target recommended products that they are interested in, and then obtain the first preference similarity between the target user and historical users; Get the time when the target user last acted on the target product and the time when the historical user last acted on the target product, and combine the total number of products and the number of target recommended products of common interest to obtain the similarity of the second preference of the target user and the historical user; Performing weighted processing on the first preference similarity and the second preference similarity to obtain the preference similarity between the target user and the historical users, and constructing a similar user set based on the historical users corresponding to the top several preference similarity scores; The first preference similarity is expressed as follows: The second preference similarity is expressed as follows: The preference similarity is expressed as follows: in, represents the first preference similarity, represents the second preference similarity, Indicates the target user and historical users Similarity of preferences, Indicates the target user The time when the last action on the target product occurred, Indicates historical users The time when the last action on the target product occurred, Indicates Target recommended products, Indicates the target user The number of target recommended products of interest, Indicates historical users The number of target recommended products of interest, Indicates the target user and historical users The number of target recommended products of common interest; The second processing module obtains the initial interest of historical users in the similar user set for the target recommended product based on the behavior-related data, obtains the final interest of the target user for the target recommended product through the preference similarity and the initial interest, and corrects the final interest to obtain the corrected interest and constructs a target recommended product library based on the corrected interest; Determine the target user's initial interest in the target recommended product through the target user's click browsing history, collection history, and purchase history; Based on the first preference similarity and the initial interest, the first interest of the target user in the target recommended product is obtained, and through the second preference similarity and the initial interest, the second interest of the target user in the target product is obtained; Based on the first interest level and the second interest level, obtaining a final interest level of the target user for the target product; The final interest degree is corrected by using the Gaussian link function to obtain the corrected interest degree, and the target products corresponding to the top several corrected interest degree scores are used to construct a target recommended product library; The initial interest degree is expressed as follows: The first interest level is expressed as follows: The second interest level is expressed as follows: The final interest level is expressed as follows: The modified interest level is expressed as follows: in, in, Indicates the final interest, represents a set of similar users, represents the user similarity obtained by the maximum likelihood function constructed based on the Gaussian link function, Indicates historical users The average interest score of Indicates the target user Recommend products to target Interest ranking, Indicates the number of target recommended products, Represent the normalized target users Recommend products to target Interest ranking and historical users Recommend products to target Interest ranking, Indicates user similarity, represents the initial interest, They represent the number of click-through records, favorite records, and purchase records respectively. Respectively represent the proportion of click browsing records, collection records and purchase records, Represent the first interest, the second interest and the final interest respectively. represents a set of similar users, represents the first preference similarity, represents the second preference similarity, Indicates the target user and historical users The number of target recommended products of common interest, Indicates Target recommended products, Represents the inverse function of the standard Gaussian distribution; A data updating module dynamically updates the target recommended product library to obtain an updated target recommended product library, and performs data preprocessing on relevant data of the target recommended products in the updated target recommended product library, wherein the relevant data of the target recommended products include historical yields, historical net capital inflows, and historical scales, to obtain a training sample set; Construct a training module, construct a risk budget pre-training model, train and optimize the risk budget pre-training model through the training sample set to obtain the risk budget model. The risk budget pre-training model includes a data input layer, a main structure layer, a risk budget layer and a result conversion layer. The main structure layer includes a pooling layer and a fully connected layer. The data input layer receives the training sample set, and performs multi-level feature extraction on the training sample set under different steps through the convolution layer and the pooling layer to obtain relevant features and perform dimensionality reduction operations to obtain reduced dimensionality features. The two features of the reduced dimensionality feature set that are in different steps are spliced ​​to obtain spliced ​​features and input into the fully connected layer; the fully connected layer receives the spliced ​​features and combines them with the activation function ReLu to obtain a set of risk budget values ​​of each target recommended product in the time series, where the sum of all risk budget values ​​is 1. The risk budget layer processes the risk budget value and obtains a convex optimization result, i.e., the optimal proportion of the target recommended product. The risk budget layer processes the risk budget value and obtains a convex optimization result. The specific process is as follows: Construct the objective function and make the result of the objective function minimize the prediction error, which is expressed as follows: The constraints are: Where N represents the number of target recommended products in the target recommended product portfolio. represents the combination weight matrix, is the target recommended product covariance matrix, represents the combined variance, Indicates The weight of the target recommended product, Indicates The risk budget value of the target recommended product, and ; Combined with the preset stability condition, the objective function is modified and transformed to obtain the transformed objective function, which is expressed as follows: The constraints are: Based on the solver, the optimal ratio of each target recommended product is obtained by normalizing the results, which is expressed as follows: in, Indicates The weight of the target recommended product after transformation under the preset stability condition, Indicates The optimal ratio of recommended products for each target; The result display module visualizes the optimal ratio of the target recommended products.

6. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

7. A financial product recommendation device based on collaborative filtering and neural network, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 4 is implemented.

Images