Financial user portrait analysis method and system based on knowledge graph

Abstract

The invention discloses a financial user portrait analysis method and system based on a knowledge graph, and the method comprises the steps: recognizing a financial entity through an FNER algorithm, carrying out the relation embedding through employing a TransFinE algorithm, integrating the time weight attenuation and risk propagation constraints, constructing a financial knowledge graph, calculating the user influence distribution through employing a FinRank algorithm, and carrying out the analysis of the financial user portrait. And finally carrying out semantic clustering and grouping to obtain a user portrait classification result. The technical problems that an existing financial user portrait analysis method cannot effectively process a multi-dimensional semantic association relationship, lacks a user network influence quantification mechanism and neglects time sequence evolution characteristics are solved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a financial user portrait analysis method and system based on knowledge graph. Background Art

[0002] Financial institutions currently primarily use traditional user profiling methods to build customer profile models. These methods typically employ statistical analysis and machine learning modeling based on structured customer profiles, transaction flow data, and credit records. Traditional methods, including rule-based expert systems, decision tree classification, and cluster analysis, analyze user characteristics such as age, income, occupation, and historical transaction behavior to categorize users into different risk levels and product preference categories. These methods play a vital role in business scenarios such as financial risk assessment, product recommendations, and precision marketing, providing data support for financial institutions' customer management and business decision-making.

[0003] However, existing technologies have significant shortcomings: data quality and integrity issues are prominent, data formats and quality standards of different data sources vary greatly, and there are noise, missing values ​​and inconsistencies, which affect the accuracy of user portrait construction; computational complexity and scalability issues are serious. With the increase in user scale and data dimension, the computational overhead of traditional algorithms grows exponentially, making it difficult to process large-scale data; dynamic and real-time performance is insufficient, and most existing methods are based on static data analysis, making it difficult to capture and respond to dynamic changes in user behavior in a timely manner; model interpretability and transparency need to be improved, and complex machine learning models are often "black box" in nature, making it difficult for financial institutions to understand and explain the decision-making process of the model.

[0004] Based on the above analysis, existing technologies face new challenges when dealing with the multi-dimensional relationships among financial users: traditional planar data analysis methods cannot effectively represent and utilize the complex semantic relationships between users, products, and risks, resulting in a lack of in-depth relationship mining in user portrait analysis; the semantic integration and entity alignment problems of multi-source heterogeneous financial data have not been effectively solved, and the data silos between different systems restrict the comprehensiveness and accuracy of user portraits; there is a lack of effective technical means for quantitative modeling of user influence propagation and risk transmission mechanisms in financial networks, and existing methods find it difficult to accurately identify key users and assess network risks; the temporal evolution characteristics and dynamic update mechanisms of user behavior patterns have not been fully considered in user portrait analysis, resulting in portrait results lagging behind actual business needs. Summary of the Invention

[0005] This application provides a financial user portrait analysis method and system based on knowledge graph, which is used to solve the technical problems that existing financial user portrait analysis methods cannot effectively handle multi-dimensional semantic association relationships, lack a mechanism to quantify user network influence, and ignore temporal evolution characteristics.

[0006] The object of the present invention is achieved like this:

[0007] In the first aspect, the present application provides a financial user portrait analysis method based on a knowledge graph, and the financial user portrait analysis method based on the knowledge graph includes: performing financial entity identification processing on bank transaction data, credit data and user association data through the FNER algorithm to obtain user entities, product entities and risk entities; performing TransFinE relationship embedding processing on multidimensional feature data according to the user entities, product entities and risk entities to obtain a relationship vector representation containing time weight decay and risk propagation constraints; performing graph structure construction processing on the relationship vector representation to obtain a financial knowledge graph containing user nodes, product nodes and risk propagation paths; performing influence weight calculation processing on the user nodes in the financial knowledge graph through the FinRank algorithm to obtain a user influence distribution that integrates risk propagation coefficients and asset weights; performing semantic clustering and grouping processing on user features according to the user influence distribution to obtain user portrait classification results containing risk tolerance, product preferences and investment potential.

[0008] In a second aspect, the present application provides a financial user portrait analysis system based on a knowledge graph, the financial user portrait analysis system based on a knowledge graph comprising:

[0009] The identification module is used to perform financial entity identification processing on bank transaction data, credit data and user association data using the FNER algorithm to obtain user entities, product entities and risk entities;

[0010] An embedding module, configured to perform TransFinE relation embedding processing on the multi-dimensional feature data according to the user entity, product entity, and risk entity, to obtain a relation vector representation including time weight decay and risk propagation constraints;

[0011] A construction module, configured to construct a graph structure of the relationship vector representation to obtain a financial knowledge graph comprising user nodes, product nodes, and risk propagation paths;

[0012] A calculation module is used to calculate the influence weight of user nodes in the financial knowledge graph using the FinRank algorithm to obtain a user influence distribution that integrates risk propagation coefficients and asset weights;

[0013] The grouping module is used to perform semantic clustering and grouping processing on user characteristics according to the user influence distribution, and obtain user portrait classification results including risk tolerance, product preference and investment potential.

[0014] In a third aspect, a knowledge graph-based financial user portrait analysis device is provided, comprising a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to enable the knowledge graph-based financial user portrait analysis device to perform the knowledge graph-based financial user portrait analysis method described above.

[0015] In a fourth aspect, a computer-readable storage medium is provided, the computer-readable storage medium storing instructions, which, when executed on a computer, enable the computer to perform the knowledge graph-based financial user portrait analysis method described above.

[0016] In the technical solutions provided in the present application, the FNER algorithm is used to perform financial entity recognition processing on bank transaction data, credit data and user association data, overcoming the problem of insufficient precision of traditional entity recognition methods in the financial field, and significantly improving the recognition accuracy of financial terms and product names, thereby obtaining more accurate user entity, product entity and risk entity basic data. The TransFinE relationship embedding processing solves the technical defects of traditional relationship embedding methods by incorporating time weight decay and risk propagation constraints, which ignores the timeliness and risk propagation characteristics of financial relationships, so that the relationship vector representation can accurately reflect the dynamic characteristics and risk correlation of financial businesses. The graph structure construction processing converts the relationship vector representation into a financial knowledge graph containing user nodes, product nodes and risk propagation paths, breaking through the limitations of traditional flat data analysis and constructing a graph structure model that can express multi-dimensional semantic associations. The FinRank algorithm innovatively solves the technical problem of user influence power in the financial network by calculating the influence weight by fusing the risk propagation coefficient and the asset weight, which better adapts to the business characteristics of the financial field compared to the traditional PageRank algorithm. The semantic clustering grouping processing clusters user features based on user influence distribution, overcoming the shortcomings of traditional clustering methods that ignore network structure information, and the generated user portrait classification results more accurately reflect the real status and behavior characteristics of users in the financial ecosystem, providing more reliable decision-making basis for precise marketing, risk control and personalized services of financial institutions.

[0017] In the field of financial user profiling, the FNER algorithm's customized design for the financial field significantly improves the accuracy of entity recognition compared to general named entity recognition methods, especially when processing professional terms such as financial product codes and risk level identifiers; the time weight decay mechanism and risk propagation constraint mechanism introduced by the TransFinE algorithm are specifically optimized for the timeliness and risk transmission characteristics of financial relationships, so that the embedding vector can more accurately encode the core characteristics of financial business; the FinRank algorithm incorporates the homogeneous propagation law of financial risks into the influence calculation process through the innovative design of the risk propagation coefficient, and better identifies key nodes and risk propagation paths in financial networks than traditional graph algorithms; the semantic clustering algorithm combined with graph connectivity constraints ensures the rationality of the clustering results in business logic, and the eight user groups generated accurately correspond to the customer stratification needs in the actual business of financial institutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 This is a schematic diagram of an embodiment of a financial user profile analysis method based on a knowledge graph in an embodiment of the present application;

[0020] Figure 2 This is a schematic diagram of an embodiment of a financial user profile analysis system based on a knowledge graph in an embodiment of the present application;

[0021] Figure 3 It is a schematic block diagram of the structure of a financial user portrait analysis device based on a knowledge graph in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The embodiments of the present application provide a method and system for analyzing financial user portraits based on a knowledge graph. The terms "first", "second", "third", "fourth", etc. (if any) 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. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of the financial user portrait analysis method based on the knowledge graph includes:

[0024] Step S101: Perform financial entity identification processing on bank transaction data, credit data, and user-related data using the FNER algorithm to obtain user entities, product entities, and risk entities;

[0025] Step S102: Perform TransFinE relation embedding processing on the multi-dimensional feature data based on the user entity, product entity, and risk entity to obtain a relation vector representation containing time weight decay and risk propagation constraints;

[0026] Step S103: constructing a graph structure based on the relationship vector representation to obtain a financial knowledge graph including user nodes, product nodes, and risk propagation paths;

[0027] Step S104: Calculate the influence weight of user nodes in the financial knowledge graph using the FinRank algorithm to obtain a user influence distribution that integrates risk propagation coefficients and asset weights.

[0028] Step S105: semantically cluster and group user features according to user influence distribution to obtain user portrait classification results including risk tolerance, product preference and investment potential.

[0029] It is understandable that the execution subject of this application can be a financial user profile analysis system based on knowledge graph, or a terminal or server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0030] Specifically, the FNER algorithm is used to identify financial entities in bank transaction data, credit information, and user-related data. The FNER algorithm first constructs a specialized dictionary for the financial sector. This dictionary contains bank product names such as "credit card," "financial product," and "mortgage," financial terms such as "overdue," "default," and "guarantee," and risk level indicators such as "high risk," "medium risk," and "low risk." The raw data is then fed into a bidirectional long short-term memory (LSTM) network, a neural network capable of processing both forward and backward sequential information. It encodes text sequences to generate feature vectors. The feature vectors are then annotated using a conditional random field (CRF) annotation algorithm based on the financial entity vocabulary. CRF is a sequence annotation algorithm that can identify entity boundaries. Ultimately, candidate entities such as user name, ID number, bank card number, product code, and risk level indicator are extracted and grouped by type into user entities, product entities, and risk entities.

[0031] Based on the identified entities, the multidimensional feature data is processed through TransFinE relational embedding. The TransFinE algorithm is an embedding method specifically designed for financial relationships. It first constructs relational triples, representing the ownership relationship between users and products, the social relationships between users, and the association between products and risk levels as a head entity-relationship-tail entity. Time weights are then calculated based on historical transaction times. The recent relationship weight is set to 1.0 and decreases by 0.05 each month to reflect the timeliness of financial relationships. Risk propagation constraints are also calculated, strengthening the association between two entities with the same risk level and reducing it when the risk levels differ significantly. Iterative optimization is performed using the objective function where the head entity vector plus the relationship vector equals the tail entity vector, generating a relationship vector representation that incorporates time decay and risk constraints.

[0032] The relationship vector representation is processed to construct a graph structure. First, a node attribute matrix is ​​constructed based on the relationship vector. This matrix contains characteristics such as the user's age, income, and credit score; characteristics such as the product's yield, risk level, and term; and characteristics such as the risk type, level, and impact range. An adjacency matrix is ​​then generated based on the relationship mapping between the head and tail entities. The adjacency matrix records the connection relationships between nodes, with a value of 1 indicating a connection and a value of 0 indicating no connection. Next, the degree centrality and betweenness centrality of each node are calculated. Degree centrality reflects the number of direct connections a node has, while betweenness centrality reflects the node's role as a bridge in the network. Based on these centrality values, the graph is optimized, removing redundant nodes and edges with low centrality to form a streamlined graph structure. Finally, the graph is organized hierarchically by user, product, and relationship layers to construct a financial knowledge graph.

[0033] The FinRank algorithm calculates the influence weight of user nodes. The FinRank algorithm first calculates edge weights by taking the logarithm of the transaction amount and multiplying it by 0.5, multiplying the transaction frequency by 0.3, and multiplying the relationship duration by 0.2. The three are summed to obtain the composite edge weight. The risk propagation coefficient is then calculated based on differences in user risk levels. The coefficient for users with the same risk level is 1.2, for users with a one-level difference, it is 1.0, and for users with a difference of two or more levels, it is 0.8. The edge weights are multiplied by the risk propagation coefficient to obtain the adjusted edge weights. The user nodes are then initialized with a damping coefficient of 0.85, giving each node equal initial influence. An iterative calculation is then performed, with each node's influence equal to 0.15 divided by the total number of nodes plus 0.85 multiplied by the weighted sum of the influences of all neighboring nodes. Iterations are continued until convergence occurs, with the difference between two consecutive calculations less than 0.000001. This yields the final user influence distribution.

[0034] Semantic clustering is performed based on user influence distribution. First, the user influence distribution is fused with the node feature vector to generate a comprehensive user feature vector containing influence weights. A user similarity matrix is ​​then constructed, and the Euclidean distance between each pair of users is calculated as the similarity metric. A spectral clustering algorithm is then applied with graph connectivity constraints to ensure strong connectivity between users of the same category in the original knowledge graph. A grid search determines the optimal number of clusters to be eight, classifying users into eight categories: conservative investors, steady investors, growth investors, aggressive investors, high-net-worth individuals, small and micro business owners, young white-collar workers, and retirees. Finally, each user category is labeled with risk tolerance, product preferences, and investment potential to form a user profile classification result. For example, the transaction data of a bank user named Zhang shows that he holds multiple wealth management products with a high transaction frequency. The FNER algorithm identifies that his user entity contains attributes such as high income and high education, and his product entity contains high-risk products such as stock funds and hybrid funds. The risk entity indicates that he has a high risk tolerance. After TransFinE relationship embedding, Zhang's association vector with high-risk products has a high weight, and the time weight indicates recent trading activity. Zhang's node in the knowledge graph has high degree centrality and betweenness centrality, indicating a prominent position within the financial network. The FinRank algorithm calculated Zhang's influence distribution value to be high, and cluster analysis categorized him as an active investor, characterized by a high risk tolerance, a preference for high-yield products, and significant investment potential.

[0035] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0036] Construct a financial domain-specific dictionary containing bank product names, financial terms, and risk level identifiers to obtain a financial entity vocabulary database;

[0037] Bank transaction data, credit data, and user association data are input into a bidirectional long short-term memory network for sequence encoding to obtain a text sequence feature vector.

[0038] Based on the financial entity vocabulary, the text sequence feature vector is annotated with conditional random fields to obtain the entity boundary label sequence;

[0039] Perform entity extraction on the original text data based on the entity boundary marker sequence to obtain a candidate entity set of user name, ID number, bank card number, product code and risk level identifier;

[0040] The candidate entity set is classified and merged according to entity type to obtain user entities, product entities and risk entities.

[0041] Specifically, the construction of a specialized financial dictionary begins with the collection of core banking vocabulary. The dictionary comprises three main categories of vocabulary. The Banking Product Name category includes specific product names such as credit cards, savings cards, wealth management products, funds, insurance, mortgages, auto loans, and consumer loans. The Financial Terminology category encompasses specialized terms such as overdue, default, guarantee, mortgage, credit reporting, risk control, credit limit, interest rate, and yield rate. The Risk Level Identifier category covers risk-related identifiers such as high risk, medium risk, low risk, no risk, risk exposure, and risk coefficient. The dictionary construction process extracts standardized vocabulary by analyzing extensive banking business documents and regulatory filings. This, combined with the field name and code mappings within the bank's internal business systems, creates a structured financial entity vocabulary library that includes a vocabulary ontology, synonym mappings, and hypernymy relationships. The bidirectional long short-term memory network performs sequence encoding on the input data by first converting bank transaction data, credit reporting data, and user-related data into a text sequence format. The bidirectional long short-term memory network (LSTM) is a variant of a recurrent neural network (RNN) capable of processing information in both forward and backward directions. Forward propagation processes information from the beginning of a sentence to the end, while backward propagation processes information from the end of a sentence to the beginning. The network's input layer receives a sequence of word vectors processed through word embeddings. The hidden layer controls the flow of information through three gating mechanisms: the forget gate, the input gate, and the output gate. The forget gate determines which information is discarded from the cell state, the input gate determines which new information is stored, and the output gate determines which information is output. The forward and backward hidden states are concatenated to generate a text sequence feature vector containing contextual semantic information. This vector is typically twice the dimension of the bidirectional hidden layer.

[0042] The conditional random field labeling process identifies entity boundaries based on the financial entity vocabulary for the text sequence feature vector. The conditional random field is a discriminative probabilistic model that captures the dependency between labels and the relationship between the observed sequence and the label sequence by defining a feature function. The labeling process uses the BIO labeling system, where B represents the start of the entity position, I represents the internal position of the entity, and O represents the non-entity position. The conditional random field determines the optimal labeling sequence by calculating the weights of the state feature function and the transition feature function. The state feature function describes the relationship between the current position observation and the label, and the transition feature function describes the transition relationship between adjacent position labels. The Viterbi algorithm is used to find the label sequence with the maximum probability under the given observation sequence. The maximum probability path of each label at each position is calculated through dynamic programming, and the final entity boundary label sequence is output.

[0043] The entity extraction process extracts specific entity content from the original text data based on the entity boundary label sequence. The extraction process traverses the label sequence, records the start of the entity when encountering a B label, continues to record the entity content when encountering a corresponding I label, and ends the current entity recording when encountering an O label or a different type of B label. The entity content is extracted from the original text through string slicing operations, while recording the entity type and position information in the text. The extraction result contains candidate entity content of different types, such as user name entities like "Zhang San" and "Li Si", ID number entities like "110101199001011234", bank card number entities like "6225881234567890", product code entities like "LCP001" and "FND002", and risk level identification entities like "high risk" and "AA level".

[0044] The classification and merging processing groups the candidate entity set according to the predefined entity type. The processing process establishes an entity type mapping table, the user entity type includes user name, ID number, mobile phone number, address and other personal identification information, the product entity type includes product code, product name, product category, yield and other product related information, and the risk entity type includes risk level, risk coefficient, risk description, risk category and other risk related information. The merging process judges the specific type of the candidate entity through string matching and regular expression matching, different identification information of the same user is aggregated through the association field, different attribute information of the same product is associated through the product code, and different descriptions of the same risk category are merged through semantic similarity calculation. For example, when a bank customer manager enters the information of a certain customer, the original data contains the text content "the customer Wang Mou Mou applies for the financial product LCP001 high-risk product and needs risk assessment", the bidirectional long short-term memory network converts the text into a feature vector sequence containing semantic information, the conditional random field identifies "Wang Mou Mou" corresponding to B-PER and I-PER labels, "LCP001" corresponding to B-PRO label, and "high risk" corresponding to B-RISK label, the entity extraction processing extracts the user name "Wang Mou Mou", the product code "LCP001", and the risk level "high risk" three candidate entities, and the classification and merging processing classifies "Wang Mou Mou" into the user entity category, "LCP001" into the product entity category, and "high risk" into the risk entity category.

[0045] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0046] Based on the user entity, the product entity and the risk entity, a relationship triple construction processing is performed to obtain original triple data of user-product holding relationship, user-user social relationship and product-risk level relationship;

[0047] According to the historical transaction time, a time weight calculation processing is performed on the original triple data to obtain a time decay weight coefficient with a recent relationship weight of 1.0 and a monthly decrease of 0.05;

[0048] The time decay weight coefficient and the risk level difference are calculated by a risk propagation constraint calculation processing to obtain a risk consistency loss function and a time loss function;

[0049] Based on the head entity vector plus the relationship vector equal to the tail entity vector, an iterative optimization processing is performed on the entity vector and the relationship vector to obtain a low-dimensional dense vector representation of the user entity, the product entity and the risk entity;

[0050] The low-dimensional dense vector representation is processed by vector combination according to the head entity, the relationship and the tail entity to obtain a relationship vector representation containing time weight decay and risk propagation constraint.

[0051] Specifically, the relationship triple construction process establishes inter-entity associations based on the identified user, product, and risk entities. A triple is the fundamental representation unit of a knowledge graph, consisting of a head entity, a relationship, and a tail entity, and is used to describe the semantic relationship between two entities. User-product holding relationship triples are constructed by analyzing a user's product purchase history and holding status. For example, when user Zhang holds wealth management product LCP001, the triple "Zhang-Hold-LCP001" is formed, with the relationship type marked as "Hold." User-user social relationship triples are constructed by analyzing inter-user transfer records, guarantee relationships, and recommendation relationships. When user Zhang guarantees a loan for user Li, the triple "Zhang-Guarantee-Li" is formed, with the relationship type marked as "Guarantee." Product-risk level relationship triples are constructed using product risk rating data. When wealth management product LCP001 is rated high risk, the triple "LCP001-Risk Level-High Risk" is formed, with the relationship type marked as "Risk Level." The construction process traverses all entity combinations, determines whether there is a specific relationship between entities through relationship pattern matching and business rules, and records the qualified entity pairs and their relationship types as raw triple data.

[0052] The time-weighted calculation process quantifies the timeliness of raw triple data based on historical transaction times. The time-decay weight coefficient reflects the timeliness of financial relationships, with recent relationships having greater importance and influence. The calculation process first determines the time of occurrence of each triple relationship. The specific time of relationship establishment is determined using time information such as transaction timestamps, account opening times, and contract signing times. The difference in months between the relationship establishment time and the current time is then calculated. Recent relationships (those occurring within the current month) are weighted 1.0, indicating the strongest influence. The weight decay rule is set to decrease by 0.05 per month: relationships from one month ago receive a weight of 0.95, relationships from two months ago receive a weight of 0.90, and so on, until the weight reaches the minimum threshold. The calculation formula is: the weight is 1.0 minus the difference in months multiplied by 0.05. The weight approaches zero when the difference in months exceeds twenty months. This time-decay mechanism ensures that the relationship weights in the knowledge graph reflect the dynamics of the financial market and the timeliness of user behavior.

[0053] The risk propagation constraint calculation process takes into account the time decay weight coefficient and the difference in risk levels. The risk propagation constraint is based on the propagation law of financial risks in the user network. The relationship between users with similar risk levels has a stronger influence, while the relationship between users with large differences in risk levels has a weaker influence. The risk consistency loss function quantifies the degree of risk inconsistency by calculating the absolute value of the difference in risk levels between related users. When two users have the same risk level, the loss value is zero, and the greater the difference in risk level, the higher the loss value. The time loss function strengthens the influence of the time factor through the square term of the time decay weight coefficient. The longer the relationship, the less influence it has on the current user profile. The two loss functions are combined through weighted summation. The weight coefficient is determined according to business needs and historical data verification results to form a comprehensive constraint term for the subsequent vector optimization process.

[0054] The iterative optimization process trains the entity and relationship vectors based on an objective function that requires the sum of the head entity vector and the relationship vector to equal the tail entity vector. This objective function is derived from the core concept of the TransE algorithm, which states that in the embedding space, relationships are represented as translation vectors from the head entity to the tail entity. The optimization process uses a stochastic gradient descent algorithm to update the vector parameters by calculating the gradient of the objective function. During each iteration, the algorithm randomly samples a batch of triples, calculates the objective function value for the current vector parameters, then calculates the partial derivatives of the objective function with respect to each vector parameter, and updates the vector parameters based on the learning rate and gradient direction. The optimization objective function consists of a weighted combination of a distance term, a risk consistency loss term, and a time loss term. The distance term ensures that the head entity vector plus the relationship vector of the correct triple is as close as possible to the tail entity vector, while the loss term ensures that the vector representation conforms to the timeliness and risk propagation characteristics of financial services. The iterative process continues until the objective function converges or the preset number of iterations is reached, ultimately obtaining low-dimensional dense vector representations of user entities, product entities, and risk entities.

[0055] The vector combination processing reorganizes the low-dimensional dense vector representation according to the structure of the head entity, the relationship, and the tail entity. The combination process traverses all original triple data, and according to the identifiers of the head entity and the tail entity in the triple, finds the corresponding vector from the vector representation dictionary, and according to the relationship type, finds the corresponding relationship vector. The head entity vector, the relationship vector, and the tail entity vector are spliced or other mathematical operations in a fixed order to form a relationship vector representation containing complete triple information. This relationship vector representation not only retains the semantic information of the original triple, but also integrates the features of time weight decay and risk propagation constraints. For example, user Liu Mou bought a medium-risk financial product MCP002 three months ago, and the original triple is "Liu Mou-holds-MCP002". The time weight calculation shows that the weight of this relationship is 0.85, and the risk level of Liu Mou is medium risk, which is consistent with the product risk level, so the risk consistency loss is zero. After iteration and optimization processing, the user vector of Liu Mou, the holding relationship vector, and the product MCP002 vector all integrate time decay and risk constraint information. Finally, the vector combination processing combines the three vectors according to the triple structure to form a relationship vector representation containing time weight decay and risk propagation constraints. This vector representation not only reflects the holding relationship between the user and the product, but also embodies the timeliness and risk propagation characteristics of the relationship.

[0056] In a specific embodiment, the process of performing step S103 can specifically include the following steps:

[0057] Based on the relationship vector representation, a node attribute matrix construction process is performed to obtain a node attribute matrix containing user features, product features, and risk features;

[0058] According to the head entity and tail entity relationship in the relationship vector representation, a mapping process is performed on the connection relationship to obtain adjacency matrices of user-product connections, user-user connections, and product-risk connections;

[0059] The node attribute matrix and the adjacency matrix are subjected to graph structure calculation processing to obtain the degree centrality and betweenness centrality values of each node;

[0060] Based on the degree centrality and betweenness centrality values, a graph optimization process is performed to obtain an optimized graph structure after deleting low-weight nodes and edges;

[0061] The optimized graph structure is subjected to hierarchical organization processing according to the user layer, the product layer, and the relationship layer to obtain a financial knowledge graph containing user nodes, product nodes, and risk propagation paths.

[0062] Specifically, the node attribute matrix construction process extracts feature information from various entities based on the relationship vector representation and organizes it into a matrix format. The node attribute matrix serves as the input data structure for the graph neural network. The rows of the matrix represent the nodes in the graph, and the columns represent the feature dimensions of the nodes. The construction process first traverses all relationship vector representations, identifying the identifiers of the head and tail entities, and extracting the corresponding vector representations of the entities as node features. User features include dimensional information such as age, income level, occupation category, education background, credit score, and risk preference parsed from the user vector representation. These features are encoded and stored in different dimensional positions of the vector. Product features include attribute information such as product type, yield rate, risk level, investment period, minimum investment amount, and issuing institution extracted from the product vector representation. Each product node corresponds to a row of data in the matrix. Risk features include risk-related attributes such as risk type, risk level, impact scope, regulatory classification, and historical loss data, obtained from the risk vector representation. The matrix construction process converts the scattered vector representation into a regular matrix format through vector indexing and feature mapping. The row index corresponds to the node number, the column index corresponds to the feature dimension, and the matrix element values ​​are derived from the corresponding dimension values ​​in the vector representation.

[0063] The connection relationship mapping process constructs the graph topology based on the head and tail entity information in the relationship vector representation. The adjacency matrix is ​​a mathematical representation of the graph structure. The matrix size is a square matrix of the number of nodes multiplied by the number of nodes, and the matrix element value indicates whether there is a connection relationship between the corresponding nodes. The mapping process traverses all relationship vector representations, extracts the head and tail entity identifiers of each relationship, finds the corresponding row and column positions in the adjacency matrix, and sets the matrix element value to 1 to indicate a connection and 0 to indicate no connection. User-product connections are determined by the relationship between users and products. When an ownership relationship exists between a user node and a product node, the connection is marked at the corresponding position in the adjacency matrix. User-user connections are determined by social relationships, guarantee relationships, and recommendation relationships between users. When any of these relationships exist between two user nodes, the connection is marked. Product-risk connections are determined by the risk level assignment of products. When a risk level relationship exists between a product node and a risk node, the connection is marked. The mapping process also needs to account for the directionality of the relationship. For directed relationships, the connection is only marked on one side of the adjacency matrix; for undirected relationships, the connection is marked at both symmetric positions in the adjacency matrix.

[0064] Graph structure calculations combine the node attribute matrix with the adjacency matrix to calculate the graph's topological characteristic indicators. Degree centrality calculates the number of edges connecting each node. The node's degree is calculated by counting the number of nonzero elements in each row or column of the adjacency matrix. A higher degree value indicates a denser connection within the graph. Betweenness centrality calculates how often a node serves as a bridge between other nodes in the graph. The calculation process first uses a breadth-first search algorithm to find all shortest paths between any two nodes in the graph. The number of times each node appears on a shortest path is then counted. A higher number of appearances indicates a more important node's role as a bridge. Degree centrality reflects the local importance of a node, while betweenness centrality reflects its global importance. The two metrics are combined to assess a node's influence and status within the financial network. The calculation process traverses all nodes, calculating both degree centrality and betweenness centrality for each node, forming a vector of node importance scores.

[0065] Graph optimization processes identify and delete redundant nodes and edges in the graph based on degree centrality and betweenness centrality values. The optimization process sets a centrality threshold. Nodes with degree centrality below the threshold are considered isolated nodes or edge nodes, and nodes with betweenness centrality below the threshold are considered non-critical nodes. The deletion strategy first identifies nodes that meet both degree centrality and betweenness centrality below the threshold, removes these nodes from the graph structure, and deletes all edge connections associated with these nodes. Edge deletion is based on the importance score of the connected nodes. When the importance scores of both connected nodes are low, the edge is deleted. When the connection strength is below the set threshold, the edge is also deleted. The optimization process needs to maintain the connectivity of the graph to avoid deleting key connections, which would cause the graph to split into multiple isolated subgraphs. The deletion operation is achieved by modifying the adjacency matrix and the node attribute matrix, setting the corresponding matrix rows and columns to zero or directly removing them from the matrix, forming an optimized graph structure that is smaller in scale but retains the core structure.

[0066] Hierarchical organization optimizes the graph structure by arranging it hierarchically according to node type. Hierarchical organization is a typical organizational approach for knowledge graphs, enhancing the readability and operability of the graph by assigning different types of entities to different levels. The user layer contains nodes of all user types, representing individual and corporate bank customers. Node attributes include basic customer information, financial status, and behavioral characteristics. The product layer contains nodes of all financial product types, including savings, investment, credit, and insurance products. Node attributes include product yield characteristics, risk characteristics, and maturity characteristics. The relationship layer is not an independent node layer, but rather a collection of edges connecting the user and product layers. The relationship layer contains risk propagation paths, which describe the direction and intensity of risk propagation within user and product networks. The hierarchical organization process assigns nodes to corresponding levels based on their entity type identifiers. Edges are then assigned to either inter-layer or intra-layer connections based on their connection type, forming a clearly structured three-layer financial knowledge graph. For example, consider a bank customer named Wang who holds a high-risk stock fund product and has a guarantee relationship with his colleague named Li. The node attribute matrix construction process encodes Wang's age, income, and risk preference as user node attributes, the stock fund's yield, risk level, and investment period as product node attributes, and the high-risk indicator as a risk node attribute. The connection mapping process marks the connection between Wang and the fund product, the guarantee connection between Wang and Li, and the hierarchical connection between the fund product and high-risk in the adjacency matrix. Graph structure calculations show that Wang's node has a degree centrality of three, indicating that it has three connections. A high betweenness centrality indicates that it occupies a key position in the guarantee network. Graph optimization retains key nodes like Wang and removes marginal customer nodes with few connections. The hierarchical organization process assigns Wang to the user layer, fund products to the product layer, and the high-risk indicator to the risk attribute. Wang's holding relationship with the fund product and the hierarchical relationship between the fund product and high-risk constitute the risk propagation path, ultimately forming a complete financial knowledge graph containing user nodes, product nodes, and risk propagation paths.

[0067] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0068] Based on the transaction amount, transaction frequency and relationship duration of user nodes in the financial knowledge graph, edge weight calculation is performed to obtain the edge weight values ​​between user nodes;

[0069] The risk propagation coefficient is calculated based on the risk level differences between user nodes to obtain the risk propagation coefficient between users with different risk levels;

[0070] The edge weight value is integrated with the risk propagation coefficient to obtain the adjusted edge weight value;

[0071] Initialize the weight distribution of user nodes based on the preset damping coefficient to obtain the initial influence value of each user node;

[0072] Iteratively update the current node influence based on the adjusted edge weight value and neighbor node influence value to obtain the updated node influence value;

[0073] The updated node influence values ​​are subjected to convergence judgment processing to obtain the user influence distribution that integrates the risk propagation coefficient and asset weight.

[0074] Specifically, edge weight calculation is based on quantitative analysis of transaction behavior data between user nodes in the financial knowledge graph. The transaction amount reflects the scale of funds between users. The calculation process extracts all historical transaction records between users, accumulates the transfer amount, investment amount, loan amount and other fund flow data between the same user pair, and obtains the total transaction amount between users. Due to the large range of transaction amount values, directly using the original amount will lead to deviations in weight calculation. Therefore, the transaction amount is logarithmically transformed. The logarithmic transformation can compress the value range and highlight relative differences. The transaction frequency counts the number of transactions between users. The number of transactions for the same user pair within the statistical period is calculated by traversing the transaction records. The higher the transaction frequency, the closer the relationship between users. The relationship duration calculates the time span from the first establishment of the relationship between users to the most recent transaction. The longer the duration, the more stable the user relationship. Edge weight calculation multiplies the logarithm of the transaction amount by a weighting factor of 0.5, the transaction frequency by a weighting factor of 0.3, and the relationship duration by a weighting factor of 0.2. The three weighted results are summed to obtain the edge weight between user nodes. The weighting factor setting reflects the dominant role of the transaction amount and the supporting roles of frequency and duration. The risk propagation coefficient calculation process quantifies the intensity of risk propagation in the network based on the difference in risk levels between user nodes. Risk level differences reflect the similarity between users' risk tolerance and investment preferences. Users with similar risk levels are more likely to experience risk propagation and behavioral influence. The calculation process first obtains the risk level identifier of each user node. Risk levels are generally divided into five levels: conservative, stable, balanced, growth, and aggressive, and the corresponding values ​​are coded from 1 to 5. The absolute value of the risk level difference between user pairs is then calculated. A difference of 0 indicates that the risk levels of the two users are exactly the same. A larger difference indicates a greater difference in risk preferences. The risk diffusion coefficient is set as 1.2 for users with the same risk level, indicating increased influence among users with similar risk profiles; 1.0 for users with a one-level difference in risk level, indicating normal influence; and 0.8 for users with two or more levels of difference in risk level, indicating decreased influence among users with different risk preferences. This setting reflects the phenomenon in financial markets where users with similar risk profiles are more likely to influence each other.

[0075] The fusion process mathematically calculates the edge weights and the risk propagation coefficient to obtain the adjusted edge weights. The fusion process uses a multiplication operation, multiplying each edge weight by the corresponding user pair's risk propagation coefficient to obtain the adjusted edge weights after incorporating the risk factors. The multiplication operation ensures that the risk propagation coefficient adjusts the edge weights. When the risk propagation coefficient is greater than 1, the edge weight increases; when it is less than 1, the edge weight decreases; and when the risk propagation coefficient is equal to 1, the edge weight remains unchanged. The adjusted edge weights reflect both the strength of the transaction relationship between users and the impact of risk propagation, providing more accurate edge weight parameters for subsequent influence calculations.

[0076] The initialization weight distribution process sets the initial influence values ​​for all user nodes based on the preset damping coefficient. The damping coefficient is a core parameter of the PageRank algorithm, which represents the probability of continuing to propagate along the link during a random walk and is usually set to 0.85. The initialization process sets the influence values ​​of all user nodes in the graph to equal initial values. The initial value calculation formula is 1 minus the damping coefficient divided by the total number of nodes, that is, the initial influence of each node is 0.15 divided by the total number of nodes. This uniform initialization ensures the fairness of the algorithm convergence and avoids biased calculation results due to deviations in the initial value setting. The initial influence value serves as the starting point for iterative calculations, and after multiple rounds of iterations, it will gradually converge to a stable influence distribution.

[0077] The iterative update process calculates the new influence value of the current node based on the adjusted edge weights and the influence values ​​of neighboring nodes. Iterative update is the core calculation process of the PageRank algorithm. In each iteration, the influence value of each node is updated based on the influence of its neighboring nodes. The update formula is: the influence of the current node is equal to 0.15 divided by the total number of nodes plus 0.85 multiplied by the sum of the influence of all neighboring nodes pointing to the node and the corresponding edge weight, divided by the out-degree of the neighboring nodes. The calculation process traverses each user node, finds all neighboring nodes pointing to the node, obtains the current influence value of each neighboring node and the adjusted weight value of the connecting edge, calculates the neighboring node influence value multiplied by the edge weight value and divided by the neighboring node out-degree, adds the contribution value of all neighboring nodes, multiplies it by the damping coefficient 0.85, and finally adds 1 minus the damping coefficient divided by the total number of nodes to obtain the updated influence value of the current node.

[0078] The convergence judgment process determines whether the algorithm has converged by comparing the difference between the results of two adjacent iterations. Convergence judgment is the termination condition of the iterative algorithm. When the algorithm reaches a stable state, the iteration stops and the final result is output. The judgment process calculates the absolute value of the difference between the influence value of all nodes in the current iteration round and the influence value of the corresponding node in the previous round, and adds the absolute value of the difference of all nodes to obtain the overall difference value. When the overall difference value is less than the preset convergence threshold of 0.000001, the algorithm is judged to have converged. The influence distribution after convergence is the user influence distribution that integrates the risk propagation coefficient and asset weight. This distribution not only reflects the importance of the user in the trading network, but also takes into account the regulatory effect of risk propagation on influence. For example, a bank customer named Zhao has large-scale financial transactions with multiple high-net-worth customers. The transaction amount has a high logarithmic value, the transaction frequency is multiple times a month, and the relationship lasts for more than two years. The calculated edge weight value is large. Zhao's risk level is growth-type, which is similar to the risk level of his main trading counterparts. The risk transmission coefficient is 1.2. The adjusted edge weight value after fusion processing is further increased. At initialization, Zhao's influence is equal to that of other nodes. After multiple rounds of iterative updates, Zhao obtains a large amount of influence transfer from neighboring nodes. His final influence value is significantly higher than the average level. After the convergence judgment confirms that the algorithm is stable, Zhao ranks high in the user influence distribution, reflecting his core position and important influence in the financial network.

[0079] In a specific embodiment, the process of performing edge weight calculation based on the transaction amount, transaction frequency, and relationship duration of the user node in the financial knowledge graph may specifically include the following steps:

[0080] Perform logarithmic transformation on the transaction amounts between user nodes in the financial knowledge graph to obtain the logarithmic value of the transaction amount;

[0081] Statistically calculate and process the transaction frequency based on the historical transaction records between user nodes to obtain the transaction frequency value per unit time;

[0082] The relationship duration is calculated based on the first transaction time and the last transaction time between user nodes to obtain the relationship duration value;

[0083] The transaction amount is multiplied by a weight coefficient of 0.5, the transaction frequency is multiplied by a weight coefficient of 0.3, and the relationship duration is multiplied by a weight coefficient of 0.2, and the weighted sum is calculated to obtain a comprehensive weight score.

[0084] The connection strength between user nodes is quantified based on the comprehensive weight score to obtain the edge weight value between user nodes.

[0085] Specifically, the logarithmic transformation process applies a mathematical transformation to the transaction amounts between user nodes in the financial knowledge graph. Logarithmic transformation is a data preprocessing technique used to compress the numerical range and reduce the impact of outliers. It is particularly important in financial data analysis because transaction amounts often span multiple orders of magnitude. The transformation process first extracts all transaction records between user nodes from the financial knowledge graph, including financial flow data such as transfer amounts, investment amounts, loan amounts, and guarantee amounts. Each transaction amount is then incremented by 1 and its natural logarithm is taken. This increment is intended to prevent logarithmic errors caused by zero values. The natural logarithmic transformation converts exponential differences in the original amounts into linear differences. For example, the original transaction amounts ranging from 1,000 to 1,000,000 yuan are transformed into a range of 6.9 to 13.8 after logarithmic transformation, significantly reducing the numerical span and facilitating subsequent calculations. The transformed values ​​maintain the relative magnitude of the original data while reducing the impact of outliers on the overall analysis. The transformed results, as the logarithmic values ​​of the transaction amounts, are stored in the intermediate data structure used for edge weight calculation.

[0086] Statistical calculations analyze transaction frequency patterns based on historical transaction records between user nodes. Transaction frequency reflects the level of activity and stability in user relationships, with higher frequencies indicating closer business dealings. The statistical process traverses all historical transaction records between user pairs, arranges them chronologically, and counts the number of transactions within a specified time window. This calculation utilizes a sliding window approach, dividing the entire observation period into equal time units, such as months or quarters. The number of transactions between user pairs within each time unit is counted, and the average transaction frequency is then calculated as the transaction frequency value for that user pair. The statistical process distinguishes between different types of transactions, including active transactions such as transfers and remittances and passive transactions such as automatic deductions. The bidirectional nature of transactions is also considered to ensure the accuracy of frequency statistics. Frequency calculations also need to address uneven time intervals. Through normalization, frequency values ​​across different observation periods are converted to standardized frequency values ​​per unit of time, eliminating the impact of differences in observation period length on frequency comparisons.

[0087] The relationship duration calculation process quantifies the stability of relationships by analyzing the transaction time span between user nodes. Relationship duration is an important indicator of the depth of user relationships; longer durations indicate more stable and reliable relationships. The calculation process first identifies the time of the first transaction between the user pair. By traversing all transaction records, the transaction with the earliest timestamp is found, which serves as the start time of the relationship establishment. The last transaction between the user pair is then identified, and the transaction with the latest timestamp is used as the most recent active time of the relationship. The relationship duration is equal to the result of subtracting the first transaction time from the last transaction time, and is expressed in days, months, or years. The calculation process needs to handle intermittent transactions. When transactions between users are interrupted for a long period of time, it is necessary to determine whether the relationship is still valid. A maximum interval threshold, such as one year, is typically set. Intervals exceeding this threshold are considered to indicate a relationship interruption, and the relationship duration is recalculated. The duration value also needs to be normalized, converting the absolute time into a relative time fraction to facilitate comparison and calculation with other weighting factors.

[0088] The weighted summation process applies mathematical calculations to the logarithmic values ​​of transaction amount, transaction frequency, and relationship duration according to preset weight coefficients. The weight coefficients are set based on an analysis of the importance of financial services. Transaction amount reflects the scale of the financial relationship between users, with the highest weight coefficient of 0.5 indicating its dominant role. Transaction frequency reflects the level of relationship activity, with a medium weight coefficient of 0.3 indicating its important supporting role. Relationship duration reflects the stability of the relationship, with a low weight coefficient of 0.2 indicating its fundamental support role. The calculation process multiplies the logarithmic value of each user pair's transaction amount by 0.5 to obtain the amount weight contribution, the transaction frequency value by 0.3 to obtain the frequency weight contribution, and the relationship duration value by 0.2 to obtain the time weight contribution. These three weight contributions are summed to obtain the user pair's overall weight score. Weighted summation ensures that values ​​of different dimensions are appropriately combined according to business importance. The sum of the weight coefficients, 1.0, ensures that the overall score remains within a reasonable range.

[0089] Quantization converts the connection strength between user nodes into standardized edge weight values ​​based on the comprehensive weight score. Connection strength quantification is the foundation of graph structure analysis, and edge weight values ​​directly influence subsequent network analysis and influence calculation results. The quantization process first analyzes the distribution of the comprehensive weight scores of all user pairs, calculating statistical characteristics such as the maximum, minimum, mean, and standard deviation of the scores. Then, a maximum-minimum normalization method is used to map the comprehensive weight scores to a standard interval of 0 to 1. The normalization formula is the difference between the current score and the minimum score, divided by the difference between the maximum score and the minimum score. The normalized value serves as the weight value of the user-pair connection edge. Quantization also includes outlier detection and processing. When the comprehensive weight scores of certain user pairs deviate significantly from the normal distribution, truncation or smoothing is used to prevent extreme values ​​from affecting the overall weight distribution. The final edge weight value reflects the true strength of the relationship between users and meets the graph algorithm's standardization requirements for weight values. For example, a long-term investment consulting relationship exists between bank customer Chen and investment consultant Sun. Chen regularly transfers funds to Sun's investment account for entrusted financial management. Historical transaction records show that the cumulative transaction amount between the two is large. After logarithmic transformation, a higher logarithmic value of the transaction amount is obtained. The transaction frequency between the two is two to three times a month, which reflects a stable business transaction pattern. The transaction frequency value obtained by statistical calculation is at a medium level. The duration of their relationship spans three years from the first transaction to the most recent transaction. The relationship duration calculation results in a longer time value. The weighted summation process multiplies the logarithmic value of the amount by 0.5, the frequency value by 0.3, and the time value by 0.2, and then adds them together to obtain a comprehensive weight score. This score ranks high among all user pairs. The quantitative process converts the comprehensive weight score into an edge weight value close to 1.0 through standardized transformation, indicating that there is a strong connection between Chen and Sun. This edge weight value accurately reflects the degree of close relationship between the two in the financial network.

[0090] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0091] The user influence distribution is integrated with the user node feature vector in the financial knowledge graph to obtain the user comprehensive feature vector including the influence weight.

[0092] Based on the user's comprehensive feature vector, the spectral clustering similarity matrix is ​​constructed to obtain the similarity matrix between users;

[0093] The number of clusters is optimized based on the similarity matrix between users and the graph connectivity constraint, and the optimal clustering parameter is obtained as 8;

[0094] The user comprehensive feature vector is input into a spectral clustering algorithm for grouping processing, obtaining eight types of user groups of conservative investors, stable investors, growth investors, active investors, high net worth customers, small and micro business owners, young white-collar workers, and retired groups.

[0095] The eight types of user groups are labeled with risk tolerance, product preference, and investment potential, obtaining user portrait classification results containing risk tolerance, product preference, and investment potential.

[0096] Specifically, the fusion processing integrates the user influence distribution with the user node feature vector in the financial knowledge graph. The user influence distribution is a one-dimensional numerical vector calculated by the FinRank algorithm, and each element corresponds to the influence of a user in the financial network. The user node feature vector is a high-dimensional vector containing user basic attributes, financial status, behavior characteristics, and other multi-dimensional information. The fusion process uses vector splicing to add the influence value of each user as a new feature dimension to the end of the original node feature vector, forming an extended feature vector with increased dimensions. The fusion operation also includes feature weight adjustment, which sets a weight factor according to the importance of the influence value, and combines the influence feature with the original feature. The weight of the influence feature is usually set to a high value to highlight its key role in the user portrait. The fused user comprehensive feature vector not only retains the user's basic attribute information, but also integrates the user's status and influence in the financial network, providing a more comprehensive user feature description for subsequent clustering analysis.

[0097] The similarity matrix construction process calculates the similarity relationship between users based on the user comprehensive feature vector. Spectral clustering is a clustering algorithm based on graph theory, which requires the construction of a similarity matrix between users as input data. The construction process iterates through all user pairs, calculating the Euclidean distance between each pair of user comprehensive feature vectors. The Euclidean distance is obtained by calculating the square sum of the difference between the corresponding dimensions of the two vectors and then taking the square root. The smaller the distance value, the more similar the users. Then the distance value is converted to a similarity value using a Gaussian kernel function. The similarity is equal to the negative distance square of the natural constant divided by the exponential of the bandwidth parameter. The bandwidth parameter controls the width of the similarity function, and the smaller the parameter value, the faster the similarity decays. The similarity matrix is a symmetric matrix with a size of the square of the number of users. The matrix element value represents the similarity of the corresponding user pair. The diagonal elements are 1, indicating that the user is completely similar to itself, and the non-diagonal element values are between 0 and 1, indicating the similarity between users.

[0098] The optimization selection process determines the optimal number of clusters based on the inter-user similarity matrix and graph connectivity constraints. The number of clusters is a key parameter in the spectral clustering algorithm. Too few clusters results in overly coarse classification, while too many clusters results in overly fine classification. Therefore, a data-driven approach is required to select the optimal number. The optimization process utilizes the eigenvalue decomposition method. The similarity matrix is ​​Laplace transformed to obtain a Laplace matrix. The eigenvalues ​​and eigenvectors of the Laplace matrix are then calculated. The distribution pattern of the eigenvalues ​​reflects the clustering structure of the data. The graph connectivity constraint requires that users within each cluster in the clustering results have strong connectivity in the original financial knowledge graph, avoiding assigning distant users to the same cluster. Constraint verification is performed by calculating the average shortest path length between users within a cluster. Clusters with excessively long paths are considered to violate the connectivity constraint. Combining eigenvalue analysis with connectivity constraints, a grid search method was used to test clustering quality under different numbers of clusters. Ultimately, a number of eight clusters was determined to achieve the optimal classification.

[0099] The grouping process inputs the user's comprehensive feature vector into the spectral clustering algorithm for automatic classification. The spectral clustering algorithm first performs eigendecomposition on the Laplace matrix, extracts the eigenvectors corresponding to the first 8 smallest eigenvalues, and combines these eigenvectors by column to form a new feature space. Each user's comprehensive feature vector is then projected into the new feature space to obtain the user representation vector after dimensionality reduction. The traditional K-means clustering algorithm is then applied to the new feature space to divide the users into 8 groups. The clustering process iteratively optimizes the category center and user allocation, minimizing the sum of squared distances between users within the category until the algorithm converges to obtain a stable clustering result. The eight user groups are named conservative investors, stable investors, growth investors, active investors, high net worth customers, small and micro business owners, young white-collar workers, and retired groups. Each group contains a collection of users with similar characteristics and behavior patterns.

[0100] The tagging process semantically describes the business characteristics of eight user groups. The tagging process analyzes the common characteristics and behavioral patterns of users within each group, extracting typical group characteristics as tag content. Risk tolerance tags are determined based on the risk preference distribution of users within the group. Conservative investors are labeled as having a low risk tolerance, aggressive investors are labeled as having a high risk tolerance, and other groups are labeled as having a corresponding medium risk tolerance. Product preference tags are determined based on the product holdings of users within the group. High-net-worth individuals prefer private banking products and high-end wealth management products, young professionals prefer internet finance products and fund investments, and retirees prefer conservative savings and insurance products. Investment potential tags are determined based on a comprehensive assessment of factors such as the asset size, income level, and age structure of users within the group. Young professionals and small and micro business owners are labeled as having high investment potential, while retirees are labeled as having low investment potential. The tagging results form a user profile classification, each containing detailed descriptions of risk tolerance, product preferences, and investment potential, providing decision support for financial institutions' targeted marketing and personalized services. For example, after the processing of the above steps, the influence distribution value of bank customer Zhou is relatively high, indicating that he has an important position in the financial network. The node feature vector shows that he is a middle-aged high-income group and holds a variety of investment products. The fusion processing combines the influence value with the feature vector to form a comprehensive feature vector. The similarity matrix construction finds that Zhou has a high similarity with other high-net-worth users. After optimizing the selection and determining the number of clusters of 8, the spectral clustering algorithm assigns Zhou to the high-net-worth customer group. Users in this group have the characteristics of high asset scale and diversified investment portfolios. The labeling processing labels the group with portrait labels of high risk tolerance, preference for high-end financial products, and high investment potential. As a member of the high-net-worth customer group, Zhou obtains an accurate user portrait classification. The bank formulates exclusive wealth management service plans and product recommendation strategies based on the characteristics of his group.

[0101] The above describes the financial user portrait analysis method based on the knowledge graph in the embodiment of the present application. The following describes the financial user portrait analysis system based on the knowledge graph in the embodiment of the present application. Figure 2 In one embodiment of the present application, a financial user profile analysis system based on a knowledge graph includes: an identification module for performing financial entity identification processing on bank transaction data, credit data, and user association data using an FNER algorithm to obtain user entities, product entities, and risk entities;

[0102] An embedding module, configured to perform TransFinE relation embedding processing on the multi-dimensional feature data according to the user entity, product entity, and risk entity, to obtain a relation vector representation including time weight decay and risk propagation constraints;

[0103] A construction module, configured to construct a graph structure of the relationship vector representation to obtain a financial knowledge graph comprising user nodes, product nodes, and risk propagation paths;

[0104] A calculation module is used to calculate the influence weight of user nodes in the financial knowledge graph using the FinRank algorithm to obtain a user influence distribution that integrates risk propagation coefficients and asset weights;

[0105] The grouping module is used to perform semantic clustering and grouping processing on user features according to the user influence distribution, and obtain user portrait classification results including risk tolerance, product preference and investment potential.

[0106] above Figure 2 The financial user profile analysis system based on knowledge graph in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The financial user profile analysis device based on knowledge graph in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0107] Reference Figure 3 In an embodiment of the present invention, a financial user portrait analysis device based on knowledge graph is also provided. The financial user portrait analysis device based on knowledge graph can be a server, and its internal structure can be as follows: Figure 3 As shown. The financial user profile analysis device based on the knowledge graph includes a processor, a memory, a display screen, an input device, a network interface and a database connected via a system bus. Among them, the computer-designed processor is used to provide computing and control capabilities. The memory of the financial user profile analysis device based on the knowledge graph includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the financial user profile analysis device based on the knowledge graph is used to store the corresponding data in this embodiment. The network interface of the financial user profile analysis device based on the knowledge graph is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0108] Those skilled in the art will understand that Figure 3 The structure shown in is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the knowledge graph-based financial user profile analysis device to which the solution of the present invention is applied.

[0109] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the financial user portrait analysis method based on the knowledge graph.

[0110] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0111] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a financial user profile analysis device based on a knowledge graph (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and other media that can store program code.

[0112] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A financial user portrait analysis method based on knowledge graph, characterized in that: The method comprises: Financial entity identification is performed on bank transaction data, credit data, and user-related data using the FNER algorithm to obtain user entities, product entities, and risk entities. Performing TransFinE relation embedding processing on the multi-dimensional feature data according to the user entity, product entity, and risk entity to obtain a relation vector representation including time weight decay and risk propagation constraints; The relationship vector representation is processed into a graph structure to obtain a financial knowledge graph including user nodes, product nodes and risk propagation paths; The influence weight of user nodes in the financial knowledge graph is calculated using the FinRank algorithm to obtain a user influence distribution that integrates risk propagation coefficients and asset weights; According to the user influence distribution, user characteristics are semantically clustered and grouped to obtain user portrait classification results including risk tolerance, product preference and investment potential.

2. The financial user profile analysis method based on knowledge graph according to claim 1 is characterized in that: The method of performing financial entity identification processing on bank transaction data, credit data and user association data through the FNER algorithm to obtain user entities, product entities and risk entities includes: Construct a financial domain-specific dictionary containing bank product names, financial terms, and risk level identifiers to obtain a financial entity vocabulary database; Inputting the bank transaction data, credit data, and user association data into a bidirectional long short-term memory network for sequence encoding processing to obtain a text sequence feature vector; Performing conditional random field labeling processing on the text sequence feature vector based on the financial entity vocabulary library to obtain an entity boundary label sequence; Perform entity extraction processing on the original text data according to the entity boundary marker sequence to obtain a candidate entity set of user name, ID number, bank card number, product code and risk level identifier; The candidate entity set is classified and merged according to entity type to obtain user entities, product entities and risk entities.

3. The financial user profile analysis method based on knowledge graph according to claim 1 is characterized in that: The multi-dimensional feature data is subjected to TransFinE relation embedding processing according to the user entity, product entity, and risk entity to obtain a relation vector representation containing time weight decay and risk propagation constraints, including: Performing relationship triple construction processing based on the user entity, product entity, and risk entity to obtain original triple data of user-product holding relationship, user-user social relationship, and product-risk level relationship; Performing time weight calculation on the original triple data based on historical transaction time to obtain a time decay weight coefficient with a recent relationship weight of 1.0 and a monthly decrease of 0.05; The time decay weight coefficient and the risk level difference are subjected to risk propagation constraint calculation processing to obtain a risk consistency loss function and a time loss function; Based on the objective function that the sum of the head entity vector and the relationship vector equals the tail entity vector, the entity vector and the relationship vector are iteratively optimized to obtain low-dimensional dense vector representations of user entities, product entities, and risk entities. The low-dimensional dense vector representation is subjected to vector combination processing according to the head entity, the relationship, and the tail entity to obtain a relationship vector representation containing time weight decay and risk propagation constraints.

4. The financial user profile analysis method based on knowledge graph according to claim 1 is characterized in that: The graph structure construction process of the relationship vector representation to obtain a financial knowledge graph including user nodes, product nodes and risk propagation paths includes: Performing node attribute matrix construction processing based on the relationship vector representation to obtain a node attribute matrix including user characteristics, product characteristics, and risk characteristics; Mapping the connection relationship according to the head entity and tail entity relationship in the relationship vector representation to obtain an adjacency matrix of user-product connection, user-user connection and product-risk connection; Performing graph structure calculation processing on the node attribute matrix and the adjacency matrix to obtain the degree centrality and betweenness centrality values ​​of each node; Performing graph optimization based on the degree centrality and betweenness centrality values ​​to obtain an optimized graph structure after deleting low-weight nodes and edges; The optimized graph structure is hierarchically organized according to the user layer, product layer and relationship layer to obtain a financial knowledge graph containing user nodes, product nodes and risk propagation paths.

5. The financial user profile analysis method based on knowledge graph according to claim 1 is characterized in that: The FinRank algorithm is used to calculate the influence weight of user nodes in the financial knowledge graph to obtain the user influence distribution that integrates the risk propagation coefficient and asset weight, including: Perform edge weight calculation based on the transaction amount, transaction frequency, and relationship duration of the user nodes in the financial knowledge graph to obtain edge weight values ​​between the user nodes; The risk propagation coefficient is calculated based on the risk level differences between user nodes to obtain the risk propagation coefficient between users with different risk levels; Fusing the edge weight value with the risk propagation coefficient to obtain an adjusted edge weight value; Initialize the weight distribution of user nodes based on the preset damping coefficient to obtain the initial influence value of each user node; Iteratively updating the current node influence according to the adjusted edge weight value and the neighbor node influence value to obtain an updated node influence value; The updated node influence values ​​are subjected to convergence judgment processing to obtain a user influence distribution that integrates risk propagation coefficients and asset weights.

6. The financial user profile analysis method based on knowledge graph according to claim 5 is characterized in that: The edge weight calculation process is performed based on the transaction amount, transaction frequency, and relationship duration of the user nodes in the financial knowledge graph to obtain the edge weight values ​​between the user nodes, including: Performing logarithmic transformation on the transaction amounts between user nodes in the financial knowledge graph to obtain logarithmic values ​​of the transaction amounts; Performing statistical calculations on the transaction frequency based on the historical transaction records between the user nodes to obtain the transaction frequency value per unit time; Calculating the relationship duration based on the first transaction time and the last transaction time between the user nodes to obtain a relationship duration value; The transaction amount logarithm is multiplied by a weight coefficient of 0.5, the transaction frequency logarithm is multiplied by a weight coefficient of 0.3, and the relationship duration logarithm is multiplied by a weight coefficient of 0.2, and the weighted summation is performed to obtain a comprehensive weight score; The connection strength between user nodes is quantified based on the comprehensive weight score to obtain the edge weight value between user nodes.

7. The financial user profile analysis method based on knowledge graph according to claim 1 is characterized in that: The semantic clustering and grouping of user characteristics according to the user influence distribution is performed to obtain user profile classification results including risk tolerance, product preference and investment potential, including: Fusing the user influence distribution with the user node feature vector in the financial knowledge graph to obtain a user comprehensive feature vector including influence weights; Performing spectral clustering similarity matrix construction processing based on the user comprehensive feature vector to obtain an inter-user similarity matrix; Optimizing the number of clusters according to the inter-user similarity matrix and the graph connectivity constraint to obtain a clustering parameter with an optimal number of clusters of 8; The user comprehensive feature vector is input into the spectral clustering algorithm for grouping processing to obtain eight user groups: conservative investors, steady investors, growth investors, aggressive investors, high net worth customers, small and micro business owners, young white-collar workers, and retired groups; The eight user groups are labeled with risk tolerance, product preference and investment potential labels to obtain user portrait classification results including risk tolerance, product preference and investment potential.

8. A financial user portrait analysis system based on knowledge graph, characterized by: A method for analyzing a financial user profile based on a knowledge graph according to any one of claims 1 to 7, wherein the system for analyzing a financial user profile based on a knowledge graph comprises: The identification module is used to perform financial entity identification processing on bank transaction data, credit data and user association data using the FNER algorithm to obtain user entities, product entities and risk entities; An embedding module, configured to perform TransFinE relation embedding processing on the multi-dimensional feature data according to the user entity, product entity, and risk entity, to obtain a relation vector representation including time weight decay and risk propagation constraints; A construction module, configured to construct a graph structure of the relationship vector representation to obtain a financial knowledge graph comprising user nodes, product nodes, and risk propagation paths; A calculation module is used to calculate the influence weight of user nodes in the financial knowledge graph using the FinRank algorithm to obtain a user influence distribution that integrates risk propagation coefficients and asset weights; The grouping module is used to perform semantic clustering and grouping processing on user features according to the user influence distribution, and obtain user portrait classification results including risk tolerance, product preference and investment potential.

9. A financial user portrait analysis device based on knowledge graph, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and the processor implements the financial user portrait analysis method based on knowledge graph described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor executes the financial user profile analysis method based on the knowledge graph as described in any one of claims 1 to 7.

Images