Target customer prediction method and device based on relational network

Abstract

The invention discloses a target customer prediction method and device based on a relational network. The method comprises the following steps: generating a data wide table based on acquired in-line data and out-line data; based on the in-row data and the out-row data, public opinion risk tags and corresponding tag values are obtained through a pre-established public opinion risk classification model; constructing a feature vector set according to the data wide table and a preset variable screening rule; based on preset fact class tags, public opinion risk tags and corresponding tag values, constructing a target customer pool; determining a target customer based on the fact class tag, the public opinion risk tag, the target customer pool, the feature vector set and a pre-constructed relationship network diagram, wherein the relationship network diagram is constructed based on a data wide table; according to the method, unified integration of internal and external data, structured modeling of an enterprise association relationship and joint quantitative analysis of customer similarity features and network structure features are realized, so that the accuracy and stability of target customer identification are improved.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method and apparatus for predicting target customers based on relational networks. Background Technology

[0002] As businesses continue to expand and financial institutions enhance their digital capabilities, customer marketing is gradually shifting from a traditional experience-driven model to a data-driven, intelligent management approach. In customer marketing, risk assessment, credit management, and investment and financing services, banks and large enterprises typically rely on structured data from multiple internal systems. However, differences in data structure, field definitions, and data quality between these systems make it difficult to directly integrate internal data into an accurate enterprise profile.

[0003] At the same time, the scale of publicly available external data has grown rapidly, providing richer information sources for business analysis, including business registration information, judicial information, publicly available financial information, risk sentiment, and supply chain data. However, due to the diverse sources, complex structures, and inconsistent update frequencies of external data, as well as the lack of a unified linkage mechanism with internal data, corporate information is fragmented, limiting its effective application in customer analysis and business decision-making.

[0004] Furthermore, businesses commonly engage in multi-layered and diverse business relationships, including equity relationships, supply chain relationships, investment relationships, transactional relationships, and group affiliations. These relationships significantly impact a company's operational stability, creditworthiness, and business potential. However, existing customer marketing methods often rely on a company's static attributes or single-dimensional characteristics, failing to reflect its position within the network of relationships and effectively uncovering the potential business value and risk transmission paths between companies.

[0005] Therefore, existing technologies generally face problems in customer marketing, such as scattered data sources inside and outside the industry, insufficient utilization of corporate relationships, a single method for measuring customer similarity, and difficulty in structuring and quantifying risk factors. These issues make it difficult to meet the actual needs of refined and intelligent customer marketing development.

[0006] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section. Summary of the Invention

[0007] This invention provides a target customer prediction method based on relationship networks, which enables unified integration of intra-industry and inter-industry data, structured modeling of enterprise relationships, and joint quantitative analysis of customer similarity features and network structure features, thereby improving the accuracy and stability of target customer identification.

[0008] This target customer prediction method based on relationship networks includes:

[0009] A wide data table is generated based on the acquired in-row and out-of-row data;

[0010] Based on the in-line and out-of-line data, a pre-established public opinion risk classification model is used to obtain public opinion risk labels and their corresponding label values.

[0011] A feature vector set is constructed based on the wide data table and the preset variable filtering rules;

[0012] A target customer pool is constructed based on preset factual tags and the aforementioned public opinion risk tags and their corresponding tag values;

[0013] Target customers are determined based on the factual tags, the public opinion risk tags, the target customer pool, the feature vector set, and the pre-built relationship network graph, which is constructed based on the data wide table.

[0014] In some embodiments, generating a wide data table based on the acquired in-row and out-of-row data includes:

[0015] Based on the in-row data and the out-of-row data, determine the primary key field, association type, and primary key matching rules;

[0016] Based on the primary key field and the primary key matching rules, perform primary key matching operations on the in-row data and the out-of-row data, and obtain the matching results based on the association type;

[0017] Based on the matching results, the data within the row and the data outside the row are merged at the field level to generate a wide data table.

[0018] In some embodiments, the method further includes:

[0019] If the primary key field of the in-row data matches multiple primary key fields of the out-of-row data, a primary key field is selected from the multiple primary key fields of the out-of-row data as the matching result according to the preset data processing rules, or the primary key fields of the multiple out-of-row data are aggregated to generate a matching result.

[0020] In some embodiments, constructing a relationship network graph based on the data wide table includes:

[0021] Enterprise information is extracted from the data wide table to construct a node set; wherein, the enterprise information includes: enterprise name, unified social credit code, enterprise registered address and legal representative;

[0022] Relationship data corresponding to preset enterprise relationship types are extracted from the data wide table, and relationship edges between nodes are established; the enterprise relationship types include: shareholder relationship, external investment relationship, supply chain relationship, capital transaction relationship and group affiliation relationship;

[0023] A relational network graph is constructed based on the set of nodes and the relational edges.

[0024] In some embodiments, the method further includes:

[0025] Select a target node in the relational network graph, and determine the neighboring nodes of the target node based on the relational edges of the target node;

[0026] The target node is added to the community where each neighbor node is located, and the modularity transformation value of each neighbor node is determined based on the first modularity of the target node in the original community and the second modularity of the target node in the community where each neighbor node is located.

[0027] Sort the modularity change values ​​of each neighboring node in a preset order and determine the maximum value of the modularity change value;

[0028] If the maximum value is greater than zero, the target node will be assigned to the community where the neighbor node corresponding to the maximum value is located.

[0029] The above steps are repeated iteratively until the communities to which all nodes in the relational network graph belong tend to stabilize, resulting in an updated relational network graph.

[0030] In some embodiments, obtaining public opinion risk labels and their corresponding label values ​​based on the in-line data and out-of-line data through a pre-established public opinion risk classification model includes:

[0031] The acquired public opinion risk data is labeled according to preset tag types to generate text tag pairs consisting of the public opinion risk data and corresponding tags; wherein, the tag types include sentiment tags and topic category tags;

[0032] The text tag pairs are input into a pre-established public opinion risk classification model, and the public opinion risk classification model is iteratively trained based on preset model parameters to obtain public opinion risk tags and their corresponding tag values.

[0033] In some embodiments, constructing a feature vector set based on the wide data table and preset variable filtering rules includes:

[0034] Target feature variables are selected from the wide data table based on preset variable selection rules; wherein, the variable selection rules include: qualitative selection rules based on expert experience and quantitative selection rules based on statistical analysis;

[0035] The target feature variables are subjected to feature engineering processing to obtain a set of feature vectors.

[0036] In some embodiments, the feature engineering process includes bucketing, wherein performing feature engineering on the target feature variable to obtain a set of feature vectors includes:

[0037] Based on the data distribution characteristics of the target feature variables, the discretization interval is determined;

[0038] The target feature variable is mapped to the discretization interval to generate the feature vector.

[0039] In some embodiments, the feature engineering process includes a smoothing process, wherein performing feature engineering on the target feature variable to obtain a set of feature vectors includes:

[0040] The target feature variable is smoothed by applying a logarithmic function or a square root function to generate the feature vector.

[0041] In some embodiments, the feature engineering process includes normalization processing, and the feature engineering process on the target feature variable to obtain a feature vector set includes:

[0042] Determine the maximum and minimum values ​​of the target feature variable;

[0043] The target feature variable is normalized based on its maximum and minimum values ​​to obtain the feature vector.

[0044] In some embodiments, constructing a target customer pool based on preset factual tags and public opinion risk tags and their corresponding tag values ​​includes:

[0045] The preset factual tags and the public opinion risk tags are sorted according to a second preset order, and weights are assigned according to the sorting results;

[0046] The labels are weighted and aggregated according to their respective weights and corresponding label values ​​to obtain the label scores for each enterprise.

[0047] A target customer pool is constructed based on the aforementioned tag scores.

[0048] In some embodiments, determining target customers based on the factual tags, the public opinion risk tags, the target customer pool, the feature vector set, and the pre-constructed relationship network graph includes:

[0049] The Jaccard similarity between customers is determined based on the aforementioned factual tags;

[0050] The cosine similarity between customers is determined based on the aforementioned public opinion risk tags and the aforementioned feature vector set;

[0051] Based on the Jaccard similarity, the cosine similarity, and the target customer pool, a recall list is generated;

[0052] Target customers are identified based on the recall list and the relationship network diagram.

[0053] This invention also provides a target customer prediction device based on a relationship network, which can realize the unified integration of intra-industry and inter-industry data, the structured modeling of enterprise relationships, and the joint quantitative analysis of customer similarity features and network structure features, thereby improving the accuracy and stability of target customer identification.

[0054] This target customer prediction device based on relationship networks includes:

[0055] The wide data table generation module is used to generate a wide data table based on the acquired in-row and out-of-row data;

[0056] The public opinion risk acquisition module is used to obtain public opinion risk labels and their corresponding label values ​​based on the in-line data and out-of-line data through a pre-established public opinion risk classification model;

[0057] The feature vector construction module is used to construct a set of feature vectors based on the wide data table and preset variable filtering rules;

[0058] The target customer pool construction module is used to construct a target customer pool based on preset factual tags and the public opinion risk tags and their corresponding tag values;

[0059] The target customer identification module is used to identify target customers based on the factual tags, the public opinion risk tags, the target customer pool, the feature vector set, and a pre-built relationship network graph, wherein the relationship network graph is constructed based on the data wide table.

[0060] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described target customer prediction method based on a relationship network.

[0061] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described target customer prediction method based on a relationship network.

[0062] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described target customer prediction method based on a relationship network.

[0063] The target customer prediction method and apparatus based on relationship networks provided in this invention effectively solves the problems of incomplete enterprise profiles caused by scattered multi-source data, inconsistent field definitions, and redundant or missing enterprise information by unifying and integrating intra-industry and inter-industry data and constructing a wide data table, thereby significantly improving the data foundation quality for customer identification. The public opinion risk tags and their corresponding quantitative tag values ​​generated based on the public opinion risk classification model enable external risk events to be mapped to the enterprise profile system in real time, achieving dynamic representation of the enterprise's external risk exposure and enhancing the responsiveness of target customer prediction to risk changes. By introducing variable screening rules and feature engineering methods to construct a feature vector set, multi-dimensional operating indicators such as enterprise operating scale, financial stability, and behavioral characteristics are structurally and quantitatively processed, improving the interpretability, comparability, and stability of model input characteristics of the enterprise profile. By combining factual tags with public opinion risk tags to construct a target customer pool, an access evaluation mechanism based on a combination of business rules and data-driven approaches is implemented, thereby improving the accuracy and rationality of target customer screening. Furthermore, the enterprise relationship network constructed based on the wide data table can quantitatively characterize key structural features such as equity linkages, supply chain relationships, and financial transactions between enterprises, providing a more comprehensive reflection of an enterprise's position and influence within the overall network environment. By jointly modeling the relationship network structural features with a feature vector system, this invention can comprehensively assess target customers across both customer ontology features and network relationship features, thereby effectively improving the accuracy, stability, and business adaptability of target customer predictions. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0065] Figure 1 This is a flowchart illustrating a target customer prediction method based on a relationship network in one embodiment of the present invention.

[0066] Figure 2 This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0067] Figure 3This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0068] Figure 4 This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0069] Figure 5 This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0070] Figure 6 This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0071] Figure 7 This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0072] Figure 8 This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0073] Figure 9 This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0074] Figure 10 This is a flowchart illustrating a target customer prediction method based on a relationship network in another embodiment of the present invention.

[0075] Figure 11 This is a schematic diagram of the target customer prediction device based on a relationship network in another embodiment of the present invention;

[0076] Figure 12 This is a schematic diagram of the physical structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments and their descriptions are used to explain the present invention, but are not intended to limit the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other. The acquisition, storage, use, and processing of data in the technical solutions of this application all comply with relevant laws and regulations. The user information in the embodiments of this application is obtained through legal and compliant means, and the acquisition, storage, use, and processing of user information have been authorized and agreed upon by the customer.

[0078] To facilitate understanding of the technical solution provided in this application, the relevant content of the technical solution in this application will be explained below.

[0079] To address the technical challenges of low target customer identification accuracy, fragmented and difficult-to-integrate internal and external data sources, insufficient enterprise relationship mining, and weak interpretability of customer recall methods based on single similarity in existing customer marketing processes, this application proposes a target customer prediction method based on relationship networks. This method unifies and integrates internal and external data to construct a wide data table, and generates an enterprise relationship network diagram based on this table to reflect the multidimensional relationships between enterprises. It utilizes a public opinion risk classification model to obtain enterprise public opinion risk labels, and combines this with a feature vector system constructed through variable screening and feature engineering to achieve a unified quantitative expression of enterprise operational characteristics, risk characteristics, and behavioral characteristics. Through joint modeling of the enterprise relationship network diagram, enterprise public opinion risk labels, and feature vector system, an interpretable, quantifiable, and scalable customer prediction system is constructed, enabling a comprehensive assessment of customer value, potential business needs, and risk levels, thereby improving the accuracy, stability, and business adaptability of target customer identification.

[0080] The aforementioned customers can be corporate customers or individual customers. This embodiment uses corporate customers as an example for illustration, but the present invention is not limited thereto.

[0081] This invention provides a target customer prediction method based on relationship networks, such as... Figure 1 As shown, it includes steps 101 to 105.

[0082] Step 101: Generate a wide data table based on the obtained in-row and out-of-row data.

[0083] Step 102: Based on internal and external data, obtain public opinion risk labels and their corresponding label values ​​through a pre-established public opinion risk classification model.

[0084] Step 103: Construct a set of feature vectors based on the wide data table and the preset variable filtering rules.

[0085] Step 104: Construct a target customer pool based on preset factual tags and public opinion risk tags and their corresponding tag values.

[0086] Step 105: Identify target customers based on factual tags, public opinion risk tags, target customer pool, feature vector set, and pre-constructed relationship network diagram. The relationship network diagram is constructed based on a wide data table.

[0087] According to the above embodiments, this invention effectively solves the problems of incomplete enterprise profiles caused by scattered multi-source data, inconsistent field definitions, and redundant or missing enterprise information by unifying and integrating internal and external data and constructing a wide data table, thereby significantly improving the data foundation quality for customer identification. The public opinion risk tags and their corresponding quantitative tag values ​​generated based on the public opinion risk classification model enable external risk events to be mapped to the enterprise profile system in real time, achieving dynamic representation of the enterprise's external risk exposure and enhancing the responsiveness of target customer prediction to risk changes. By introducing variable screening rules and feature engineering methods to construct a feature vector set, multi-dimensional operating indicators such as enterprise operating scale, financial stability, and behavioral characteristics are structurally and quantitatively processed, improving the interpretability, comparability, and stability of model input characteristics of the enterprise profile. By combining factual tags with public opinion risk tags to construct a target customer pool, an access evaluation mechanism based on a combination of business rules and data-driven approaches is implemented, thereby improving the accuracy and rationality of target customer screening. Furthermore, the enterprise relationship network constructed based on the wide data table can quantitatively characterize key structural features such as equity linkages, supply chain relationships, and financial transactions between enterprises, providing a more comprehensive reflection of an enterprise's position and influence within the overall network environment. By jointly modeling the relationship network structural features with a feature vector system, this invention can comprehensively assess target customers across both customer ontology features and network relationship features, thereby effectively improving the accuracy, stability, and business adaptability of target customer predictions.

[0088] In some embodiments, such as Figure 2 As shown, step 101 includes steps 201 to 203.

[0089] Step 201: Based on inline and outline data, determine the primary key field, association type, and primary key matching rules.

[0090] Step 202: Perform primary key matching operations on in-row and out-of-row data based on the primary key field and primary key matching rules, and obtain the matching results based on the association type.

[0091] Step 203: Based on the matching results, perform field-level merging processing on the in-row data and out-of-row data to generate a wide data table.

[0092] In this embodiment of the invention, internal bank data is obtained based on the bank's internal business system to characterize the operational and transaction characteristics of corporate clients within the banking system. This internal bank data includes customer attribute data and account attribute data. The customer attribute data originates from multiple dimensions, including basic corporate information, financial statement data, deposit and loan information, and settlement transaction data. The account attribute data comes from core business forms such as transaction logs, account customer details, and account combination query tables, and reflects information such as corporate account structure, fund flows, and account usage behavior.

[0093] External data related to corporate customer marketing is obtained from the bank's data integration platform. This external data comes from third-party data services, public data resources, or external regulatory data, covering a variety of data themes such as historical development, management capabilities, business experience, corporate relationships, asset and liability status, fund flow characteristics, legal litigation information, innovation capability evaluation, and risk warning information. It is used to supplement the external operating characteristics and risk characteristics of enterprises beyond the bank's internal data.

[0094] In some embodiments, the target customer prediction method based on relationship networks further includes: if the primary key field of the in-row data matches the primary key fields of multiple out-of-row data, selecting one primary key field from the primary key fields of the multiple out-of-row data as the matching result according to a preset data processing rule, or performing aggregation processing on the primary key fields of the multiple out-of-row data to generate a matching result.

[0095] In this embodiment of the invention, after the in-row data and out-of-row data are acquired, the associated fields and matching rules of the two are first defined.

[0096] Specifically, the primary key field for intra-industry data is defined as either the customer's unique identification number or the unified credit granting number, while the primary key field for inter-industry data is defined as the unified social credit code and the enterprise owner's ID number. These primary key fields are used to uniquely identify data from different sources. The customer's unique identification number in intra-industry data can be correlated with the enterprise owner's ID number in inter-industry data, and the unified credit granting number in intra-industry data can be correlated with the unified social credit code in inter-industry data, thus achieving a one-to-one primary key correspondence between intra-industry and inter-industry data.

[0097] Furthermore, association types are defined based on the characteristics of intra-industry and inter-industry data, including: a one-to-one association type where one customer corresponds to one external record, a one-to-many association type where one customer corresponds to multiple external records (e.g., multiple credit records), and a many-to-many association type where multiple customers correspond to multiple external records. Primary key matching rules are configured based on the above association types to facilitate the structured integration of data from heterogeneous sources.

[0098] Primary key matching rules are used to indicate the specific matching method between data within a row and data outside a row, and data merging is achieved through field-level matching based on the primary key field.

[0099] For example, a JOIN operation can be performed on in-row and out-of-row data using SQL statements to complete a relational data association based on the primary key field. Alternatively, Python scripts or data processing tools (such as Pandas or Tableau) can be used to perform primary key matching on in-row and out-of-row data.

[0100] During the primary key matching process, if the primary key field of the data within a row does not match the primary key field of the data outside the row, the primary key fields of the data within the row and the data outside the row are marked as missing values ​​to indicate that the matching record cannot be effectively associated.

[0101] When the primary key field of intra-row data matches multiple primary key fields of inter-row data simultaneously, the target primary key field is selected as the final matching result from multiple candidate primary key fields according to data priority rules and field aggregation rules. Alternatively, the primary key fields of multiple inter-row data can be aggregated, and a unified matching result can be generated through field merging, record merging, or statistical summarization.

[0102] By using the missing value identification and multiple matching mechanisms described above, the integrity and consistency of the primary key matching process can be ensured, avoiding data anomalies caused by duplicate or invalid matching, and improving the accuracy and reliability of wide data table construction.

[0103] In some embodiments, step 203 specifically includes: based on the primary key matching result, merging the successfully matched in-row data and out-of-row data according to the field-level correspondence to generate a wide data table.

[0104] For example, internal data may include fields such as customer ID, customer name, credit amount, and transaction records. External data may include fields such as unified social credit code, company name, credit score, and tax records. Based on the above field correspondence, integrating internal and external data at the field level can generate a wide data table containing multi-dimensional attributes such as customer ID, customer name, credit amount, transaction records, unified social credit code, company name, credit score, and tax records. This achieves a unified structured expression of corporate customer attributes, business behavior, and external risk information.

[0105] Furthermore, to improve the integrity and effectiveness of the wide data table, data quality processing is performed on missing values ​​in the wide data table.

[0106] For example, two strategies can be employed: missing value removal and missing value imputation. When the missing percentage of a certain feature field exceeds a preset threshold (e.g., 90%), that feature can be deleted. If the missing percentage does not exceed the preset threshold, missing values ​​can be imputed using methods such as mode imputation, median imputation, or mean imputation, based on expert experience. This missing value handling process effectively improves the reliability of the wide data table and the stability of subsequent modeling stages.

[0107] In some embodiments, such as Figure 3As shown, constructing a relational network graph based on the above wide data table includes steps 301 to 303.

[0108] Step 301: Extract enterprise information from the wide data table and construct a node set. Enterprise information includes: enterprise name, Unified Social Credit Identifier (USCI), registered address, and legal representative.

[0109] Step 302: Extract the relationship data corresponding to the preset enterprise relationship types from the data wide table, and establish the relationship edges between nodes. Enterprise relationship types include: shareholder relationships, external investment relationships, supply chain relationships, capital transaction relationships, and group affiliation relationships.

[0110] Step 303: Construct a relational network graph based on the node set and relational edges.

[0111] In this embodiment of the invention, basic information corresponding to the enterprise entity is extracted based on the generated wide data table and used as node attributes in the relationship network graph. The enterprise information includes fields such as enterprise name, unified social credit code, registered address, and legal representative, which can be used to uniquely identify the enterprise and represent its basic attributes. Subsequently, based on the enterprise owner information or shareholder information in the wide data table, individual nodes are further constructed, allowing enterprise nodes and individual nodes to participate simultaneously in the structural representation of the relationship network.

[0112] Furthermore, various enterprise relationship types are pre-defined based on business analysis needs to describe the structural relationships between enterprises. These enterprise relationship types include: shareholder relationships, external investment relationships, supply chain relationships, capital flow relationships, and group affiliations. Among them, shareholder relationships reflect the equity holding relationship between shareholders and enterprises; external investment relationships reflect the investment structure of enterprises investing in other enterprises; supply chain relationships describe the upstream and downstream business transactions of enterprises in the industry chain; capital flow relationships record the flow of funds between enterprises; and group affiliations characterize the organizational structure relationships of enterprises within the same enterprise group.

[0113] After the node information is determined, relationship edges between enterprise nodes are established according to the preset enterprise relationship types to construct a relationship network graph.

[0114] For example, for shareholder relationships, directional edges can be established between shareholder nodes and enterprise nodes based on shareholder information, recording attributes such as shareholding ratio and subscribed amount on the edges. For external investment relationships, directional edges can be established between investing and invested enterprises based on investment records, recording attributes such as investment amount and investment ratio on the edges. For supply chain relationships, edges can be established between supplier and customer nodes based on supply chain transaction data, recording attributes such as transaction amount and transaction frequency. For fund transfer relationships, fund transfer edges can be established between relevant enterprise nodes based on fund flow records, recording fields such as fund amount and transaction date. For group affiliation relationships, association edges can be established between enterprise nodes within the same group based on group account information to reflect the group's structural hierarchy. Through the above node construction and edge establishment process, a relationship network diagram depicting the basic attributes and multidimensional relationships of enterprises can be generated.

[0115] For example, taking fund transfer transactions as an example, the transaction relationships between enterprises can be extracted based on transaction flow data. The payer and payee enterprises participating in the transaction are identified from the transaction flow table, and the two enterprises are respectively treated as nodes in the relationship network diagram.

[0116] To ensure the stability and validity of transaction relationships, transaction frequency can be filtered, retaining only enterprise pairs with more than a preset threshold (e.g., three transactions) that reflect continuous business dealings. If the transaction count condition is met, a financial relationship is considered to exist between the enterprise pairs, thus establishing a relationship edge between the two corresponding enterprise nodes. Attribute information such as transaction amount, transaction frequency, or transaction time can be recorded on this relationship edge to characterize the strength and characteristics of the financial relationship.

[0117] For enterprise relationships with many-to-many characteristics, such as supply chain relationships or financial transactions, the association fields in the wide data table are traversed to ensure a complete and accurate representation of each real enterprise relationship in the relationship network diagram. For enterprise group relationships composed of group account information, to avoid the loss or confusion of hierarchical relationships, multi-level associations in the group structure are modeled hierarchically, so that the organizational structure of the enterprise group has clear hierarchy and traceability in the relationship network diagram.

[0118] After the relationship edges are constructed, the constructed relationship network is validated to ensure the accuracy of the network structure. Specifically, the source of association corresponding to each edge can be verified to determine whether it correctly reflects the actual enterprise relationships. For duplicate relationship edges, deduplication is performed to eliminate data redundancy, ensuring the uniqueness and consistency of the relationship network. In addition, node and edge attributes can be adjusted as needed according to subsequent modeling requirements, such as supplementing node classification attributes, supplementing edge weight attributes, or standardizing field formats, thereby improving the structured quality and data usability of the relationship network.

[0119] After completing the relationship network construction and the above verification process, the generated enterprise relationship network is stored in a graph database format, such as a Neo4j graph database, to support subsequent queries, calculations, and visualization analysis. Graph computing or graph visualization tools (such as NetworkX) can be used to visualize the enterprise relationship network, presenting the relationships between enterprises in a graph structure.

[0120] In some embodiments, the target customer prediction method based on the relationship network further includes: determining the node centrality of each node in the relationship network graph based on the node degree and the total number of nodes in the relationship network graph.

[0121] This invention uses Degree Centrality as an indicator to measure the connection strength and structural influence of enterprise nodes in a relational network. It uses the association edges formed by enterprise relation types as the basic elements of the relational network to reflect the structural role of enterprise nodes in different range networks or different complex relationships.

[0122] Specifically, for each node in the relational network graph, the total number of nodes in the relational network graph is calculated based on the number of its connecting edges (i.e., node degree). The centrality value of each node is calculated according to the following formula, and then a list of nodes and their corresponding centrality values ​​are generated.

[0123] The formula for calculating degree centrality is as follows:

[0124]

[0125] in, Represents a node The degree, This represents the total number of nodes in the relational network graph.

[0126] In some embodiments, such as Figure 4 As shown, the target customer prediction method based on relationship networks also includes steps 401 to 405.

[0127] Step 401: Select a target node in the relational network graph and determine the target node's neighboring nodes based on the relational edges of the target node.

[0128] Step 402: Add the target node to the community where each neighbor node is located, and determine the modularity transformation value of each neighbor node based on the first modularity of the target node in the original community and the second modularity of the target node in the community where each neighbor node is located.

[0129] Step 403: Sort the modularity change values ​​of each neighboring node in a preset order and determine the maximum value of the modularity change value.

[0130] Step 404: If the maximum value is greater than zero, assign the target node to the community where the neighbor node corresponding to the maximum value is located.

[0131] Step 405: Iterate through the above steps until all the communities of the nodes in the relational network graph have stabilized, and obtain the updated relational network graph.

[0132] In this embodiment of the invention, the Louvain algorithm is used to partition the constructed relational network graph into communities. The Louvain algorithm is a community partitioning method based on modularity optimization. Modularity measures the difference between the actual connections within a community and the expected connections in a random network; its value typically ranges from [value missing]. .

[0133] Initialize each node in the relational network graph as an independent community. For any target node in the relational network graph... Iterate through the communities to which all its neighboring nodes belong, and then transfer the target node to each community. Hypothetically join the community where each neighbor node is located, and calculate the module degree increment based on the change in module degree before and after joining. .

[0134] At the target node During the process of moving from the original community C to the candidate community D, the target node Modularity increment It can be represented as:

[0135]

[0136] in, For the target node Within the original community C, the degree of connectivity. For the target node Connectivity within candidate community D This represents the total number of all relation edges in the relational network graph.

[0137] If the module degree increment of a certain neighboring community If it is greater than zero, it means that the target node will be... Assigning a node to this candidate community can improve the overall quality of community partitioning. In this case, the candidate neighbor community with the largest increase in modularity is selected as the target node. The new community. If the modularity of all communities increases... If none of them are greater than zero, then keep the target node. The community divisions remain unchanged.

[0138] The above steps are iterated through all nodes in the relational network graph until the community division of all nodes no longer changes, thus obtaining a stable community structure.

[0139] Nodes belonging to the same community are merged and compressed to form new nodes. The above modularity optimization process is repeated in the compressed relational network graph until the modularity value of the overall relational network no longer changes, thus obtaining the community division result of the relational network graph.

[0140] For example, the calculation results of node centrality can be used to identify nodes with high structural importance in a relational network. Nodes with high centrality usually have strong connectivity and influence in the relational network, and often serve as core nodes in the community discovery process, significantly affecting the community division results and the overall community structure.

[0141] Node centrality can be used to further analyze the structure of a divided community, such as identifying core nodes within the community, evaluating the tightness of connections within the community, and judging the structural stability of the community. Node centrality can also serve as an auxiliary indicator for evaluating community quality, measuring the connection characteristics between nodes within the community.

[0142] For example, high-quality communities typically contain a large number of central nodes with strong connections, exhibiting greater density and consistency in their internal structure. By combining node centrality with community partitioning results, the reliability and interpretability of relationship networks in target customer discovery and association structure insights can be enhanced.

[0143] In some embodiments, such as Figure 5 As shown, step 102 includes steps 501 to 502.

[0144] Step 501: Perform data annotation processing on the acquired public opinion risk data according to the preset tag types to generate text tag pairs consisting of public opinion risk data and corresponding tags. The tag types include sentiment tags and topic category tags.

[0145] Step 502: Input the text tag pairs into the pre-established public opinion risk classification model, and iteratively train the public opinion risk classification model based on the preset model parameters to obtain public opinion risk tags and their corresponding tag values.

[0146] In this embodiment of the invention, corporate public opinion risk data is acquired and then annotated. Data annotation includes two categories: sentiment annotation and topic annotation. Sentiment annotation is used to determine the sentiment tendency of the corporate risk public opinion data, while topic annotation is used to identify the topic type to which the public opinion text belongs. Sentiment categories include positive, negative, and neutral. Topic types include preset categories such as product type, service type, brand type, and environmental protection type. After completing the data annotation, the public opinion text and its corresponding tags are constructed into text tag pairs.

[0147] The above text label pairs are converted into the input format for the Natural Language Processing (BERT) model. This input format includes model input vectors such as Token ID, Segment ID, and Mask ID.

[0148] After completing the above data formatting process, the dataset is divided into training set, validation set and test set, for example, in a ratio of 70%:15%:15%.

[0149] During the model training phase, the training set is input into a pre-built BERT model, and model training parameters, including learning rate, batch size, and number of training epochs, are set. The cross-entropy loss function is used as the loss function for the classification task, and the Adam optimizer is used to iteratively update the model parameters. The classifier in the fully connected layer of the BERT model performs classification prediction on the input data and outputs the corresponding public opinion risk label and its quantified label value.

[0150] In some embodiments, such as Figure 6 As shown, step 103 includes steps 601 to 602.

[0151] Step 601: Select target feature variables from the wide data table based on preset variable selection rules. These variable selection rules include: qualitative selection rules based on expert experience and quantitative selection rules based on statistical analysis.

[0152] Step 602: Perform feature engineering on the target feature variables to obtain a set of feature vectors.

[0153] In this embodiment of the invention, a qualitative analysis method based on expert experience is first used to filter the fields in the wide data table to obtain candidate feature variables.

[0154] Specifically, business experts assess the importance of key performance indicators (KPIs) in corporate clients' marketing efforts, and candidate KPIs are selected by combining external data such as industry reports and professional research literature. Candidate KPIs may include: sales revenue, debt-to-equity ratio, profit margin, industry category, credit rating, total debt, and brand influence.

[0155] Sales revenue is used to characterize the company's operating scale; the debt-to-asset ratio and total liabilities are used to characterize the company's financial soundness and solvency, and are important reference indicators for risk warning; profit margin is used to characterize the company's profitability; industry category reflects the company's operating environment and demand characteristics; credit rating is used to characterize the company's credit status and is a core variable in risk assessment; brand influence reflects the company's operating scale and market competitiveness to a certain extent.

[0156] After completing the qualitative screening above, a quantitative screening method based on statistical analysis is further used to conduct correlation analysis on the candidate feature variables in order to evaluate the degree of association between the candidate feature variables.

[0157] For example, the linear correlation between candidate feature variables (brand influence) and candidate feature variables (such as credit rating) can be calculated based on the Pearson correlation coefficient to identify redundant or strongly collinear features.

[0158] For example, when the correlation between brand influence and credit rating is significantly higher, it can be determined that the brand influence variable is redundant and thus it can be removed.

[0159] According to the above embodiments, by combining qualitative and quantitative screening, characteristic variables with high repetition, weak correlation, or irrelevance to customer marketing goals can be effectively excluded.

[0160] In some embodiments, the feature engineering process includes bucketing. For example... Figure 7 As shown, step 602 includes steps 701 to 702.

[0161] Step 701: Determine the discretization interval based on the data distribution characteristics of the target feature variable.

[0162] Step 702: Map the target feature variables to the discretized interval to generate feature vectors.

[0163] In this embodiment of the invention, some continuous target feature variables can be discretized (bucketed) to generate discretized feature vectors.

[0164] Specifically, select the target feature variable that needs to be discretized, such as continuous operating indicators like sales revenue and profit margin. Determine its discretization interval based on the actual data distribution characteristics of the target feature variable. This discretization interval can be divided based on the natural quantile method, the equal interval segmentation method, or a custom interval based on business rules.

[0165] After determining the discretization interval, the continuous target feature variables are converted into corresponding discretized feature variables according to the interval mapping rules.

[0166] For example, if the target feature variable is sales revenue, the discretization range can be determined based on the actual data distribution characteristics of sales revenue. For instance, 0–5 million yuan corresponds to the low sales revenue range, 5 million–10 million yuan corresponds to the medium sales revenue range, and over 10 million yuan corresponds to the high sales revenue range.

[0167] In some embodiments, the feature engineering process includes a smoothing process. Step 602 further includes: performing a smoothing transformation on the target feature variable based on a logarithmic function or a square root function to generate a feature vector.

[0168] In this embodiment of the invention, for continuous target feature variables whose data distribution is skewed, long-tailed, or has extreme values, a smoothing transformation can be used to reduce data volatility.

[0169] Specifically, select the target feature variables that need to be smoothed, such as sales revenue and debt-to-equity ratio, or other operational or financial indicators. Use a logarithmic transformation function or a square root transformation function to smooth the target feature variables, obtaining the smoothed feature vector.

[0170] After the smoothing process is completed, the smoothed data can be verified by data distribution visualization methods (such as histograms or probability density maps) to ensure the effectiveness of the smoothing process.

[0171] For example, if the target feature variable is sales revenue, a logarithmic transformation can be used to smooth it out. The smoothed sales revenue can be calculated according to the following formula (3).

[0172]

[0173] In some embodiments, the feature engineering process includes a normalization process. For example... Figure 8 As shown, step 602 also includes steps 801 to 802.

[0174] Step 801: Determine the maximum and minimum values ​​of the target feature variable.

[0175] Step 802: Normalize the target feature variables based on their maximum and minimum values ​​to obtain the feature vector.

[0176] In this embodiment of the invention, some target characteristic variables (such as sales revenue, debt-to-equity ratio, profit margin, etc.) are normalized.

[0177] Specifically, the Min-Max normalization method can be used to linearly scale the target feature variable X across a given interval, transforming the target feature variable X to... Interval.

[0178] The normalized eigenvectors are calculated using the following formula (4). .

[0179]

[0180] in, For target feature variables, The minimum value of the target feature variable. This represents the maximum value of the target feature variable.

[0181] For example, if the target feature variable is sales revenue, and its value ranges from 3 million to 10 million yuan, it can be scaled down using Min-Max normalization. The normalized value for Company A, with sales of 5 million yuan, is approximately (500-300) / (1000-300) ≈ 0.2857. For Company B, with sales of 10 million yuan, the normalized value is 1.0. For Company C, with sales of 3 million yuan, the normalized value is 0.0.

[0182] In some alternative implementations, the Z-score standardization method can be used to transform the target feature variables into a distribution with a mean of 0 and a standard deviation of 1, thereby eliminating dimensional differences and enhancing the robustness of the model.

[0183] According to the above embodiments, the comparability of feature variables can be improved through the above normalization or standardization processing, and the stability and effectiveness of feature vectors in the model training process can be enhanced.

[0184] In some embodiments, when setting factual labels and their corresponding label values, the enterprise can be quantitatively evaluated based on multiple dimensions such as its historical development, management structure, business model, industry characteristics, related relationships, asset and liability structure, cash flow, and risk warning information from external data, thereby generating corresponding factual labels and quantitative label values.

[0185] In terms of historical development, based on the company's establishment date, business registration change records, and disclosure records of major events, the system can calculate the company's years of establishment, number of change records, and number of major events, thereby generating a "years of establishment" tag or a "frequent change" tag and its corresponding tag value. In terms of management, based on the company's organizational structure and management data, the system can calculate the average tenure of management, the proportion of management with higher education levels, and the number of departments, thus generating a "stable management" tag and a "highly educated management" tag and their corresponding tag values. In terms of business model, based on the company's revenue source distribution, market share, and product category data, the system can calculate the number of revenue source types, market share ranking, and number of product types, generating a "diversified revenue" tag and a "market leadership" tag and their corresponding tag values.

[0186] In terms of industry characteristics, based on industry rankings, the number of major competitors, and industry market trend data, the system can calculate a company's market share, the number of competitors, and the market share growth rate, generating industry leader and highly competitive labels and their corresponding values. In terms of related-party relationships, based on shareholder, subsidiary, and partner information, the system can calculate the number of related companies, the largest shareholder's shareholding ratio, and the number of key partners, generating frequent related-party transaction and dispersed shareholding labels and their corresponding values. In terms of asset and liability structure, based on the company's financial statement data, the system can calculate the debt-to-equity ratio, current ratio, and quick ratio, generating high debt-to-equity ratio and good liquidity labels and their corresponding values.

[0187] In terms of cash flow, based on cash flow data from a company's operating, investing, and financing activities, net cash flow from operating activities, net cash flow from investing activities, and net cash flow from financing activities can be calculated, thereby generating labels for sufficient cash flow and active investment, along with their corresponding label values. In terms of risk warning, by combining operating indicators such as changes in profit margin and market share, the magnitude of profit margin decline and market share decline can be calculated, generating labels for high financial risk and high market risk, along with their corresponding label values.

[0188] In some embodiments, such as Figure 9 As shown, step 104 includes steps 901 to 903.

[0189] Step 901: Sort the preset factual tags and public opinion risk tags according to the second preset order, and assign weights according to the sorting results.

[0190] Step 902: Based on the weights of each factual label and public opinion risk label and their corresponding label values, perform weighted summation processing on each label to obtain the label score for each enterprise.

[0191] Step 903: Construct a target customer pool based on tag scores.

[0192] In this embodiment of the invention, the label values ​​of each factual label and each public opinion risk label are sorted according to a preset sorting rule (e.g., from largest to smallest), and corresponding weights are assigned to each label based on the sorting results. The weight of each factual label is multiplied by its corresponding label value, and the weight of each public opinion risk label is multiplied by its corresponding label value. The results of these multiplications are then weighted and summed to obtain the comprehensive label score for each enterprise under the corresponding label system.

[0193] After obtaining the tag scores of all enterprises, the enterprises can be sorted in descending order of tag scores, and enterprises that meet the preset threshold or quantity conditions can be selected based on the sorting results to generate a target customer pool.

[0194] According to the above embodiments, the labeling system based on weight allocation and weighted scoring can achieve a comprehensive quantitative evaluation of multiple dimensions of enterprise characteristics.

[0195] In some embodiments, such as Figure 10 As shown, step 105 includes steps 1001 to 1004.

[0196] Step 1001: Determine the Jaccard similarity between customers based on factual labels.

[0197] Step 1002: Determine the cosine similarity between customers based on the set of public opinion risk labels and feature vectors.

[0198] Step 1003: Generate a recall list based on Jaccard similarity, cosine similarity, and the target customer pool.

[0199] Step 1004: Identify target customers based on the recall list and relationship network diagram.

[0200] In this embodiment of the invention, Jaccard similarity is used to quantitatively evaluate the similarity between customers. Jaccard similarity is suitable for calculating the degree of overlap of binary tags and can effectively measure the similarity between two customers in the factual tag dimension.

[0201] For any two customers in the target customer pool, their factual label sets are denoted as set A and set B, respectively. Their Jaccard similarity can be calculated using the following formula (5).

[0202]

[0203] Where A represents the set of factual tags for one customer (or enterprise), and B represents the set of factual tags for another customer (or enterprise). This indicates the number of tags shared by the two customers. This represents the total number of all different tags for two customers. The Jaccard similarity score ranges from [value missing]. The closer the Jaccard similarity value is to 1, the higher the similarity between two customers on factual labels.

[0204] For example, suppose the fact-type label sets for customer A and customer B are as follows:

[0205] Client A: {Years of Establishment, Stable Management Team, Diversified Revenue Streams, Industry Leader, Ample Cash Flow}

[0206] Client B: {Years of Establishment Tag, Frequent Changes Tag, Diversified Revenue Stream Tag, Industry Leader Tag, High Financial Risk Tag}

[0207] The two clients share the following tags: {Years of Establishment Tag, Diversified Revenue Tag, Industry Leader Tag}, i.e. , .

[0208] According to the Jaccard similarity formula, the similarity between the two customers is as follows:

[0209]

[0210] Cosine similarity is used to calculate the similarity between text tag vectors or feature vectors of customers. Cosine similarity measures the cosine of the angle between two vectors in vector space, reflecting the consistency of vector directions, and is suitable for measuring the similarity of text feature vectors such as public opinion risk tags.

[0211] For any two customers in the target customer pool, their public opinion risk labels are vectorized and denoted as text label vector A and text label vector B, respectively. The cosine similarity can be calculated according to the following formula (6):

[0212]

[0213] Where A represents the text tag vector of one customer (or enterprise), and B represents the text tag vector of another customer (or enterprise). This represents the dot product of text label vector A and text label vector B. This represents the magnitude of the text label vector A. This represents the magnitude of the text label vector B. The range of cosine similarity values ​​is... The closer the value is to 1, the more consistent the vector direction is, and the higher the similarity of the text tags between the two customers.

[0214] For example, suppose the text tag vectors of customer A and customer B are as follows:

[0215] Customer A's text tag vector ;

[0216] Customer B's text tag vector B .

[0217] The text tag similarity between customer A and customer B can be calculated based on the cosine similarity formula mentioned above.

[0218] Text label vector The modulus length is as follows:

[0219]

[0220] The magnitude of the text label vector B is as follows:

[0221]

[0222] The dot product of text label vector A and text label vector B is as follows:

[0223]

[0224] The above text label vector Substituting the magnitude of text label vector A, the magnitude of text label vector B, and the dot product of text label vector A and text label vector B into the above formula (6), the cosine similarity between text label vector A and text label vector B is calculated as follows:

[0225]

[0226] For any two enterprise customers in the target customer pool, their feature vectors are denoted as feature vector A and feature vector B, respectively. The same formula can be used to calculate the cosine similarity to measure the degree of similarity between the enterprises in the dimension of business characteristics.

[0227] The Jaccard similarity and cosine similarity of each customer (or enterprise) in the target customer pool are sorted in descending order to obtain the ranking results. Customers whose ranking results are below a preset threshold are selected as the recall targets to generate a recall list.

[0228] For any customer (or enterprise) in the target customer pool, calculate the Jaccard similarity and cosine similarity between each customer according to the above formulas (5) and (6), and sort the above Jaccard similarity and cosine similarity in descending order to obtain the similarity ranking result between customers (or enterprises).

[0229] After obtaining the ranking results, based on a preset threshold range (such as a similarity ranking threshold or a similarity score threshold), customers in the ranking results that fall within the threshold range are selected and identified as recall targets, thereby generating the corresponding recall list.

[0230] Based on the candidate companies already selected in the recall list, locate the corresponding company nodes in the established relationship network graph, and extract the structural indicators of the company nodes in the relationship network graph, including node centrality, community affiliation, node degree, neighbor node properties, and cross-community connectivity.

[0231] By jointly analyzing the aforementioned network structure characteristics with the similarity scores of candidate companies in the recall list, target enterprise customers with high business value or potential marketing value can be further screened out.

[0232] For example, node centrality can be used as an indicator to assess the influence of candidate companies in the relationship network. If a company in the recall list has a high centrality, a large number of neighboring nodes, or occupies a core position in a key community, it indicates that the company has strong business connections in the industry chain, supply chain structure, or related network, thus giving it priority in inclusion as a target customer. Furthermore, candidate companies can be grouped based on community segmentation results. If the nodes within a candidate company's community are closely connected, and the company has a close connection with core nodes in the community, its priority in being selected as a target customer can be further increased.

[0233] For example, a comprehensive score for each company can be generated by weighting the similarity scores calculated by the recall model based on the relationship edge weights of the candidate companies (such as the amount of funds transferred, transaction frequency, shareholding ratio, etc.). Companies in the recall list are then ranked according to their comprehensive scores, and those with scores exceeding a preset threshold are selected as target customers.

[0234] According to the above embodiments, by simultaneously introducing multi-source label similarity calculation and relational network structure feature analysis in the customer prediction process, high customer identification accuracy and business interpretability are achieved. Jaccard similarity, constructed based on factual labels, can quantify the similarity of enterprises in basic attributes and operational characteristics. Cosine similarity calculation based on text labels and feature vectors can characterize the vectorized proximity of enterprises in terms of public opinion risk and operational behavior characteristics, enabling the recall stage to comprehensively identify candidate enterprises highly similar to the target characteristics from multiple dimensions. By combining structural indicators such as node centrality, community affiliation, and edge weights in the relational network graph, customers with higher business potential can be further identified from aspects such as the enterprise's association structure, supply chain position, and network influence, achieving secondary screening and optimization of the recall results. Through the joint judgment of similarity ranking and network structure features, this invention can effectively improve the accuracy, stability, and completeness of target customer identification.

[0235] This application provides a target customer prediction device based on a relationship network, applied to the aforementioned target customer prediction method based on a relationship network. This target customer prediction device based on a relationship network is based on the same inventive concept as the target customer prediction method based on a relationship network in one embodiment of this application, and the principle of solving the problem is similar. Therefore, the implementation of the target customer prediction device based on a relationship network is the same as the implementation of the target customer prediction method based on a relationship network in one embodiment of this application, and repeated details will not be described again. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0236] like Figure 11 As shown, the target customer prediction device 1100 based on relationship networks includes:

[0237] The data wide table generation module 1101 is used to generate a data wide table based on the acquired in-row data and out-of-row data.

[0238] The public opinion risk acquisition module 1102 is used to obtain public opinion risk labels and their corresponding label values ​​based on the in-line data and out-of-line data through a pre-established public opinion risk classification model.

[0239] The feature vector construction module 1103 is used to construct a set of feature vectors based on the wide data table and preset variable filtering rules.

[0240] The target customer pool construction module 1104 is used to construct a target customer pool based on preset factual tags and the public opinion risk tags and their corresponding tag values.

[0241] The target customer identification module 1105 is used to identify target customers based on the factual tags, the public opinion risk tags, the target customer pool, the feature vector set, and a pre-built relationship network diagram, wherein the relationship network diagram is constructed based on the data wide table.

[0242] According to the above embodiments, this invention effectively solves the problems of incomplete enterprise profiles caused by scattered multi-source data, inconsistent field definitions, and redundant or missing enterprise information by unifying and integrating internal and external data and constructing a wide data table, thereby significantly improving the data foundation quality for customer identification. The public opinion risk tags and their corresponding quantitative tag values ​​generated based on the public opinion risk classification model enable external risk events to be mapped to the enterprise profile system in real time, achieving dynamic representation of the enterprise's external risk exposure and enhancing the responsiveness of target customer prediction to risk changes. By introducing variable screening rules and feature engineering methods to construct a feature vector set, multi-dimensional operating indicators such as enterprise operating scale, financial stability, and behavioral characteristics are structurally and quantitatively processed, improving the interpretability, comparability, and stability of model input characteristics of the enterprise profile. By combining factual tags with public opinion risk tags to construct a target customer pool, an access evaluation mechanism based on a combination of business rules and data-driven approaches is implemented, thereby improving the accuracy and rationality of target customer screening. Furthermore, the enterprise relationship network constructed based on the wide data table can quantitatively characterize key structural features such as equity linkages, supply chain relationships, and financial transactions between enterprises, providing a more comprehensive reflection of an enterprise's position and influence within the overall network environment. By jointly modeling the relationship network structural features with a feature vector system, this invention can comprehensively assess target customers across both customer ontology features and network relationship features, thereby effectively improving the accuracy, stability, and business adaptability of target customer predictions.

[0243] In some embodiments, the data wide table generation module 1101 includes:

[0244] The primary key rule determination submodule is used to determine the primary key field, association type, and primary key matching rule based on the in-row data and the out-of-row data.

[0245] The primary key matching submodule is used to perform primary key matching operations on the in-row data and the out-of-row data according to the primary key field and the primary key matching rules, and obtain the matching result based on the association type.

[0246] The field merging submodule is used to perform field-level merging processing on in-row data and out-of-row data based on the matching results, and generate a wide data table.

[0247] In some embodiments, the apparatus further includes a multi-record processing module, configured to, if the primary key field of the in-row data matches multiple primary key fields of the out-of-row data, select one primary key field from the multiple primary key fields of the out-of-row data as the matching result according to a preset data processing rule, or perform aggregation processing on the primary key fields of the multiple out-of-row data to generate a matching result.

[0248] In some embodiments, the apparatus further includes the relationship network construction module, the relationship network construction module comprising:

[0249] The node set construction submodule is used to extract enterprise information from the data wide table and construct the node set. The enterprise information includes: enterprise name, unified social credit code, enterprise registered address, and legal representative.

[0250] The relationship edge generation submodule is used to extract relationship data corresponding to preset enterprise relationship types from the data wide table and establish relationship edges between nodes. The enterprise relationship types include: shareholder relationships, external investment relationships, supply chain relationships, capital transaction relationships, and group affiliation relationships.

[0251] The relation network generation submodule is used to construct a relation network graph based on the set of nodes and the relation edges.

[0252] In some embodiments, the apparatus further includes a community division module, the community division module comprising:

[0253] The target node determination submodule is used to select a target node in the relational network graph and determine the neighbor nodes of the target node based on the relational edges of the target node.

[0254] The modularity calculation submodule is used to add the target node to the community where each neighbor node is located, and to determine the modularity transformation value of each neighbor node based on the first modularity of the target node in the original community and the second modularity of the target node in the community where each neighbor node is located.

[0255] The modularity sorting submodule is used to sort the modularity change values ​​of each neighboring node according to a preset order and determine the maximum value of the modularity change value.

[0256] The community allocation submodule is used to allocate the target node to the community where the neighbor node corresponding to the maximum value is located if the maximum value is greater than zero.

[0257] The community stability judgment submodule is used to iteratively execute the above steps until the communities of all nodes in the relationship network graph tend to be stable, thus obtaining the updated relationship network graph.

[0258] In some embodiments, the public opinion risk acquisition module 1102 includes:

[0259] The tag generation submodule is used to annotate the public opinion risk data according to preset tag types, generating text tag pairs consisting of the public opinion risk data and corresponding tags. The tag types include sentiment tags and topic category tags.

[0260] The model training submodule is used to input the text label pairs into a pre-established public opinion risk classification model, and to iteratively train the public opinion risk classification model based on preset model parameters to obtain public opinion risk labels and their corresponding label values.

[0261] In some embodiments, the feature vector construction module 1103 includes:

[0262] The feature filtering submodule is used to filter target feature variables from the data wide table based on preset variable filtering rules. These variable filtering rules include: qualitative filtering rules based on expert experience and quantitative filtering rules based on statistical analysis.

[0263] The feature engineering submodule is used to perform feature engineering processing on the target feature variables to obtain a set of feature vectors.

[0264] In some embodiments, the feature engineering process includes bucketing, and the feature engineering submodule includes:

[0265] The discretization interval determination unit is used to determine the discretization interval based on the data distribution characteristics of the target feature variable.

[0266] The discretization mapping unit is used to map the target feature variable to the discretization interval to generate the feature vector.

[0267] In some embodiments, the feature engineering process includes a smoothing process, and the feature engineering submodule specifically includes: performing a smoothing transformation on the target feature variable based on a logarithmic function or a square root function to generate the feature vector.

[0268] In some embodiments, the feature engineering process includes normalization processing, and the feature engineering submodule includes:

[0269] An extreme value determination unit is used to determine the maximum and minimum values ​​of the target feature variable.

[0270] The normalization transformation unit is used to perform a normalization transformation on the target feature variable based on the maximum value and the minimum value of the target feature variable to obtain the feature vector.

[0271] In some embodiments, the target customer pool construction module 1104 includes:

[0272] The tag sorting submodule is used to sort the preset fact tags and the public opinion risk tags according to a second preset order, and to assign weights according to the sorting results.

[0273] The tag score generation submodule is used to perform weighted summation processing on each of the factual tags and the public opinion risk tags and their corresponding tag values ​​to obtain the tag score of each enterprise.

[0274] The customer pool generation submodule is used to construct a target customer pool based on the tag scores.

[0275] In some embodiments, the target customer identification module 1105 includes:

[0276] The Jaccard similarity calculation module is used to determine the Jaccard similarity between customers based on the factual class labels.

[0277] The cosine similarity calculation module is used to determine the cosine similarity between customers based on the public opinion risk label and the feature vector set.

[0278] The recall list generation module is used to generate a recall list based on the Jaccard similarity, the cosine similarity, and the target customer pool.

[0279] The target customer screening module is used to determine target customers based on the recall list and the relationship network diagram.

[0280] According to the above embodiments, by simultaneously introducing multi-source label similarity calculation and relational network structure feature analysis in the customer prediction process, high customer identification accuracy and business interpretability are achieved. Jaccard similarity, constructed based on factual labels, can quantify the similarity of enterprises in basic attributes and operational characteristics. Cosine similarity calculation based on text labels and feature vectors can characterize the vectorized proximity of enterprises in terms of public opinion risk and operational behavior characteristics, enabling the recall stage to comprehensively identify candidate enterprises highly similar to the target characteristics from multiple dimensions. By combining structural indicators such as node centrality, community affiliation, and edge weights in the relational network graph, customers with higher business potential can be further identified from aspects such as the enterprise's association structure, supply chain position, and network influence, achieving secondary screening and optimization of the recall results. Through the joint judgment of similarity ranking and network structure features, this invention can effectively improve the accuracy, stability, and completeness of target customer identification.

[0281] Figure 12 This is a schematic diagram of the physical structure of a computer device provided in an embodiment of the present invention, such as... Figure 12 As shown, the computer device includes a processor 1201, a memory 1202, and a bus 1203.

[0282] The processor 1201 and the memory 1202 communicate with each other via the bus 1203.

[0283] The processor 1201 is used to call program instructions in the memory 1202 to execute the methods provided in the above-described method embodiments.

[0284] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described target customer prediction method based on a relationship network.

[0285] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described target customer prediction method based on a relationship network.

[0286] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0287] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0288] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0289] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0290] In the description of this specification, the references to terms such as "an embodiment," "a specific embodiment," "some embodiments," "for example," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0291] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A target customer prediction method based on relationship networks, characterized in that, include: A wide data table is generated based on the acquired in-row and out-of-row data; Based on the in-line and out-of-line data, a pre-established public opinion risk classification model is used to obtain public opinion risk labels and their corresponding label values. A feature vector set is constructed based on the wide data table and the preset variable filtering rules; A target customer pool is constructed based on preset factual tags and the aforementioned public opinion risk tags and their corresponding tag values; Target customers are determined based on the factual tags, the public opinion risk tags, the target customer pool, the feature vector set, and the pre-built relationship network graph, which is constructed based on the data wide table.

2. The method according to claim 1, characterized in that, The process of generating a wide data table based on the acquired in-row and out-of-row data includes: Based on the in-row data and the out-of-row data, determine the primary key field, association type, and primary key matching rules; Based on the primary key field and the primary key matching rules, perform primary key matching operations on the in-row data and the out-of-row data, and obtain the matching results based on the association type; Based on the matching results, the data within the row and the data outside the row are merged at the field level to generate a wide data table.

3. The method according to claim 2, characterized in that, Also includes: If the primary key field of the in-row data matches multiple primary key fields of the out-of-row data, a primary key field is selected from the multiple primary key fields of the out-of-row data as the matching result according to the preset data processing rules, or the primary key fields of the multiple out-of-row data are aggregated to generate a matching result.

4. The method according to claim 1, characterized in that, Constructing a relationship network graph based on the aforementioned wide data table includes: Enterprise information is extracted from the data wide table to construct a node set; wherein, the enterprise information includes: enterprise name, unified social credit code, enterprise registered address and legal representative; Relationship data corresponding to preset enterprise relationship types are extracted from the data wide table, and relationship edges between nodes are established; the enterprise relationship types include: shareholder relationship, external investment relationship, supply chain relationship, capital transaction relationship and group affiliation relationship; A relational network graph is constructed based on the set of nodes and the relational edges.

5. The method according to claim 1 or 4, characterized in that, Also includes: Select a target node in the relational network graph, and determine the neighboring nodes of the target node based on the relational edges of the target node; The target node is added to the community where each neighbor node is located, and the modularity transformation value of each neighbor node is determined based on the first modularity of the target node in the original community and the second modularity of the target node in the community where each neighbor node is located. Sort the modularity change values ​​of each neighboring node in a preset order and determine the maximum value of the modularity change value; If the maximum value is greater than zero, the target node will be assigned to the community where the neighbor node corresponding to the maximum value is located. The above steps are repeated iteratively until the communities to which all nodes in the relational network graph belong tend to stabilize, resulting in an updated relational network graph.

6. The method according to claim 1, characterized in that, The process of obtaining public opinion risk labels and their corresponding label values ​​based on the in-line and out-of-line data through a pre-established public opinion risk classification model includes: The acquired public opinion risk data is labeled according to preset tag types to generate text tag pairs consisting of the public opinion risk data and corresponding tags; wherein, the tag types include sentiment tags and topic category tags; The text tag pairs are input into a pre-established public opinion risk classification model, and the public opinion risk classification model is iteratively trained based on preset model parameters to obtain public opinion risk tags and their corresponding tag values.

7. The method according to claim 1, characterized in that, The step of constructing a feature vector set based on the data wide table and preset variable filtering rules includes: Target feature variables are selected from the wide data table based on preset variable selection rules; wherein, the variable selection rules include: qualitative selection rules based on expert experience and quantitative selection rules based on statistical analysis; The target feature variables are subjected to feature engineering processing to obtain a set of feature vectors.

8. The method according to claim 7, characterized in that, The feature engineering process includes bucketing, and the feature engineering process performed on the target feature variables to obtain a set of feature vectors includes: Based on the data distribution characteristics of the target feature variables, the discretization interval is determined; The target feature variable is mapped to the discretization interval to generate the feature vector.

9. The method according to claim 7, characterized in that, The feature engineering process includes smoothing processing. The feature engineering process performed on the target feature variables to obtain a set of feature vectors includes: The target feature variable is smoothed by applying a logarithmic function or a square root function to generate the feature vector.

10. The method according to claim 7, characterized in that, The feature engineering process includes normalization processing. The feature engineering process performed on the target feature variables to obtain a set of feature vectors includes: Determine the maximum and minimum values ​​of the target feature variable; The target feature variable is normalized based on its maximum and minimum values ​​to obtain the feature vector.

11. The method according to claim 1, characterized in that, The target customer pool is constructed based on preset factual tags and public opinion risk tags and their corresponding tag values, including: The preset factual tags and the public opinion risk tags are sorted according to a second preset order, and weights are assigned according to the sorting results; The labels are weighted and aggregated according to their respective weights and corresponding label values ​​to obtain the label scores for each enterprise. A target customer pool is constructed based on the aforementioned tag scores.

12. The method according to claim 11, characterized in that, The process of determining target customers based on the factual tags, the public opinion risk tags, the target customer pool, the feature vector set, and the pre-constructed relationship network graph includes: The Jaccard similarity between customers is determined based on the aforementioned factual tags; The cosine similarity between customers is determined based on the aforementioned public opinion risk tags and the aforementioned feature vector set; Based on the Jaccard similarity, the cosine similarity, and the target customer pool, a recall list is generated; Target customers are identified based on the recall list and the relationship network diagram.

13. A target customer prediction device based on a relationship network, characterized in that, include: The wide data table generation module is used to generate a wide data table based on the acquired in-row and out-of-row data. The public opinion risk acquisition module is used to obtain public opinion risk labels and their corresponding label values ​​based on the in-line data and out-of-line data through a pre-established public opinion risk classification model; The feature vector construction module is used to construct a set of feature vectors based on the wide data table and preset variable filtering rules; The target customer pool construction module is used to construct a target customer pool based on preset factual tags and the public opinion risk tags and their corresponding tag values; The target customer identification module is used to identify target customers based on the factual tags, the public opinion risk tags, the target customer pool, the feature vector set, and a pre-built relationship network graph, wherein the relationship network graph is constructed based on the data wide table.

14. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 12.

16. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 12.

Images