Knowledge graph-based intelligent analysis method for digital traffic industry chain

Abstract

The embodiment of the invention discloses an intelligent analysis method of a digital traffic industry chain based on a knowledge graph, and relates to the technical field of information. The method comprises the following steps: collecting enterprise data and ontology data of a digital traffic industry chain; constructing a knowledge graph of the digital traffic industry chain; the text description of the map nodes is enhanced, and special identifiers are enhanced before the BERT model is input; performing feature coding on each node by using a BERT encoder; converting the feature codes into importance scores by using a scoring network; spreading importance scores by using a hierarchical graph neural network, and updating feature representation and importance scores of industrial link nodes at the same time; and calculating the importance score of the industrial link by using the node importance score. According to the method, an effective method is provided for predicting the strong and weak links of the digital traffic industry chain by combining knowledge graph construction, BERT coding, a scoring network and a hierarchical graph neural network.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an intelligent analysis method for the digital transportation industry chain based on knowledge graphs. Background Technology

[0002] Against the backdrop of the rapid development of the digital transportation industry, a deep understanding of the structure, strengths and weaknesses, and development trends of each link in the digital transportation industry chain has become crucial. Traditional industry analysis methods mainly rely on manual labor, which is inefficient and costly. Furthermore, given the rapid development of the digital transportation industry, current industry analysis and survey methods suffer from significant lag and information obsolescence. Therefore, it is necessary to leverage information technology to improve the process of digital transportation industry analysis.

[0003] Knowledge graph technology, as a structured knowledge representation method, helps integrate and express multi-layered and multi-domain information. In the digital transportation industry, knowledge graphs can cover multi-layered information, from enterprise data to ontology data, including key information such as industry segments, enterprise types, and market size. However, current applications of knowledge graphs in digital transportation mainly focus on information presentation and relationship representation, lacking in-depth analytical tools for the development of strong and weak links in the digital transportation industry chain. The accuracy of related information is very low, therefore, in practical applications, the analysis of the strengths and weaknesses of the digital transportation industry chain still relies on manual methods.

[0004] It is evident that the current application of knowledge graph technology in the digital transportation industry is mainly limited to the analysis and summarization of superficial key information such as industry links, enterprise types, and market size. More in-depth analysis processes, such as the analysis of the strengths and weaknesses of the digital transportation industry chain, still require manual work, resulting in low efficiency and significant delays in analysis results. Summary of the Invention

[0005] The embodiments of the present invention provide an intelligent analysis method for the digital transportation industry chain based on knowledge graphs, which can improve the automation level in the analysis process of the strong and weak links of the digital transportation industry chain, thereby improving efficiency and the timeliness of analysis results.

[0006] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0007] A knowledge graph-based intelligent analysis method for the digital transportation industry chain includes:

[0008] S1. Acquire enterprise data and ontology data of the digital transportation industry chain, and construct a knowledge graph of the digital transportation industry chain;

[0009] S2. Extract the text descriptions of nodes from the constructed knowledge graph;

[0010] S3. Input the text description of the node into the BERT encoder and obtain the feature representation of the node;

[0011] S4. Obtain the importance score of the node using the feature representation of the node, and update it through a hierarchical graph neural network;

[0012] S5. Use the updated importance score to determine the strength or weakness of each link in the industrial chain;

[0013] S6. Generate visual information based on the strength of each link in the determined industrial chain and send it to the display device.

[0014] The focus of this embodiment is on obtaining the relative strengths and weaknesses of each link in the industry chain. Visualized information is merely an information carrier used to display the relative strengths and weaknesses of each link in the industry chain. Examples include charts, explanatory text, and topology diagrams. These visualized information can be generated using existing technologies, such as AI. The display devices can be user terminals, such as smartphones and computers, or large display screens in data centers.

[0015] Specifically, S1 includes: extracting operational data of enterprises related to the digital transportation industry chain from the database, and generating enterprise data based on the operational data. The enterprise data is represented as e, where e = {type, name, scale, products_list}. `type` represents the industry category to which enterprise e belongs, corresponding to a specific industry segment in the ontology data of the digital transportation industry chain. `name` represents the name of the enterprise. `scale` represents the market size of the enterprise; the value of `scale` corresponds to at least the enterprise's total annual sales. `products_list` represents a list of all products produced by the enterprise. The ontology data is represented as O = {O node O edge}, O node O represents the set of nodes that make up the industrial links included in the industrial chain. edge This represents the edges between these nodes. The industry links include membership and supply relationships within the industry chain. Specifically, the enterprise data of the digital transportation industry chain is defined as e∈E, where E is the set of all digital transportation enterprises. Each enterprise's data is represented as e={type,name,scale,products_list}. `type` represents the industry category to which enterprise e belongs, corresponding to a specific industry link in the ontology data of the digital transportation industry chain. `name` represents the enterprise's name. `scale` represents the enterprise's market size, such as total sales in a given year. `products_list` represents a list of all products produced by the enterprise. The ontology data of the digital transportation industry chain is defined as O={O node Oedge}, O node O represents the set of nodes comprising all industrial links included in the industrial chain. edge This represents the edges between these nodes, and these industry links include both membership and supply relationships. The ontology data of the digital transportation industry chain was obtained by experts in the field of digital transportation.

[0016] In S1, constructing a knowledge graph of the digital transportation industry chain includes: constructing entities in the knowledge graph based on entity types, where the entity types include: [industry segment, enterprise, product]; and constructing relationships in the knowledge graph based on relationship types, where the relationship types include [belong to, supply, include enterprise, production]; wherein, the "industry segment" node is represented by the node set O in the ontology model O. node The knowledge graph is structured as follows: "Enterprise" nodes are identified by their names in the enterprise data; "Product" nodes are identified by their product lists (products_list) in the enterprise data; "Belong to" and "Supply" relationships are identified by annotations in the ontology model O; "Containing Enterprises" relationships are identified by each enterprise's type data; and "Production" relationships are identified by their names and product lists (products_list) in the enterprise data. The knowledge graph uses description logic for knowledge construction, with a structure similar to triples like <head entity, relation, tail entity>. "Head entity" and "tail entity" are the nodes in the graph, and relations are the edges. Entity types include [industry segment, enterprise, product], and relation types include [belong to, supply, containing enterprise, production]. "Industry segment" nodes are identified by the node set O in the ontology model O. node The "Enterprise" node is determined by the name in the enterprise data. The "Product" node is determined by the products list in the enterprise data. The "Belong To" and "Supply" relationships are determined by expert annotations in the ontology model O. The "Contains Enterprises" relationship is determined by the type data for each enterprise. The "Produces" relationship is determined by the name and products list in the enterprise data.

[0017] Specifically, to facilitate subsequent encoding, [CLS] and [SEP] are also added to the text description in the text concatenation part, where [CLS] and [SEP] represent the beginning and end of the text, respectively. In addition, special identifiers [type] and [\type] are used to represent the start and end identifiers of entity types, and [name] and [\name] are used to represent the start and end of entity names. S2 includes: using the entity type and name of the node in the knowledge graph to obtain the text description of the node, where the text description of the "Industry Link" node is: {[CLS],[type],Industry Link,[\type],[name],Industry Name,[\name],[SEP]};

[0018] The text description of the “Enterprise” node is: {[CLS],[type],Enterprise,[\type],[name],Enterprise Name,[\name],[SEP]};

[0019] The text description of the “Product” node is: {[CLS],[type],Product,[\type],[name],Product Name,[\name],[SEP]}; [CLS] and [SEP] represent the beginning and end of the text, respectively; [type] and [\type] represent the beginning and end identifiers of the entity type; and [name] and [\name] represent the beginning and end of the entity name.

[0020] In the process of acquiring the text description of each node in the knowledge graph of the digital transportation industry chain, start identifiers [CLS] and end identifiers [SEP] are added to the beginning and end of the text, start identifiers [type] and end identifiers [\type] are added to the beginning and end of the entity type, and start identifiers [name] and end identifiers [\name] are added to the beginning and end of the entity name, thus expanding the BERT model dictionary and obtaining the concatenated text.

[0021] {[CLS],[type],Industry Segment,[\type],[name],Industry Name,[\name],[SEP]}.

[0022] Specifically, S3 includes: a BERT encoder that takes the text description of a node as input, and the BERT encoder runs a BERT model to obtain the features of the text.

[0023]

[0024] Where 'a' is the number of tokens for node category name, 'b' is the number of tokens for node name, and 'T' is the number of tokens for node name. [CLS] T represents the classification label in the BERT model.[type] T represents the marker indicating the starting position of a category in the BERT model. [\type] This represents a marker in the BERT model that indicates the end position of a category. The first token represents the node category name. T represents the a-th token of the node category name. [name] T represents a marker in the BERT model indicating the starting position of a node name. [\name] This represents a marker in the BERT model that indicates the end position of a node name. The first token representing the node name. T represents the b-th token of the node name. [SEP] This represents the sentence delimiter in the BERT model. The BERT encoder encodes the feature representation of each node, and the text description of each node is input into the BERT model to obtain the feature representation of the text sentence. The feature representation T uses the special identifier [CLS]. [CLs] To represent the characteristics of a node, we will use [type], [\type],

[0025] Special identifiers such as [name] and [\name] are also encoded. During training, the parameters of these special identifiers in the BERT embedding layer are updated. The text description of each node is input into the BERT model to obtain the feature representation of the text sentence. The feature representation T uses the special identifier [CLS]. [CLS] To represent the feature representation of a node. Taking the "Product" node as an example, its text description feature representation is as follows:

[0026]

[0027] Where 'a' represents the number of tokens for "product" and 'b' represents the number of tokens for "product name". For clarity, the following text will use a non-bold 'T'. 产品 The feature representation of a single product is T. [CLS] Use bold Represents multiple product features T [CLS] The matrix formed.

[0028] In this embodiment, step S4 includes: obtaining the importance score S of a node using the feature representation of the node. 节点 =MLP(T 节点 MLP stands for Fully Connected Network; it updates the representations and corresponding importance scores of nodes in the knowledge graph through information propagation in a hierarchical graph neural network. Taking the "product" node as an example, its textual description feature representation is T. 产品Construct a fully connected neural network to convert text description features into importance scores S. 产品 The conversion. For ease of reference, the S without bold will be used in the following text. 产品 To indicate the score of an individual product, use bold text. This represents a vector composed of scores from multiple products. A scoring network is constructed to transform the textual description features of nodes in the knowledge graph of the digital transportation industry chain into importance scores. A fully connected neural network is used to achieve the transformation from textual description features to importance scores S. Taking a product as an example, the feature representation of the textual description of a product node is T. 产品 Importance score S 产品 The calculation formula is: S 产品 =MLP(T 产品 ), where MLP stands for fully connected network.

[0029] The method of updating the feature representations and corresponding importance scores of nodes in the knowledge graph through information propagation via hierarchical graph neural networks includes: for the "product" level of the graph neural network, updating the feature representations of "product" nodes using graph attention techniques. And importance score Where L represents the number of layers in the graph neural network, and the hierarchical graph neural network is divided according to entity type, including at least three levels: "product" level, "enterprise" level, and "industry link" level, where T 产品,L-1 This represents the feature representation of a single "product" node at layer L-1. S represents the feature representation of all "enterprise" nodes at level L-1. 产品,L-1 This represents the importance score of a single "product" node at level L-1. This represents the importance score of all "enterprise" nodes in layer L-1. For the "enterprise" level of the graph neural network, the feature representation of the "enterprise" nodes is updated using graph attention techniques. And the score of the "enterprise" node Among them, T 企业,L-1 This represents the feature representation of a single "enterprise" node at layer L-1. S represents the feature representation of all "product" nodes at level L-1. 企业,L-1 This represents the importance score of a single "enterprise" node at level L-1. This represents the importance score of all "enterprise" nodes in layer L-1. For the "industry segment" level of the graph neural network, the feature representation of the "industry segment" nodes is updated using graph attention techniques. The importance score of the "industry link" node Among them, T 产业环节,L-1 This represents the feature representation of a single "industry link" node at layer L-1. S represents the feature representation of all "enterprise" nodes at level L-1. 产业环节,L-1 This represents the importance score of a single "industry link" node in layer L-1. This represents the importance score of all "enterprise" nodes at layer L-1. The attention weights in the "product" level attention module consist of the node's feature attention and the enterprise's market size attention; the attention weights in the "enterprise" level attention module consist of the node's feature attention and the node's degree attention; and the attention weights in the "industry segment" level attention module consist of the node's feature attention and the enterprise's market size attention.

[0030] After information is disseminated through hierarchical graph neural networks at the "product" level, "enterprise" level, and "industry link" level, such as... Figure 2 As shown, the importance scores of all industry chain nodes in the graph are updated. The hierarchical graph neural network can be stacked k times, such as... Figure 3 As shown, the importance score of the final "industry link" node is S. 产业环节,k The feature and importance scores of nodes are updated through information propagation in a hierarchical graph neural network. Graph attention technology is used for this update. The attention weights in the product's attention module consist of two parts: the feature attention of the nodes and the market size attention of the enterprise. The feature update calculation formula is as follows:

[0031]

[0032] Q = T 产品,L-1 W Q

[0033]

[0034]

[0035]

[0036]

[0037] Product T, L = Softmax(W) feature ·W scale V

[0038] The formula for updating the score is:

[0039]

[0040]

[0041] Among them W Q W KW V Here, Q, K, and V are the weight matrix, and Q, K, and V are three matrices used as intermediate variables in the attention algorithm. These matrices can be different for different nodes. Softmax is the activation function, d is the feature dimension, MarketScale is the market size function, returning the market size scale for each firm, and W... feature and W scale , where are the feature attention score and the market size attention score, respectively, and L is the number of layers in the graph neural network.

[0042] The attention weights in the enterprise's attention module consist of two parts: the feature attention of the node and the degree attention of the node. The feature update calculation formula is as follows:

[0043]

[0044] Q = T 企业,L-1 W Q

[0045]

[0046]

[0047]

[0048]

[0049] T 企业,L =Softmax(W feature ·W degree V

[0050] The formula for updating the score is:

[0051]

[0052]

[0053] Among them W Q W K W V Here, W is the weight matrix, Softmax is the activation function, d is the feature dimension, degree is the degree function used to calculate the degree of each node, and W is the feature matrix. featire and W degree , where are the feature attention and the degree attention score of the node, respectively, and L is the number of layers in the graph neural network.

[0054] The attention weights in the Attention module of the industry chain consist of two parts: the feature attention of the nodes and the market size attention of the enterprises. The feature update calculation formula is as follows:

[0055]

[0056] Q = T 产业环节,L-1 W Q

[0057]

[0058]

[0059]

[0060]

[0061] T industry segment, L =Softmax(W feature ·W scale V

[0062] The formula for updating the score is:

[0063]

[0064]

[0065] Among them W Q W K W V Here, W is the weight matrix, Softmax is the activation function, d is the feature dimension, MarketScale is the market size function, and it returns the market size scale for each firm. feature and W scale , where are the feature attention score and the market size attention score, respectively, and L is the number of layers in the graph neural network.

[0066] Specifically, in S5, the following are included: the updated importance score is used as the label for the industry link node, wherein the average importance score of the sub-category "industry link" is used as the value of the strength of the importance of the corresponding parent category node; and the links in the industry chain are sorted according to the importance score of the industry link node.

[0067] This involves obtaining the importance score of each node. The importance scores of the nodes after propagation are used to calculate the scores S∈[0,1] for each corresponding enterprise and industry segment. Based on statistical information from the digital transportation industry, the importance score of each industry segment can be obtained, normalized to the range [0,1], and these representative scores S... goldThis serves as the label for each node in the industry chain. Finally, the mean squared error loss function is used to optimize the parameters of the entire model. The importance score of the parent industry chain is calculated because entities in the "industry chain" category that do not have a parent node directly connected to the "enterprise" node cannot have their scores directly calculated. The average score of all connected sub-categories of "industry chains" is used as the strength score of the parent node. Based on the importance score of each industry chain, the strength of the industry chain is predicted according to the importance score. Specifically, the importance score of each node is obtained, and the updated node importance score is used to progressively calculate the industry environment score, S. 产业环节,k The score S∈[0,1]. The mean squared error loss function is used to optimize the parameters of the entire model, and its calculation formula is: loss=(SS gold ) 2 Where S is the score calculated by the model, S gold The score is the actual market size after being normalized to [0,1] by BERT and a scoring network.

[0068] This invention provides an intelligent analysis method for the digital transportation industry chain based on knowledge graphs. The method collects enterprise data and ontology data from the digital transportation industry chain; constructs a knowledge graph of the digital transportation industry chain; enhances the text descriptions of graph nodes and adds special identifiers before inputting them into the BERT model; uses a BERT encoder to encode the features of each node; uses a scoring network to convert the feature codes into importance scores; uses a hierarchical graph neural network to propagate the importance scores, while simultaneously updating the feature representations and importance scores of industry link nodes; and uses the node importance scores to calculate the importance score of the industry link. This method, by combining knowledge graph construction, BERT encoding, scoring networks, and hierarchical graph neural networks, provides an effective automated analysis process for predicting the strength and weakness of links in the digital transportation industry chain, thereby improving the automation level, efficiency, and timeliness of the analysis results. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments 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.

[0070] Figure 1 This is a schematic diagram of the main process provided for an embodiment of the present invention;

[0071] Figure 2 This is a schematic diagram of a hierarchical graph neural network structure provided in an embodiment of the present invention;

[0072] Figure 3 This is a schematic diagram of a hierarchical graph neural network structure with k stacks provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the overall process flow provided in the embodiments of the present invention. Detailed Implementation

[0073] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Embodiments of the present invention will be described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of the present invention means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0074] This embodiment addresses the complexity of the digital transportation industry chain by proposing a knowledge graph-based method for predicting strong and weak links. By constructing a deep knowledge graph of the digital transportation industry chain and integrating enterprise data and ontology information, a comprehensive and in-depth analysis of industry chain relationships is achieved. A BERT encoder is used to extract node features, considering entity importance, thus solving the problem of encoding difficult-to-encode knowledge graphs. A scoring network is implemented to convert node text encoding into importance scores. A hierarchical graph neural network comprehensively propagates the importance scores of each level of the industry chain, deriving the score for each industry link. This method comprehensively considers textual semantics and structural information, achieving information propagation and feature learning for each link of the industry chain. By introducing prior knowledge from professional fields and key factors such as market size, this method can uncover potential relationships and development opportunities in the knowledge graph, thereby providing comprehensive and accurate reference information for decision-makers in the digital transportation industry chain.

[0075] This invention provides an intelligent analysis method for the digital transportation industry chain based on knowledge graphs, such as... Figure 1 , 4 As shown, it includes:

[0076] Given enterprise data and an ontology model of the digital transportation industry chain, a knowledge graph of the digital transportation industry chain is constructed. Based on the text description of each graph node, a BERT encoder is used to encode the features of each node. These features are then transformed by a scoring network to obtain an importance score. The score is then propagated through a hierarchical graph neural network to update the node's importance score. Using the updated node scores, the importance score of each industry segment is calculated and optimized using a mean squared error loss function. Furthermore, for industry segment entities without direct connections to enterprise nodes, their scores are obtained by calculating the average score of their connected sub-industry segment nodes.

[0077] The specific implementation method is described below.

[0078] (1) Data collection

[0079] Acquire enterprise data and ontological data of the digital transportation industry chain.

[0080] The enterprise data in the digital transportation industry chain is defined as e∈E, where E is a set containing all digital transportation enterprises. Each enterprise's data is represented as e={type,name,scale,products_list}. `type` represents the industry category to which enterprise e belongs, corresponding to a specific industry segment in the digital transportation industry chain ontology data. `name` represents the enterprise's name. `scale` represents the enterprise's market size, such as total sales in a given year. `products_list` represents a list of all products produced by the enterprise. The enterprise data in the digital transportation industry chain is extracted from the corresponding enterprise's website introduction. The extracted text is segmented and tokenized using the NLTK toolkit. Named entity annotations are used to extract the enterprise's name (`name`), industry category (`name`), market size (`scale`), and the list of all products (`products_list`) from the text.

[0081] The ontology data of the digital transportation industry chain is defined as O = {O node O edge}, O node O represents the set of nodes comprising all industrial links included in the industrial chain. edge This represents the edges between these nodes, and these industry links include both membership and supply relationships. The ontology data of the digital transportation industry chain was obtained by experts in the field of digital transportation.

[0082] Collect market size and product market share for each link in the digital transportation industry chain. Based on authoritative research and statistical reports in the digital transportation industry, the market size and product market share of each link can be obtained, normalized to the range [0,1], and these scores S... gold As a label for the corresponding node.

[0083] (2) Constructing a knowledge graph

[0084] A knowledge graph for the digital transportation industry chain is constructed based on enterprise data and ontology data from the digital transportation industry chain. The knowledge graph uses description logic for knowledge construction, with the specific structure of the description logic being a triple similar to <head entity, relation, tail entity>. "Head entity" and "tail entity" are nodes in the graph, and relations are edges. Entity types include [industry link, enterprise, product], and relation types include [belong to, supply, include enterprise, production]. The "industry link" node is derived from the node set O in the ontology model O. nodeThe "Enterprise" node is determined by the name in the enterprise data. The "Product" node is determined by the products list in the enterprise data. The "Belong To" and "Supply" relationships are determined by expert annotations in the ontology model O. The "Contains Enterprise" relationship is determined by the type data of each enterprise. The "Produces" relationship is determined by the name and products list in the enterprise data. The relationships between entities are formally defined using description logic, including multiple relationship types. All relationship types and their associated entity types are shown in the following formula:

[0085] {Industry segment, belonging to, industry segment}

[0086] {Industry chain, supply chain, industry chain}

[0087] {Industry chain, including enterprises}

[0088] {Enterprise, Production, Product}

[0089] (3) Extract node text descriptions

[0090] Obtain the text description of each node in the knowledge graph of the digital transportation industry chain. The text description of each node is obtained by concatenating its entity type and name in the knowledge graph. For ease of subsequent encoding, [CLS] and [SEP] are also added to the text description during the text concatenation process, where [CLS] and [SEP] represent the beginning and end of the text, respectively. In addition, the special identifiers [type] and [\type] are used to represent the start and end identifiers of the entity type, and [name] and [\name] are used to represent the start and end identifiers of the entity name.

[0091] The text description of the "Industry Link" node is as follows:

[0092] {[CLS],[type],Industry Segment,[\type],[name],Industry Name,[\name],[SEP]}

[0093] The text description of the "Enterprise" node is as follows:

[0094] {[CLS],[type],Company,[\type],[name],CompanyName,[\name],[SEP]}

[0095] The text description for the "Product" node is as follows:

[0096] {[CLS],[type],Product,[\type],[name],Product Name,[\name],[SEP]}

[0097] For the text description of each node in the knowledge graph of the digital transportation industry chain, add start identifier [CLS] and end identifier [SEP] to the beginning and end of the text, add start identifier [type] and end identifier [\type] to the beginning and end of the entity type, and add start identifier [name] and end identifier [\name] to the beginning and end of the entity name, thus expanding the BERT model dictionary and obtaining the concatenated text.

[0098] (4) Encoding node features

[0099] The BERT encoder encodes the feature representation of each node. The text description of each node is input into the BERT model to obtain the feature representation of the text sentence. The feature representation T uses the special identifier [CLS]. [CLS] To represent the feature representation of a node. Taking the "Product" node as an example, its text description feature representation is as follows:

[0100]

[0101] Its formal definition is:

[0102]

[0103]

[0104] Where 'a' is the number of tokens for 'product' and 'b' is the number of tokens for 'product name'. [CLS] The dimension is 768. For ease of representation, the following text will use a non-bold T. 产品 The feature representation of a single product is T. [CLS] Use bold Represents multiple product features T [CLS] The matrix formed.

[0105] When using the BERT encoder to encode the feature representation of each node, special identifiers such as [type], [\type], [name], and [\name] are also encoded. During training, the parameters of these special identifiers in the BERT embedding layer are updated.

[0106] (5) Construct a scoring network to transform the textual description features of each node in the knowledge graph of the digital transportation industry chain into importance scores. Taking the "product" node as an example, its textual description features are represented as T. 产品 Construct a fully connected neural network to convert text description features into importance scores S. 产品 The conversion. For convenience, S is used. 产品This represents the product's importance score. The calculation formula is:

[0107] S-product = MLP(T-product)

[0108] MLP stands for Fully Connected Network.

[0109] (6) Constructing a hierarchical graph neural network

[0110] Construct a hierarchical graph neural network. Update the feature representations and importance scores of nodes in the industry chain through information propagation within the hierarchical graph neural network, thus aggregating information about relevant enterprises and products. For the "product" level of the graph neural network, the attention weights in the product's attention module consist of two parts: the feature attention of the node and the market size attention of the enterprise. The feature update calculation formula is as follows:

[0111]

[0112] Q = T 产品,L-1 W Q

[0113]

[0114]

[0115]

[0116]

[0117] Product T L =Softmax(W feature ·W scale V

[0118] The formula for updating the score is:

[0119]

[0120]

[0121] Among them W Q W K W V Here, W is the weight matrix, Softmax is the activation function, d is the feature dimension, MarketScale is the market size function, and it returns the market size scale for each firm. feature and W scale , where are the feature attention score and the market size attention score, respectively, and L is the number of layers in the graph neural network.

[0122] For the "enterprise" level of a graph neural network, the attention weights in the enterprise's attention module consist of two parts: the feature attention of the nodes and the degree attention of the nodes. The feature update calculation formula is as follows:

[0123]

[0124] Q = T 企业,L-1 W Q

[0125]

[0126]

[0127]

[0128]

[0129] T 企业,L =Softmax(W feature ·W degree V

[0130] The formula for updating the score is:

[0131]

[0132]

[0133] Among them W Q W K W V Here, W is the weight matrix, Softmax is the activation function, d is the feature dimension, degree is the degree function used to calculate the degree of each node, and W is the feature matrix. feature and W degree , where are the feature attention and the degree attention score of the node, respectively, and L is the number of layers in the graph neural network.

[0134] For the "industry segment" level of graph neural networks, the attention weights in the Attention module of the industry segment consist of two parts: the feature attention of the nodes and the market size attention of the enterprises. The feature update calculation formula is as follows:

[0135]

[0136] Q = T 产业环节,L-1 W Q

[0137]

[0138]

[0139]

[0140]

[0141] T industry segment, L =Softmax(W feature ·W scale V

[0142] The formula for updating the score is:

[0143]

[0144]

[0145] Among them W Q W K W V Here, W is the weight matrix, Softmax is the activation function, d is the feature dimension, MarketScale is the market size function, and it returns the market size scale for each firm. feature and W scale , where are the feature attention score and the market size attention score, respectively, and L is the number of layers in the graph neural network.

[0146] The specific parameters for the three attention modules mentioned above are: d = 768, W ... Q W K W V It is a parameter matrix of 768*256.

[0147] (7) Importance score of propagation nodes

[0148] Obtain the importance score for each link. Calculate the score S for the "Industry Link" node step-by-step using the updated node scores. 产业环节,k S∈[0,1]. Based on statistical information from the digital transportation industry, the market size of each industry segment can be obtained, normalized to the range of [0,1], and these scores S representing market size can be used to determine the market size. gold This serves as a label for each node in the industry chain. Finally, the mean squared error loss function is used to optimize the parameters of the entire model.

[0149] The mean squared error loss function is used to optimize the parameters of the entire model, and its calculation formula is as follows:

[0150] Loss=(SS gold ) 2

[0151] Where S is the score calculated by the model, S goldThe score is obtained by normalizing the actual market size to [0,1]. The Adam optimizer is used during model optimization, with 10 optimization rounds.

[0152] (8) Predicting the score of the parent industry segment

[0153] Obtain the strength score of the parent category's industry segment. For entities in the "Industry Segment" category, those parent nodes not directly connected to the "Enterprise" node cannot have their scores calculated directly. The strength score of the parent node is calculated by averaging the scores of all its connected subcategories of "Industry Segments". The calculation formula is as follows:

[0154]

[0155] Where s 父类 s represents the strength score of the parent industry predicted by the model. 子类 The time model calculates the strength score for each industry link directly connected to the enterprise node. "Subclass" represents the set of nodes connected to the parent class node, and |subclass| represents the size of the set.

[0156] (9) Prediction of the strength and weakness of the industrial chain

[0157] Based on the predicted scores for each industry segment node, they are ranked from highest to lowest importance, representing the ranking of each industry segment's importance in the knowledge graph. Based on the ranking results, we can conduct a quantitative analysis of the strength and weakness of relationships within the industry chain.

[0158] This invention implements a knowledge graph-based method for predicting the strength and weakness of digital transportation industry chains and their links. Based on the final industry link ranking results, an evaluated ranking is performed using labeled industry link rankings. As shown in Table 1, the predicted industry rankings are completely consistent with the standard labels, with a mean squared error of 1.3454e-06.

[0159] Table 1 Experimental Evaluation Results

[0160]

[0161]

[0162] This embodiment constructs a deep digital transportation industry chain knowledge graph. This method comprehensively considers enterprise data, ontology information, and link relationships to achieve a comprehensive and in-depth prediction of the entire industry chain, improving the accuracy and comprehensiveness of the prediction. By adding entity identifiers to the BERT model dictionary, the model can better highlight the importance of entities in triples, thereby better learning the entity features in triple text and improving the ability to model entities. A scoring network is used to map text features to importance scores, evaluating the importance scores of products, etc., from text data, thus easily deriving a more reasonable importance judgment from a large number of text feature representations. A hierarchical graph neural network is used to comprehensively propagate the importance scores of each level of the industry chain, realizing the dynamic updating of node correlation. This helps to more comprehensively understand the relationships between nodes at different levels and improves the accuracy of predicting the strength and weakness of links in the industry chain.

[0163] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above descriptions are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A knowledge graph-based intelligent analysis method for the digital transportation industry chain, characterized in that, include: S1. Acquire enterprise data and ontology data of the digital transportation industry chain, and construct a knowledge graph of the digital transportation industry chain; S2. Extract the text descriptions of nodes from the constructed knowledge graph; S3. Input the text description of the node into the BERT encoder and obtain the feature representation of the node; S4. Obtain the importance score of the node using the feature representation of the node, and update it through a hierarchical graph neural network; S5. Use the updated importance score to determine the strength or weakness of each link in the industrial chain; S6. Generate visual information based on the strength of each link in the determined industrial chain and send it to the display device.

2. The method according to claim 1, characterized in that, S1 includes: Operational data of enterprises related to the digital transportation industry chain are extracted from the database, and enterprise data is generated based on the operational data. The enterprise data is represented as e, where e = {type, name, scale, products_list}. `type` represents the industry category to which enterprise e belongs, corresponding to a specific industry segment in the ontology data of the digital transportation industry chain. `name` represents the name of the enterprise. `scale` represents the market size of the enterprise; the value of `scale` corresponds to at least the enterprise's total annual sales. `products_list` represents a list of all products produced by the enterprise. The ontology data is represented as O = {O} node O edge }, O node O represents the set of nodes that make up the industrial links included in the industrial chain. edge These represent the edges between these nodes, and the industrial links include belonging relationships and supply relationships in the industrial chain.

3. The method according to claim 2, characterized in that, In S1, the construction of the knowledge graph for the digital transportation industry chain includes: The entities in the knowledge graph are constructed based on entity types, including [industry segment, enterprise, product]. The relationships in the knowledge graph are constructed based on relation types, including [belong to, supply, include enterprise, production]. Among them, the "industry link" node is represented by the node set O in the ontology model O. node The "Enterprise" node is determined by the name in the enterprise data, the "Product" node is determined by the products list in the enterprise data, the "belong to" relationship and the "supply" relationship are determined by the annotations in the ontology model O, the "contains enterprises" relationship is determined by the type data of each enterprise, and the "production" relationship is determined by the name and products list in the enterprise data.

4. The method according to claim 1, characterized in that, S2 includes: Using the entity type and name of the node in the knowledge graph, we obtain the text description of the node. The text description of the "Industry Link" node is as follows: {[CLS],[type],Industry Segment,[\type],[name],Industry Name,[\name],[SEP]}; The text description of the "Enterprise" node is: {[CLS],[type],Enterprise,[\type],[name],Enterprise Name,[\name],[SEP]}; The text description for the "Product" node is: {[CLS],[type],Product,[\type],[name],Product Name,[\name],[SEP]}; [CLS] and [SEP] represent the beginning and end of the text, respectively; [type] and [\type] represent the beginning and end identifiers of the entity type; and [name] and [\name] represent the beginning and end of the entity name.

5. The method according to claim 1, characterized in that, S3 includes: The text description of the node is input into the BERT encoder, which runs the BERT model to obtain the features of the text: Where 'a' is the number of tokens for node category name, 'b' is the number of tokens for node name, and 'T' is the number of tokens for node name. [CLS] T represents the classification label in the BERT model. [type] T represents the marker indicating the starting position of a category in the BERT model. [\type] This represents a marker in the BERT model that indicates the end position of a category. The first token represents the node category name. T represents the a-th token of the node category name. [name] T represents a marker in the BERT model indicating the starting position of a node name. [\name] This represents a marker in the BERT model that indicates the end position of a node name. The first token representing the node name. T represents the b-th token of the node name. [SEP] This represents the sentence separator tag in the BERT model.

6. The method according to claim 1, characterized in that, In S4, it includes: Obtain the importance score S of a node using its feature representation. 节点 =MLP(T 节点 MLP stands for Fully Connected Network; The representations of nodes and their corresponding importance scores in the knowledge graph are updated through information propagation in a hierarchical graph neural network.

7. The method according to claim 6, characterized in that, The method of updating the feature representations and corresponding importance scores of nodes in the knowledge graph through information propagation via a hierarchical graph neural network includes: For the "product" level of a graph neural network, the feature representation T of the "product" node is updated using graph attention techniques. 产品,L =Attention(T) 产品,L-1 ,T 企业,L-1 ), and importance score S 产品,L =Attention(T) 产品,L-1 ,T 企业,L-1 ,S 产品,L-1 ,S 企业,L-1 ), where L is the number of layers in the graph neural network, and the hierarchical graph neural network is divided according to entity type, including at least three levels: "product" level, "enterprise" level, and "industry link" level, where T 产品,L-1 T represents the feature representation of a single "product" node at layer L-1. 企业,L-1 S represents the feature representation of all "enterprise" nodes at level L-1. 产品,L-1 S represents the importance score of a single "product" node at level L-1. 企业,L-1 This represents the importance score of all "enterprise" nodes at level L-1; For the "enterprise" level of the graph neural network, the feature representation T of the "enterprise" node is updated using graph attention techniques. 企业,L =Attention(T) 企业,L-1 ,T 产品,L-1 ), and the score S of the "Enterprise" node. 企业,L =Attention(T) 企业,L-1 ,T 产品,L-1 ,S 企业,L-1 ,S 产品,L-1 ), where T 企业,L-1 T represents the feature representation of a single "enterprise" node at layer L-1. 产品,L-1 S represents the feature representation of all "product" nodes at level L-1. 企业,L-1 S represents the importance score of a single "enterprise" node at level L-1. 产品,L-1 This represents the importance score of all "enterprise" nodes at level L-1; For the "industry chain" level of graph neural networks, the feature representation T of the "industry chain" nodes is updated using graph attention techniques. 产业环节,L =Attention(T) 产业环节,L-1 ,T 企业,L-1 The importance score S for the "industry link" node. 产业环节,L =Attention(T) 产业环节,L-1 ,T 企业,L-1 ,S 产业环节,L-1 ,S 企业,L-1 ), where T 产业环节,L-1 T represents the feature representation of a single "industry link" node at layer L-1. 企业,L-1 S represents the feature representation of all "enterprise" nodes at level L-1. 产业环节,L-1 S represents the importance score of a single "industry link" node in layer L-1. 企业,L-1 This represents the importance score of all "enterprise" nodes in layer L-1.

8. The method according to claim 7, characterized in that, The attention weights in the "product" level attention module consist of the node's feature attention and the enterprise's market size attention; the attention weights in the "enterprise" level attention module consist of the node's feature attention and the node's degree attention; and the attention weights in the "industry segment" level attention module consist of the node's feature attention and the enterprise's market size attention.

9. The method according to claim 7, characterized in that, In S5, it includes: The updated importance score serves as the label for industry segment nodes, where the average importance score of the subclass "industry segment" serves as the value of the strength of importance of the corresponding parent class node. The links in the industrial chain are ranked according to their importance scores.

Images