Supplier portrait visual display method and system based on knowledge graph

Abstract

The application discloses a supplier portrait visualization display method and system based on a knowledge graph, relates to the technical field of data processing, and comprises the following steps: performing information entropy driven compression on a pre-constructed supplier knowledge graph through a graph information bottleneck algorithm to generate a supplier compressed subgraph; obtaining a supplier feature set by positioning a target node in the supplier compressed subgraph; screening a target portrait template from a portrait template library according to the supplier feature set; inputting the supplier compressed subgraph, the supplier feature set and the target portrait template into a double-channel collaborative embedding architecture to generate a supplier portrait embedding vector; and associating the supplier portrait embedding vector with the target node for visualization display. The application is used to solve the problems of insufficient expression of supplier features, lack of adaptability of portrait templates and insufficient interpretability of visualization results in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, more particularly, the present application relates to a supplier portrait visualization display method and system based on a knowledge graph. BACKGROUND

[0002] With the continuous expansion of the scale of the supply chain network and the increasing diversification of the types of suppliers, enterprises face problems such as complex supplier evaluation dimensions, information update not in time, and insufficient visualization display in the procurement and management process. In order to realize the fine management and risk control of suppliers, the supplier portrait method based on the knowledge graph has gradually become a research hotspot, which can depict the multi-dimensional relationship between suppliers through the graph structure and provide assistance for supplier evaluation and decision-making.

[0003] However, the existing supplier portrait method still has the following shortcomings: on the one hand, in the knowledge graph data processing, most methods only rely on the original full graph for modeling, and cannot effectively distinguish the information amount of different nodes and edges, resulting in a large graph size and too much redundant information, which not only affects the operation efficiency, but also makes the extracted supplier features have large noise and are difficult to accurately represent the characteristics of the target supplier; on the other hand, in the portrait generation and display, the existing methods mostly rely on fixed templates or simple feature splicing, lack of dynamic matching and collaborative embedding mechanism between supplier features and portrait templates, resulting in that the finally generated supplier portrait is disconnected with business needs, has insufficient explainability, and is difficult to truly reflect the comprehensive ability and industry characteristics of the supplier in the visualization level. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a supplier portrait visualization display method and system based on a knowledge graph, which introduces a graph information bottleneck algorithm, a portrait template matching mechanism and a double-channel collaborative embedding architecture to solve the problems of insufficient expression of supplier features, lack of adaptability of portrait templates and insufficient explainability of visualization results.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0006] The supplier portrait visualization display method based on the knowledge graph comprises the following steps: performing information entropy driven compression on the pre-constructed supplier knowledge graph through a graph information bottleneck algorithm to generate a supplier compressed subgraph; obtaining a supplier feature set by locating a target node in the supplier compressed subgraph; screening a target portrait template from a portrait template library according to the supplier feature set; inputting the supplier compressed subgraph, the supplier feature set and the target portrait template into a double-channel collaborative embedding architecture to generate a supplier portrait embedding vector; and associating the supplier portrait embedding vector with the target node for visualization display.

[0007] In a preferred embodiment, the information entropy driven compression of the pre-constructed supplier knowledge graph by the graph information bottleneck algorithm generates a supplier compressed subgraph, specifically: obtaining the attribute field and structural connection degree of all nodes of the supplier knowledge graph; for each node, calculating the attribute information entropy value according to the probability distribution of the attribute field value category, and calculating the structural information entropy value according to the structural connection degree; combining the attribute information entropy value and the structural information entropy value by weighting to obtain the comprehensive information entropy value of the node; determining a screening threshold according to the statistical distribution of the comprehensive information entropy values of all nodes; screening nodes with a comprehensive information entropy value greater than or equal to the screening threshold to obtain a high information content node set; and performing an edge selection strategy on the supplier knowledge graph based on the high information content node set to obtain the supplier compressed subgraph.

[0008] In a preferred embodiment, the edge selection strategy specifically includes: screening edges with both end nodes belonging to the high information content node set from the edge set of the supplier knowledge graph to obtain a reserved edge set; and merging multiple edges in the reserved edge set that meet a pre-set similarity threshold or type consistency condition into a single weighted edge, wherein if the edges to be merged are non-weighted edges, a default weight value is assigned, and the weights of the edges to be merged are added or weighted averaged to obtain the weight value of the single weighted edge.

[0009] In a preferred embodiment, the obtaining of the supplier feature set by locating the target node in the supplier compressed subgraph specifically includes: locating the target node in the supplier compressed subgraph and retrieving the neighborhood nodes of the target node; obtaining the structural feature vectors of the target node and the neighborhood nodes based on the topological structure of the supplier compressed subgraph, and aggregating the structural feature vectors of the target node and the neighborhood nodes to obtain a structural feature vector set; extracting the behavior data fields corresponding to the target node and the neighborhood nodes from the supplier data, and performing weighted aggregation on the behavior data fields to generate a behavior feature vector; and performing feature alignment and fusion of the structural feature vector set and the behavior feature vector in a pre-set field index order, and combining with the node classification label of the target node to generate the supplier feature set.

[0010] In a preferred embodiment, the screening of the target profiling template from the profiling template library according to the supplier feature set specifically includes: loading a plurality of candidate profiling templates from the profiling template library, and calculating the structural similarity, behavior similarity and label similarity between the supplier feature set and each candidate profiling template; performing weighted summation on the structural similarity, behavior similarity and label similarity between the supplier feature set and each candidate profiling template to obtain a multi-modal similarity score of each candidate profiling template; selecting the candidate profiling template with the highest multi-modal similarity score as the target profiling template; and performing cropping and field processing on the target profiling template to generate an optimized target profiling template.

[0011] In a preferred embodiment, the target portrait template is cropped and field processed, specifically: based on the node classification label of the target node, fields in the target portrait template irrelevant to the industry to which the target node belongs are deleted; fields present in the supplier feature set but missing in the target portrait template are added to the target portrait template; fields present in the target portrait template but missing in the supplier feature set are deleted from the target portrait template; and the arrangement order of the fields in the target portrait template is adjusted according to a preset field priority.

[0012] In a preferred embodiment, the dual-channel collaborative embedding architecture comprises a structural feature processing channel, a behavioral feature processing channel, and a collaborative fusion unit.

[0013] In a preferred embodiment, the supplier compressed subgraph, the supplier feature set, and the target portrait template are input into the dual-channel collaborative embedding architecture to generate a supplier portrait embedding vector, specifically: in the structural feature processing channel, an adjacency relationship is constructed based on the supplier compressed subgraph, and graph convolution operation is performed on the target node and the neighborhood nodes to obtain a structural semantic vector; in the behavioral feature processing channel, a multi-layer perceptron model is constructed for the behavioral feature vector in the supplier feature set to obtain a behavioral semantic vector; and in the collaborative fusion unit, the structural semantic vector and the behavioral semantic vector are respectively mapped to the template field space based on the field order and the field type of the target portrait template, and are aligned and weightedly fused to generate the supplier portrait embedding vector.

[0014] In a preferred embodiment, the supplier portrait embedding vector is associated with the target node, specifically: the supplier portrait embedding vector is subjected to normalization processing and is mapped to a two-dimensional or three-dimensional coordinate space through a dimension reduction projection algorithm to obtain a set of mapping coordinates; and the set of mapping coordinates is bound to the target node to generate a corresponding visualized node in a knowledge graph visualization interface.

[0015] The supplier portrait visualization display system based on a knowledge graph comprises: a graph compression module configured to perform information entropy-driven compression on a pre-constructed supplier knowledge graph through a graph information bottleneck algorithm to generate a supplier compressed subgraph; a supplier feature construction module configured to obtain a supplier feature set by locating a target node in the supplier compressed subgraph; a portrait template selection module configured to select a target portrait template from a portrait template library according to the supplier feature set; a dual-channel collaborative embedding module configured to input the supplier compressed subgraph, the supplier feature set, and the target portrait template into a dual-channel collaborative embedding architecture to generate a supplier portrait embedding vector; and a visualization module configured to associate the supplier portrait embedding vector with the target node and perform visualization display.

[0016] The technical effects and advantages of the supplier portrait visualization display method and system based on a knowledge graph are as follows:

[0017] The embodiment introduces a graph information bottleneck algorithm to perform information entropy driven compression on the pre-constructed supplier knowledge graph, which can effectively filter low-value redundant node and edge structures, highlight key suppliers and their associated elements, reduce feature noise interference on the profiling results, and thus improve the accuracy of feature extraction and the stability of profiling modeling; by combining the dual-channel collaborative embedding architecture of structural features and behavioral features in the supplier compressed subgraph, and introducing the field order and type constraints of the target profiling template for collaborative alignment and weighted fusion, the structural relationship and business behavior characteristics of the supplier can be accurately described, and the integrity and explainability of the profiling results can be improved; by introducing a multi-modal similarity scoring mechanism in the template selection stage, the structural, behavioral and label information are integrated for evaluation, which can ensure high matching degree between the selected template and the target supplier, and realize dynamic adaptation of the template and the actual features in the cutting and field processing link, avoiding irrelevant information interference. Thus, the problems of insufficient feature expression of suppliers, lack of adaptability of profiling templates and insufficient explainability of visualized results in the prior art are effectively solved. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A knowledge graph based supplier profiling visual display method is provided for the embodiment of the present application.

[0019] Figure 2 A structure diagram of a dual-channel collaborative embedding architecture of a knowledge graph based supplier profiling visual display method is provided for the embodiment of the present application.

[0020] Figure 3 A system structure diagram of a knowledge graph based supplier profiling visual display method is provided for the embodiment of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0022] Embodiment 1, Figure 1 A knowledge graph based supplier profiling visual display method is provided, including the following steps:

[0023] S1, performing information entropy driven compression on the pre-constructed supplier knowledge graph by a graph information bottleneck algorithm to generate a supplier compressed subgraph;

[0024] S2, obtain a supplier feature set by locating a target node in a supplier compressed subgraph;

[0025] S3, screen a target portrait template from a portrait template library according to the supplier feature set;

[0026] S4, input the supplier compressed subgraph, the supplier feature set and the target portrait template into a double-channel collaborative embedding architecture to generate a supplier portrait embedding vector;

[0027] S5, associate the supplier portrait embedding vector with the target node and perform visual display.

[0028] The embodiment can effectively filter low-value redundant nodes and edge structures, highlight key suppliers and their associated elements, reduce feature noise interference on the portrait result, thereby improving the accuracy of feature extraction and the stability of portrait modeling, by introducing a graph information bottleneck algorithm to perform information entropy-driven compression on a pre-constructed supplier knowledge graph; the double-channel collaborative embedding architecture combining structural features and behavioral features in the supplier compressed subgraph, and the introduction of the field order and type constraints of the target portrait template for collaborative alignment and weighted fusion can accurately depict the structural relationship and business behavior characteristics of the supplier, and improve the completeness and interpretability of the portrait result; the introduction of a multi-modal similarity scoring mechanism in the template selection stage can integrate structural, behavioral and label information for evaluation, ensuring high matching degree between the selected template and the target supplier, and realizing dynamic adaptation of the template to the actual features in the trimming and field processing link, avoiding irrelevant information interference. Thus, the problems of insufficient expression of supplier features, lack of adaptability of portrait templates and insufficient interpretability of visual results in the prior art are effectively solved.

[0029] S1, performing information entropy-driven compression on a pre-constructed supplier knowledge graph by a graph information bottleneck algorithm to generate a supplier compressed subgraph.

[0030] In the embodiment, the information entropy-driven compression on the pre-constructed supplier knowledge graph by the graph information bottleneck algorithm to generate the supplier compressed subgraph is specifically:

[0031] Obtaining attribute fields and structural connection degrees of all nodes of the supplier knowledge graph;

[0032] For each node, calculating an attribute information entropy value according to the probability distribution of the attribute field value category, and calculating a structural information entropy value according to the structural connection degree;

[0033] Weighting and combining the attribute information entropy value and the structural information entropy value to obtain a comprehensive information entropy value of the node;

[0034] Determining a screening threshold according to the statistical distribution of the comprehensive information entropy values of all nodes;

[0035] Screening nodes with comprehensive information entropy values greater than or equal to a screening threshold value to obtain a high-information-node set;

[0036] Performing an edge selection strategy on the supplier knowledge graph based on the high-information-node set to obtain a supplier compressed subgraph.

[0037] It should be noted that the graph information bottleneck algorithm is a graph compression method based on information entropy driving, which maximizes the core information related to the target task while minimizing the redundant and noise information while compressing the scale of the graph structure.

[0038] Further, for the pre-constructed supplier knowledge graph, the execution process of the graph information bottleneck algorithm includes: calculating the attribute information entropy value and the structure information entropy value of the node, and obtaining the comprehensive information entropy value by weighted combination; determining the screening threshold value, screening the high-information-node; performing an edge selection strategy on the supplier knowledge graph to obtain a supplier compressed subgraph.

[0039] Further, by introducing the graph information bottleneck algorithm, the pre-constructed supplier knowledge graph is subjected to information entropy driven compression processing, which can effectively retain key information nodes and important semantic relationships in the knowledge graph while reducing redundant data, thereby reducing the data scale while ensuring information integrity, and improving overall processing efficiency; The introduction of the algorithm not only significantly optimizes the quality of the input features, but also provides a dense and pure input environment for subsequent portrait template selection and double-channel collaborative embedding, so that the accuracy of template matching and the expressiveness of embedding vectors are enhanced; On this basis, the visual display module can present clearer industry clustering relationships and supplier portrait differences based on the optimized embedding results, thereby realizing the comprehensive technical effects of noise reduction, quality improvement and enhanced explainability, and forming support and collaborative optimization for the entire supplier portrait generation process.

[0040] It should be noted that the basic structure of the pre-constructed supplier knowledge graph includes nodes, edges and attribute fields, wherein the nodes include suppliers and their related business elements, such as supplier enterprise ontology, industry affiliation, geographic location, main products, historical transaction records, financial indicators, risk events, partners, customer evaluations, etc.

[0041] The edges represent semantic relationships between nodes, such as "supplier-product supply", "supplier-industry affiliation", "supplier-transaction record", "supplier-partner", "supplier-risk event", etc.

[0042] The attribute field is specific descriptive data attached to the node or edge, such as the registered capital, credit rating, operating period, product category, cooperation frequency, default rate, etc.

[0043] It should be noted that the structural connectivity specifically reflects the number of connections between the node and other nodes, connection strength, and connection diversity, etc.

[0044] Further, the structural connectivity can be calculated by one or a combination of the following methods:

[0045] (1) Degree: the number of edges connected to the node;

[0046] (2) Weighted degree: in the case of edges with weights, the weights of the connected edges are added up to represent the connection strength of the node;

[0047] (3) Connection diversity index: if a node is connected to multiple nodes of different categories, its structural connectivity is relatively higher;

[0048] (4) Centrality index: such as betweenness centrality, eigenvector centrality, etc., used to measure the propagation effect or influence of the node in the entire knowledge graph.

[0049] It should be noted that the probability distribution of the value category of the attribute field refers to the distribution of the value category of the attribute field of the node in the supplier knowledge graph, and the probability distribution is obtained by frequency statistics or normalization. For example, the attribute fields include industry, credit rating, region, and order status. The frequency of each value category of the industry in the supplier knowledge graph is counted: manufacturing industry = 30 times, financial industry = 10 times, service industry = 20 times. The frequency of each category is divided by the total frequency to obtain the probability distribution of the value category of the industry: manufacturing industry 0.5, financial industry 0.167, service industry 0.333.

[0050] It should be noted that the information entropy value of the attribute is calculated as follows:

[0051] Obtain the probability distribution value of the node in the actual value category of the attribute field (for example, if a node belongs to the manufacturing industry, the probability distribution of the manufacturing industry is 0.5);

[0052] Substitute the probability distribution value of the node in the actual value category of the attribute field into the information entropy formula:

[0053]

[0054] Wherein, is the information entropy value of the node in the current attribute field, is the probability distribution value of the node in the actual value category of the attribute field (for example, if a node belongs to the manufacturing industry, the information entropy value of the industry of the node = -log(0.5));

[0055] Calculate the information entropy value of each attribute field of the node, and perform a weighted sum to obtain the attribute information entropy value of the node.

[0056] It should be noted that the calculation of the structural information entropy value based on structural connectivity is specifically as follows:

[0057] Normalize the distribution of the structural connectivity of a node among all nodes to obtain the structural probability value of that node:

[0058]

[0059] in, For nodes The structural probability value, For nodes Structural connectivity, It is a set of nodes in the supplier knowledge graph. For nodes Structural connectivity, The sum of the structural connectivity of all nodes in the supplier knowledge graph;

[0060] Calculate the structural information entropy value of this node based on the structural probability value:

[0061]

[0062] in, For nodes The structural information entropy value, For nodes The structural probability value.

[0063] It should be noted that the weighted combination of attribute information entropy values ​​and structural information entropy values ​​to obtain the comprehensive information entropy value of a node is as follows:

[0064] Based on attribute information entropy and structural information entropy, define the dynamic weight parameters of the nodes:

[0065]

[0066] in, For nodes The dynamic weight parameters, and , For nodes The structural information entropy value, For nodes The entropy value of the attribute information; when the node's attribute information is richer, The tendency is to increase the value of attribute information, thereby increasing its weight in the overall entropy value; conversely, when the structural information of a node is more distinctive, It tends to decrease, thereby highlighting the contribution of structural information;

[0067] The attribute information entropy value and the structure information entropy value are combined by weighting to obtain a comprehensive information entropy value of the node:

[0068]

[0069] wherein, the comprehensive information entropy value of the node, is a dynamic weight parameter of the node, is a structure information entropy value of the node, is an attribute information entropy value of the node. It should be noted that the screening threshold is determined according to the statistical distribution of the comprehensive information entropy values of all nodes, and specifically: The comprehensive information entropy values of all nodes are statistically analyzed to obtain distribution characteristics such as mean, standard deviation, quantile, etc., and the screening threshold is determined in combination with the distribution characteristics.

[0070] Further, the determination of the screening threshold can be specifically:

[0071] (1) Dynamic threshold based on mean and standard deviation:

[0072]

[0073] wherein,

[0074] the screening threshold, is the mean, is the standard deviation,

[0075] is an adjustment parameter (such as 0.5-1.0), used to control the strictness of the screening threshold. (2) Proportional threshold based on quantile: the upper quartile is directly taken as the screening threshold. (3) Adaptive hybrid strategy: the mean+standard deviation and quantile methods are combined to dynamically select a better threshold.

[0076] In this embodiment, the edge selection strategy is specifically:

[0077] From the edge set of the supplier knowledge graph, edges with both ends belonging to the high information node set are screened to obtain a reserved edge set.

[0078]

[0079]

[0080] ​​​​For the multiple edges in the reserved edge set that meet the preset similarity threshold or the type consistency condition, a single weighted edge is merged, wherein if the edges to be merged are non-weighted edges, a default weight value is given, and the weights of the edges to be merged are accumulated or weighted averaged to obtain the weight value of the single weighted edge.

[0081] It should be noted that the preset similarity threshold means that when the similarity between the attribute features of the multiple edges is greater than or equal to the preset threshold, it is considered that these edges have semantic redundancy and have merging potential, wherein the similarity can be calculated by cosine similarity, Jaccard coefficient or Euclidean distance.

[0082] It should be noted that the type consistency condition means that when the semantic types of the multiple edges are the same (for example, all are "cooperation relationship", "transaction relationship" or "supply relationship"), it is considered that the types are consistent, and further checking the time sequence, data source or credibility is performed to determine whether to merge.

[0083] It should be noted that the single weighted edge is specifically:

[0084] If the edges to be merged are non-weighted edges, a default weight value is given;

[0085] When all the edges to be merged have weight values, the following methods are adopted according to application requirements:

[0086] Cumulative method: suitable for transaction frequency, interaction times and other scenarios, the weight values of all the edges to be merged are directly added to obtain the weight value of the single weighted edge;

[0087] Weighted average method: suitable for evaluation classification or probability type index, the weight values of all the edges to be merged are weighted averaged according to the source credibility to obtain the weight value of the single weighted edge; wherein the source credibility is the credibility weight of the data source of the edge, for example, the official database source credibility = 0.9, the third-party industry data source credibility = 0.7.

[0088] S2, obtaining a supplier feature set by locating a target node in the supplier compression subgraph.

[0089] In this embodiment, the obtaining of the supplier feature set by locating the target node in the supplier compression subgraph is specifically:

[0090] Locating the target node in the supplier compression subgraph, retrieving the neighbor nodes of the target node;

[0091] Obtaining the structural feature vectors of the target node and the neighbor nodes based on the topological structure of the supplier compression subgraph, and aggregating the structural feature vectors of the target node and the neighbor nodes to obtain a structural feature vector set;

[0092] extracting behavior data fields corresponding to the target node and the neighbor nodes from the supplier data, performing weighted aggregation on the behavior data fields, and generating a behavior feature vector;

[0093] aligning and fusing the structure feature vector set and the behavior feature vector according to a preset field index order, combining with a node classification label of the target node, and generating a supplier feature set.

[0094] It should be noted that the target node refers to the node corresponding to the target supplier to be profiled and visualized. The user can locate the target node in the supplier compressed subgraph by inputting the unique identifier of the target supplier.

[0095] It should be noted that the neighbor node mainly refers to the first-order adjacent node of the target node, and can also be extended to the second-order or multi-order adjacent node.

[0096] It should be noted that the structure feature vector set is specifically:

[0097] Based on the topology structure of the supplier compressed subgraph, an adjacency matrix is established.

[0098] For the target node and the neighbor nodes, basic structure attributes are extracted as initial structure feature vectors, and the basic structure attributes include node degree, weighted degree, local clustering coefficient, node centrality index in the graph, and node type code.

[0099] The initial structure feature vectors of the target node and the neighbor nodes are input into the adjacency matrix for feature propagation. During the propagation process, the features of the neighbor nodes are transmitted to the target node after being attenuated or amplified according to the edge weight.

[0100] The aggregated operation is performed on the neighbor feature vectors received by the target node, and the aggregation method can be weighted average, max pooling or attention weighted summation.

[0101] Finally, the structure feature vectors of the target node and its neighbor nodes are combined to form a structure feature vector set.

[0102] It should be noted that the behavior data field refers to data reflecting the actual dynamic performance and interactive behavior of the supplier, including transaction data, contract and cooperation records, credit and compliance behavior, dynamic interactive behavior, and industry-related dynamic indicators.

[0103] It should be noted that the structure feature vector set and the behavior feature vector are aligned and fused according to the preset field index order, and combined with the node classification label of the target node to generate a supplier feature set, which is specifically:

[0104] First, the field index order is predefined, which can be set according to the field arrangement standard in the supplier portrait template library or industry specification, for example, the field index order is in turn: basic attribute field, transaction behavior field, financial risk field, and performance capability field.

[0105] Each component in the structural feature vector set corresponding to the target node and the neighbor node is aligned with the corresponding component in the behavior feature vector according to the preset field index order, so as to ensure that the features from different sources are consistent in semantics and position.

[0106] After alignment, the structural features and the behavior features under the same index are fused, and the fusion mode can be vector splicing, weighted summation or linear combination based on correlation coefficient, thereby generating a comprehensive feature vector corresponding to the field index order.

[0107] Finally, the above comprehensive feature vector and the node classification label of the target node are combined to form a supplier feature set.

[0108] S3, according to the supplier feature set, screening the target portrait template from the portrait template library.

[0109] In the embodiment, the target portrait template is selected from the portrait template library according to the supplier feature set, specifically:

[0110] Load several candidate portrait templates from the portrait template library, calculate the structural similarity, behavior similarity and label similarity between the supplier feature set and each candidate portrait template;

[0111] Weighted sum the structural similarity, behavior similarity and label similarity between the supplier feature set and each candidate portrait template to obtain a multi-modal similarity score of each candidate portrait template;

[0112] Select the candidate portrait template with the highest multi-modal similarity score as the target portrait template;

[0113] Crop and process the fields of the target portrait template to generate an optimized target portrait template.

[0114] It should be noted that the portrait template library has the following sources: first, based on industry expert experience and business rules, manually design standard portrait templates suitable for different industry scenarios, such as supplier portrait templates for manufacturing, logistics or financial services; second, automatically mine representative field combinations and feature patterns through clustering analysis and statistical modeling of historical supplier data, and solidify them as candidate portrait templates; third, support enterprise users to extend and customize portrait templates according to business needs in actual use, and the newly added templates can also be included in the portrait template library.

[0115] It should be noted that the candidate portrait template is composed of a plurality of portrait fields, each portrait field including three elements of field name, field type and field weight, wherein the field name is used to identify the business meaning corresponding to the field, such as “enterprise basic information”, “financial status”, “contract performance situation”, “logistics performance efficiency”, “industry classification label” and the like; the field type is used to distinguish the data source or characteristic category carried by the field, such as structure characteristic field, behavior characteristic field, industry classification label; and the field weight is used to represent the importance of the field in the template, and the system can perform differential processing based on the field weight when calculating the multi-modal similarity.

[0116] It should be noted that the structure similarity, behavior similarity and label similarity between the calculated supplier feature set and each candidate portrait template are as follows:

[0117] The template structure vector set, the template behavior vector set and the industry classification label of each candidate portrait template are extracted by NLP;

[0118] For each candidate portrait template, the vector similarity measure, such as cosine similarity, is used to calculate the structure similarity between the template structure vector set and the structure characteristic vector set, and to calculate the behavior similarity between the template behavior vector set and the behavior characteristic vector set;

[0119] For the label similarity, if the node classification similarity and the industry classification label are the same, the label similarity is set to 1, and if they are different, the label similarity can be calculated based on the industry hierarchical tree:

[0120]

[0121] wherein, is the label similarity, is the node classification label, is the industry classification label, is and the depth of the common ancestor node on the industry hierarchical tree, for example, = “manufacturing > electronic components > semiconductors”, = “manufacturing > electronic components > circuit boards”, and the common ancestor is “manufacturing > electronic components”, is 2; is the depth on the industry hierarchical tree, is the depth on the industry hierarchical tree.

[0122] It should be noted that the weighted sum, the weighted coefficients of the structural similarity, the behavior similarity and the label similarity can be obtained by expert experience setting, learning optimization based on historical data or dynamic self-adaptive allocation.

[0123] In this embodiment, the target portrait template is cropped and field processed, specifically:

[0124] Based on the node classification label of the target node, the fields in the target portrait template that are irrelevant to the industry to which the target node belongs are deleted;

[0125] Fields that exist in the supplier feature set but are missing in the target portrait template are added to the target portrait template;

[0126] Fields that exist in the target portrait template but are missing in the supplier feature set are deleted from the target portrait template;

[0127] According to the preset field priority, the arrangement order of the fields in the target portrait template is adjusted.

[0128] It should be noted that the preset field priority can be set based on industry standards or general specifications, or can be set in combination with the needs of specific business scenarios. In addition, the preset field priority can also be manually configured by experts during system initialization, or the priority ranking can be dynamically generated by statistical analysis of historical usage data, for example, the higher the query frequency of a certain field or the frequency of use in portrait display, the higher the priority of the field is automatically promoted.

[0129] S4, inputting the supplier compressed subgraph, the supplier feature set and the target portrait template into a double-channel collaborative embedding architecture to generate a supplier portrait embedding vector.

[0130] In this embodiment, the double-channel collaborative embedding architecture includes a structural feature processing channel, a behavior feature processing channel and a collaborative fusion unit.

[0131] Figure 2 A structural diagram of the double-channel collaborative embedding architecture of the supplier portrait visualization display method based on the knowledge graph is given.

[0132] It should be noted that the two parallel channels in the double-channel collaborative embedding architecture extract the relationship structure information and the business behavior information of the supplier respectively, so that the feature expression is more targeted and complete. The target portrait template is introduced as a fusion reference standard to ensure that the features from different sources can maintain consistency with the portrait display requirements during fusion. The template field space constraint is introduced in the fusion stage to ensure that the embedding result can directly correspond to the portrait display field, thereby shortening the semantic distance from the embedding vector to the visual display.

[0133] In the embodiment, the supplier compressed subgraph, the supplier feature set and the target portrait template are input into the double-channel collaborative embedding architecture to generate a supplier portrait embedding vector, specifically:

[0134] In the structural feature processing channel, an adjacency relationship is constructed based on the supplier compressed subgraph, and a graph convolution operation is performed on the target node and the neighbor nodes to obtain a structural semantic vector.

[0135] In the behavior feature processing channel, a multi-layer perceptron modeling is performed on the behavior feature vector in the supplier feature set to obtain a behavior semantic vector.

[0136] In the fusion collaborative unit, the structural semantic vector and the behavior semantic vector are respectively mapped to the template field space based on the field order and the field type of the target portrait template, and are aligned and weighted fused to generate a supplier portrait embedding vector.

[0137] S5, the supplier portrait embedding vector is associated with the target node, and visualized display is performed.

[0138] In the embodiment, the supplier portrait embedding vector is associated with the target node, specifically:

[0139] The supplier portrait embedding vector is normalized and mapped to a two-dimensional or three-dimensional coordinate space through a dimension reduction projection algorithm to obtain a mapping coordinate set.

[0140] The mapping coordinate set is bound with the target node to generate a corresponding visual node in the knowledge graph visualization interface.

[0141] It should be noted that the dimension reduction projection algorithm includes but is not limited to a principal component analysis (PCA) algorithm, a t-distributed stochastic neighbor embedding (t-SNE) algorithm, and a uniform manifold approximation and projection (UMAP) algorithm.

[0142] Embodiment 2, Figure 3 The system of the supplier portrait visualization display method based on the knowledge graph is given, which comprises:

[0143] A graph compression module is configured to perform information entropy driven compression on a pre-constructed supplier knowledge graph through a graph information bottleneck algorithm to generate a supplier compressed subgraph.

[0144] A supplier feature construction module is configured to obtain a supplier feature set by locating a target node in the supplier compressed subgraph.

[0145] A portrait template selection module is configured to select a target portrait template from a portrait template library according to the supplier feature set.

[0146] The double-channel collaborative embedding module is configured to input the supplier compressed subgraph, the supplier feature set and the target portrait template into a double-channel collaborative embedding architecture to generate a supplier portrait embedding vector.

[0147] The visualization module is configured to associate the supplier portrait embedding vector with the target node and perform visual display.

[0148] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the nearest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0149] The above embodiments can be realized wholly or partially by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product.

[0150] Those skilled in the art can realize that the modules and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solutions. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0151] In addition, the functional modules in each embodiment of the present application can be integrated in one processing module, or each module can exist physically, or two or more modules can be integrated in one module.

[0152] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0153] Finally, the above is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for visualizing and displaying a supplier portrait based on a knowledge graph, characterized in that, The method comprises the following steps: information entropy driven compression is performed on the pre-constructed supplier knowledge graph through a graph information bottleneck algorithm to generate a supplier compressed subgraph; a supplier feature set is obtained by locating a target node in the supplier compressed subgraph; a target profiling template is selected from a profiling template library according to the supplier feature set; the supplier compressed subgraph, the supplier feature set and the target profiling template are input into a double-channel collaborative embedding architecture to generate a supplier profiling embedding vector; the supplier profiling embedding vector is associated with the target node for visual display; the double-channel collaborative embedding architecture comprises a structural feature processing channel, a behavior feature processing channel and a collaborative fusion unit; in the structural feature processing channel, graph convolution operation is performed on the target node and the neighboring nodes based on the adjacency relationship of the supplier compressed subgraph to obtain a structural semantic vector; in the behavior feature processing channel, a multi-layer perceptron model is established for the behavior feature vector in the supplier feature set to obtain a behavior semantic vector; in the collaborative fusion unit, the structural semantic vector and the behavior semantic vector are respectively mapped to the template field space based on the field order and the field type of the target profiling template, and are aligned and weightedly fused to generate the supplier profiling embedding vector. 2.The knowledge graph-based supplier portrait visualization display method according to claim 1, characterized in that, The information entropy driven compression is performed on the pre-constructed supplier knowledge graph through the graph information bottleneck algorithm to generate the supplier compressed subgraph, and specifically comprises the following steps: attribute fields and structural connection degrees of all nodes of the supplier knowledge graph are obtained; for each node, attribute information entropy is calculated according to the probability distribution of the attribute field value category, and structural information entropy is calculated according to the structural connection degree; the attribute information entropy and the structural information entropy are combined by weighting to obtain a comprehensive information entropy value of the node; a screening threshold is determined according to the statistical distribution of the comprehensive information entropy values of all nodes; nodes with a comprehensive information entropy value greater than or equal to the screening threshold are screened to obtain a high information amount node set; an edge selection strategy is performed on the supplier knowledge graph based on the high information amount node set to obtain the supplier compressed subgraph. 3.The knowledge graph-based supplier portrait visualization display method according to claim 2, characterized in that, The edge selection strategy specifically comprises the following steps: from the edge set of the supplier knowledge graph, edges with both end nodes belonging to the high information amount node set are screened to obtain a reserved edge set; if the edges to be merged are non-weighted edges, a default weight value is assigned, and the weights of the edges to be merged are added or weightedly averaged to obtain the weight value of the single weighted edge. 4.The knowledge graph-based supplier portrait visualization display method according to claim 3, characterized in that, The supplier feature set is obtained by locating the target node in the supplier compressed subgraph, and specifically comprises the following steps: the target node is located in the supplier compressed subgraph, and the neighboring nodes of the target node are retrieved; structural feature vectors of the target node and the neighboring nodes are obtained based on the topological structure of the supplier compressed subgraph, and the structural feature vectors of the target node and the neighboring nodes are aggregated to obtain a structural feature vector set; behavior data fields corresponding to the target node and the neighboring nodes are extracted from the supplier data, and the behavior data fields are weightedly aggregated to generate a behavior feature vector; According to the preset field index order, the structural feature vector set and the behavior feature vector are aligned and fused, and are combined with the node classification label of the target node to generate a supplier feature set. 5.The knowledge graph-based supplier portrait visualization display method according to claim 4, characterized in that, The target portrait template is selected from a portrait template library according to the supplier feature set, and specifically comprises: A plurality of candidate portrait templates are loaded from the portrait template library, and structural similarity, behavior similarity and label similarity between the supplier feature set and each candidate portrait template are calculated. The structural similarity, behavior similarity and label similarity between the supplier feature set and each candidate portrait template are weighted and summed to obtain a multi-modal similarity score of each candidate portrait template. The candidate portrait template with the highest multi-modal similarity score is selected as the target portrait template. The target portrait template is cropped and field-processed to generate an optimized target portrait template. 6.The knowledge graph-based supplier portrait visualization display method according to claim 5, characterized in that, The target portrait template is cropped and field-processed, and specifically comprises: Based on the node classification label of the target node, fields in the target portrait template that are irrelevant to the industry to which the target node belongs are deleted. Fields that exist in the supplier feature set but are missing in the target portrait template are added to the target portrait template. Fields that exist in the target portrait template but are missing in the supplier feature set are deleted from the target portrait template. According to a preset field priority, the arrangement order of the fields in the target portrait template is adjusted. 7.The knowledge graph-based supplier portrait visualization display method according to claim 6, characterized in that, The supplier portrait embedding vector is associated with the target node, and specifically comprises: Normalization processing is performed on the supplier portrait embedding vector, and a dimension reduction projection algorithm is used to map it to a two-dimensional or three-dimensional coordinate space to obtain a set of mapping coordinates. The set of mapping coordinates is bound to the target node to generate a corresponding visual node in the knowledge graph visualization interface.

8. A system using the knowledge graph-based supplier portrait visualization display method according to any one of claims 1-7, comprising: a graph compression module configured to perform information entropy-driven compression on a pre-constructed supplier knowledge graph by a graph information bottleneck algorithm to generate a supplier compressed subgraph; a supplier feature construction module configured to obtain a supplier feature set by locating a target node in the supplier compressed subgraph; a portrait template selection module configured to select a target portrait template from a portrait template library according to the supplier feature set; a dual-channel collaborative embedding module configured to input the supplier compressed subgraph, the supplier feature set and the target portrait template into a dual-channel collaborative embedding architecture to generate a supplier portrait embedding vector; the dual-channel collaborative embedding architecture comprises a structural feature processing channel, a behavior feature processing channel and a collaborative fusion unit; in the structural feature processing channel, an adjacency relationship is constructed based on the supplier compressed subgraph, and graph convolution operations are performed on the target node and the neighboring nodes to obtain a structural semantic vector; in the behavior feature processing channel, a multi-layer perceptron model is constructed for the behavior feature vector in the supplier feature set to obtain a behavior semantic vector; in the fusion collaborative unit, the structural semantic vector and the behavior semantic vector are respectively mapped to the template field space based on the field order and the field type of the target portrait template, and are aligned and weighted fused to generate the supplier portrait embedding vector. A visualization module is configured to associate the supplier profile embedding vector with the target node for visualization.

Images