User portrait determination method and device based on knowledge graph

Abstract

The embodiment of the invention provides a user portrait determination method and device based on a knowledge graph, and the method comprises the steps: collecting the data of a target user in N dimensions from a plurality of data sources, and obtaining an initial data set; performing data cleaning and preprocessing operation on the initial data set to obtain a target data set; converting a plurality of data included in the target data set into a triple data format to obtain a triple set; constructing an initial knowledge graph of the target data set; mapping the triple set into the initial knowledge graph to obtain a target knowledge graph; and inputting the target knowledge graph into the target classification network to obtain a user portrait of the target user output by the target classification network. Through the method and the device, the problem of insufficient user portrait resolution in related technologies is solved, and the effect of increasing the user portrait resolution is achieved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of data, and specifically, to a method and device for determining a user portrait based on a knowledge graph. Background Art

[0002] With the rapid development of the Internet and big data technology, personalized services have become a key factor in improving user experience and corporate competitiveness. In the field of financial services, banks and financial institutions pay particular attention to the in-depth understanding of user behavior and preferences in order to provide more refined financial product recommendations. However, traditional methods lack effective integration and association when building user portraits and making personalized recommendations, making it difficult to fully understand users. Based only on limited user attributes and historical behavior data, the depth and breadth of user preferences and needs are ignored, resulting in insufficient resolution of user portraits. Summary of the invention

[0003] The embodiments of the present application provide a method and device for determining a user portrait based on a knowledge graph, so as to at least solve the problem of insufficient resolution of user portraits in related technologies.

[0004] According to one embodiment of the present application, a method for determining a user portrait based on a knowledge graph is provided, comprising: collecting data of a target user in N dimensions from multiple data sources to obtain an initial data set, wherein N is a natural number greater than 1; performing data cleaning and preprocessing operations on the initial data set to obtain a target data set; converting multiple data included in the target data set into a triple data format to obtain a triple set, wherein each triple in the triple set includes an entity of the target user, a relationship type of a product transaction performed by the entity, and a product transaction attribute corresponding to the entity; constructing an initial knowledge graph of the target data set, wherein the nodes in the initial knowledge graph are used to represent the target user and the attributes of the target user, and the edges in the initial knowledge graph are used to represent the relationship between the multiple data included in the target data set; mapping the triple set to the initial knowledge graph to obtain a target knowledge graph; inputting the target knowledge graph into a target classification network to obtain a user portrait of the target user output by the target classification network.

[0005] In an exemplary embodiment, data of the target user in N product dimensions are collected from multiple data sources to obtain an initial data set, including at least one of the following: collecting data including user information and income information of the target user from the multiple data sources to obtain the initial data set; collecting historical transaction record information of the target user from the multiple data sources to obtain the initial data set; collecting investment data of the target user from the multiple data sources to obtain the initial data set, wherein the investment data includes investment type, investment amount and investment period; collecting risk data of the target user from the multiple data sources to obtain the initial data set, wherein the risk data includes risk assessment results, risk type and historical investment risk tolerance data.

[0006] In an exemplary embodiment, the multiple data included in the target data set are converted into a triple data format to obtain a triple set, including: parsing the data attributes, data types and data content of each of the data; identifying the entities in the data according to the user names included in the data attributes; determining the product transaction attributes corresponding to the entity according to the data type; extracting the relationship type of the product transaction performed by the entity from the data content; and generating a triple for each of the data according to the triple data format to obtain the triple set.

[0007] In an exemplary embodiment, the triple set is mapped to the initial knowledge graph to obtain a target knowledge graph, including: for each triple in the triple set, the triple is mapped to the initial knowledge graph through the following steps to obtain the target knowledge graph: parsing the triple, converting the data format of the triple into the data format of the initial knowledge graph to obtain a target triple; adding nodes in the initial knowledge graph that match both the target entity in the target triple and the product transaction attributes corresponding to the target entity; and creating a connection relationship between the matching nodes according to the relationship type of the product transaction performed by the target entity.

[0008] In an exemplary embodiment, the target classification network is a classification network obtained by training an initial classification network using multiple groups of sample training data, wherein the sample training data includes a sample knowledge graph, and the initial classification network includes a feature extraction subnetwork, a feature encoding subnetwork, a message passing subnetwork, and an output subnetwork, wherein for each group of the training data, the feature extraction subnetwork is used to extract local features and global features of sample nodes, sample edges, and connection relationships from the sample knowledge graph, wherein the local features and the global features both include features of sample users corresponding to the sample nodes and attribute features of the sample users, and relationship features between multiple sample data corresponding to the sample edges; the feature encoding subnetwork is used to encode the local features and the global features to obtain local feature vectors and global feature vectors; the message passing subnetwork is used to extract feature vectors of neighbor nodes of each sample node from the local feature vectors and the global feature vectors, and pass the feature vectors of the neighbor nodes to each sample node to obtain a target feature vector of each sample node; the output subnetwork is used to perform portrait classification on each sample node based on the target feature vector of each sample node, and output a user portrait of the sample user.

[0009] In an exemplary embodiment, after the target knowledge graph is input into the target classification network and the user portrait of the target user output by the target classification network is obtained, the method further includes: extracting the key feature vector of the target user from the user portrait, wherein the key feature vector includes the user characteristics, behavioral characteristics, investment characteristics and risk tolerance characteristics of the target user; constructing a product rating matrix, wherein the product rating matrix includes the ratings of the target product by N users, wherein N is a natural number greater than 1; calculating the similarity between the target user and the N users using the key feature vector; searching for K users from the N users based on the N similarities, wherein the similarity between the K users and the target user is greater than a preset similarity; predicting the rating of the target user for the target product according to the K ratings corresponding to the K users in the product rating matrix; and generating first recommendation information for recommending the target product to the target user according to the rating of the target product by the target user.

[0010] In an exemplary embodiment, after generating first recommendation information for recommending the target product to the target user according to the target user's rating of the target product, the method further includes: constructing a product feature vector for the target product, wherein the product feature vector includes type features, yield features, and risk level features of the target product; calculating a user feature vector for the target user using the user portrait; generating second recommendation information for recommending the target product to the target user based on the similarity between the user feature vector and the product feature vector; and generating target recommendation information for the target product using the first recommendation information and the second recommendation information.

[0011] In an exemplary embodiment, after generating the target recommendation information of the target product using the first recommendation information and the second recommendation information, the method further includes: collecting user behavior data, wherein the user behavior data includes behavior data acting on the target product; converting the user behavior data into a feedback signal, and inputting the feedback signal into the target classification network to adjust the feature weight of the user portrait of the target user in the target classification network according to the feedback signal.

[0012] According to another embodiment of the present application, a device for determining a user portrait based on a knowledge graph is provided, comprising: a first acquisition module, used to collect data of a target user in N dimensions from multiple data sources to obtain an initial data set, wherein N is a natural number greater than 1; a first processing module, used to perform data cleaning and preprocessing operations on the initial data set to obtain a target data set; a first conversion module, used to convert multiple data included in the target data set into a data format of triples to obtain a triple set, wherein each triple in the triple set includes an entity of the target user, a product executed by the entity, and a user's computer program product. The relationship type of the transaction and the product transaction attributes corresponding to the entity; a first construction module, used to construct an initial knowledge graph of the target data set, wherein the nodes in the initial knowledge graph are used to represent the target user and the attributes of the target user, and the edges in the initial knowledge graph are used to represent the relationship between the multiple data included in the target data set; a first mapping module, used to map the triple set to the initial knowledge graph to obtain a target knowledge graph; a first input module, used to input the target knowledge graph into a target classification network to obtain a user portrait of the target user output by the target classification network.

[0013] According to another embodiment of the present application, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in any of the above method embodiments are implemented.

[0014] According to another embodiment of the present application, a computer-readable storage medium is provided, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0015] According to another embodiment of the present application, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0016] Through this application, the data of the target user in N dimensions are collected from multiple data sources to obtain an initial data set, and data cleaning and preprocessing operations are performed on the initial data set to obtain a target data set; the multiple data included in the target data set are converted into a triple data format to obtain a triple set, and the initial knowledge graph of the target data set is constructed, and the triple set is mapped to the initial knowledge graph to obtain a target knowledge graph; the target knowledge graph is input into the target classification network to obtain a user portrait of the target user output by the target classification network. The user portrait determination method based on the knowledge graph can significantly improve the accuracy and comprehensiveness of the user portrait. Therefore, the problem of insufficient user portrait resolution in the related art can be solved, and the effect of increasing the user portrait resolution can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a hardware structure block diagram of a server device of a method for determining a user portrait based on a knowledge graph in an embodiment of the present application;

[0018] Figure 2 is a flowchart of a method for determining a user portrait based on a knowledge graph according to an embodiment of the present application;

[0019] Figure 3 It is a flow chart of a method for generating a personal portrait for financial product recommendation for bank users based on a knowledge graph according to an embodiment of the present application;

[0020] Figure 4 It is a specific flow chart of the method for generating a personal portrait for financial product recommendation of a bank user based on a knowledge graph according to an embodiment of the present application;

[0021] Figure 5 It is a structural block diagram of a device for determining a user portrait based on a knowledge graph according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] The embodiments of the present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0023] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0024] The method embodiments provided in the embodiments of the present application can be executed in a server device or a similar computing device. Taking running on a server device as an example, Figure 1 1 is a hardware structure block diagram of a server device of a method for determining a user profile based on a knowledge graph in an embodiment of the present application. Figure 1 As shown, the server device may include one or more ( Figure 1 Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data, wherein the server device may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above server device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0025] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the method for determining the user portrait based on the knowledge graph in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the server device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0026] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the server device. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0027] In this embodiment, a method for determining a user portrait based on a knowledge graph is provided. Figure 2 is a flowchart of a method for determining a user portrait based on a knowledge graph according to an embodiment of the present application, such as Figure 2 As shown, the process includes the following steps:

[0028] Step S202, collecting data of the target user in N dimensions from multiple data sources to obtain an initial data set, where N is a natural number greater than 1;

[0029] Optionally, this embodiment includes but is not limited to applications in the following scenarios: financial risk control and credit assessment, e-commerce personalized recommendation and advertising targeting.

[0030] Optionally, the target user in this embodiment refers to a specific user who needs personalized recommendation, and the user has a unique identifier.

[0031] Optionally, N in this embodiment represents the number of dimensions, and the data of N dimensions refers to information describing the user from different angles, including basic information, historical transaction records, investment preferences, risk tolerance, etc., where N is a natural number greater than 1, indicating the multi-dimensional nature of the data.

[0032] Optionally, the original data set of the target user in N dimensions collected from multiple data sources in this embodiment may contain noise, duplicate or missing values.

[0033] Step S204, performing data cleaning and preprocessing operations on the initial data set to obtain a target data set;

[0034] Optionally, after data cleaning and preprocessing of the initial data set, the high-quality, structured data set obtained in this embodiment is used for subsequent knowledge graph construction and user portrait generation. The specific operations include:

[0035] Deduplication: Since multi-source data collection may cause the same information to appear repeatedly in different data sources, deduplication is required to ensure the uniqueness of each piece of data information. You can set a unique data identifier (such as a combination of user ID and transaction timestamp) to compare and remove duplicate records and retain the latest or most detailed version.

[0036] Outlier detection and processing: Outliers may be caused by data entry errors, system failures, or extreme events and need to be identified and processed. Statistical methods (such as standard deviation, quartiles) or machine learning models (such as the isolation forest algorithm) can be used to detect outliers. Detected outliers can be corrected, deleted, or processed using interpolation techniques.

[0037] Missing value filling: The data may contain missing values, which will affect the accuracy and completeness of subsequent analysis. Filling methods include simple filling (such as using the mean or median), filling based on predictions of other attributes, or intelligent filling using deep learning technology (such as the GAIN algorithm).

[0038] Data format standardization: Ensure that data obtained from different data sources have the same format and scale to facilitate subsequent data processing and analysis. For example, standardize dates to the YYYY-MM-DD format and convert monetary amounts to the same monetary unit.

[0039] Data type conversion: Convert unstructured data such as investment preferences and risk assessment results described in text into structured data for easier data processing and storage. This can be achieved through natural language processing techniques such as entity recognition and relationship extraction.

[0040] Data quality check: Perform data integrity and consistency checks to ensure data reliability and accuracy. For example, check whether data fields are complete, data types are consistent, and whether the logical relationships between data are reasonable.

[0041] Privacy protection and compliance checks: When processing user data, you need to comply with data protection regulations such as GDPR or CCPA to ensure that data usage complies with privacy protection requirements. This may include data anonymization, using only data fields that have been consented to by the user, and setting access rights control during data processing.

[0042] Data integration and association: Associating data from different sources to build a comprehensive multi-dimensional data view of the user. For example, associating the user's basic information with transaction records to identify the user's investment pattern and risk preference.

[0043] Feature engineering: Based on the preprocessed data set, feature selection and feature construction are performed to provide high-quality input features for subsequent user profile generation and recommendation algorithms.

[0044] Optionally, in the knowledge graph, a triple is represented as (entity, relationship, attribute), which is used to describe the relationship between entities and the attributes of entities. For example, (user A, invests in, low-risk funds) describes that user A has investment behavior in low-risk funds.

[0045] Step S206, converting the multiple data included in the target data set into a triple data format to obtain a triple set, wherein each triple in the triple set includes an entity of the target user, a relationship type of a product transaction performed by the entity, and a product transaction attribute corresponding to the entity;

[0046] Optionally, the triple set is a set formed by converting the data in the target data set into a triple format, which is used to construct a knowledge graph.

[0047] Step S208, constructing an initial knowledge graph of the target data set, wherein the nodes in the initial knowledge graph are used to represent the target user and the attributes of the target user, and the edges in the initial knowledge graph are used to represent the relationship between the multiple data included in the target data set;

[0048] Optionally, the initial knowledge graph is an initial version of the knowledge graph constructed based on the target dataset, including user entities, attribute entities and the relations between them.

[0049] Step S210, mapping the triple set to the initial knowledge graph to obtain a target knowledge graph;

[0050] Optionally, the target knowledge graph is an optimized and updated version of the knowledge graph after mapping the triple set to the initial knowledge graph, which can more accurately represent the data and relationships of the target user.

[0051] Step S212: input the target knowledge graph into the target classification network to obtain the user portrait of the target user output by the target classification network.

[0052] Optionally, the target classification network refers to a neural network model used for classification or prediction, which may be based on a graph neural network (GNN), a convolutional neural network (CNN), or other types of deep learning models, and is used to extract user features from the target knowledge graph and generate user portraits.

[0053] In this embodiment, the execution subject of the above steps can be a terminal, a server, a specific processor set in the terminal or server, or a processor or processing device set relatively independently from the terminal or server, but is not limited thereto.

[0054] For example, assuming the target user is Xiao Li, N=4, that is, the collected data dimensions include: basic information, historical transaction records, investment preferences and risk tolerance.

[0055] S302 collects Xiao Li’s personal information (such as age, occupation), transaction records (such as deposits and withdrawals, transfers, consumption, and investments), investment preferences (such as preferred investment type, term, and amount), and risk tolerance (risk assessment results) from multiple data sources such as the bank’s customer management system, transaction database, investment intention survey, and risk assessment system.

[0056] S304: Perform data cleaning on the collected initial data set, including deduplication, anomaly detection and missing value processing, to ensure data quality.

[0057] S306, converting each record after data cleaning into a triple format, for example, Xiao Li's transaction record may be converted into: (Xiao Li, investing in, short-term and medium-term bond funds), (Xiao Li, purchasing, low-risk financial products), (Xiao Li, depositing, time deposits), etc.

[0058] S308, constructing an initial knowledge graph, taking user Xiao Li and his attributes as nodes, and taking the relationships and attributes in the above triples as edges and attributes on the edges, to form a graph structure that describes the multi-dimensional characteristics of Xiao Li.

[0059] S302, map the converted triple set to the initial knowledge graph, update or add relevant nodes and edges, optimize the graph structure, and obtain the target knowledge graph.

[0060] S310, input the target knowledge graph into the pre-trained target classification network, such as the neural network in the figure. The network analyzes Xiao Li's position, relationship and attributes in the graph, and outputs a user profile containing multi-dimensional information such as Xiao Li's behavior characteristics, investment preferences, and risk tolerance. For example, the user profile may indicate that Xiao Li prefers conservative investment, prefers medium- and short-term financial products, and has a medium risk tolerance.

[0061] Through the above steps, the data of the target user in N dimensions are collected from multiple data sources to obtain an initial data set, and data cleaning and preprocessing operations are performed on the initial data set to obtain a target data set; the multiple data included in the target data set are converted into a triple data format to obtain a triple set, and the initial knowledge graph of the target data set is constructed, and the triple set is mapped to the initial knowledge graph to obtain a target knowledge graph; the target knowledge graph is input into the target classification network to obtain a user portrait of the target user output by the target classification network. The user portrait determination method based on the knowledge graph can significantly improve the accuracy and comprehensiveness of the user portrait. Therefore, the problem of insufficient user portrait resolution in the related technology can be solved, and the effect of increasing the user portrait resolution can be achieved.

[0062] In an exemplary embodiment, data of the target user in N product dimensions are collected from multiple data sources to obtain an initial data set, including at least one of the following: collecting data including user information and income information of the target user from the multiple data sources to obtain the initial data set; collecting historical transaction record information of the target user from the multiple data sources to obtain the initial data set; collecting investment data of the target user from the multiple data sources to obtain the initial data set, wherein the investment data includes investment type, investment amount and investment period; collecting risk data of the target user from the multiple data sources to obtain the initial data set, wherein the risk data includes risk assessment results, risk type and historical investment risk tolerance data.

[0063] Optionally, the historical transaction record information in this embodiment is used to record all transaction details of the user, such as the goods purchased on the e-commerce platform, the purchase date, the purchase quantity, the payment method, etc., which is the basic data for analyzing user preferences and behavior patterns.

[0064] Optionally, the investment data in this embodiment relates to the user's investment behavior in other financial products or services, such as investment type, investment amount and investment period of stocks, bonds, funds, etc. These data are helpful to understand the user's financial management habits and risk preferences.

[0065] Optionally, the risk data in this embodiment includes the user's risk assessment results, risk type, and historical investment risk tolerance data, which are important bases for evaluating the user's risk tolerance and formulating personalized financial product recommendation strategies.

[0066] For example, assuming that the target user is Xiaogang, the following steps describe how to collect multi-dimensional information from the e-commerce platform and related financial data sources, generate an initial data set, and obtain a high-quality target data set through data cleaning and preprocessing:

[0067] S402, collecting user information and income information: collecting information such as Xiaogang's age (28 years old), gender (male), occupation (software engineer), educational background (Master of Computer Science), and monthly income (15,000 yuan) from the user database and financial information records of the e-commerce platform.

[0068] S404, collect historical transaction record information: collect all purchase records of Xiaogang in the past year from the historical order database of the e-commerce platform, including product type (such as electronic equipment, sports equipment, books), purchase time, purchase frequency, purchase amount, etc.

[0069] S406, collect investment data: collect Xiaogang's investment details from the associated financial data sources, including his preference for investing in stocks and bonds, the average investment amount is around 100,000 yuan, and the investment period is mostly medium to long term.

[0070] S408, Collect risk data: From Xiaogang’s risk assessment results, we know that he has a medium risk tolerance and prefers a conservative investment strategy, and is less involved in high-risk investment projects.

[0071] S410, constructing an initial data set: integrating the above-collected user information, historical transaction records, investment data and risk data to form an initial data set containing Xiaogang's information in multiple dimensions.

[0072] S412, Data cleaning and preprocessing: Perform operations such as deduplication, outlier detection, missing value filling, and data standardization on the initial data set to ensure data accuracy and consistency. For example, remove duplicate product purchase records, fill in missing transaction time, and unify the classification standards of product attributes.

[0073] Through the above steps, the e-commerce platform can build a high-quality target dataset for Xiaogang, which contains comprehensive information about him, his transactions, investments, and risk tolerance. Next, the platform can convert this data into triple format, build an initial knowledge graph, and generate user portraits through machine learning models for personalized product recommendations and customized financial services. For example, the platform may find that Xiaogang prefers to buy the latest technology products, while his investment tendencies prefer stable financial products. This information will be used to recommend the latest electronic products and stable financial products to him, while considering his risk tolerance and avoiding recommending overly risky investment options. Through such personalized recommendations, the e-commerce platform can provide services that better meet Xiaogang's needs and improve user satisfaction and loyalty.

[0074] In an exemplary embodiment, the multiple data included in the target data set are converted into a triple data format to obtain a triple set, including: parsing the data attributes, data types and data content of each of the data; identifying the entities in the data according to the user names included in the data attributes; determining the product transaction attributes corresponding to the entity according to the data type; extracting the relationship type of the product transaction performed by the entity from the data content; and generating a triple for each of the data according to the triple data format to obtain the triple set.

[0075] Optionally, in this embodiment, data attributes refer to descriptive features of data, such as data type, data source, data format, etc. According to the data attributes, the type of data can be determined, such as numeric type, text type, timestamp type, etc. The data type is crucial for data processing and conversion.

[0076] Optionally, data content refers to the actual information contained in the data, such as the user's purchase history, evaluation information, investment details, etc., which is the key basis for generating triples.

[0077] Optionally, in a knowledge graph, an entity refers to an object with a unique identifier, such as a user, product, brand, etc., which is the subject and object of a triple.

[0078] Optionally, product transaction attributes are used to describe characteristics of an entity related to a product transaction, such as transaction amount, transaction time, transaction type, product category, etc.

[0079] The relationship type of the product transaction performed by the entity represents the relationship between the entity and the product transaction, such as purchase, browse, evaluate, invest, etc., which is the relationship part in the triple.

[0080] Through the above steps, the data in the target dataset can be converted into a set of triples, providing a structured data foundation for the subsequent construction of the knowledge graph and the generation of user portraits. For example, for Xiaohua's purchase history, a series of triples describing the relationship between it and the product can be generated. These triples can be further used to identify Xiaohua's purchase preferences, brand preferences, consumption frequency and other characteristics, providing accurate input for personalized recommendation algorithms. For evaluation information and investment records, the corresponding triples are also generated and integrated into the knowledge graph to enhance the richness and depth of user portraits.

[0081] In an exemplary embodiment, the triple set is mapped to the initial knowledge graph to obtain a target knowledge graph, including: for each triple in the triple set, the triple is mapped to the initial knowledge graph through the following steps to obtain the target knowledge graph: parsing the triple, converting the data format of the triple into the data format of the initial knowledge graph to obtain a target triple; adding nodes in the initial knowledge graph that match both the target entity in the target triple and the product transaction attributes corresponding to the target entity; and creating a connection relationship between the matching nodes according to the relationship type of the product transaction performed by the target entity.

[0082] Optionally, the triple data is mapped into a knowledge graph to construct a target knowledge graph. This process includes three key steps: data format conversion, node addition, and relationship creation.

[0083] Data format conversion is to parse the data in the triple set and convert it into the data format of the initial knowledge graph. This process includes: Data model matching: The initial knowledge graph may adopt a specific data model, such as (Resource Description Framework, referred to as RDF) or a property graph model, while the triple set may come from different sources and have different formats. Data format conversion ensures that the triple is consistent with the data model of the knowledge graph. For example, the "entity", "relationship" and "attribute" in the triple are mapped to the nodes, edges and node attributes in the graph respectively. Entity alignment: For the entity in the triple, the corresponding node needs to be found in the knowledge graph (if it exists). Entity alignment is achieved through entity resolution technology, which identifies the corresponding relationship between the triple entity and the node in the graph by comparing the identifier, attribute or description of the entity. If there is no matching entity node in the initial knowledge graph, a new node will be created based on the entity information in the triple. Relationship type mapping: The relationship type in the triple needs to match the edge type in the knowledge graph to ensure that the relationship between entities is correctly represented in the graph. Relationship type mapping may require comparison against a predefined list of relationship types to ensure consistent terminology and definitions are used.

[0084] Add nodes to the knowledge graph that match the entities and attributes in the target triple, including: Entity uniqueness: Each entity in the knowledge graph has a unique identifier to distinguish different entities. When adding a node, you need to check whether this entity already exists in the graph to avoid adding the same node repeatedly. Attribute extension: The attributes in the target triple may be used to extend the attributes of existing nodes, or to create new nodes with these attributes. For example, if the target triple contains the occupation attribute of user Xiao Li, after finding the Xiao Li node in the knowledge graph, this attribute can be added to the node. Automatic entity creation: If the entity in the target triple cannot find a matching node in the initial knowledge graph, a new node needs to be created to represent this entity. When creating a new node, you need to determine its type (such as user, product, category, etc.) and attributes.

[0085] According to the relationship type of the product transaction performed by the target entity, a connection relationship is created between the matching nodes, including: Relationship definition: In the knowledge graph, the relationship (i.e., edge) defines the connection between entities. According to the relationship type in the target triple, such as purchase, evaluation, or investment, the corresponding edge is created between the corresponding entity nodes. The creation of an edge may require the definition of the edge type, direction, and possible edge attributes. Directionality of the edge: In the knowledge graph, the edge can be undirected or directed, depending on the nature of the relationship. For example, the "purchase" relationship is usually directed, indicating the user's purchase behavior of the product. Dynamic update of relationships and attributes: If the relationship in the target triple describes a new interaction between entities or an update of an attribute, the relationship and attributes in the knowledge graph should be updated to reflect the latest status. This may involve the addition, deletion, or modification of edges.

[0086] For example, for the purchase relationship between user Xiao Li and books, if there is no corresponding node in the initial knowledge graph, Xiao Li and book nodes are created, and the edge of the purchase relationship is added to connect the Xiao Li node and the book node. If the Xiao Li node already exists but there is no book node, only the book node is created and a relationship is created between the two. If the user's behavioral attributes (such as the number of purchases and the amount) are updated, the corresponding attributes and relationships will also be updated in the knowledge graph to ensure the timeliness and accuracy of the graph. This process is the key to building a dynamic knowledge graph that responds to real-time needs.

[0087] Through the above steps, the triple set is mapped to the initial knowledge graph, and the target knowledge graph is constructed, which can more accurately and comprehensively represent the user's behavior and preferences, as well as the relationship between entities. This target knowledge graph is an important data foundation for the subsequent generation of user portraits and optimization of personalized recommendation algorithms, which can help the system understand user needs and market dynamics more deeply and provide more intelligent and personalized services.

[0088] In an exemplary embodiment, the target classification network is a classification network obtained by training an initial classification network using multiple groups of sample training data, wherein the sample training data includes a sample knowledge graph, and the initial classification network includes a feature extraction subnetwork, a feature encoding subnetwork, a message passing subnetwork, and an output subnetwork, wherein for each group of the training data, the feature extraction subnetwork is used to extract local features and global features of sample nodes, sample edges, and connection relationships from the sample knowledge graph, wherein the local features and the global features both include features of sample users corresponding to the sample nodes and attribute features of the sample users, and relationship features between multiple sample data corresponding to the sample edges; the feature encoding subnetwork is used to encode the local features and the global features to obtain local feature vectors and global feature vectors; the message passing subnetwork is used to extract feature vectors of neighbor nodes of each sample node from the local feature vectors and the global feature vectors, and pass the feature vectors of the neighbor nodes to each sample node to obtain a target feature vector of each sample node; the output subnetwork is used to perform portrait classification on each sample node based on the target feature vector of each sample node, and output a user portrait of the sample user.

[0089] Optionally, the target classification network is a deep learning model. It trains the initial classification network using multiple sets of sample training data, so that it can extract and encode the features of entities (sample nodes) and relationships (sample edges) from the knowledge graph, thereby generating a multi-dimensional profile of the user. The network consists of four main sub-networks: feature extraction, feature encoding, message passing, and output sub-networks, which perform the following tasks in sequence:

[0090] Feature extraction subnetwork: The main task of this subnetwork is to extract various types of features from the sample knowledge graph, including local features and global features. Local features usually refer to the features of edges (relationships) and nodes (attributes) directly connected to a specific node (or user), while global features involve the structure of the entire graph and indirect connections between entities. These features may include basic information about users (such as age and gender), user behavior patterns (such as purchase history and browsing history), product attributes (such as type and price), and complex relationship networks between entities (such as the purchase relationship between users and products, the similarity between products, etc.). The feature extraction subnetwork can extract these multi-dimensional features from the knowledge graph through graph convolution, graph attention mechanism and other technologies, providing input for subsequent feature encoding.

[0091] The feature encoding subnetwork is used to encode the extracted local features and global features into dense feature vectors, namely local feature vectors and global feature vectors. The encoding process may use technologies such as deep neural networks, graph neural networks (GNNs), and self-attention mechanisms to convert the attributes and relationships of entities into numerical representations for subsequent calculations and analysis.

[0092] The message passing subnetwork extracts feature vectors from the neighbor nodes of each sample node through the message passing mechanism in the graph neural network, and passes this information back to the sample node to form a target feature vector. This process allows the network to capture the interactions and relationships between entities, such as the similarity between users, the indirect connection between users and products, etc. By passing the feature vectors of neighbor nodes, the target feature vector of each sample node can more comprehensively reflect its position, relationship and role in the knowledge graph, thereby providing a richer user representation.

[0093] The output subnetwork classifies the user profile based on the target feature vector of each sample node and generates a user profile. This is usually achieved by connecting a classifier or regressor to the output of the message passing subnetwork to predict the user's behavioral characteristics, interest preferences, risk tolerance, etc. The output subnetwork may use components such as fully connected layers and softmax layers to convert the target feature vector into a probability distribution or a specific behavior prediction, and finally output a user profile.

[0094] The training process of the entire target classification network relies on multiple sets of sample training data, including sample knowledge graphs, which contain user behavior and attribute information, as well as the relationship between entities. By repeatedly iterating on these sample data, the network can learn the best strategy for generating user profiles from the knowledge graph, in order to achieve more accurate prediction and classification on new data.

[0095] For example, for a target user named Xiaohong on an e-commerce platform, the network will extract features such as purchase behavior, evaluation history, and browsing history from her knowledge graph, encode them into feature vectors, and integrate the behavioral features of her friends in the social network through the message passing mechanism, and finally generate a detailed user profile that includes Xiaohong's interest preferences, consumption habits, and user attributes. This user profile can be used to improve personalized recommendations to ensure that the recommended products are more in line with Xiaohong's actual needs and preferences.

[0096] In the training process of this embodiment, the network also adjusts its parameters by optimizing loss functions (such as cross entropy loss, mean square error, etc.) to improve the accuracy and personalization of user portrait generation. The ultimate goal of network design and training is to generate high-quality user portraits even when facing users that have never been seen before or complex and changeable behavior patterns, providing strong support for applications such as personalized services, product recommendations, and risk management.

[0097] In an exemplary embodiment, after the target knowledge graph is input into the target classification network and the user portrait of the target user output by the target classification network is obtained, the method further includes: extracting the key feature vector of the target user from the user portrait, wherein the key feature vector includes the user characteristics, behavioral characteristics, investment characteristics and risk tolerance characteristics of the target user; constructing a product rating matrix, wherein the product rating matrix includes the ratings of the target product by N users, wherein N is a natural number greater than 1; calculating the similarity between the target user and the N users using the key feature vector; searching for K users from the N users based on the N similarities, wherein the similarity between the K users and the target user is greater than a preset similarity; predicting the rating of the target user for the target product according to the K ratings corresponding to the K users in the product rating matrix; and generating first recommendation information for recommending the target product to the target user according to the rating of the target product by the target user.

[0098] Optionally, this embodiment extracts a key feature vector from the user portrait, which contains multi-dimensional information such as user characteristics, behavioral characteristics, investment characteristics, and risk tolerance characteristics. Among them, user characteristics may involve basic personal information (such as age, gender, occupation), behavioral characteristics include user behavior data on the platform (such as browsing history, purchase history, evaluation behavior), investment characteristics cover user investment preferences and historical investment data, and risk tolerance characteristics reflect the user's risk preference and historical risk tolerance.

[0099] For example, for user Xiao Wang, the key feature vectors may include: age 30, gender male, occupation IT engineer, preference for technology products, average consumption amount of 1,000 yuan, investment type preference for stable funds, etc. These features will be used for subsequent similarity calculations and product rating predictions.

[0100] Optionally, the product rating matrix is ​​to organize the rating information of N users on the target product into a matrix, where N represents the number of users, including the target user and other users. The construction of the product rating matrix helps to understand the evaluation of the same product by different users and the preference similarity between users.

[0101] For example, a product rating matrix may contain 1,000 users' ratings on a smartwatch, ranging from 1 (dislike very much) to 5 (like very much). Through the product rating matrix, you can intuitively see which users gave high ratings, which users gave low ratings, and the central tendency of the rating distribution.

[0102] Optionally, the key feature vector is used to calculate the similarity between the target user and N users, which can be achieved through metrics such as cosine similarity, Euclidean distance or Pearson correlation coefficient. The purpose of similarity calculation is to find a set of users who are similar to the target user in behavior, preference and risk tolerance, so as to refer to their product evaluations to predict the target user's potential rating of the product.

[0103] For example, by calculating the cosine similarity between key feature vectors, we can find user groups with similar interests and preferences as user Xiao Wang. These users may have a common preference for technology products or have similar tendencies in investing in conservative funds.

[0104] Optionally, according to the N similarities, K users (K users) whose similarities are greater than a preset threshold are found from the N users, and based on their ratings in the product rating matrix, the rating of the target user to the target product is predicted. The preset threshold is used to screen out a group of users most similar to the target user, and is usually set based on experience or through cross-validation.

[0105] For example, if the similarity between the key feature vector of user Xiao Wang and some users in the user set exceeds 0.8 (preset threshold), these users are defined as similar users (K users). Assuming that the average rating of a smartwatch by K users is 4.5, it is predicted that the rating of user "Xiao Wang" for this smartwatch is also close to 4.5.

[0106] Optionally, first recommendation information is generated to recommend the target product to the target user based on the predicted score of the target user for the target product. If the predicted score is higher than the average score of the user on the platform or a preset recommendation threshold, it is considered that the product may meet the preferences of the target user and is added to the recommendation list.

[0107] In e-commerce or financial service scenarios, this process may involve the optimization of recommendation algorithms, such as collaborative filtering, content-based recommendation, or hybrid recommendation techniques, to improve the accuracy and personalization of recommendations.

[0108] For example, if user Xiao Wang's predicted rating for a smart watch is 4.8, which is higher than the platform average rating, the system will recommend the smart watch as a product and generate the first recommendation information to recommend the smart watch to user "Xiao Wang".

[0109] In this embodiment, by extracting and encoding the user's key feature vectors from the knowledge graph and combining them with the product rating matrix for similarity calculation and rating prediction, the recommendation system can more accurately understand user needs and provide more appropriate recommendation services based on user behavior and preferences.

[0110] In an exemplary embodiment, after generating first recommendation information for recommending the target product to the target user according to the target user's rating of the target product, the method further includes: constructing a product feature vector for the target product, wherein the product feature vector includes type features, yield features, and risk level features of the target product; calculating a user feature vector for the target user using the user portrait; generating second recommendation information for recommending the target product to the target user based on the similarity between the user feature vector and the product feature vector; and generating target recommendation information for the target product using the first recommendation information and the second recommendation information.

[0111] Optionally, after generating the first recommendation information based on user behavior and preference, in order to further improve the accuracy and personalization of the recommendation, the system can combine the user portrait and product features to generate the second recommendation information, and finally integrate and generate the target recommendation information of the target product.

[0112] Optionally, the system constructs a product feature vector for the target product, which includes key attributes such as product type features, yield features, and risk level features. Product type features may include the type or category of the product, such as stocks, bonds, and mixed funds; yield features reflect the expected return level of the product, which may be the historical average yield, the expected maximum yield, etc.; and risk level features identify the level of product risk, such as low risk, medium risk, and high risk.

[0113] For example, for a smart watch, its product feature vector may include: the type feature is a smart device, the yield feature is "non-financial product, no yield", and the risk level feature is "zero risk".

[0114] Optionally, the system further calculates the user feature vector of the target user using the generated user portrait. The user portrait contains rich user information, such as age, gender, occupation, historical transaction records, investment preferences, etc. The user feature vector extracts dimensions related to product features in the user portrait, such as the user's preference for product type, expectation of rate of return, and acceptance of risk level.

[0115] For example, for user Xiao Ming, his user feature vector may reflect his preference for technology products, his expectation for a medium rate of return, and his acceptance of medium risk.

[0116] Optionally, the system generates second recommendation information based on the similarity between the user feature vector and the product feature vector. This process may use, for example, cosine similarity, Euclidean distance, or a specially designed similarity metric to evaluate the degree of match between user characteristics and product attributes.

[0117] For example, if the similarity between the feature vector of user Xiao Ming and the product feature vector of the smart watch is high, this may be because Xiao Ming has a strong interest in technology products, and the zero-risk feature of the smart watch is consistent with Xiao Ming's risk preference. The system will generate a second recommendation information, suggesting that the smart watch may meet Xiao Ming's preferences.

[0118] Optionally, the system generates target recommendation information of the target product by combining the first recommendation information and the second recommendation information. This step may include adjusting the weights of the first recommendation information and the second recommendation information to reflect the relative importance of user behavior and product attributes in the recommendation decision.

[0119] For example, the system may weight the first recommendation information based on user behavior and the second recommendation information based on product attributes, and the weight coefficient (λ) reflects the degree of influence of user behavior on the recommendation result. If user Xiao Ming's score prediction for smart watches (first recommendation information) is 4.5, and the similarity between the smart watch and Xiao Ming's user feature vector (second recommendation information) is 0.9, then the target recommendation information may take these two points into consideration and give a higher recommendation score, suggesting that Xiao Ming may have a strong interest in smart watches.

[0120] The target recommendation information can also include product description, recommendation reasons, user evaluation summary, etc. to enhance the persuasiveness of the recommendation and user experience. For example, when recommending a smartwatch, the system may include the main functions of the product, features that match the preferences in the user profile, and summaries of favorable reviews from other similar users to increase the attractiveness and credibility of the recommendation.

[0121] This embodiment can generate more comprehensive and personalized recommendation information by comprehensively evaluating the degree of match between target users and target products from multiple dimensions, thereby improving the accuracy of recommendations and user satisfaction, thereby providing users with more intelligent and considerate financial management or consumption advice.

[0122] In an exemplary embodiment, after generating the target recommendation information of the target product using the first recommendation information and the second recommendation information, the method further includes: collecting user behavior data, wherein the user behavior data includes behavior data acting on the target product; converting the user behavior data into a feedback signal, and inputting the feedback signal into the target classification network to adjust the feature weight of the user portrait of the target user in the target classification network according to the feedback signal.

[0123] Optionally, after generating the target recommendation information of the target product, in order to further optimize the recommendation system, improve the accuracy of the user portrait and the personalization of the recommendation, this embodiment will continue to collect the user's behavior data on the target product, including but not limited to clicks, browsing, purchases, evaluations, sharing and other behaviors. These behavior data are direct feedback from the user's response to the recommendation information, which is crucial for evaluating the recommendation effect and optimizing the user portrait. For example, if user Xiao Zhang browses the recommended smart watch for a long time and finally buys it, this shows that the recommendation is highly matched with the user's interests, and the system should record this behavior as positive feedback.

[0124] Optionally, in this embodiment, the collected user behavior data needs to be converted into feedback signals that can be understood and processed by the machine learning model. This process involves quantifying and encoding the behavior data to adapt to the input format of the model. For example, purchase behavior can be converted into high-value feedback signals, while browsing behavior may be converted into different levels of feedback signals depending on the length of browsing time. The strength and type of feedback signals reflect the user's acceptance and preference for the recommended products.

[0125] Optionally, this embodiment inputs the converted feedback signal into the target classification network used previously. The network adjusts the weights of the features in the user profile according to the feedback signal through the back-propagation mechanism to optimize the future product recommendations for user Xiao Zhang. For example, if the feedback signal of purchasing a smartwatch indicates that Xiao Zhang has a strong interest in technology products, the target classification network will increase the weights of the features in Xiao Zhang's user profile that are related to technology product preferences to improve the probability and accuracy of future recommendations of technology products.

[0126] Optionally, in this embodiment, the feature weight adjustment in the target classification network is performed based on the feedback signal, with the goal of making the user portrait more accurately reflect the user's real preferences and behavior patterns, thereby optimizing the performance of the recommendation algorithm.

[0127] This process can be achieved through the following steps:

[0128] 1) Calculate the correlation between the feedback signal and the features in the user profile to determine which features have an impact on the user's decision;

[0129] 2) Adjust the weights of the affected features based on the strength of the feedback signal;

[0130] 3) Update weight parameters through network optimization algorithms (such as gradient descent) to enable the model to more accurately predict user behavior.

[0131] When adjusting feature weights, the system needs to consider the diversity and instability of user behavior and update the model in a timely manner to reflect the dynamic changes in user preferences and behavior patterns. For example, if user Xiao Zhang recently becomes interested in fitness products, the feedback signal will reflect this change, and the network will adjust the feature weights related to fitness products accordingly to ensure that future recommendations can more accurately meet Xiao Zhang's new needs.

[0132] This embodiment collects user behavior data and converts it into feedback signals, which are then input into the target classification network to adjust the weights of user portrait features. The recommendation system can continuously learn and optimize, and can provide personalized recommendations that are more in line with user needs and interests.

[0133] The present application is described below in conjunction with specific embodiments:

[0134] The purpose of this specific embodiment is to propose a method for generating a personal portrait of a bank user's financial product recommendation based on a knowledge graph, such as Figure 3 As shown, the following steps are included:

[0135] S1, collects users’ basic personal information, historical transaction records, investment preferences and risk tolerance from multiple data sources to build a multi-dimensional data set of users; Figure 4 As shown, in this embodiment, S1 includes the following sub-steps:

[0136] S11, collecting basic personal information including the user's name, age, gender, occupation and income level from multiple data sources D personal ;

[0137] S12, collecting historical transaction records D including the user's deposit and withdrawal records, transfer records, consumption records and investment records from multiple data sources transaction ;

[0138] S13, collecting investment preferences including user's investment type preference, investment amount preference and investment term preference from multiple data sources D preference ;

[0139] S14, collect risk tolerance D including the user's risk assessment results, risk preferences and historical investment risk tolerance from multiple data sources risk ;

[0140] S15, integrating the above-collected personal basic information, historical transaction records, investment preferences and risk tolerance to construct a multi-dimensional user data set D user :D user ={D personal , D transaction , D preference , D risk}.

[0141] S2, cleaning and preprocessing the collected user multi-dimensional data set, including deduplication, outlier processing, and missing value filling; Figure 4 As shown, in this embodiment, S2 includes the following sub-steps:

[0142] S21, for the user multi-dimensional data set D user For duplicate data records, the most complete and accurate data is retained. The user multi-dimensional data set after deduplication is recorded as D user_unique ;

[0143] S22, after deduplication, the multi-dimensional user dataset D user_unique Perform outlier processing, use statistical methods to detect and mark outliers in the data set, correct or delete outliers, and the user multi-dimensional data set after outlier processing is recorded as D user_clean ;

[0144] S23, the user multi-dimensional data set D after outlier processing user_clean For missing data items, the missing values ​​are filled by using the mean value filling method, median filling method or nearest neighbor filling method. The filled user multidimensional data set is recorded as D user_completed ;

[0145] S24, the user multi-dimensional data set D after deduplication, outlier processing and missing value filling user_completed Standardization is performed to make data of different dimensions comparable. The standardized user multi-dimensional data set is recorded as D user_final .

[0146] S3 stores the cleaned and preprocessed user multi-dimensional data set in the form of triples and constructs a knowledge graph; the knowledge graph includes nodes and edges, where nodes represent users and their related attributes, and edges represent the relationship between user data.

[0147] The specific triple generation includes: for the user's investment preferences and risk tolerance described in text, a customized natural language processing (NLP) algorithm can be used to extract entities and relationships. The specific process includes: first, using a pre-trained word vector model to convert unstructured text into an embedded representation, then extracting key entities (such as investment type and risk level) through named entity recognition (NER), and using dependency syntactic analysis to determine the relationship between entities, and mapping this information into a triple in the form of (subject, predicate, object). For example, "User A prefers low-risk investment" is converted to ("User A", "preference", "low-risk investment").

[0148] Triple generation rules: Customized mapping rules are used to ensure that information extracted from different data sources can be uniformly mapped into the knowledge graph. For example, consistency check algorithms are used to ensure that attribute types match and avoid data conflicts.

[0149] like Figure 4 As shown, in this implementation, S3 includes the following sub-steps:

[0150] S31, the cleaned and preprocessed user multi-dimensional data set D user_final Converted into triples for storage, each triple is represented as (subject, predicate, object), and the triple generation rule is: Generate triples T(d i )=(s i , p i , o i ), where d i Represents each data record in the user's multidimensional data set, T(d i ) represents the generated triple set, s i Represents a user or user-related entity subject, p i Represents the relationship between entities predicate, o i Represents the attribute value object associated with the entity;

[0151] For the user's basic personal information, the user ID is used as the subject, the personal attribute is used as the predicate, and the attribute value is used as the object: personal (u) = {(u, "hasAge", a), (u, "hasGender", g), ...}; where u is the user ID, a is age, and g is gender;

[0152] For historical transaction records, use the transaction record identifier as the subject, the transaction type as the predicate, and the transaction details as the object: transaction (t) = {(t, "hasAmount", amt), (t, "hasDate", date), ...}; where t is the transaction record identifier, amt is the transaction amount, and date is the transaction date;

[0153] For investment preference, use the user ID as the subject, the investment preference type as the predicate, and the preference value as the object: preference (p) = {(u, "prefersInvestmentType", type), ...}; where type is the investment type preference;

[0154] For risk tolerance, use the user ID as the subject, the risk attribute as the predicate, and the attribute value as the object: risk (r) = {(u, "hasRiskTolerance", tolerance), ...}; where tolerance is the risk tolerance;

[0155] S32, construct an initial knowledge graph G0. The knowledge graph consists of nodes and edges. Nodes represent users and their related attributes, and edges represent the relationship between user data. The knowledge graph is expressed as: G0 = (V, E); where V is a node set, including user nodes and attribute nodes: E is an edge set, including the relationship between users and attribute nodes:

[0156] S33, maps the triple (subject, predicate, object) to the knowledge graph G0:

[0157] For each subject and object in a triple, if there is no corresponding node in the node set V, a new node is created:

[0158] if Then V←V∪{s};

[0159] if Then V←V∪{o};

[0160] Create new edges between nodes according to the predicate and update the edge set E: E←E∪{(s, p, o)}.

[0161] S4, perform semantic analysis on the user's unstructured data to extract key information; use natural language processing technology to perform semantic analysis on the user's unstructured data, extract key information, and enhance the semantic level of the knowledge graph. The NLP technologies used in this step include semantic parsing, text classification, and entity recognition. Through semantic parsing, the implicit meaning in the user's text is extracted, and related entities and relationships are added to the knowledge graph, for example, the description of "high-risk investment" is associated with "risk tolerance". An improved version of the deep learning-based language model (such as BERT) is used to enhance the accuracy of text processing and enhance the semantic level.

[0162] like Figure 4 As shown, in this implementation mode, S4 includes the following sub-steps:

[0163] S41, semantic relationship enhancement: Use natural language processing technology to perform semantic analysis on users' unstructured data, identify and add implicit semantic relationships:

[0164] Unstructured data sets, using NLP techniques to generate relational R semantic ={(u, "relatedTo", t j )|t j is a semantically related entity of u}; where u is a user node, t1, t2, …, t n is the relevant entity in the unstructured data, R semantic is a set of semantic relations;

[0167] S42, Redundant Edge Removal: Detect and remove duplicate or unnecessary edges in the knowledge graph:

[0168] E optimized =E\{e i |e i Repeated in the spectrum}; among them, E optimized is the optimized edge set, e i is the edge in the knowledge graph;

[0169] S43, Frequent subgraph mining: Use frequent subgraph mining algorithms to identify and extract subgraph patterns that frequently appear in user data: G sub =(V sub , E sub ) is a frequent subgraph pattern, then G optimized =G0∪{G sub |G sub Frequently appears in user data}; among them, G sub is the frequent subgraph pattern, V sub is the set of nodes in the frequent subgraph pattern, E sub is the edge set in the frequent subgraph pattern, G optimized It is the optimized knowledge graph.

[0170] S5, stores the constructed knowledge graph in the graph database; and improves the access efficiency of graph data through indexing and query optimization technology. Among them, the indexing and query optimization strategy of the knowledge graph: establish specific indexing strategies for nodes and edges, such as establishing cache indexes for frequently accessed nodes and relationships. At the same time, pre-computation and parallel processing technology are used to accelerate query performance. For example, a distributed graph computing framework is used to process large-scale graph data. Figure 4 As shown, in this implementation, S5 includes the following sub-steps:

[0171] S51, the optimized knowledge graph G optimizedStored in a graph database, which supports graph structure storage and query operations;

[0172] S52, index the knowledge graph data stored in the graph database, the index includes node index and edge index, to improve the speed and efficiency of data query. Node index: For each node v∈V optimized , establish an index based on node attributes; edge index: for each edge e∈E optimized , establish an index based on the relationship type;

[0173] S53, apply query optimization technology to optimize the data access efficiency of the knowledge graph through pre-computation, caching and parallel processing:

[0174] Pre-computation: For common queries, pre-calculate and store query results to reduce the cost of real-time calculations;

[0175] Cache: Cache frequently accessed data in memory to improve query response speed;

[0176] Parallel processing: For complex queries, parallel computing technology is used to improve query efficiency using multi-core processors or distributed computing environments;

[0177] S54 dynamically adjusts index and cache strategies based on user query requirements to ensure efficient data access performance in different query scenarios:

[0178] Dynamic indexing: dynamically adjust the structure of node index and edge index according to query frequency and access pattern;

[0179] Dynamic caching: Adjusts cached data and capacity based on real-time query load to ensure that frequently accessed data is always in the cache.

[0180] S6, based on the knowledge graph, through graph algorithms and machine learning models, conduct in-depth analysis of user data to generate user portraits including user personal information, behavioral characteristics, investment preferences and risk tolerance; the algorithms used in this step include: using community discovery algorithms (such as Louvain algorithm) to identify user groups, combining path analysis algorithms and centrality indicators to mine users' key behaviors. The generation process of user portraits is combined with deep learning models (models based on improved graph neural networks) to predict users' multi-dimensional characteristics, including behavioral characteristics, investment preferences, etc. Figure 4 As shown, in this implementation, S6 includes the following sub-steps:

[0181] S61, based on knowledge graph G optimized , apply community discovery algorithm, path analysis algorithm and node importance calculation algorithm to conduct in-depth analysis of user data;

[0182] S62, using the community discovery algorithm to identify user groups in the knowledge graph and generate user group labels: C = {c1, c2, ..., c n}; where C is the set of identified user groups, each c i represents a user group; for each user group c i , analyze their common features and generate user group labels L i ;

[0183] S63, apply the path analysis algorithm to calculate the user's behavior path in the knowledge graph and identify key behavior nodes. For each user node u∈V optimized , calculate the shortest path P(u, t) from u to the target node, identify the key behavior nodes on the path, and record them as K(u);

[0184] S64, using the node importance calculation algorithm, evaluates the importance of the user node in the knowledge graph, generates user behavior features, calculates the centrality index of each user node u, including degree centrality, closeness centrality and betweenness centrality, expressed as C(u), and generates user behavior features F according to the centrality index behavior (u);

[0185] S65, combined with the user's basic personal information D personal 、Historical transaction records D transaction 、Investment Preference D preference and risk tolerance D risk , use the machine learning model to generate user portraits and construct input feature vectors:

[0186] X u ={D personal , D transaction , D preference , D risk , L i , K(u), C(u)};

[0187] Use the pre-trained machine learning model M to predict the input feature vector and generate the user profile P u , User portrait P u Including multi-dimensional information such as users’ personal information, behavioral characteristics, investment preferences and risk tolerance.

[0188] S7, using the generated user portrait, combined with collaborative filtering and content recommendation technology, recommend the most suitable financial products to users; Among them, the specific algorithms of collaborative filtering and content recommendation technology: The solution of this application mentions the combination of collaborative filtering and content recommendation technology. This includes using user-based collaborative filtering (User-based CF) and content recommendation algorithms to construct a user-product rating matrix. Collaborative filtering mainly predicts users' ratings of unrated products by calculating the similarity between user portraits (such as cosine similarity or Pearson correlation coefficient). Content recommendation recommends financial products by calculating the similarity between user feature vectors and product feature vectors.

[0189] Dynamic adjustment of weight coefficients: In the scheme applied for this time, the comprehensive score of the recommendation result is obtained by weighted summation of collaborative filtering score and content recommendation similarity. The formula is S p =α·R uj +(1-α).sim(X u , Y p ), where the weight coefficient α can be adjusted according to actual needs. α can be dynamically adjusted based on user feedback to better balance the impact of collaborative filtering and content recommendation. Figure 4 As shown, in this implementation, S7 includes the following sub-steps:

[0190] S71, using the generated user portrait P u , extract the user's key feature vector X u ,The key feature vectors include the user’s personal information, behavioral characteristics, investment preferences, and risk tolerance;

[0191] S72, recommend financial products through collaborative filtering algorithm:

[0192] Construct the user-product rating matrix R, where R ij Represents user u i For product P j Ratings;

[0193] Using User Profile u Calculate the similarity between users sim(u i ,u j ), commonly used similarity calculation methods include cosine similarity and Pearson correlation coefficient:

[0194]

[0195] According to the similarity calculation results, find the k users most similar to the target user u, denoted as N u ;

[0196] Based on similar users N uRating records of target user u, predict the unrated product P j Rating R uj :

[0197]

[0198] S73, recommending financial products through content recommendation technology:

[0199] Construct feature vector Y for each financial product p , including product type, rate of return, and risk level attributes;

[0200] Using User Profile u Calculate user feature vector X u and the product feature vector Y p The similarity sim(X u , Y p ):

[0201]

[0202] Based on the similarity calculation results, recommend a set of financial products that best matches the user's characteristics;

[0203] S74, combining the recommendation results of collaborative filtering and content recommendation technology to generate the final financial product recommendation list P final , for the final recommended list of financial products P final For each recommended product p, calculate its comprehensive score S p , the comprehensive score is determined by the collaborative filtering score R uj and content recommendation similarity sim(X u , Y p )Weighted summation yields:

[0204] S p =α·R uj +(1-α)·sim(X u , Y p ); where α is the weight coefficient, which is adjusted according to actual needs;

[0205] According to the comprehensive score S p Sort the recommended products and generate the final recommendation list P final ;

[0206] S75, the final financial product recommendation list P final Push to users and record user feedback information.

[0207] S8, monitors changes in user behavior data in real time, and uses stream processing technology to dynamically update knowledge graphs and user portraits; optional implementations of stream processing technology include using Apache Flink or Spark Streaming, both of which support efficient real-time data processing. Through stream processing technology, changes in user behavior can be dynamically monitored, and the knowledge graph can be updated in a timely manner after the data is updated to ensure that the system can reflect the user's latest status. Frequency and rules for data updates: The rules for data updates can be set according to the frequency of user interactions. For example, for users who interact frequently, the window period for data updates can be shortened to ensure the timeliness of the portrait.

[0208] S9, based on the latest user portrait, the recommendation algorithm is continuously optimized using the machine learning model to ensure that the recommended financial products always meet the latest needs of users. Among them, this embodiment uses user feedback to optimize the recommendation algorithm. The user's clicks on the recommended results, browsing time and other behavioral data can be used as feedback, and the user portrait is continuously updated. Through the feedback collection mechanism, the model can be iteratively trained to ensure that the model can be adjusted as user needs change.

[0209] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0210] In this embodiment, a device for determining a user portrait based on a knowledge graph is also provided, which is used to implement the above-mentioned embodiments and preferred implementation modes, and will not be repeated hereafter. As used below, the term "module" can implement a combination of software and / or hardware for a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0211] Figure 5 is a structural block diagram of a device for determining a user profile based on a knowledge graph according to an embodiment of the present application, such as Figure 5 As shown, the device comprises:

[0212] A first collection module 52 is used to collect data of target users in N dimensions from multiple data sources to obtain an initial data set, where N is a natural number greater than 1;

[0213] A first processing module 54 is used to perform data cleaning and preprocessing operations on the initial data set to obtain a target data set;

[0214] A first conversion module 56 is used to convert the multiple data included in the target data set into a triple data format to obtain a triple set, wherein each triple in the triple set includes an entity of the target user, a relationship type of a product transaction performed by the entity, and a product transaction attribute corresponding to the entity;

[0215] A first construction module 58 is used to construct an initial knowledge graph of the target data set, wherein the nodes in the initial knowledge graph are used to represent the target user and the attributes of the target user, and the edges in the initial knowledge graph are used to represent the relationship between the multiple data included in the target data set;

[0216] A first mapping module 510 is used to map the triple set to the initial knowledge graph to obtain a target knowledge graph;

[0217] The first input module 512 is used to input the target knowledge graph into the target classification network to obtain the user portrait of the target user output by the target classification network.

[0218] In an exemplary embodiment, the above-mentioned first acquisition module includes at least one of the following: a first acquisition unit, used to collect data including user information and income information of the target user from multiple data sources to obtain the initial data set; a second acquisition unit, used to collect historical transaction record information of the target user from multiple data sources to obtain the initial data set; a third acquisition unit, used to collect investment data of the target user from multiple data sources to obtain the initial data set, wherein the investment data includes investment type, investment amount and investment period; a fourth acquisition unit, used to collect risk data of the target user from multiple data sources to obtain the initial data set, wherein the risk data includes risk assessment results, risk type and historical investment risk tolerance data.

[0219] In an exemplary embodiment, the above-mentioned first conversion module includes: a first parsing unit, used to parse the data attributes, data types and data content of each of the data; a first identification unit, used to identify the entities in the data according to the user names included in the data attributes; a first determination unit, used to determine the product transaction attributes corresponding to the entity according to the data type; a first extraction unit, used to extract the relationship type of the product transactions performed by the entity from the data content; and a first generation unit, used to generate a triple of each of the data according to the data format of the triple, thereby obtaining the triple set.

[0220] In an exemplary embodiment, the above-mentioned first mapping module includes: a first processing unit, used to map each triple in the triple set to the initial knowledge graph through the following steps to obtain the target knowledge graph: parsing the triple, converting the data format of the triple into the data format of the initial knowledge graph to obtain the target triple; adding nodes that match the target entity in the target triple and the product transaction attributes corresponding to the target entity to the initial knowledge graph; and creating a connection relationship between the matching nodes according to the relationship type of the product transaction performed by the target entity.

[0221] In an exemplary embodiment, the target classification network is a classification network obtained by training an initial classification network using multiple groups of sample training data, wherein the sample training data includes a sample knowledge graph, and the initial classification network includes a feature extraction subnetwork, a feature encoding subnetwork, a message passing subnetwork, and an output subnetwork, wherein for each group of the training data, the feature extraction subnetwork is used to extract local features and global features of sample nodes, sample edges, and connection relationships from the sample knowledge graph, wherein the local features and the global features both include features of sample users corresponding to the sample nodes and attribute features of the sample users, and relationship features between multiple sample data corresponding to the sample edges; the feature encoding subnetwork is used to encode the local features and the global features to obtain local feature vectors and global feature vectors; the message passing subnetwork is used to extract feature vectors of neighbor nodes of each sample node from the local feature vectors and the global feature vectors, and pass the feature vectors of the neighbor nodes to each sample node to obtain a target feature vector of each sample node; the output subnetwork is used to perform portrait classification on each sample node based on the target feature vector of each sample node, and output a user portrait of the sample user.

[0222] In an exemplary embodiment, the above-mentioned device includes: a first extraction module, which is used to input the target knowledge graph into the target classification network, obtain the user portrait of the target user output by the target classification network, and then extract the key feature vector of the target user from the user portrait, wherein the key feature vector includes the user characteristics, behavioral characteristics, investment characteristics and risk tolerance characteristics of the target user; a second construction module, which is used to construct a product rating matrix, wherein the product rating matrix includes the ratings of N users on the target product, wherein N is a natural number greater than 1; a first calculation module, which is used to calculate the similarity between the target user and the N users using the key feature vector; a first search module, which is used to search K users from the N users according to the N similarities, wherein the similarity between the K users and the target user is greater than a preset similarity; a first prediction module, which is used to predict the rating of the target user for the target product according to the K ratings corresponding to the K users in the product rating matrix; a first generation module, which is used to generate first recommendation information for recommending the target product to the target user according to the rating of the target user on the target product.

[0223] In an exemplary embodiment, the above-mentioned device also includes: a third construction module, which is used to generate first recommendation information for recommending the target product to the target user according to the target user's score for the target product, and then construct a product feature vector for the target product, wherein the product feature vector includes type characteristics, yield characteristics and risk level characteristics of the target product; a first calculation module, which is used to calculate the user feature vector of the target user using the user portrait; a second generation module, which is used to generate second recommendation information for recommending the target product to the target user based on the similarity between the user feature vector and the product feature vector; and a third generation module, which is used to generate target recommendation information for the target product using the first recommendation information and the second recommendation information.

[0224] In an exemplary embodiment, the above-mentioned device also includes: a first collection module, which is used to collect user behavior data after generating target recommendation information of the target product using the first recommendation information and the second recommendation information, wherein the user behavior data includes behavior data acting on the target product; a first input module, which is used to convert the user behavior data into a feedback signal, and input the feedback signal into the target classification network, so as to adjust the feature weight of the user portrait of the target user in the target classification network according to the feedback signal.

[0225] It should be noted that the above modules can be implemented by software or hardware. For the latter, it can be implemented in the following ways, but not limited to: the above modules are all located in the same processor; or the above modules are located in different processors in any combination.

[0226] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0227] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0228] An embodiment of the present application further provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0229] In an exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0230] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any one of the above method embodiments are implemented.

[0231] An embodiment of the present application further provides another computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any of the above method embodiments are implemented.

[0232] The embodiments of the present application also provide a computer program, which includes computer instructions, which are stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps of any one of the above method embodiments.

[0233] For specific examples in this embodiment, reference may be made to the examples described in the above embodiments and exemplary implementation modes, and this embodiment will not be described in detail herein.

[0234] Obviously, those skilled in the art should understand that the above modules or steps of the present application can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present application is not limited to any specific combination of hardware and software.

[0235] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for determining a user profile based on a knowledge graph, characterized in that: include: Collect data of target users in N dimensions from multiple data sources to obtain an initial data set, where N is a natural number greater than 1; Performing data cleaning and preprocessing operations on the initial data set to obtain a target data set; Converting the multiple data included in the target data set into a triple data format to obtain a triple set, wherein each triple in the triple set includes an entity of the target user, a relationship type of a product transaction performed by the entity, and a product transaction attribute corresponding to the entity; Constructing an initial knowledge graph of the target data set, wherein the nodes in the initial knowledge graph are used to represent the target user and the attributes of the target user, and the edges in the initial knowledge graph are used to represent the relationship between the multiple data included in the target data set; Mapping the triple set to the initial knowledge graph to obtain a target knowledge graph; The target knowledge graph is input into a target classification network to obtain a user profile of the target user output by the target classification network.

2. The method according to claim 1, characterized in that Collect data of target users in N product dimensions from multiple data sources to obtain an initial data set, including at least one of the following: Collecting data including user information and income information of the target user from a plurality of the data sources to obtain the initial data set; Collecting historical transaction record information of the target user from the plurality of data sources to obtain the initial data set; Collecting the target user's investment data from a plurality of data sources to obtain the initial data set, wherein the investment data includes investment type, investment amount and investment period; The risk data of the target user is collected from the plurality of data sources to obtain the initial data set, wherein the risk data includes risk assessment results, risk types, and historical investment risk tolerance data.

3. The method according to claim 1, characterized in that Convert the multiple data included in the target data set into a triple data format to obtain a triple set, including: Analyze the data attributes, data types and data contents of each of the data; identifying an entity in the data based on a user name included in an attribute of the data; Determine the product transaction attribute corresponding to the entity according to the data type; extracting, from the data content, a relationship type of a product transaction performed by the entity; The entity, the product transaction attribute and the relationship type of the product transaction performed by the entity in each of the data are combined according to the data format of the triplet to generate a triplet for each of the data to obtain the triplet set.

4. The method according to claim 1, characterized in that Mapping the triple set to the initial knowledge graph to obtain a target knowledge graph includes: For each triple in the triple set, the triple is mapped to the initial knowledge graph through the following steps to obtain the target knowledge graph: Parsing the triples, converting the data format of the triples into the data format of the initial knowledge graph, and obtaining target triples; Adding a node that matches both the target entity in the target triple and the product transaction attribute corresponding to the target entity to the initial knowledge graph; A connection relationship is created between the matching nodes according to the relationship type of the product transaction performed by the target entity.

5. The method according to claim 1, characterized in that The target classification network is a classification network obtained by training an initial classification network using multiple sets of sample training data, wherein the sample training data includes a sample knowledge graph, and the initial classification network includes a feature extraction subnetwork, a feature encoding subnetwork, a message passing subnetwork, and an output subnetwork, wherein for each set of the training data, The feature extraction subnetwork is used to extract local features and global features of sample nodes, sample edges, and connection relationships from the sample knowledge graph, wherein the local features and the global features both include features of sample users corresponding to the sample nodes and attribute features of the sample users, and relationship features between multiple sample data corresponding to the sample edges; The feature encoding subnetwork is used to perform encoding operations on both the local features and the global features to obtain a local feature vector and a global feature vector; The message passing subnetwork is used to extract the feature vectors of the neighboring nodes of each of the sample nodes from the local feature vectors and the global feature vectors, and pass the feature vectors of the neighboring nodes to each of the sample nodes to obtain the target feature vector of each of the sample nodes; The output subnetwork is used to perform profile classification on each of the sample nodes based on the target feature vector of each of the sample nodes, and output a user profile of the sample user.

6. The method according to claim 1, characterized in that After inputting the target knowledge graph into a target classification network to obtain a user portrait of the target user output by the target classification network, the method further includes: Extracting a key feature vector of the target user from the user portrait, wherein the key feature vector includes user characteristics, behavior characteristics, investment characteristics, and risk tolerance characteristics of the target user; Constructing a product rating matrix, wherein the product rating matrix includes ratings of N users on the target product, wherein N is a natural number greater than 1; Calculating the similarity between the target user and the N users using the key feature vector; Searching for K users from the N users according to the N similarities, wherein the similarity between the K users and the target user is greater than a preset similarity; Predicting the target user's rating of the target product according to the K ratings corresponding to the K users in the product rating matrix; According to the target user's rating of the target product, first recommendation information is generated to recommend the target product to the target user.

7. The method according to claim 6, characterized in that After generating first recommendation information for recommending the target product to the target user according to the target user's rating of the target product, the method further includes: Constructing a product feature vector of the target product, wherein the product feature vector includes type features, yield features, and risk level features of the target product; Calculating a user feature vector of the target user using the user portrait; generating second recommendation information for recommending the target product to the target user according to the similarity between the user feature vector and the product feature vector; The target recommendation information of the target product is generated by using the first recommendation information and the second recommendation information.

8. The method according to claim 7, characterized in that After generating target recommendation information of the target product using the first recommendation information and the second recommendation information, the method further includes: Collecting user behavior data, wherein the user behavior data includes behavior data acting on the target product; The user behavior data is converted into a feedback signal, and the feedback signal is input into the target classification network, so as to adjust the feature weight of the user portrait of the target user in the target classification network according to the feedback signal.

9. A device for determining a user profile based on a knowledge graph, characterized in that: include: A first collection module is used to collect data of target users in N dimensions from multiple data sources to obtain an initial data set, where N is a natural number greater than 1; A first processing module, configured to perform data cleaning and preprocessing operations on the initial data set to obtain a target data set; A first conversion module is used to convert the multiple data included in the target data set into a triple data format to obtain a triple set, wherein each triple in the triple set includes an entity of the target user, a relationship type of a product transaction performed by the entity, and a product transaction attribute corresponding to the entity; A first construction module is used to construct an initial knowledge graph of the target data set, wherein the nodes in the initial knowledge graph are used to represent the target user and the attributes of the target user, and the edges in the initial knowledge graph are used to represent the relationship between the multiple data included in the target data set; A first mapping module, used for mapping the triple set to the initial knowledge graph to obtain a target knowledge graph; The first input module is used to input the target knowledge graph into the target classification network to obtain the user portrait of the target user output by the target classification network.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program implements the steps of the method described in any one of claims 1 to 8 when executed by a processor.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method described in any one of claims 1 to 8 are implemented.

Images