Heterogeneous graph-based bank potential corporate user recommendation method
By constructing a heterogeneous graph-based corporate customer relationship network, dividing it into multiple subgraphs and performing similarity fusion to generate a quantitative graph, the problem of poor marketing effectiveness in existing technologies is solved, enabling more accurate recommendations of potential corporate customers and improving marketing results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING ZHONGYI ANTU TECH CO
- Filing Date
- 2022-08-10
- Publication Date
- 2026-06-19
AI Technical Summary
Among existing bank marketing methods, rule-based methods rely on expert experience and have limited effectiveness, while traditional machine learning methods are insufficient in representing relational features in bank financial data, resulting in poor marketing results.
A heterogeneous graph-based corporate user relationship network is constructed, which is divided into multiple subgraphs based on application scenarios. A similarity matrix is calculated and fused to generate a quantized graph. This graph is then used to recommend potential corporate users based on the input bank user data.
It improves the accuracy and effectiveness of bank marketing by accurately recommending potential corporate clients through heterogeneous graph technology, thereby enhancing marketing results.
Smart Images
Figure CN115455308B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of heterogeneous graph analysis and corporate marketing, and in particular to a method for recommending potential corporate customers of banks based on heterogeneous graphs. Background Technology
[0002] Currently, banks use big data for corporate marketing, primarily employing rule-based methods and traditional machine learning approaches. Rule-based methods involve experts designing rules tailored to specific targets or scenarios for marketing purposes. While rule-based rule development relies on expert experience, domain knowledge requires accumulated experience, and simple conditional rules often result in low marketing effectiveness.
[0003] Traditional machine learning methods involve data mining from large amounts of banking data, constructing datasets through feature engineering, and then training machine learning models. On the one hand, traditional methods require a large number of labeled samples for model validation. On the other hand, feature engineering is also an experience-based process, requiring the definition and processing of feature combinations specific to a particular problem. However, banking financial data contains a wealth of relational information, and traditional feature engineering cannot adequately represent user relationship characteristics, resulting in low marketing effectiveness of machine learning in certain specific marketing scenarios. Summary of the Invention
[0004] This application provides a method for recommending potential corporate clients of banks based on heterogeneous graphs, aiming to accurately recommend potential corporate clients and improve marketing effectiveness.
[0005] This application provides a method for recommending potential corporate customers of banks based on heterogeneous graphs, including:
[0006] A corporate user relationship network is constructed based on corporate user data, wherein the corporate user relationship network is a heterogeneous graph, and the heterogeneous graph includes multiple corporate user relationships;
[0007] Based on the application scenario, the various public user relationships are divided into multiple public user relationship subgraphs, and the similarity matrix of the multiple public user relationship subgraphs is calculated.
[0008] By fusing multiple similarity matrices, a quantitative graph of public user relationships is obtained;
[0009] Based on the quantitative graph of corporate user relationships and the input bank users, potential corporate users of the input bank users are recommended.
[0010] In one embodiment, fusing multiple similarity matrices to obtain a quantitative graph of public user relationships includes:
[0011] Normalize each of the similarity matrices to obtain each state matrix;
[0012] The local similarity matrices are obtained by measuring the local relationships of each similarity matrix according to a preset algorithm;
[0013] The local similarity matrices are normalized to obtain the kernel matrices.
[0014] For each different state matrix, based on its corresponding kernel matrix, and through a predetermined number of iterations using a cross-diffusion method, the cross-state matrix is obtained;
[0015] The fusion similarity matrix is calculated based on the state matrix after each different crossover, and the public user relationship quantization graph is obtained based on the fusion similarity matrix.
[0016] The step of obtaining the quantitative graph of public user relationships based on the fused similarity matrix includes:
[0017] Determine the diagonal matrix of the fusion similarity matrix, and determine the adjacency matrix based on the fusion similarity matrix and its diagonal matrix;
[0018] A similarity graph is generated based on the adjacency matrix, and the similarity graph is determined as the quantitative graph of the public user relationship.
[0019] The method of recommending potential corporate clients of the input bank users based on the corporate user relationship quantification graph and the input bank users includes:
[0020] The relevance of all corporate users to the input bank users is calculated based on the corporate user relationship quantification graph, and the relevance is sorted by a preset algorithm.
[0021] Based on the relevance after sorting, all target corporate users that meet the preset relevance threshold are recommended, and all target corporate users are identified as potential corporate users of the input bank user.
[0022] The application scenarios include fund transfer transactions; the process of dividing the various corporate user relationships according to the application scenarios yields multiple corporate user relationship sub-graphs, including:
[0023] Based on the aforementioned transfer transaction scenario, determine each corporate user and their transfer transaction data;
[0024] Based on each corporate user and its transfer transaction data, the various corporate user relationships are divided into transfer transaction relationships, resulting in a transfer transaction relationship subgraph;
[0025] In the transfer transaction relationship subgraph, the nodes are each corporate user, the nodes are transfer transaction data, and the edge attributes are transaction time information and transaction amount information.
[0026] The application scenarios include industry scenarios; the process of dividing the various corporate user relationships according to the application scenarios yields multiple corporate user relationship sub-graphs, including:
[0027] Based on the industry scenarios described, determine the information of each corporate user and their industry.
[0028] Based on the information of each corporate user and its industry, the various corporate user relationships are divided into industry relationships, resulting in an industry relationship sub-graph.
[0029] In the industry relationship subgraph, the nodes represent each corporate user and industry information, and the nodes' edges represent the industries to which each corporate user belongs and their upstream and downstream partners.
[0030] The application scenarios include regional scenarios; the process of dividing the various corporate user relationships according to the application scenarios yields multiple corporate user relationship sub-graphs, including:
[0031] Based on the aforementioned regional scenario, determine the information of each corporate user and their respective regions;
[0032] Based on the information of each corporate user and its region, the various corporate user relationships are divided into regional relationships, resulting in a regional relationship subgraph.
[0033] In the regional relationship subgraph, the nodes represent the regional information of each corporate user, and the nodes' edges represent the regions where each corporate user is located and their proximity relationships.
[0034] The heterogeneous graph-based method for recommending potential corporate bank users provided in this application divides various corporate user relationships in the heterogeneous graph of the corporate user relationship network into corporate user relationship subgraphs according to the application scenario. Then, it combines the corporate user relationship subgraphs to obtain a quantitative corporate user relationship graph. Finally, it accurately recommends potential corporate users of the input bank user through the quantitative corporate user relationship graph, thereby improving marketing effectiveness. Attached Figure Description
[0035] To more clearly illustrate the technical solutions of this application, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0036] Figure 1 This is a schematic diagram of the heterogeneous graph-based recommendation method for potential corporate bank users provided in this application;
[0037] Figure 2 This is a schematic diagram of a corporate user relationship network for an application scenario provided in this application;
[0038] Figure 3 This application provides a sub-diagram of transfer transaction relationships;
[0039] Figure 4 The industry relationship sub-diagram provided in this application;
[0040] Figure 5 The regional relationship sub-graph provided in this application. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0042] Furthermore, combined Figures 1 to 5 This application describes the method for recommending potential corporate customers of banks based on heterogeneous graphs. Figure 1 This is a schematic diagram of the heterogeneous graph-based recommendation method for potential corporate bank users provided in this application; Figure 2 This is a schematic diagram of a corporate user relationship network for an application scenario provided in this application; Figure 3 This application provides a sub-diagram of transfer transaction relationships; Figure 4 The industry relationship sub-diagram provided in this application; Figure 5 The regional relationship sub-graph provided in this application.
[0043] This application provides an embodiment of a method for recommending potential corporate bank users based on heterogeneous graphs. It should be noted that although the logical order is shown in the flowchart, under certain data conditions, the steps shown or described may be performed in a different order than that shown here.
[0044] This application uses an electronic device as an example to illustrate the execution of the embodiments. This application uses a recommendation system as one of the manifestations of the electronic device, and does not impose any limitations.
[0045] like Figure 1 , Figure 1 This is a schematic diagram of the heterogeneous graph-based method for recommending potential corporate bank customers provided in this application. The heterogeneous graph-based method for recommending potential corporate bank customers provided in this application includes:
[0046] Step S10: Construct a corporate user relationship network based on corporate user data, wherein the corporate user relationship network is a heterogeneous graph, and the heterogeneous graph includes multiple corporate user relationships.
[0047] It should be noted that the heterogeneous graph-based method for recommending potential corporate bank users provided in this application can be applied to recommending potential corporate bank users. Therefore, the users in this application refer to corporate bank users.
[0048] Specifically, we first define a heterogeneous graph G, which can be represented as G = {V, E, X}. V ,X E},in, Represents a set of nodes. Denotes the set of edges. Represents the set of node types and This represents the set of edge types. Further, the set of node types... and edge type set satisfy X V Represents the set of node attributes, X E This represents the set of edge attributes. Secondly, using corporate user data from banks, relationships between corporate users are constructed. These relationships can be categorized into simple relationships (direct association between corporate users) and complex relationships (associations between corporate users through other objects), resulting in a corporate user relationship network.
[0049] It should be noted that the corporate user relationship network is essentially a heterogeneous graph containing multiple corporate user relationships.
[0050] In one application scenario, the node set includes corporate users, industry information, and regional information; the node attribute set is latitude and longitude; the edge set includes transfer transactions, the industry to which the corporate user belongs, upstream and downstream industries, the corresponding user's region, and neighboring region relationships; the edge attribute set includes transaction time information and transaction amount information. The corporate user relationship network constructed based on the node set, node attribute set, edge set, and edge attribute set is as follows: Figure 2 As shown, Figure 2 This is a schematic diagram of a corporate user relationship network for an application scenario provided in this application.
[0051] Step S20: Divide the various public user relationships according to the application scenario to obtain multiple public user relationship subgraphs, and calculate the similarity matrix of the multiple public user relationship subgraphs.
[0052] This application embodiment divides various corporate user relationships in the corporate user relationship network according to application scenarios. The corporate user relationship network is essentially a heterogeneous graph containing various corporate user relationships. Therefore, it can be understood as dividing the various corporate user relationships in the heterogeneous graph into multiple corporate user relationship subgraphs according to application scenarios. Further, a similarity matrix of multiple corporate user relationship subgraphs is calculated.
[0053] It should be noted that the application scenarios in this application embodiment can be diverse. Commonly used scenarios include, but are not limited to, transfer transaction scenarios, industry scenarios, and regional scenarios. Therefore, the integration and application of corresponding corporate user relationships include transfer transaction relationships, industry relationships, and regional relationships. Further, it can be understood that the corporate user relationship subgraph includes transfer transaction relationship subgraphs, industry relationship subgraphs, and regional relationship subgraphs. It should also be noted that, in addition to the aforementioned transfer transaction relationship subgraphs, industry relationship subgraphs, and regional relationship subgraphs, there are also subgraphs that, such as transaction relationships and industry relationships, constitute upstream and downstream relationship subgraphs in the industrial chain.
[0054] Furthermore, for different relational subgraphs, different quantization strategies are used to calculate the relational subgraph similarity matrix and the relational weights between public user nodes, and the application scenarios of transfer transactions, industry scenarios, and regional scenarios are elaborated in detail:
[0055] For an application scenario focused on fund transfer transactions, the specific steps are as follows: First, identify each corporate user and their fund transfer transaction data based on the scenario. Then, based on these data, categorize the various corporate user relationships into fund transfer transaction relationships, resulting in a subgraph of these relationships. In this subgraph, nodes represent individual corporate users, edges represent fund transfer transaction data, and edge attributes include transaction time and transaction amount information. (Refer to...) Figure 3 , Figure 3 The transfer transaction relationship sub-diagram provided in this application, Figure 3 Part (a) in the diagram represents the transfer transaction relationship subgraph. It should be noted that the transfer transaction relationship subgraph in this embodiment is a multi-path attribute homogeneous graph. Furthermore, the similarity calculation of the transfer transaction relationship subgraph is performed using relationship quantification. The specific relationship quantification process for similarity calculation is as follows:
[0056] First, an aggregation function F is defined: V0×V0×T→V0×V0, where V0 represents the set of public user nodes and T represents a given transaction time range. This function is used to aggregate different transaction edges between two public user nodes, thereby transforming the original multi-way homogeneous graph into a simple homogeneous graph. The aggregation function learning method in this application includes, but is not limited to, rule-based methods, statistical methods (scorecards, manually defined indicators), and machine learning. Furthermore, this application embodiment can use multiple aggregation functions to generate multiple transaction relationship aggregation subgraphs, such as aggregation based on transaction amount, transaction frequency, or a combination of both.
[0057] In one embodiment, reference is made to Figure 3 , Figure 3 Part (b) in the text represents the transfer transactions between corporate users. Figure 3 The document illustrates the transfer transaction relationship between corporate user A and corporate user B.
[0058] Furthermore, for the homogeneous aggregate subgraph G0=(V0,E0), where E0={e ij |i∈V0,j∈V0} represents the aggregate edge weight. Choosing an appropriate similarity calculation method, we calculate the relationship weight matrix W between public users. For example, we can directly use the aggregate edge weight as the relationship weight:
[0059] W ij =e ij
[0060] The homogeneous graph similarity calculation method in this application embodiment also includes the Adar exponent method, the Katz exponent method, the SimRank algorithm, and graph embedding algorithms. Furthermore, if there are multiple aggregated subgraphs, the same calculation process can be used to obtain multiple relation weight matrices W. (n) .
[0061] For industry-specific application scenarios: determine each corporate user and industry information based on the industry scenario; divide the various corporate user relationships into industry relationships based on each corporate user and industry information to obtain an industry relationship subgraph; the nodes in the industry relationship subgraph are each corporate user and industry information, and the nodes' edges are the industries to which each corporate user belongs and the upstream and downstream of those industries.
[0062] Reference Figure 4 , Figure 4 The industry relationship sub-graph provided in this application Figure 4 (a) in the diagram represents the industry relationship subgraph. It should be noted that the industry relationship subgraph is an attribute-free heterogeneous graph. For attribute-free heterogeneous graphs, this embodiment of the application can use a graph structure similarity measurement method to process the attribute-free heterogeneous graph.
[0063] Furthermore, a meta-path approach is employed to identify relevant corporate users through industry relationships. This application provides, but is not limited to, the following meta-paths:
[0064]
[0065] In one embodiment, reference is made to Figure 4 , Figure 4 Part (b) describes several industry-based relationships between corporate users. Figure 4 The document illustrates several industry-based relationships between corporate users A, B, and C.
[0066] Furthermore, the similarity calculation of the industry relationship subgraph is performed by relation quantification. The specific relation quantification process for similarity calculation is as follows:
[0067] It should be noted that the quantization methods include, but are not limited to, the PathSim method, the PTE method, the ESim method, and graph embedding methods. This application uses the PathSim method as an example. Specifically, for public user i and public user j and a set of symmetric metapaths... The similarity is calculated as follows:
[0068]
[0069] in, Represents the metapath connecting node x and node y. The number of metapath instances.
[0070] For application scenarios with a regional focus: Based on the regional scenario, determine each corporate user and its regional information; based on each corporate user and its regional information, divide the various corporate user relationships into regional relationships, resulting in a regional relationship subgraph; where the nodes in the regional relationship subgraph represent the regional information of each corporate user, and the node edges represent the regions where each corporate user resides and their proximity relationships, as shown in the reference... Figure 5 , Figure 5 The regional relationship sub-graph provided in this application.
[0071] Figure 5 Part (a) in the diagram represents the regional relationship subgraph. It should be noted that the regional relationship subgraph is an attribute heterogeneous graph, where nodes contain latitude and longitude coordinates but not edge attributes. This application embodiment performs two relationship measurements: similarity calculation of node attributes and similarity calculation of the graph structure.
[0072] For calculating the similarity of node attributes, the method used in this application embodiment is a distance-based metric, specifically: assuming the geographical location information of public user i and public user j are represented as (a i ,b i ) and (a j ,b j ), calculate the actual distance between the two according to the Haversine formula:
[0073]
[0074] Where r represents the Earth's radius. Attribute similarity is calculated based on the distance metric and the RBF kernel function:
[0075]
[0076] Where σ is an adjustable parameter. For other types of attributes, other vector similarity calculation methods such as p-norm distance, cosine similarity, and Jaccard similarity can be selected according to the actual situation.
[0077] For similarity calculation of graph structures, quantification methods include, but are not limited to, the PathSim method, the PTE method, the ESim method, and graph embedding methods. This application embodiment uses the PathSim method as an example. Specifically, firstly, a meta-path method is used to determine relevant public users through regional relationships. This application embodiment provides, but is not limited to, the following meta-paths:
[0078]
[0079] In one embodiment, reference is made to Figure 5 , Figure 5 Part (b) describes several region-based relationships between public users. Figure 5 The document illustrates several region-based relationships between corporate user A and corporate user B.
[0080] Step S30: Merge multiple similarity matrices to obtain a quantitative graph of public user relationships.
[0081] Further, step S20 yields the similarity matrices of the transfer transaction relationship subgraph, the industry relationship subgraph, and the region relationship subgraph. Further, a similarity fusion method based on cross-diffusion is used to fuse these matrices to obtain a quantitative graph of corporate user relationships, as described in steps S301 to S305.
[0082] Furthermore, steps S301 to S305 are described as follows:
[0083] Step S301 normalizes each similarity matrix to obtain each state matrix;
[0084] Step S302: Measure the local relationships of each similarity matrix according to a preset algorithm to obtain each local similarity matrix;
[0085] Step S303: Normalize each of the local similarity matrices to obtain each kernel matrix;
[0086] Step S304: For each different state matrix, according to its corresponding kernel matrix, it is iterated a preset number of times through cross-diffusion to obtain the cross-state matrix;
[0087] Step S305: Calculate the fusion similarity matrix based on the state matrices after each different crossover, and obtain the public user relationship quantization graph based on the fusion similarity matrix.
[0088] First, calculate the state matrix P. (n) Specifically, given a similarity matrix W(n) For the similarity matrix W (n) After row normalization, the state matrix P is obtained. (n) :
[0089]
[0090] Furthermore, the kernel matrix Q is calculated. (n) Specifically, the similarity matrix W is measured according to a preset algorithm (K-nearest neighbor algorithm). (n) Calculate the local similarity matrix based on the local relationships.
[0091]
[0092] Here, KNN(i) represents the set of the K nodes with the highest similarity to node i. Similarly, for the local similarity matrix... After row normalization, the kernel matrix is obtained.
[0093]
[0094] Furthermore, the similarity matrix is fused, which specifically includes the following three steps: initializing the state matrix P. (n) As the initial state matrix:
[0095]
[0096] Furthermore, different state matrices are iterated through cross-diffusion based on the corresponding kernel matrices to obtain the cross-state matrices. The specific calculation formula is as follows:
[0097]
[0098] Where, α i Indicates the weighted adjustment coefficient and You can set it based on experience (generally take...). ), where η represents the regularization parameter, and the iteration stops after reaching a specified number of steps.
[0099] Furthermore, the fusion similarity is calculated, and the final state matrix is used as the fusion similarity matrix P based on the state matrices after different crossovers. (c) :
[0100]
[0101] Depending on the actual application requirements, other similarity fusion methods such as simple aggregation algorithms (maximum value, average value) may also be used in the embodiments of this application.
[0102] Furthermore, based on the fusion similarity matrix P (c) A quantitative graph of relationships between public users is obtained, as described in steps S3051 to S3052.
[0103] This application embodiment accurately obtains a quantitative graph of corporate user relationships based on a cross-diffusion similarity fusion method, thereby accurately recommending potential corporate users of input bank users and improving marketing effectiveness.
[0104] Furthermore, steps S3051 to S3052 are described as follows:
[0105] Step S3051: Determine the diagonal matrix of the fused similarity matrix, and determine the adjacency matrix based on the fused similarity matrix and its diagonal matrix;
[0106] Step S3052: Generate a similarity graph based on the adjacency matrix, and determine the similarity graph as the public user relationship quantification graph.
[0107] Specifically, determine the diagonal matrix diag(P) composed of the diagonal elements of the fusion similarity matrix. (c) According to the fusion similarity matrix P (c) and its diagonal matrix diag(P) (c) Calculate the adjacency matrix L, where the formula for calculating the adjacency matrix L is L = P. (c) -diag(P (c) Furthermore, a similarity graph G′=(V′,E′) is generated based on the adjacency matrix L, where V′ represents the set of public user nodes, E′ represents the set of directed edges, and the edge weights represent the quantized relationship weights. The similarity graph G′=(V′,E′) is then determined as the quantized graph of public user relationships.
[0108] This application embodiment accurately obtains a quantitative graph of corporate user relationships by fusing a similarity matrix with its corresponding diagonal matrix. This quantitative graph of corporate user relationships is then used to accurately recommend potential corporate users to input bank users, thereby improving marketing effectiveness.
[0109] Step S40: Based on the corporate user relationship quantification graph and the input bank user, recommend potential corporate users of the input bank user.
[0110] Identify the core user (input bank user) and input the input bank user into the corporate user relationship quantification graph. Recommend potential corporate users of the input bank user through the corporate user relationship quantification graph, as described in steps S401 to S402.
[0111] The heterogeneous graph-based method for recommending potential corporate bank users provided in this application constructs a corporate user relationship network using heterogeneous graphs and user data. Then, based on the application scenario, various corporate user relationships within this network are divided into corporate user relationship subgraphs. These subgraphs are then combined to obtain a quantitative corporate user relationship graph. Finally, the quantitative graph accurately recommends potential corporate users from input bank users, thereby improving marketing effectiveness.
[0112] Further, steps S401 to S402 are described as follows:
[0113] Step S401: Calculate the relevance of all corporate users relative to the input bank users based on the corporate user relationship quantification graph, and sort the relevance using a preset algorithm;
[0114] Step S402: Recommend all target corporate users that meet the preset relevance threshold based on the sorted relevance, and determine all target corporate users as potential corporate users of the input bank user.
[0115] Specifically, given a core set of user nodes, the relevance of all nodes relative to the given node is calculated on the corporate relationship quantification graph. Further, the relevance is sorted from high to low using a preset algorithm to obtain a relevance ranking of potential corporate users. The input core user can be one or more. The preset algorithm in this embodiment can be the PageRank algorithm or other ranking algorithms. Further, a preset relevance threshold is combined with the relevance ranking of potential corporate users to output a set of nodes that meet the preset relevance threshold as the list of potential corporate users to be recommended. The preset relevance threshold can be understood as a ranking threshold. In one embodiment, the preset relevance threshold can be a top 15 threshold, meaning the final output is a set of the top 15 nodes as the list of potential corporate users to be recommended.
[0116] This application's embodiments start with core users and calculate node rankings as a list of potential corporate users. This makes the users on the list considered to have a strong correlation with core users in multiple aspects, making them suitable as corporate marketing targets. In other words, it accurately recommends potential corporate users to input bank users, thereby improving marketing effectiveness.
[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for recommending potential corporate clients of banks based on heterogeneous graphs, characterized in that, include: A corporate user relationship network is constructed based on corporate user data, wherein the corporate user relationship network is a heterogeneous graph, and the heterogeneous graph includes multiple corporate user relationships; Based on the application scenario, the various public user relationships are divided into multiple public user relationship subgraphs, and the similarity matrix of the multiple public user relationship subgraphs is calculated. The process involves fusing multiple similarity matrices to obtain a quantitative graph of user relationships. This fusing includes: normalizing each similarity matrix to obtain a state matrix; measuring the local relationships of each similarity matrix using a preset algorithm to obtain local similarity matrices; normalizing each local similarity matrix to obtain a kernel matrix; iterating each different state matrix a preset number of times using a cross-diffusion method based on its corresponding kernel matrix to obtain a cross-diffusion state matrix; calculating a fused similarity matrix based on the different cross-diffusion state matrices, and obtaining the quantitative graph of user relationships based on the fused similarity matrix. The process of obtaining the quantitative graph of user relationships based on the fused similarity matrix includes: determining the diagonal matrix of the fused similarity matrix; determining an adjacency matrix based on the fused similarity matrix and its diagonal matrix; generating a similarity graph based on the adjacency matrix; and determining the similarity graph as the quantitative graph of user relationships. Based on the corporate user relationship quantification graph and the input bank user, potential corporate users of the input bank user are recommended; the step of recommending potential corporate users of the input bank user based on the corporate user relationship quantification graph and the input bank user includes: calculating the relevance of all corporate users relative to the input bank user according to the corporate user relationship quantification graph, and sorting the relevance according to a preset algorithm; recommending all target corporate users that meet the preset relevance threshold according to the sorted relevance, and determining all target corporate users as potential corporate users of the input bank user.
2. The method for recommending potential corporate bank users based on heterogeneous graphs according to claim 1, characterized in that, The application scenarios include money transfer transactions; The process involves dividing the various corporate user relationships according to application scenarios to obtain multiple corporate user relationship sub-graphs, including: Based on the aforementioned transfer transaction scenario, determine each corporate user and their transfer transaction data; Based on each corporate user and its transfer transaction data, the various corporate user relationships are divided into transfer transaction relationships, resulting in a transfer transaction relationship subgraph; In the transfer transaction relationship subgraph, the nodes are each corporate user, the nodes are transfer transaction data, and the edge attributes are transaction time information and transaction amount information.
3. The method for recommending potential corporate bank users based on heterogeneous graphs according to claim 1, characterized in that, The application scenarios include industry scenarios; The process involves dividing the various corporate user relationships according to application scenarios to obtain multiple corporate user relationship sub-graphs, including: Based on the industry scenarios described, determine the information of each corporate user and their industry. Based on the information of each corporate user and its industry, the various corporate user relationships are divided into industry relationships, resulting in an industry relationship sub-graph. In the industry relationship subgraph, the nodes represent each corporate user and industry information, and the nodes' edges represent the industries to which each corporate user belongs and their upstream and downstream partners.
4. The method for recommending potential corporate bank users based on heterogeneous graphs according to claim 1, characterized in that, The application scenarios include regional scenarios; The process involves dividing the various corporate user relationships according to application scenarios to obtain multiple corporate user relationship sub-graphs, including: Based on the aforementioned regional scenario, determine the information of each corporate user and their respective regions; Based on the information of each corporate user and its region, the various corporate user relationships are divided into regional relationships, resulting in a regional relationship subgraph. In the regional relationship subgraph, the nodes represent the regional information of each corporate user, and the nodes' edges represent the regions where each corporate user is located and their proximity relationships.