Financial product data pushing method, device and equipment and storage medium

Abstract

The application provides a financial product data pushing method, device and equipment and a storage medium, and belongs to the technical field of big data. The method comprises the following steps: obtaining user basic data, obtaining product attribute data, obtaining graph feature data, determining graph modeling data, determining ontology domain objects, and determining ontology domain objects. Ontology extraction processing is performed on the graph modeling data according to the ontology domain objects, and ontology concept data is obtained; instance data is obtained from pre-stored product data; resource description framework data is obtained; a first knowledge graph is generated; a target learning model is constructed according to a pre-built multi-relation co-occurrence matrix and a pre-stored graph convolutional neural network model; a second knowledge graph is generated; a target recommendation model is generated; target user data is input into the target recommendation model, and target financial product data is generated; and the target financial product data is pushed to a target user terminal. The application improves the accuracy and interpretability of the pushed financial product data.

Description

Technical Field

[0001] This application relates to the field of big data technology, and in particular to a method, apparatus, device and storage medium for pushing financial product data. Background Technology

[0002] Currently, with the rapid development of the financial industry, a wide variety of financial products have emerged to meet people's growing financial management needs. Faced with this diverse array of financial products, how can we improve the accuracy of the financial products recommended to users to meet their requirements?

[0003] In existing technologies, the push of financial products generally begins by collecting information on users' product preferences through questionnaires, and then analyzing the collected information to determine whether the user is conservative, stable, or aggressive, and then pushes the corresponding financial products based on the user's type.

[0004] However, the inventors have discovered that the prior art has at least the following technical problems: when using the existing financial product data push method to push financial product data, because users have too little understanding of most financial products, the matching degree between the pushed financial products and the users' actual needs is low, resulting in problems with low accuracy and interpretability of the pushed financial product data. Summary of the Invention

[0005] This application provides a method, apparatus, device, and storage medium for pushing financial product data, in order to solve the problem of low accuracy and interpretability of the pushed financial product data.

[0006] On the one hand, this application provides a method for pushing financial product data, including:

[0007] Extract and process the pre-stored user data to obtain basic user data, and extract and process the pre-stored product data to obtain product attribute data;

[0008] Feature processing is performed on the product attribute data and the user basic data to obtain graph feature data;

[0009] Based on the spectral feature data and the pre-stored product data, determine the spectral modeling data;

[0010] Based on the graph modeling data, determine the ontology domain objects;

[0011] The graph modeling data is processed by ontology extraction based on the ontology domain objects to obtain ontology concept data.

[0012] Start the pre-stored ontology construction tool and obtain instance data from the pre-stored product data according to the ontology concept data;

[0013] The instance data is converted to obtain resource description framework data.

[0014] The pre-stored import tool is activated to import the resource description framework data into the preset graph database, generating the first knowledge graph.

[0015] Based on the pre-built multi-relation co-occurrence matrix and the pre-stored graph convolutional neural network model, construct the target learning model;

[0016] A second knowledge graph is generated based on the first knowledge graph and the target learning model;

[0017] The second knowledge graph is introduced into the pre-stored recommendation model to generate the target recommendation model;

[0018] Acquire target user data and input the target user data into the target recommendation model to generate target financial product data;

[0019] The target financial product data is pushed to the target user's terminal.

[0020] In one possible implementation, the step of extracting and processing pre-stored user data to obtain user basic data, and extracting and processing pre-stored product data to obtain product attribute data, includes: executing a pre-stored preprocessing script to extract product type data, product price change data, product rating data, and product source data from the pre-stored product data, and determining the product type data, product price change data, product rating data, and product source data as product attribute data; executing the pre-stored preprocessing script to extract user description data and user preference data from the user basic data, and determining the user description data and user preference data as user basic data.

[0021] In one possible implementation, the step of performing feature processing on the product attribute data and the user basic data to obtain graph feature data includes: classifying the product attribute data and the user basic data to obtain discrete data and identity identification data; normalizing the discrete data to obtain normalized feature data; embedding the identity identification data to obtain identity identification feature data; and obtaining graph feature data based on the normalized feature data and the identity identification feature data.

[0022] In one possible implementation, determining the graph modeling data based on the graph feature data and the pre-stored product data includes: determining graph entity data based on the graph feature data, and determining graph attribute data based on the graph entity data and the graph feature data; activating a pre-stored semantic analysis tool to perform semantic analysis on the pre-stored product data and the graph entity data to generate entity relationship data; performing knowledge fusion processing on the pre-stored product data to determine standard graph entity data; and storing the standard graph entity data, the graph attribute data, and the entity relationship data as graph modeling data.

[0023] In one possible implementation, determining the target learning model based on a pre-built multi-relation co-occurrence matrix and a pre-stored graph convolutional neural network model includes: determining multiple nodes based on the pre-stored graph convolutional neural network model; determining the importance coefficient and relationship type data of the relationship between every two nodes based on the pre-built multi-relation co-occurrence matrix and the multiple nodes; determining the importance value of one node to another in every pair of nodes based on all the relationship importance coefficients and relationship type data; performing message passing on the pre-stored training data at each node based on all the importance values ​​and the pre-stored graph convolutional neural network model to generate the encoding layer node representation data corresponding to each node; and when training the pre-stored graph convolutional neural network model, using a pre-stored block diagonal matrix to weight the pre-stored model weight matrix W. r Regularization is performed to generate a target weight matrix for each node; based on the node representation data of the encoding layer, the target weight matrix of each node, the pre-stored graph convolutional neural network model, and the pre-stored knowledge graph, the optimization parameters of the knowledge graph are determined; the pre-stored graph convolutional neural network model is trained according to the optimization parameters of the knowledge graph to construct a target learning model.

[0024] In one possible implementation, the formula for determining the importance coefficient and relationship type data between every two nodes based on the pre-built multi-relation co-occurrence matrix and the multiple nodes is as follows:

[0025]

[0026] In the formula, Let R be a pre-built multi-relation co-occurrence matrix of size R×R with 1s on the diagonal, where R is a positive integer and r is a positive integer. i Let r be the relationship type between the i-th batch of nodes of node e. j Let r and j be the relationship type between nodes in the j-th batch of node e, where r and j are both positive integers, and sum(e, r) = r(j, j). i Let ) be the relation type of the neighboring nodes of node e. i The sum of the number of sides, sum(e,r) jLet ) be the relation type of the neighboring nodes of node e. j The sum of the number of sides, For relation type r i For relation type r j The importance coefficient, |R|, is the total number of all relation types surrounding node e.

[0027] In one possible implementation, the importance value of one node to the other in every pair of nodes is determined based on all relationship importance coefficients and node relationships. The formula for this calculation is as follows:

[0028]

[0029] In the formula, For node e j In relation r i Next, for target node e h Importance value, For node e j The number of surrounding nodes.

[0030] In one possible implementation, the pre-stored graph convolutional neural network model includes multiple encoding layers; correspondingly, the calculation formula for generating encoding layer node representation data for each node by message passing the pre-stored training data at each node based on all importance values ​​and the pre-stored graph convolutional neural network model is as follows:

[0031]

[0032] In the formula, e h (l+1) For a node e in the (l+1)th layer of the pre-stored encoding layer of a convolutional neural network h The coding layer nodes represent data. For node e h The number of surrounding nodes, For different relation types r, the weight matrix that needs to be learned is... This represents the weight matrix of nodes that have no relation to other nodes. For pre-stored nodes e in the l-th layer of the encoding layer of a convolutional neural network h For node e j The importance value.

[0033] In one possible implementation, the formula for calculating the target weight matrix for each node by performing a regularization operation on the pre-stored model weight matrix using a pre-stored block diagonal matrix is ​​as follows:

[0034]

[0035] In the formula, b∈[1,B), then This is a block diagonal matrix of the pre-stored model weight matrix.

[0036] In one possible implementation, the pre-stored knowledge graph includes a valid set of triples, and the pre-stored graph convolutional neural network model includes multiple convolutional kernels;

[0037] Accordingly, the calculation formula for determining the knowledge graph optimization parameters based on the encoding layer node representation data, the target weight matrix of each node, the pre-stored graph convolutional neural network model, and the pre-stored knowledge graph is as follows:

[0038] f(v) = concat((e h ,e r ,e t )*ω m ))·W

[0039] In the formula, v = (e h ,e r ,e t ), v∈V, V is the set of valid triples in the pre-stored knowledge graph, V′ represents the set of negative samples, * is the convolution operation, ω m denoted as the m-th convolutional kernel in the pre-stored graph convolutional neural network model, where m is a positive integer and W is the target weight matrix for each node.

[0040] In one possible implementation, the formula for training the pre-stored graph convolutional neural network model based on the knowledge graph optimization parameters is as follows:

[0041]

[0042] In the formula, It is 1 or -1, when When, it represents a triple (e h ,e r ,e t ) is a positive sample, when When, it represents a triple (e h ,e r ,e t ) represents a negative sample, L represents the loss function, and λ is a preset loss constant.

[0043] In one possible implementation, the pre-stored recommendation model includes a bipartite graph; correspondingly, the step of introducing the second knowledge graph into the pre-stored recommendation model to generate the target recommendation model includes: fusing the bipartite graph and the second knowledge graph to generate a collaborative knowledge graph; splitting the collaborative knowledge graph into multiple isomorphic subgraphs according to a preset meta-path; determining the feature parameters of each isomorphic subgraph based on the multiple isomorphic subgraphs; performing aggregation and fusion processing on each isomorphic subgraph to obtain knowledge graph embedding representation data; and importing the knowledge graph embedding representation data into the pre-stored recommendation model to generate the target recommendation model.

[0044] Secondly, this application provides a financial product data push device, comprising:

[0045] The data processing module is used to extract and process pre-stored user data to obtain basic user data, and to extract and process pre-stored product data to obtain product attribute data.

[0046] The data processing module is also used to perform feature processing on the product attribute data and the user basic data to obtain map feature data;

[0047] The product knowledge graph construction module is used to determine the graph modeling data based on the graph feature data and the pre-stored product data.

[0048] The product knowledge graph construction module is also used to determine ontology domain objects based on the graph modeling data;

[0049] The product knowledge graph construction module is also used to perform ontology extraction processing on the graph modeling data according to the ontology domain objects to obtain ontology concept data;

[0050] The product knowledge graph construction module is also used to launch a pre-stored ontology construction tool and obtain instance data from the pre-stored product data according to the ontology concept data;

[0051] The product knowledge graph construction module is also used to perform format conversion processing on the instance data to obtain resource description framework data;

[0052] The product knowledge graph construction module is also used to start the pre-stored import tool to import the resource description framework data into the preset graph database and generate the first knowledge graph;

[0053] The push module is used to construct a target learning model based on a pre-built multi-relation co-occurrence matrix and a pre-stored graph convolutional neural network model;

[0054] The push module is further configured to generate a second knowledge graph based on the first knowledge graph and the target learning model;

[0055] The push module is also used to introduce the second knowledge graph into the pre-stored recommendation model to generate the target recommendation model;

[0056] The push module is also used to acquire target user data and input the target user data into the target recommendation model to generate target financial product data;

[0057] The push module is also used to push the target financial product data to the target user terminal.

[0058] Thirdly, this application provides a financial product data push device, including: a processor, and a memory communicatively connected to the processor;

[0059] The memory stores computer-executed instructions;

[0060] The processor executes computer execution instructions stored in the memory to implement the financial product data push method as described in the first aspect.

[0061] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the financial product data push method as described in the first aspect.

[0062] Fifthly, this application provides a computer program product, including computer instructions, which, when executed by a processor, implement the financial product data push method described in the first aspect.

[0063] The financial product data push method, apparatus, equipment, and storage medium provided in this application preprocess pre-stored user data and pre-stored product data on the backend server, and then perform feature processing to obtain graph feature data, which is used to construct a knowledge graph. This reduces the impact of data unrelated to the financial product push on the financial product data results, and initially improves accuracy. Furthermore, the process of constructing the first knowledge graph is reduced to four steps: determining the ontology domain, extracting ontology concept data and starting a pre-stored ontology construction tool to obtain instance data from pre-stored product data according to the ontology concept data, formatting and converting the instance data to obtain resource description framework data, and starting a pre-stored import tool to import the resource description framework data into a preset graph database to generate the first knowledge graph. This reduces the steps in the knowledge graph construction process, achieves rapid and batch data import, and improves the efficiency of knowledge graph construction. Furthermore, a target learning model is constructed based on a pre-built multi-relation co-occurrence matrix and a pre-stored graph convolutional neural network model. A second knowledge graph is then generated based on this target learning model and the first knowledge graph. Since the target learning model can more clearly distinguish the importance of different relationships surrounding an entity, the second knowledge graph can provide higher-quality training data for subsequent target recommendation models. Finally, the second knowledge graph is used to pre-store a recommendation model to generate the target recommendation model. This improves the accuracy and interpretability of the target financial product data generated by the target recommendation model. Attached Figure Description

[0064] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0065] Figure 1 This is a schematic diagram illustrating an application scenario of the financial product data push method provided in the embodiments of this application.

[0066] Figure 2 A flowchart illustrating the financial product data push method provided in this application embodiment;

[0067] Figure 3 This is a schematic diagram of a target learning model architecture provided in an embodiment of this application;

[0068] Figure 4 A schematic diagram of a target recommendation model architecture provided in an embodiment of this application;

[0069] Figure 5 This is a schematic diagram of the structure of the financial product data push device provided in the embodiments of this application;

[0070] Figure 6 This is a schematic diagram of the structure of the financial product data push device provided in the embodiments of this application.

[0071] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0072] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0073] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0074] Currently, with the rapid development of the financial industry, various financial products have emerged to meet people's growing financial management needs. Faced with a wide variety of financial products, how can we improve the accuracy of financial product recommendations to users to meet their requirements? In existing technologies, financial product recommendations typically begin by collecting information on users' product preferences through questionnaires, then analyzing the collected information to determine whether the user is conservative, moderate, or aggressive, and then recommending corresponding financial products based on that user type. The inventors have discovered that due to the rapid development and fast-paced changes in the financial market, recommending financial products solely based on user type can lead to a low match between the recommended financial product data and the user's chosen financial products, resulting in low accuracy and interpretability of the recommended financial product data.

[0075] To address the aforementioned technical problems, this application provides the following technical concept: A knowledge graph is constructed using an improved four-step method based on user information and financial product information. Then, an improved knowledge graph representation learning model, GCN-CM, is used to extract knowledge graph information, providing higher-quality knowledge graph training features for the recommendation model. The collaborative knowledge graph, first fused by the recommendation model, is then divided into multiple isomorphic subgraphs. In each subgraph, an attention mechanism is used to aggregate multi-hop neighbor information, outputting recommended financial product information. Since the knowledge graph contains interconnected graph information, the accuracy and interpretability of the recommended financial product data can be improved based on this graph information.

[0076] The data transmission method provided in this application is intended to solve the above-mentioned technical problems of the prior art.

[0077] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0078] Please refer to Figure 1 , Figure 1 This is a schematic diagram illustrating an application scenario of the financial product data push method provided in the embodiments of this application, such as... Figure 1 As shown, it includes: client 101 and server 102.

[0079] The client 101 can be used to display and allow customers to view financial products, as well as to allow users to input corresponding parameters and select information. The client 101 can be a server, such as a computer, or a mobile terminal, such as a tablet, mobile phone, or laptop. The server 102 can be used to access real-time product data of relevant financial products in the financial market, interact with the client 101, and perform corresponding operations based on the information sent by the client 101 to push corresponding financial product data to the client 101. The server 102 can be a physical server or a cloud server.

[0080] Please refer to Figure 2 , Figure 2 This is a flowchart illustrating the financial product data push method provided in this application embodiment. The execution entity in this embodiment can be... Figure 1 The server 102 in the illustrated embodiment can also be other computer-related devices, and this embodiment does not impose any special restrictions on it. Figure 2 As shown, the data push method for this financial product includes:

[0081] S201: Extract and process the pre-stored user data to obtain basic user data, and extract and process the pre-stored product data to obtain product attribute data.

[0082] In this embodiment, the pre-stored user data can be user-related data with user permission or user data collected through questionnaires. Extraction processing can be performed using data extraction software for automatic data extraction. Basic user data can be data describing basic user information, such as the user's age, gender, or occupation. Pre-stored product data can be financial product-related data stored on the server side or financial product-related data retrieved by the server from the financial market. Product attribute data can be data used to analyze the attribute requirements of financial products. For example, the type of financial product, its price change over the past year, and its net asset value.

[0083] S202: Perform feature processing on product attribute data and user basic data to obtain graph feature data.

[0084] In this embodiment, feature processing may include normalization processing and embedding processing. The graph feature data may be feature data used to construct a knowledge graph.

[0085] Specifically, in an optional embodiment of this application, step S202 includes:

[0086] S202a: Classify and process product attribute data and user basic data to obtain discrete data and identity identification data.

[0087] In this embodiment, discrete data can be data that can only be calculated using natural numbers or integer units, such as the number of financial products a user holds, the risk level of the financial products, and the user's age. Identification data can be a unique identifier representing a user or financial product, such as the product number and the user's number.

[0088] S202b: Normalize discrete data to obtain normalized feature data.

[0089] S202c: Embed the identity data to obtain identity feature data.

[0090] In this embodiment, normalization can be the process of limiting numerical data or data that can be represented numerically to a certain range, such as the range being the interval 0 to 1 or the interval -1 to 1. Embedding processing can refer to the process of converting discrete variables into continuous vectors. For example, the user's gender is male, which corresponds to the number 1, and female, which corresponds to the number 0. The user is an adult, which corresponds to the number 1, and a minor, which corresponds to the number 0. Then, user A, who is an adult, will obtain a vector (1, 0) after embedding processing.

[0091] In this embodiment, the normalized feature data can be numerical data that is restricted to a certain numerical range, and the identity feature data can be vector data used to identify identity.

[0092] S202d: Obtain the map feature data based on the normalized feature data and the identity feature data.

[0093] In this embodiment, the graph feature data can be feature class data used to extract the original data for constructing the knowledge graph. For example, the graph feature data can be vectors and numerical segments representing semantics.

[0094] S203: Determine the map modeling data based on the map feature data and the pre-stored product data.

[0095] In this embodiment, the graph modeling data can be the component data that makes up the knowledge graph, for example, the graph feature data can be triple data. The entire process of obtaining the graph feature data can include attribute extraction, relation extraction, and entity extraction. Among them, attribute extraction and entity extraction belong to named entity recognition. For example, after relation extraction and entity extraction, information entities such as the salesperson in charge of the financial product and the company to which the financial product belongs are obtained.

[0096] Specifically, in an optional embodiment of this application, step S203 includes: determining map entity data based on map feature data, and determining map attribute data based on map entity data and map feature data;

[0097] S203a: Activate the pre-stored semantic analysis tool to perform semantic analysis on the pre-stored product data and graph entity data, and generate entity relationship data.

[0098] In this embodiment, the pre-stored semantic analysis tool can be used to extract semantic relationships between two or more entities. This pre-stored semantic analysis tool can be a pre-packaged tool library. For example, it could be NLTK (Natural Language Toolkit) or Stanford CoreNLP (Stanford Core Natural Language Processing). Entity relationship data can be the semantic relationships between two or more entities. For example, "The product manager of User A's No. 1 financial product is Zhang San, and Zhang San works for Company X." Then, the semantic relationship between entity "Zhang San" and entity "No. 1 financial product" is "product manager is," and the semantic relationship between entity "Zhang San" and entity "Company X" is "works for."

[0099] S203b: Perform knowledge fusion processing on pre-stored product data to determine standard map entity data.

[0100] S203c: Store standard graph entity data, graph attribute data, and entity relationship data as graph modeling data.

[0101] In this embodiment, knowledge fusion processing can unify the expression of entities and eliminate contradictions and ambiguities in pre-stored product data to filter similar entities in different texts. Knowledge fusion processing can adopt a manually predefined standard approach; for example, if "product manager" and "manager" appear simultaneously, then such information can be uniformly referred to as "product manager".

[0102] S204: Determine the ontology domain objects based on the graph modeling data.

[0103] In this embodiment, the ontology domain object can be a general term describing the professional field or category of the ontology to be constructed. For example, the determined ontology domain object can be a financial product. The process of determining the ontology domain object can be done manually or automatically by software.

[0104] S205: Perform ontology extraction processing on the graph modeling data according to the ontology domain objects to obtain ontology concept data.

[0105] In this embodiment, ontology extraction can be performed by indexing and extracting data from graph modeling data, specifically using extraction tools such as Python scripts. Ontology concept data can be data related to an ontology domain. For example, if the ontology domain is financial products, then the ontology concept data can include the scope of purchase or sale of financial products, the release time of financial products, the scale of financial products, the custodians and managers of financial products, etc.

[0106] S206: Start the pre-stored ontology construction tool and obtain instance data from the pre-stored product data according to the ontology concept data.

[0107] In this embodiment, the pre-stored ontology building tool can be an ontology modeling tool. For example, the pre-stored ontology building tool can be the Cellfie plugin, a built-in module in the ontology editing and knowledge acquisition software Protégé, which is used to import spreadsheet data into the OWL ontology. Specifically, the process of using the pre-stored ontology building tool can be as follows: first, click "Create axioms from Excel" in the toolbar to import financial product data in spreadsheet format to form instance data.

[0108] S207: Perform format conversion processing on the instance data to obtain resource description framework data.

[0109] In this embodiment, the format conversion process can be carried out by using a format conversion tool to convert data from one data format to another. For example, by using the rdf2rdf package, a resource framework definition tool, and then entering the command line, the generated OWL (Ontology Web Language) data can be converted into RDF (Resource Description Framework) data.

[0110] S208: Start the pre-stored import tool to import the resource description framework data into the preset graph database and generate the first knowledge graph.

[0111] In this embodiment, the pre-stored import tool can be a script or an executable statement, such as a Cypher statement querying a graph database. The RDF data obtained from the conversion process in step S207 is imported into the pre-stored graph database. The pre-stored graph data can be a network-structure-oriented database, for example, the pre-stored graph database can be a Neo4j graph database.

[0112] S209: Construct a target learning model based on a pre-built multi-relation co-occurrence matrix and a pre-stored graph convolutional neural network model.

[0113] In this embodiment, the multi-relation co-occurrence matrix can be a matrix with the same number of rows and columns, and a diagonal of 1. The pre-stored graph convolutional neural network model (GCN) can be a graph neural network employing convolution operations. The target learning model can be a knowledge graph representation learning model.

[0114] Specifically, in an optional embodiment of this application, step S209 includes:

[0115] S209a: Determine multiple nodes based on the pre-stored graph convolutional neural network model.

[0116] S209b: Based on the pre-built multi-relation co-occurrence matrix and multiple nodes, determine the importance coefficient and relationship type data of the relationship between every two nodes.

[0117] In this embodiment, the input to the pre-stored graph convolutional neural network model can be the nodes and edges of the neural network, and the output can be multiple nodes, each including feature information of its own nodes. The pre-built multi-relation co-occurrence matrix can be a manually set co-occurrence matrix used for multi-relation data modeling to distinguish the importance of different relations surrounding an entity.

[0118] Based on the above embodiments, as an optional embodiment of this application, the calculation formula for step S209b is as follows:

[0119]

[0120] In the formula, Let R be a pre-built multi-relation co-occurrence matrix of size R×R with 1s on the diagonal, where R is a positive integer and r is a positive integer. i Let r be the relationship type between the i-th batch of nodes of node e. j Let r and j be the relationship type between nodes in the j-th batch of node e, where r and j are both positive integers, and sum(e, r) = r(j, j). i Let ) be the relation type of the neighboring nodes of node e. i The sum of the number of sides, sum(e,r) j Let ) be the relation type of the neighboring nodes of node e. j The sum of the number of sides, For relation type r i For relation type r j The importance coefficient, |R|, is the total number of all relation types surrounding node e.

[0121] In this embodiment, the relationship type can be a representation describing the relationship between two nodes, such as "Zhang San's financial product is product A". Based on the context information, the relationship type can be expressed as "user, financial product, product A".

[0122] For example, when node e in a group of multiple nodes is connected through relation r i Establish connections with nodes e1 and e2 through relation r j If a connection is established with node e3, then according to the matrix calculation method, C[r] i ][r j ]=C[r j ][r i ] = 2, and so on, using the formula in the above embodiments. Calculate the largest value in the entire multi-relation co-occurrence matrix.

[0123] S209c: Based on all relation importance coefficients and relation type data, determine the importance value of one node to the other in every pair of nodes.

[0124] In this embodiment, the relation importance coefficient represents the importance of relation y, which is connected to the target node via relation x. Relationship y is assigned a higher weight because it frequently co-occurs with both neighboring nodes and relation x. Relationship type data can include multiple relation types, such as relation type r. i and relation type r j .

[0125] Based on the above embodiments, in an optional embodiment of this application, in step S209, the formula for calculating the importance value of one node to the other in every two nodes is as follows:

[0126]

[0127] In the formula, For node e j In relation r i Next, for target node e h Importance value, For node e j The number of surrounding nodes.

[0128] S209d: Based on all importance values ​​and the pre-stored graph convolutional neural network model, the pre-stored training data is passed through messages at each node to generate the encoding layer node representation data corresponding to each node.

[0129] In this embodiment, the pre-stored training data can be a matrix composed of feature data that has undergone pre-processing data extraction. For example, the pre-stored training data may include entity feature data such as usernames, real-time financial product data, and product managers who manage financial products.

[0130] In an optional embodiment of this application, the pre-stored graph convolutional neural network model includes multiple encoding layers. In step S209d, the calculation formula for generating the encoding layer node representation data corresponding to each node is as follows:

[0131]

[0132] In the formula, e h (l+1) For a node e in the (l+1)th layer of the pre-stored encoding layer of a convolutional neural network h The coding layer nodes represent data. For node e h The number of surrounding nodes, For different relation types r, the weight matrix that needs to be learned is... This represents the weight matrix of nodes that have no relation to other nodes. For pre-stored nodes e in the l-th layer of the encoding layer of a convolutional neural network h For node e j The importance value.

[0133] In this embodiment, the type with no relationship with other nodes refers to self-connection. Multiple coding layers can be used to transmit multi-hop information. When the first layer represents the input features and a linear transformation is used when processing node self-connections, the input features of the first layer are one-hot encoded.

[0134] S209e: When training a pre-stored graph convolutional neural network model, use the diagonal matrix of the pre-stored block to adjust the weight matrix W of the pre-stored model. r Perform regularization to generate the target weight matrix for each node.

[0135] In this embodiment, the pre-stored block diagonal matrix can be a block matrix composed of square diagonal blocks. Regularization refers to the process of transforming a matrix into an invertible matrix.

[0136] As an optional embodiment of this application, the calculation formula for generating the target weight matrix for each node in step S209e is as follows:

[0137]

[0138] In the formula, b∈[1,B), then This is a block diagonal matrix of the pre-stored model weight matrix.

[0139] S209f: Determine the optimization parameters of the knowledge graph based on the representation data of the encoding layer nodes, the target weight matrix of each node, the pre-stored graph convolutional neural network model, and the pre-stored knowledge graph.

[0140] In this embodiment, the knowledge graph optimization parameters can be the knowledge graph unit data that best meets the requirements for the correct composition of the knowledge graph. For example, the higher the score of the triple calculated by the scoring optimization parameters, the more likely it is to be a correct triple that can be used to construct the knowledge graph.

[0141] In an optional embodiment of this application, the calculation formula for determining the knowledge graph optimization parameters in step S209f is as follows:

[0142] f(v) = concat((e h ,e r ,e t )*ω m ))·W

[0143] In the formula, v = (e h ,e r ,e t ), v∈V, V is the set of valid triples in the pre-stored knowledge graph, V′ represents the set of negative samples, * is the convolution operation, ω m denoted as the m-th convolutional kernel in the pre-stored graph convolutional neural network model, where m is a positive integer and W is the target weight matrix for each node.

[0144] In this embodiment, V′, relative to V, refers to some triples that should not be associated in the knowledge graph during model training, which are marked as negative samples. This is used to determine the calculation of the loss function in the formula for optimizing the knowledge graph parameters and to iteratively optimize the model parameters.

[0145] S209g: Train a pre-stored graph convolutional neural network model based on the optimized parameters of the knowledge graph to construct a target learning model.

[0146] Please refer to Figure 3 , Figure 3 This is a schematic diagram of a target learning model architecture provided in an embodiment of this application. Figure 3 As shown, the target learning model includes an input layer, an encoding layer, a hidden layer, and a decoding layer. In this embodiment, the target learning model is obtained by continuously training a pre-stored graph convolutional neural network model.

[0147] Based on the above embodiments, in an optional embodiment of this application, the calculation formula for training the pre-stored graph convolutional neural network model in step S209g is as follows:

[0148]

[0149] In the formula, It is 1 or -1, when When, it represents a triple (e h ,e r ,e t ) is a positive sample, when When, it represents a triple (e h ,e r ,e t ) represents a negative sample, L represents the loss function, and λ is a preset loss constant.

[0150] In this embodiment, the soft-margin loss function can be used to train the pre-stored graph convolutional neural network model. Positive samples refer to sample data belonging to the same class, and negative samples refer to sample data not belonging to a certain class. The preset loss constant can be a manually set value, for example, the preset loss constant can be 1 or 0.5.

[0151] S2010: Generate a second knowledge graph based on the first knowledge graph and the target learning model.

[0152] In this embodiment, the second knowledge graph is composed of all the graph units in the first knowledge graph after being filtered and optimized by the above embodiments.

[0153] S2011: Introduce the second knowledge graph into the pre-stored recommendation model to generate the target recommendation model.

[0154] In this embodiment, the pre-stored recommendation model can be a recommendation model that integrates a collaborative knowledge graph, such as the CKG-HAN (Collaborative Knowledge Graph Hierarchical Attention Network) model. The target recommendation model can be a recommendation model trained using an external knowledge graph and the pre-stored recommendation model.

[0155] S2012: Obtain target user data and input the target user data into the target recommendation model to generate target financial product data.

[0156] In this embodiment, target user data refers to relevant data of the target users to be pushed to, such as basic descriptive data related to users in a questionnaire for evaluating financial products before purchase. After the target user data is input into the target recommendation model, the model calculates and outputs target financial product data. This target financial product data can be personalized financial product data tailored to the target users. For example, the target financial product data may include a recommendation list of descriptive data such as basic information of multiple financial products, purchase time, purchase price, sale time and sale price, and expected returns.

[0157] S2013: Push the target financial product data to the target user's terminal.

[0158] In this embodiment, the target financial product can be pushed via communication transmission, and the target user terminal can include the target user's mobile phone, computer, tablet or smartwatch, etc., terminal devices capable of displaying the target financial product data.

[0159] In summary, the financial product data push method provided in this application first preprocesses the pre-stored user data and pre-stored product data on the backend server, and then performs feature processing to obtain graph feature data, which is used to construct a knowledge graph. This reduces the impact of data unrelated to the financial product push on the financial product data results, and initially improves accuracy. Furthermore, the process of constructing the first knowledge graph is reduced to four steps: determining the ontology domain, extracting ontology concept data and starting a pre-stored ontology construction tool to obtain instance data from the pre-stored product data according to the ontology concept data, formatting and converting the instance data to obtain resource description framework data, and starting a pre-stored import tool to import the resource description framework data into a preset graph database to generate the first knowledge graph. This reduces the steps in the knowledge graph construction process, achieves rapid and batch data import, and improves the efficiency of knowledge graph construction. Furthermore, a target learning model is constructed based on a pre-built multi-relation co-occurrence matrix and a pre-stored graph convolutional neural network model. Then, a second knowledge graph is generated based on the target learning model and the first knowledge graph. Because this target learning model can more clearly distinguish the importance of different relationships surrounding an entity, the second knowledge graph can be used to provide higher-quality training data for subsequent target recommendation models. Finally, a pre-stored recommendation model from the second knowledge graph is introduced to generate the target recommendation model. The target financial product data generated using this target recommendation model has higher accuracy and interpretability.

[0160] Based on the above embodiments, as an optional embodiment of this application, step S201 specifically includes:

[0161] S201a: Execute the pre-stored preprocessing script to extract product type data, product price change data, product rating data, and product source data from the pre-stored product data, and identify the product type data, product price change data, product rating data, and product source data as product attribute data.

[0162] S201b: Execute the pre-stored preprocessing script to extract user description data and user preference data from the user basic data, and determine the user description data and user preference data as the user basic data.

[0163] In this embodiment, the pre-stored preprocessing script can be a data scraping Python script. Product type data can be the type obtained according to the industry classification of the product. For example, product types can include machinery, liquor, services, and semiconductors. Product price change data can be the price increase and decrease of the product within a fixed time period. Product rating data can be the level of whether the product is worth purchasing. Product source data can include the company that owns the product, the product manager in charge of the product, and the market size of the product. User base data can include: age, income, occupation, and personal preferences for purchasing financial products.

[0164] Please refer to Figure 4 , Figure 4 This is a schematic diagram of a target recommendation model architecture provided in an embodiment of this application. Figure 3 As shown, the target recommendation model includes an input layer, a node attention layer, a higher-order information aggregation layer, a relation attention layer, and an output layer. The input layer comprises a collaborative knowledge graph formed by fusing a preprocessed bipartite graph and a preprocessed second knowledge graph. This collaborative knowledge graph is divided into multiple isomorphic subgraphs based on manually constructed meta-paths and input to the node attention layer, then passes through the higher-order information aggregation layer, and finally outputs through the relation attention layer.

[0165] Specifically, based on the above embodiments, in an optional embodiment of this application, step S2011 includes:

[0166] Step a: Merge the bipartite graph and the second knowledge graph to generate a collaborative knowledge graph.

[0167] Step b: Divide the collaborative knowledge graph into multiple isomorphic subgraphs according to the preset meta-path.

[0168] Step c: Determine the feature parameters of each isomorphic subgraph based on multiple isomorphic subgraphs.

[0169] Step d: Aggregate and merge each isomorphic subgraph to obtain knowledge graph embedding representation data.

[0170] Step e: Import the knowledge graph embedding representation data into the pre-stored recommendation model to generate the target recommendation model.

[0171] In this embodiment, step a is performed in the input layer, and step b is performed in the node attention layer. The calculation process of the node attention layer is as follows: A: Since the nodes in the knowledge graph are heterogeneous, and different types of nodes have different feature spaces, a transformation matrix is ​​first needed to project different types of LeakyReLU features into the same feature space. The calculation formula for the transformation process is:

[0172]

[0173] Where: h i 'Indicates projection features; Represents the transformation matrix; h i This represents the original features. Through transformation operations, the node attention layer can handle different types of nodes.

[0174] B: When considering the importance of neighboring nodes to the target node, the attention mechanism not only accelerates the model training process but also effectively captures local information, making it well-suited for learning the importance of neighboring nodes to the target node. The formula for calculating the attention coefficient is as follows:

[0175]

[0176] In the formula: Indicates the attention coefficient. Let represent a learnable vector used to adjust the shape of a vector, Φ represent the metapath, || represent the concatenation operation, and h' j The characteristics of neighbor node j can be represented by the activation function LeakyReLU in this embodiment.

[0177] like Figure 3 As shown, multiple high-order aggregation layers can be stacked. Within a high-order aggregation layer, multiple layers are used... express Perform normalization. Based on the coefficients... The information of neighbor nodes under the same meta-path is aggregated selectively. The above process only aggregates information about one-hop neighbors, failing to capture deeper, potential relationship information. A higher-order information aggregation layer is needed; by stacking multiple layers, message passing can be extended from one hop to multiple hops. Furthermore, due to the scale-free nature of heterogeneous graphs, the connections between nodes are unevenly distributed: some nodes are associated with many nodes, while others are associated with very few. To address these issues, the following steps are also required:

[0178] C: Repeat the attention coefficient calculation several times to obtain more information about neighboring nodes. The formula for step C is:

[0179]

[0180] In the formula: The node features represent the multi-head attention mechanism; || represents the concatenation operation. This indicates that node j is a neighbor of node i in the same subgraph.

[0181] In this embodiment, the multiple calculations of attention coefficients can be performed using a multi-head attention mechanism. Step C only aggregates information from one-hop neighbors and does not acquire deeper, potential relationship information. Therefore, a higher-order information aggregation layer is also needed.

[0182] D: The higher-order information aggregation layer uses an aggregation function. To achieve this, considering the interaction between the two features, step D uses the following calculation formula:

[0183]

[0184] In the formula: W1 and W2 are linear transformation matrices; ⊙ represents the multiplication of corresponding positions of the two matrices. Passing through multiple high-order information aggregation layers is equivalent to expanding outwards with multiple hops. Because starting from the second layer, the input to each aggregation layer is the output of the previous layer, the specific number of layers is used as a hyperparameter.

[0185] Finally, each node may appear in multiple isomorphic subgraphs, and the nodes have different characteristics in these isomorphic subgraphs. In this embodiment, the following can be used: It means that, where {Φ1,…,Φ p} represents the original path set.

[0186] E: Define a learning vector and a fully connected layer to obtain the attention coefficients. The formula for calculating step E is:

[0187]

[0188] In the formula: ν represents the attention coefficient, ν indicates that node i appears in different subgraphs, and q represents the attention coefficient. T Let represent a learnable vector, mainly used to adjust the shape of the vector, and b represent the bias term. In this embodiment, the activation function tanh can be used.

[0189] F: Normalize the attention coefficient to obtain the attention coefficient for learning.

[0190] G: Features of aggregate nodes in different isomorphic subgraphs based on the attention coefficients learned.

[0191] In this implementation, the loss function BPR can be used to obtain the output value. The calculation formula for step G is:

[0192]

[0193] In the formula: Ω={(u,i,j)|(u,i)∈I + ,(u,j)∈I-} represents the training set, I + I represents a positive sample, and I- represents a negative sample; y(u,i) represents the probability that user u interacts with product-related event i; σ represents a non-linear activation function.

[0194] In summary, the financial product data push method provided in this application uses a transformation matrix to project the features of different types of nodes into the same feature space, and uses an attention mechanism to accelerate the training process of the model's target recommendation, better capturing local information. A multi-head attention mechanism is then used to stabilize the model's learning process to acquire more information from neighboring nodes. In the high-order information aggregation layer, multiple layers are stacked to extend message passing from one hop to multiple hops. The BPR loss function is also used to maximize the difference in scores between positive and negative samples, assigning higher scores to known interactions between users and financial products, thereby obtaining more accurate and interpretable financial product data for push.

[0195] Please refer to Figure 5 , Figure 5 This is a schematic diagram of the structure of the financial product data push device provided in an embodiment of this application. Figure 5 As shown, the financial product data push device includes: a data processing module 51, a product knowledge graph construction module 52, and a push module 53.

[0196] The data processing module 51 is used to extract and process pre-stored user data to obtain basic user data, and to extract and process pre-stored product data to obtain product attribute data.

[0197] The data processing module 51 is also used to perform feature processing on product attribute data and user basic data to obtain map feature data.

[0198] Product knowledge graph construction module 52 is used to determine graph modeling data based on graph feature data and pre-stored product data.

[0199] The product knowledge graph construction module 52 is also used to determine ontology domain objects based on graph modeling data.

[0200] The product knowledge graph construction module 52 is also used to perform ontology extraction processing on the graph modeling data according to ontology domain objects to obtain ontology concept data.

[0201] The product knowledge graph construction module 52 is also used to launch the pre-stored ontology construction tool and obtain instance data from the pre-stored product data according to the ontology concept data.

[0202] The product knowledge graph construction module 52 is also used to perform format conversion processing on instance data to obtain resource description framework data.

[0203] The product knowledge graph construction module 52 is also used to start the pre-stored import tool to import the resource description framework data into the preset graph database and generate the first knowledge graph.

[0204] The push module 53 is used to construct a target learning model based on a pre-built multi-relation co-occurrence matrix and a pre-stored graph convolutional neural network model.

[0205] The push module 53 is also used to generate a second knowledge graph based on the first knowledge graph and the target learning model.

[0206] The push module 53 is also used to introduce the second knowledge graph into the pre-stored recommendation model to generate the target recommendation model.

[0207] The push module 53 is also used to acquire target user data and input the target user data into the target recommendation model to generate target financial product data.

[0208] The push module 53 is also used to push target financial product data to the target user terminal.

[0209] In an optional embodiment of this application, the data processing module 51 is specifically configured to: execute a pre-stored preprocessing script to extract product type data, product price change data, product rating data, and product source data from pre-stored product data, and determine the product type data, product price change data, product rating data, and product source data as product attribute data; and execute a pre-stored preprocessing script to extract user description data and user preference data from user basic data, and determine the user description data and user preference data as user basic data.

[0210] In an optional embodiment of this application, the data processing module 51 is further specifically used for: classifying product attribute data and user basic data to obtain discrete data and identity identification data; normalizing the discrete data to obtain normalized feature data; embedding the identity identification data to obtain identity identification feature data; and obtaining graph feature data based on the normalized feature data and identity identification feature data.

[0211] In an optional embodiment of this application, the product knowledge graph construction module 52 is specifically used for: determining graph entity data based on graph feature data, and determining graph attribute data based on graph entity data and graph feature data; launching a pre-stored semantic analysis tool to perform semantic analysis on pre-stored product data and graph entity data to generate entity relationship data; performing knowledge fusion processing on the pre-stored product data to determine standard graph entity data; and storing the standard graph entity data, graph attribute data, and entity relationship data as graph modeling data.

[0212] In an optional embodiment of this application, the product knowledge graph construction module 52 is further specifically used for: determining multiple nodes based on a pre-stored graph convolutional neural network model; determining the importance coefficient and relationship type data of the relationship between every two nodes based on a pre-built multi-relation co-occurrence matrix and multiple nodes; determining the importance value of one node to another in every pair of nodes based on all the relationship importance coefficients and relationship type data; and performing message passing on the pre-stored training data at each node based on all the importance values ​​and the pre-stored graph convolutional neural network model to generate the encoding layer node representation data corresponding to each node. When training the pre-stored graph convolutional neural network model, the pre-stored block diagonal matrix is ​​used to weight the pre-stored model matrix W. r Regularization is performed to generate the target weight matrix for each node. Based on the encoding layer node representation data, the target weight matrix for each node, the pre-stored graph convolutional neural network model, and the pre-stored knowledge graph, the optimization parameters of the knowledge graph are determined. The pre-stored graph convolutional neural network model is trained using the knowledge graph optimization parameters to construct the target learning model.

[0213] In an optional embodiment of this application, the product knowledge graph construction module 52 is used to determine the importance coefficient and relationship type data between every two nodes, and the calculation formula is as follows:

[0214]

[0215] In the formula, Let R be a pre-built multi-relation co-occurrence matrix of size R×R with 1s on the diagonal, where R is a positive integer and r is a positive integer. i Let r be the relationship type between the i-th batch of nodes of node e. j Let r and j be the relationship type between nodes in the j-th batch of node e, where r and j are both positive integers, and sum(e, r) = r(j, j). i Let ) be the relation type of the neighboring nodes of node e. i The sum of the number of sides, sum(e,r) j Let ) be the relation type of the neighboring nodes of node e. j The sum of the number of sides, For relation type r i For relation type rj The importance coefficient, |R|, is the total number of all relation types surrounding node e.

[0216] In an optional embodiment of this application, the product knowledge graph construction module 52 is used to generate the encoding layer node representation data corresponding to each node, and the calculation formula is as follows:

[0217]

[0218] In the formula, e h (l+1) For a node e in the (l+1)th layer of the pre-stored encoding layer of a convolutional neural network h The coding layer nodes represent data. For node e h The number of surrounding nodes, For different relation types r, the weight matrix that needs to be learned is... This represents the weight matrix of nodes that have no relation to other nodes. For pre-stored nodes e in the l-th layer of the encoding layer of a convolutional neural network h For node e j The importance value.

[0219] In an optional embodiment of this application, the product knowledge graph construction module 52 is used to generate the target weight matrix for each node, and the calculation formula is as follows:

[0220]

[0221] In the formula, b∈[1,B), then This is a block diagonal matrix of the pre-stored model weight matrix.

[0222] In an optional embodiment of this application, the pre-stored knowledge graph includes a valid set of triples, and the pre-stored graph convolutional neural network model includes multiple convolutional kernels. Accordingly, the product knowledge graph construction module 52 is used to determine the knowledge graph optimization parameters, and the calculation formula is as follows:

[0223] f(v) = concat((e h ,e r ,e t )*ω m ))·W

[0224] In the formula, v = (e h ,e r ,e t ), v∈V, V is the set of valid triples in the pre-stored knowledge graph, V′ represents the set of negative samples, * is the convolution operation, ω mdenoted as the m-th convolutional kernel in the pre-stored graph convolutional neural network model, where m is a positive integer and W is the target weight matrix for each node.

[0225] In an optional embodiment of this application, the product knowledge graph construction module 52 is used to train a pre-stored graph convolutional neural network model, and the calculation formula is as follows:

[0226]

[0227] In the formula, It is 1 or -1, when When, it represents a triple (e h ,e r ,e t ) is a positive sample, when When, it represents a triple (e h ,e r ,e t ) represents a negative sample, L represents the loss function, and λ is a preset loss constant.

[0228] In an optional embodiment of this application, the pre-stored recommendation model includes a bipartite graph. Correspondingly, the push module 53 is specifically used for: introducing a second knowledge graph into the pre-stored recommendation model to generate a target recommendation model, including: fusing the bipartite graph and the second knowledge graph to generate a collaborative knowledge graph; splitting the collaborative knowledge graph into multiple isomorphic subgraphs according to a preset meta-path; determining the feature parameters of each isomorphic subgraph based on the multiple isomorphic subgraphs; performing aggregation and fusion processing on each isomorphic subgraph to obtain knowledge graph embedding representation data; and importing the knowledge graph embedding representation data into the pre-stored recommendation model to generate the target recommendation model.

[0229] The financial product data push device provided in this embodiment can be used to execute the technical solution of the above method embodiment. Its implementation principle and technical effect are similar, and will not be described again in this embodiment.

[0230] Please refer to Figure 6 , Figure 6 This is a schematic diagram of the structure of the financial product data push device provided in the embodiments of this application, as shown below. Figure 6 As shown, the device includes at least one processor 601 and a memory 602.

[0231] The processor 601 is used to store computer execution instructions.

[0232] The memory 602 is used to execute computer execution instructions stored in the memory to implement the various steps involved in the above method embodiments. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0233] Alternatively, the memory 602 can be either standalone or integrated with the processor 601.

[0234] When the memory 602 is set up independently, the device also includes a bus 603 for connecting the memory 602 and the processor 601.

[0235] This application also provides a computer-readable storage medium storing computer-executable instructions. When the processor executes the computer-executable instructions, the above-mentioned financial product data push method is implemented.

[0236] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described financial product data push method.

[0237] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between devices or modules, and may be electrical, mechanical, or other forms.

[0238] The modules described above as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0239] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0240] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several commands to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute partial steps of the methods of the various embodiments of this application.

[0241] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0242] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0243] 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. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0244] The aforementioned storage medium can be implemented from 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 storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0245] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0246] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by program commands to related hardware. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0247] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0248] It should be noted that the financial product data push method, apparatus, device, storage medium, and product provided in this application embodiment can be used in the field of big data technology. They can also be used in any field other than big data technology. The application fields of the financial product data push method, apparatus, device, storage medium, and product provided in this application embodiment are not limited.

[0249] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for pushing financial product data, characterized in that, include: Extract and process the pre-stored user data to obtain basic user data, and extract and process the pre-stored product data to obtain product attribute data; Feature processing is performed on the product attribute data and the user basic data to obtain graph feature data; Based on the spectral feature data and the pre-stored product data, determine the spectral modeling data; Based on the graph modeling data, determine the ontology domain objects; The graph modeling data is processed by ontology extraction based on the ontology domain objects to obtain ontology concept data. Start the pre-stored ontology construction tool and obtain instance data from the pre-stored product data according to the ontology concept data; The instance data is converted to obtain resource description framework data. The pre-stored import tool is activated to import the resource description framework data into the preset graph database, generating the first knowledge graph. Based on the pre-built multi-relation co-occurrence matrix and the pre-stored graph convolutional neural network model, construct the target learning model; A second knowledge graph is generated based on the first knowledge graph and the target learning model; The second knowledge graph is introduced into the pre-stored recommendation model to generate the target recommendation model; Acquire target user data and input the target user data into the target recommendation model to generate target financial product data; Push the target financial product data to the target user's terminal; The step of determining the target learning model based on the pre-built multi-relation co-occurrence matrix and the pre-stored graph convolutional neural network model includes: Based on the pre-stored graph convolutional neural network model, multiple nodes are determined; Based on the pre-built multi-relation co-occurrence matrix and the multiple nodes, determine the importance coefficient and relationship type data of the relationship between every two nodes; Based on all relationship importance coefficients and relationship type data, determine the importance value of one node to the other in every pair of nodes; Based on all importance values ​​and the pre-stored graph convolutional neural network model, the pre-stored training data is passed to each node to generate the encoding layer node representation data corresponding to each node. When training the pre-stored graph convolutional neural network model, the pre-stored block diagonal matrix is ​​used to adjust the pre-stored model weight matrix. Perform regularization operations to generate the target weight matrix for each node; Based on the node representation data of the encoding layer, the target weight matrix of each node, the pre-stored graph convolutional neural network model, and the pre-stored knowledge graph, the optimization parameters of the knowledge graph are determined. The pre-stored graph convolutional neural network model is trained based on the optimized parameters of the knowledge graph to construct the target learning model.

2. The method according to claim 1, characterized in that, The process of extracting and processing pre-stored user data to obtain basic user data, and extracting and processing pre-stored product data to obtain product attribute data, includes: Execute the pre-stored preprocessing script to extract product type data, product price change data, product rating data, and product source data from the pre-stored product data, and determine the product type data, product price change data, product rating data, and product source data as product attribute data; The pre-stored preprocessing script is executed to extract user description data and user preference data from the user basic data, and the user description data and user preference data are determined as user basic data.

3. The method according to claim 1, characterized in that, The step of performing feature processing on the product attribute data and the user basic data to obtain graph feature data includes: The product attribute data and the user basic data are classified and processed to obtain discrete data and identity identification data; The discrete data is normalized to obtain normalized feature data; The identity data is embedded to obtain identity feature data; Based on the normalized feature data and the identity feature data, the map feature data is obtained.

4. The method according to claim 1, characterized in that, The step of determining the map modeling data based on the map feature data and the pre-stored product data includes: Based on the spectral feature data, determine the spectral entity data, and based on the spectral entity data and the spectral feature data, determine the spectral attribute data; The pre-stored semantic analysis tool is activated to perform semantic analysis on the pre-stored product data and the graph entity data to generate entity relationship data. The pre-stored product data is subjected to knowledge fusion processing to determine the standard map entity data; The standard graph entity data, the graph attribute data, and the entity relationship data are stored as graph modeling data.

5. The method according to claim 1, characterized in that, The formula for determining the importance coefficient and relationship type data between every two nodes based on the pre-built multi-relation co-occurrence matrix and the multiple nodes is as follows: In the formula, For size is Furthermore, a pre-built multi-relation co-occurrence matrix with diagonal lines of 1, where R is a positive integer, Let i be the relationship type between the i-th batch of nodes of node e. Let be the relationship type between nodes in the j-th batch of nodes of node e, where j is a positive integer. The relationship type among the adjacent nodes of node e is The sum of the number of sides, The relationship type among the adjacent nodes of node e is The sum of the number of sides, For relation type For relation types Importance coefficient, This represents the total number of all relation types surrounding node e.

6. The method according to claim 5, characterized in that, Based on all relationship importance coefficients and relationship type data, the formula for determining the importance of one node to the other in every pair of nodes is as follows: In the formula, For nodes In relationship Next target node Importance value, For nodes The number of surrounding nodes, Represents a node The set of surrounding neighboring nodes.

7. The method according to claim 6, characterized in that, The pre-stored graph convolutional neural network model includes multiple encoding layers; Accordingly, the calculation formula for generating the encoding layer node representation data for each node by passing messages through the pre-stored training data at each node based on all importance values ​​and the pre-stored graph convolutional neural network model is as follows: In the formula, For activation function, For a node in the (l+1)th layer of the encoding layer of a pre-stored convolutional neural network The coding layer nodes represent data. For nodes The number of surrounding nodes, for The set of neighboring nodes under relation r For different relation types r, the weight matrix that needs to be learned is... This represents the weight matrix of nodes that have no relation to other nodes. For pre-stored nodes in the l-th layer of the encoding layer of a convolutional neural network For nodes The importance value.

8. The method according to claim 7, characterized in that, The formula for calculating the target weight matrix for each node by performing regularization on the pre-stored model weight matrix using the diagonal matrix of the pre-stored block is as follows: In the formula, ,but B is a block diagonal matrix of the pre-stored model weight matrix, where B represents the total number of blocks for block diagonal regularization of the weight matrix.

9. The method according to claim 1, characterized in that, The pre-stored knowledge graph includes a valid set of triples, and the pre-stored graph convolutional neural network model includes multiple convolutional kernels; Accordingly, the calculation formula for determining the knowledge graph optimization parameters based on the encoding layer node representation data, the target weight matrix of each node, the pre-stored graph convolutional neural network model, and the pre-stored knowledge graph is as follows: In the formula, V is the set of valid triples in the pre-stored knowledge graph. For convolution operations, This represents the first element in the pre-stored graph convolutional neural network model. There are n convolutional kernels, where m is a positive integer. The target weight matrix for each node.

10. The method according to claim 9, characterized in that, The calculation formula for training the pre-stored graph convolutional neural network model based on the knowledge graph optimization parameters is as follows: In the formula, It is 1 or -1, when =1 indicates a triplet For positive samples, when =-1 indicates a triplet For negative samples, L represents the loss function. This is a preset loss constant.

11. The method according to claim 1, characterized in that, The pre-stored recommendation model includes a bipartite graph; Accordingly, the step of introducing the second knowledge graph into the pre-stored recommendation model to generate the target recommendation model includes: The bipartite graph and the second knowledge graph are fused together to generate a collaborative knowledge graph; The collaborative knowledge graph is split into multiple isomorphic subgraphs according to a preset meta-path; Based on the plurality of isomorphic subgraphs, determine the feature parameters of each isomorphic subgraph; Each isomorphic subgraph is aggregated and fused to obtain knowledge graph embedding representation data; The knowledge graph embedding representation data is imported into the pre-stored recommendation model to generate the target recommendation model.

12. A financial product data push device, characterized in that, include: The data processing module is used to extract and process pre-stored user data to obtain basic user data, and to extract and process pre-stored product data to obtain product attribute data. The data processing module is also used to perform feature processing on the product attribute data and the user basic data to obtain map feature data; The product knowledge graph construction module is used to determine the graph modeling data based on the graph feature data and the pre-stored product data. The product knowledge graph construction module is also used to determine ontology domain objects based on the graph modeling data; The product knowledge graph construction module is also used to perform ontology extraction processing on the graph modeling data according to the ontology domain objects to obtain ontology concept data; The product knowledge graph construction module is also used to launch a pre-stored ontology construction tool and obtain instance data from the pre-stored product data according to the ontology concept data; The product knowledge graph construction module is also used to perform format conversion processing on the instance data to obtain resource description framework data; The product knowledge graph construction module is also used to start the pre-stored import tool to import the resource description framework data into the preset graph database and generate the first knowledge graph; The push module is used to construct a target learning model based on a pre-built multi-relation co-occurrence matrix and a pre-stored graph convolutional neural network model; The push module is further configured to generate a second knowledge graph based on the first knowledge graph and the target learning model; The push module is also used to introduce the second knowledge graph into the pre-stored recommendation model to generate the target recommendation model; The push module is also used to acquire target user data and input the target user data into the target recommendation model to generate target financial product data; The push module is also used to push the target financial product data to the target user terminal; The push module is specifically used to determine multiple nodes based on the pre-stored graph convolutional neural network model; Based on the pre-built multi-relation co-occurrence matrix and the multiple nodes, determine the importance coefficient and relationship type data of the relationship between every two nodes; Based on all relation importance coefficients and relation type data, determine the importance value of one node to the other in every pair of nodes; based on all importance values ​​and the pre-stored graph convolutional neural network model, perform message passing on the pre-stored training data at each node to generate the encoding layer node representation data corresponding to each node; when training the pre-stored graph convolutional neural network model, use the pre-stored block diagonal matrix to adjust the pre-stored model weight matrix. Perform regularization to generate the target weight matrix for each node; determine the knowledge graph optimization parameters based on the node representation data of the encoding layer, the target weight matrix for each node, the pre-stored graph convolutional neural network model, and the pre-stored knowledge graph. The pre-stored graph convolutional neural network model is trained based on the optimized parameters of the knowledge graph to construct the target learning model.

13. A financial product data push device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the financial product data push method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the financial product data push method as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the financial product data push method according to any one of claims 1 to 11.

Images