An enterprise industry track dynamic mapping method and system based on a knowledge graph

Abstract

The application relates to the technical field of knowledge graph and industry analysis, and discloses an enterprise industry track dynamic mapping method and system based on a knowledge graph, wherein the method comprises the following steps: collecting multi-source enterprise original data and generating enterprise comprehensive description texts; performing entity recognition and relationship extraction to generate structured knowledge triples; constructing an enterprise knowledge graph and generating enterprise semantic vectors; constructing an industry track knowledge graph and generating track semantic vectors; calculating semantic similarity initial mapping scores and generating a candidate track set; and generating enhanced embedding vectors through graph attention network neighborhood aggregation. Through an event-driven dynamic updating mode and an emerging track automatic discovery mode based on HDBSCAN clustering, the technical problem that a static track system cannot adapt to changes in industrial development is solved, and the technical effect that both the enterprise track mapping relationship and the track system can be continuously updated with the evolution of the industry is achieved.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph and industry analysis technology, specifically to a method and system for dynamic mapping of enterprise industry tracks based on knowledge graph. Background Technology

[0002] In industry sector analysis, determining the affiliation between a company and an industry sector is a core step in building an industry map. Existing technologies typically employ the following methods to determine a company's sector affiliation: text semantic similarity calculation based on a single data source, or static classification methods based on manually defined rules and keyword matching.

[0003] The aforementioned existing technologies suffer from the following drawbacks: First, a single data source cannot comprehensively reflect a company's technological direction, leading to biases in determining industry sector affiliation. Second, relying solely on textual semantic similarity fails to capture implicit connections within the industry ecosystem; for example, the indicative role of supply chain or investment relationships in determining industry sector affiliation is overlooked. Third, the static industry sector system cannot dynamically expand with the emergence of new sectors, nor can it continuously update mapping relationships as a company's business changes. These shortcomings collectively result in insufficient accuracy in determining a company's industry sector affiliation and a technical problem of being unable to adapt to the dynamic evolution of the industry. Summary of the Invention

[0004] This invention provides a method and system for dynamic mapping of enterprise industry tracks based on knowledge graphs, which solves the technical problems in related technologies such as the difficulty in automatically generating the mapping relationship between enterprises and industry tracks, the lack of dynamic updating capability of mapping results, and the insufficient mapping accuracy caused by a single feature dimension.

[0005] This invention discloses a method for dynamic mapping of enterprise industry tracks based on knowledge graphs, comprising the following steps: collecting original enterprise data from multi-source databases; performing standardized preprocessing and deduplication on the collected text data; concatenating the business scope text, patent abstract text, recruitment description text, and product introduction text into a comprehensive enterprise description text; performing named entity recognition and relation extraction on the comprehensive enterprise description text to generate structured knowledge triples; and performing multi-strategy fusion entity alignment on the identified entities; constructing an enterprise knowledge graph based on the structured knowledge triples; inputting the comprehensive enterprise description text into a pre-trained language model; taking the output classification position hidden state vector and performing L2 normalization to obtain the enterprise semantic vector; constructing a three-level industry track knowledge graph; concatenating the track description text with key technology tags and inputting it into a pre-trained language model; taking the classification position hidden state vector and performing L2 normalization to obtain the track semantic vector; and calculating the enterprise semantic vector. The cosine similarity between the vector and the track semantic vector is used to obtain a semantic similarity score, and a candidate track set is generated. A graph attention network is used on the enterprise knowledge graph to perform neighborhood aggregation on enterprise nodes, generating enhanced embedding vectors. The graph neural network enhanced mapping score is calculated based on these enhanced embedding vectors. Two rule constraint sub-scores are calculated based on the number of IPC classification number hits and the weighted number of keyword hits, and then merged into a rule constraint score. The semantic similarity score, the graph neural network enhanced mapping score, and the rule constraint score are weighted and fused to obtain the final mapping score. Tracks with a final mapping score not lower than the mapping decision threshold are determined as the enterprise's affiliated tracks, generating a mapping relationship between enterprises and tracks and writing it into the graph database. Incremental data events in multi-source data streams are monitored in real time, triggering data re-collection, semantic vector updates, graph attention network forward inference, and mapping score recalculation for the corresponding enterprise nodes, updating the mapping relationship in the graph database.

[0006] Furthermore, the generation of the comprehensive enterprise description text also includes: performing deduplication using the combination of the enterprise's unified social credit code and the data release timestamp as the deduplication key, marking and filtering records with missing key fields, inserting special delimiters between text fragments; and setting the corresponding text fragment to an empty string when a certain type of data source cannot collect valid data.

[0007] Furthermore, the multi-strategy fusion entity alignment includes: performing exact matching with the unified social credit code as the primary key; generating a candidate alignment set by using a weighted combination of edit distance and Jaccard similarity for entities that cannot be matched by the primary key; inputting the candidate aligned entity pairs into a pre-trained language model to obtain semantic vectors and calculating cosine similarity; and determining that they are the same entity when both the rule matching result and the semantic similarity result meet the threshold condition.

[0008] Furthermore, the graph attention network adopts a two-layer stacked structure, with each layer using multiple independent attention heads. The outputs of each attention head are concatenated and then subjected to linear transformation for dimensionality reduction. When aggregating the neighborhood, only the relationship edges between enterprise nodes are considered, including competitive, cooperative, supply, and investment relationship edges. The attention weights are obtained by concatenating the transformation vectors of adjacent nodes, applying a linear rectified activation function with a negative slope, and then normalizing them using Softmax.

[0009] Furthermore, the graph attention network uses known enterprise and track affiliation labeling data as supervision signals. The training objective is to minimize the distance between the enterprise's enhanced embedding vector and its corresponding track semantic vector, while maximizing the distance between the enterprise and the semantic vector of the enterprise that does not belong to the track. The loss function is triplet loss, and the training uses the Adam optimization algorithm.

[0010] Furthermore, the IPC classification number constraint score is calculated using an exponential saturation function with the number of patents hit by the IPC classification number as the independent variable; the keyword constraint score is calculated using an exponential saturation function with the sum of the weighted number of hits of each keyword in the track keyword dictionary as the independent variable; the two scores are weighted and summed according to a fusion weight coefficient that satisfies the sum of weights equal to 1 to obtain the rule constraint score.

[0011] Furthermore, in the three-tiered industry track knowledge graph, the three-tiered tracks are dynamically expanded with industry development; the track nodes include a track popularity index attribute, which is obtained by weighted summation of four sub-scores: financing popularity score, patent popularity score, news popularity score, and recruitment popularity score, followed by Min-Max normalization. Each sub-score is obtained by summing the relevant quantitative indicators of the enterprises within the statistical time window and then taking the logarithmic transformation.

[0012] Furthermore, the mapping relationship generation also includes: setting upper limit constraints on the number of tracks belonging to each enterprise at each level; after the three-level track mapping is completed, the mapping relationship is propagated upward along the hierarchical subordinate relationship edge of the track body, and the corresponding second-level tracks and first-level tracks are included in the enterprise's track belonging set, and the highest final mapping score in each level is used as the score of the corresponding level mapping edge.

[0013] Furthermore, the dynamic update also includes an adaptive evolution step for the track system: periodically perform density clustering analysis on the enhanced embedding vectors of all enterprises; for each cluster found, calculate the centroid vector of the cluster; retrieve the track node with the highest cosine similarity to the centroid vector in the track semantic vector space; if the highest cosine similarity is lower than a preset threshold, mark the cluster as a potential emerging track candidate and push it to the domain expert review queue; after confirmation by the domain experts, create new track nodes in the graph database and update the mapping relationship.

[0014] Furthermore, the adaptive evolution steps of the track system also include declining track detection: periodically detect the decline rate of the number of companies belonging to each track and the decline rate of the track popularity index within the statistical period. When both indicators exceed the corresponding thresholds, the track is marked as a shrinking track, a processing suggestion report is generated, and after review and confirmation by domain experts, track merging or archiving operations are performed. The archived track nodes will no longer participate in new mapping calculations, and historical mapping relationship data will be retained.

[0015] This invention addresses the technical problem of incomplete information from a single data source leading to biased track attribution determination by fusing four types of heterogeneous data—business scope text, patent abstracts, recruitment descriptions, and product introductions—into a comprehensive enterprise description text and encoding it into a unified semantic vector. This achieves a more comprehensive semantic representation of the enterprise's technological direction. Furthermore, by introducing a graph attention network to perform neighborhood aggregation on the enterprise knowledge graph, this invention solves the problem that relying solely on textual semantic similarity cannot capture implicit connections within the industrial ecosystem. This achieves the technical effect that enterprises with enhanced embedding vectors incorporating relational structure information, and even those with insufficient business information disclosure, can still identify track attribution. By introducing rule-based scoring constrained by IPC classification numbers and keyword dictionaries, this invention addresses the insufficient accuracy of purely data-driven methods in scenarios with clearly defined technical features. This achieves the technical effect that expert knowledge participates in mapping score calculation in the form of structured rules. Finally, by employing an event-driven dynamic update method and an automatic discovery method for emerging tracks based on HDBSCAN clustering, this invention solves the technical problem that static track systems cannot adapt to changes in industrial development. This achieves the technical effect that both the enterprise track mapping relationship and the track system can be continuously updated with industrial evolution. Attached Figure Description

[0016] Figure 1 This is a flowchart of the enterprise industry track dynamic mapping method based on knowledge graph provided in the embodiments of the present invention; Figure 2 This is a sub-flowchart of the enterprise comprehensive description text entity recognition and relation extraction in the enterprise industry track dynamic mapping method based on knowledge graph provided in the embodiments of the present invention. Figure 3 This is a sub-flowchart of calculating rule-based track affiliation constraint scores in the knowledge graph-based enterprise industry track dynamic mapping method provided in this embodiment of the invention; Figure 4 This is a sub-flowchart of the enterprise industry track dynamic mapping method based on knowledge graph provided in the embodiments of the present invention, which is based on incremental data triggering dynamic updates. Detailed Implementation Example

[0017] In the field of industry track analysis, determining the affiliation between enterprises and industry tracks faces the following technical challenges: First, an enterprise's business scope often spans multiple tracks, and a single data source cannot fully reflect the enterprise's technological direction, leading to inaccurate track affiliation determination. Second, relying solely on textual semantic similarity cannot capture the implicit connections between enterprises within the industry ecosystem, such as the indicative role of supply chain or investment relationships in track affiliation. Third, industry tracks themselves are in a dynamic evolutionary process, with emerging tracks constantly appearing, and a static track system cannot adapt to changes in industry development. To address these challenges, this implementation provides a knowledge graph-based dynamic mapping method for enterprise-industry tracks. It constructs enterprise and track knowledge graphs by integrating multi-source data and employs a three-dimensional fusion mapping score calculation method, combining semantic similarity, graph neural network neighborhood aggregation enhancement, and rule constraints, to automatically calculate and continuously update the mapping relationship between enterprises and industry tracks.

[0018] Before implementation, the following environment is required: a graph database is deployed to store enterprise knowledge graphs and track knowledge graphs; a vector database is deployed to store enterprise semantic vectors and track semantic vectors and support near nearest neighbor retrieval; a message queue is deployed to buffer multi-source data streams; and a pre-trained language model inference service is deployed. The pre-trained language model adopts an industry-specific pre-trained model based on the BERT Large architecture, which is pre-trained twice on corpora in the fintech field and fine-tuned under supervision on samples labeled with five types of entities: enterprises, products, technologies, people, and institutions.

[0019] A knowledge graph-based dynamic mapping method for enterprise industry tracks according to an embodiment of the present invention includes the following steps: Step 1: Collect and clean raw data from multiple enterprise sources to generate comprehensive enterprise description text.

[0020] Raw enterprise data is collected from business registration databases, patent databases, recruitment platforms, and financing databases. Specific fields collected include: enterprise name, unified social credit code, and business scope description from business registration information; patent abstracts and IPC classification numbers from patent data; job titles and job descriptions from recruitment data; and financing rounds and financing amounts from financing data. All collected raw data is buffered in a message queue and synchronously persisted to a distributed file storage system.

[0021] Standardized preprocessing is performed on the collected raw text data, including removing HTML tags and special characters, unifying the conversion between simplified and traditional Chinese characters, and unifying full-width and half-width characters. Deduplication is performed using the combination of the enterprise's unified social credit code and the data release timestamp as the deduplication key, and records missing key fields are marked and filtered.

[0022] Perform text concatenation on the cleaned data and generate a comprehensive enterprise description text using the following formula. : ; in, The document containing the company's business scope; This is a compilation of patent abstracts published by the company within the past three years, arranged in descending order of patent publication time, with the first 20 patent abstracts selected. This is a compilation of currently valid job postings from the company, sorted in descending order of posting time, with the top 10 results selected. This is product description text collected from the company's official website; This indicates a text concatenation operation, inserting a special separator [SEP] between adjacent text segments.

[0023] It should be noted that the source of the above-mentioned comprehensive enterprise description text refers to the combination of four types of text: business scope, patent abstract, recruitment description, and product introduction, selected based on the different data sources' ability to represent the enterprise's technological direction. When a certain type of data source cannot collect valid data, the corresponding text fragment is set to an empty string, which does not affect the concatenation and subsequent processing of other text fragments.

[0024] In this embodiment of the application, in order to improve the timeliness of data collection, for news and public opinion data, the RSS subscription sources of mainstream financial media are connected and a news aggregation crawler based on keyword matching is deployed to collect news reports related to the target company in real time. The collected fields include news title, full text, source media, publication time and original URL. The news data stream is connected to a message queue for buffer management to support the calculation of the track popularity index in the future.

[0025] Step 2: Perform entity recognition and relation extraction on the comprehensive description text of the enterprise to generate structured knowledge triples.

[0026] Comprehensive description text of the enterprise The input is a pre-trained language model fine-tuned for domain adaptation. Named entity recognition is performed, and the output is a sequence of entity annotations in BIO format. The identified entity types include enterprise entities (ORG), product entities (PROD), technology entities (TECH), person entities (PER), and organization entities (INST). For text exceeding the maximum input length of the pre-trained language model, a sliding window strategy is used for segmentation to avoid entities being truncated at window boundaries.

[0027] Based on entity identification, a text fragment containing two target entity tags is input into the relation extraction model. The model outputs a probability distribution over a predefined set of relation types, extracting the semantic relationships between entity pairs. The predefined set of relation types includes: production relationships (PRODUCES) and provision relationships (PROVIDES) between enterprises and products; ownership relationships (OWNS), research and development relationships (DEVELOPS), and application relationships (APPLIES) between enterprises and technologies; investment relationships (INVESTS), competition relationships (COMPETES_WITH), cooperation relationships (COOPERATES_WITH), acquisition relationships (ACQUIRES), and supply relationships (SUPPLIES_TO) between enterprises; and investment relationships (FUNDED_BY) between enterprises and investment institutions. Relationship extraction results with a confidence level below 0.5 are marked as pending review and are not directly written into the knowledge graph.

[0028] Perform multi-strategy fusion entity alignment on the identified entities to resolve name discrepancies for the same enterprise across different data sources. Step 2 may specifically include: First, exact matching is performed using the Unified Social Credit Code as the primary key. For entities that cannot be matched by the primary key, a candidate alignment set is generated using the string similarity of the enterprise name (a weighted combination of edit distance and Jaccard similarity). Second, the names of the candidate aligned entity pairs and their contextual descriptions are input into a pre-trained language model to obtain the semantic vector representation of each entity and calculate the cosine similarity between the candidate aligned entity pairs. Finally, when both the rule matching result and the semantic similarity result meet the threshold condition, they are determined to be the same entity, and multiple designations are uniformly mapped to a standardized entity node identified by the Unified Social Credit Code.

[0029] After the above processing, a structured knowledge triple is generated in the format of (subject entity, relation type, object entity). Each triple is accompanied by metadata such as source identifier, confidence score and timestamp.

[0030] Step 3: Construct an enterprise knowledge graph based on structured knowledge triples and generate enterprise semantic vectors.

[0031] The structured knowledge triples generated in step 2 are written into a graph database to construct an enterprise knowledge graph. The enterprise knowledge graph is modeled using an attribute graph model, containing the following node types: Company, Product, Technology, Talent, and Investor; and the following edge types: PRODUCES (Company → Product), DEVELOPS (Company → Technology), APPLIES (Company → Technology), FUNDED_BY (Company → Investor), COMPETES_WITH (Company → Company), COOPERATES_WITH (Company → Company), SUPPLIES_TO (Company → Company), and INVESTS_IN (Company → Company). Each edge type carries common attributes such as confidence level, source data identifier, and creation timestamp.

[0032] Comprehensive description text of the enterprise Input pre-trained language model The hidden state vector at the [CLS] position of the pre-trained language model output is taken as the initial semantic vector of the enterprise. : ; in, For pre-trained language models The hidden state vector output at position [CLS] after encoding. This is a pre-trained language model. When the length of the enterprise's comprehensive description text exceeds the maximum input length of the pre-trained language model, the text is segmented into several segments, and the [CLS] vector of each segment is obtained. The average of all segment [CLS] vectors is taken as the initial semantic vector.

[0033] The initial semantic vector is subjected to L2 normalization to obtain the final enterprise semantic vector. : ; in, For vectors The L2 norm. The enterprise semantic vector. It is stored in the semantic vector attribute of the enterprise node and synchronously written to the vector database.

[0034] Step 4: Construct an industry track knowledge graph and generate track semantic vectors.

[0035] A three-tiered industry track system is constructed, with each level defined as follows: Level 1 track (L1 Track) corresponds to the macro-industry direction, with representative examples including new energy, artificial intelligence, biomedicine, and semiconductors; Level 2 track (L2 Track) corresponds to the sub-industry direction and is a subordinate concept of Level 1 track, with representative examples including power batteries and large language models; Level 3 track (L3 Track) corresponds to the specific technology or product track and is a subordinate concept of Level 2 track, which expands dynamically with industry development.

[0036] Track nodes contain the following main attributes: track unique identifier (track_id), track name (track_name), track level (track_level), track description text (track_description), key technology tag list (key_tech_tags), track heat index (track_heat_index), track maturity index (track_maturity_index), list of IPC code mappings strongly related to the track (ipc_code_mapping), track semantic vector (semantic_vector), and number of companies currently affiliated with the track (company_count). Tracks are structurally linked through the following edge types: IS_SUBTOPIC_OF (lower track → higher track, indicating a hierarchical relationship), IS_ADJACENT_TO (track ↔ track, indicating adjacent tracks at the same level), and CONVERGES_WITH (track ↔ track, indicating a trend of technological convergence between the two tracks).

[0037] For each track node, the track description text Text concatenated with key technology tags Combine them to form a comprehensive description text of the track. :

[0038] in, Describe the track text. Concatenate text for key technology tags. This indicates a text concatenation operation. It combines the track description text... Input pre-trained language model The hidden state vector at position [CLS] is taken and normalized by L2 to obtain the track semantic vector. All track semantic vectors are stored uniformly in a vector database, forming a track semantic vector space.

[0039] It should be noted that the above-mentioned track popularity index The calculation of is a quantitative indicator that comprehensively reflects the market attention of a track within the current time window. Its calculation formula is as follows: ; in, This is the index for the popularity category, with a value range from 1 to 4; For the first The weighting coefficients of the heat index sub-items satisfy the constraints. ; For the first The score for each category based on its popularity. The formula for calculating the financing popularity score is as follows: ; in, For the group of companies belonging to this track, For enterprises Corporate elements in For enterprises The total amount of financing within the statistical time window. For enterprises The weighting of financing rounds. The formula for calculating the patent popularity score is as follows: ; in, For enterprises The number of patents related to this track published within the statistical time window. The formula for calculating the news popularity score is as follows: ; in, For enterprises The number of news reports related to this track within the statistical time window. The average sentiment polarity score for this batch of news articles is output by the sentiment analysis model, with 1 for positive sentiment, 0.5 for neutral sentiment, and 0 for negative sentiment. The formula for calculating the recruitment popularity score is as follows: ; in, For enterprises The number of job postings related to this track published within the statistical time window. After the logarithmic transformation described above, each sub-score is then normalized to the [0, 100] interval using Min-Max normalization to obtain the final track popularity index. .

[0040] Step 5: Calculate the initial semantic similarity mapping score between enterprises and tracks to generate a candidate track set.

[0041] For businesses With the track Calculate the cosine similarity between the normalized semantic vectors of the two vectors, and use it as the initial semantic similarity mapping score. : ; in, For enterprises Normalized semantic vector, For the track The normalized semantic vectors of both have been processed by L2 normalization. The dot product is the cosine similarity, which takes the value range of [-1, 1]. The larger the value, the more semantically similar the company and the track are.

[0042] Utilizing the approximate nearest neighbor retrieval function of vector databases, the semantic vectors of enterprises are analyzed. Retrieve the track with the highest cosine similarity in the track semantic vector space. For each track, a candidate track set is generated, and subsequent refined mapping score calculations are performed only on the tracks in the candidate track set.

[0043] Step 6: Perform graph attention network neighborhood aggregation on enterprise nodes to generate enhanced embedding vectors.

[0044] On the enterprise knowledge graph, a graph attention network (GAT) is used to perform neighborhood aggregation enhancement on the embedded representation of enterprise nodes to capture implicit relationship information of enterprises in the industrial ecosystem. The graph attention network adopts a two-layer stacked structure, with each layer using multiple independent attention heads. The outputs of each attention head are concatenated and then subjected to linear transformation for dimensionality reduction.

[0045] Graph Attention Network The aggregation update formula for the layer is as follows: ; in, For enterprise nodes In the Hidden layer embedding vectors of the layer; The initial embedding of the enterprise node, i.e., the enterprise semantic vector. ; For enterprise nodes The set of one-hop neighbor nodes, considering only the relationship edges between enterprise nodes, including edge types such as COMPETES_WITH, COOPERATES_WITH, SUPPLIES_TO, and INVESTS_IN; For the first The learnable weight matrix of the layer; ELU is the activation function; AGG is the weighted sum aggregation function; For nodes Its neighboring nodes The attention weight is calculated using the following formula: ; in, For the first The learnable attention weight vector of the layer; Indicates transpose; This represents a vector concatenation operation; For the set of neighbor nodes in the normalized denominator The traversal index; represents an exponential function; LeakyReLU is a linear rectified activation function with a negative slope.

[0046] After aggregation through a two-layer graph attention network, enterprise nodes are obtained. Enhanced embedding vector Enterprises based on enhanced embedded vector computing With the track Graph Neural Network Enhanced Mapping Score : ; in, For enterprises The enhanced embedding vector after aggregation through two layers of the graph attention network, For the track The normalized semantic vector.

[0047] It should be noted that the training of the graph attention network described above uses known enterprise track affiliation labeling data as supervision signals, and the input is the initial embedding of each enterprise node on the enterprise knowledge graph. The graph structure outputs enhanced embeddings for each enterprise node. The training objective is to minimize the distance between the enterprise's augmented embedding and its corresponding semantic vector within its own track, while maximizing the distance with semantic vectors from other tracks. The loss function used is the triplet loss. ; in, For enterprises The positive example track (known as the track it belongs to). Negative example tracks (randomly sampled non-belonging tracks). For Euclidean distance, is the margin hyperparameter. Training uses the Adam optimization algorithm.

[0048] Step 7: Calculate the rule-based track ownership constraint score.

[0049] For each track in the candidate track set The IPC classification number constraint score and keyword constraint score are calculated separately and then merged into a rule constraint score. .

[0050] Step 7 may specifically include: Calculate the IPC classification number constraint score : Pre-build a mapping dictionary from IPC classification numbers to track identifiers, and statistically analyze enterprises. The patent portfolio belongs to the track Number of patents associated with IPC classification numbers Calculate the IPC classification number constraint score using the following formula: ; in, The attenuation coefficient is... For enterprises Hit the track The number of patents associated with IPC classification numbers; this function in The value is 0 at time, and increases monotonically with the number of patents, approaching 1.

[0051] Calculate keyword constraint scores : Pre-build a keyword dictionary for each track, containing the core terms, product names, and technology names of that track, and provide comprehensive descriptions of companies. Perform keyword hit detection and calculate the keyword constraint score using the following formula: ; in, For the track Keyword dictionary Keyword dictionary Keyword elements in For indicator functions, Keywords The weight, This is the attenuation coefficient.

[0052] Combining the above two points, calculate the rule constraint score: ; in, and To integrate the weighting coefficients, satisfy the following conditions: .

[0053] Step 8: Integrate the 3D mapping scores and execute mapping decisions to generate enterprise track mapping relationships.

[0054] For each track in the candidate track set The semantic similarity is initially mapped to a score. Graph Neural Network Enhanced Mapping Score Score constrained by rules By performing weighted fusion, we can obtain the enterprise With the track Final mapping score : ; Among them, weight Satisfying constraints and , , The weight vectors are learned through supervised training on a labeled dataset, with the training objective being to maximize the F1 score of the enterprise track affiliation category.

[0055] Perform mapping decisions to determine the enterprise Track ownership set : ; in, This is the threshold for the mapping decision. When... At that time, the enterprise was judged Belonging track It supports multi-label mapping scenarios where a single enterprise can be assigned to multiple tracks simultaneously. To ensure the accuracy of the mapping results, an upper limit constraint is set on the number of tracks each enterprise can be assigned to: no more than 3 L1 tracks, no more than 5 L2 tracks, and no more than 10 L3 tracks.

[0056] The mapping relationships are written into the graph database as directed weighted edges MAPPED_TO. The attributes of the edges include: final mapping score, initial semantic similarity mapping score, graph neural network-enhanced mapping score, rule constraint score, mapping relationship creation time, and most recent update time.

[0057] In this embodiment of the application, to support the linkage mapping of multi-level tracks, after the L3 track mapping is completed, the mapping relationship is automatically propagated upwards along the IS_SUBTOPIC_OF edge of the track body: if the enterprise If an item is mapped to a specific L3 track, then the L2 and L1 tracks to which that L3 track belongs will be automatically included in the enterprise's database. The track affiliation set, and the highest-ranking track in each level. The value serves as the score for the corresponding level of track mapping edge.

[0058] Step 9: Trigger dynamic updates based on incremental data to generate updated enterprise track mapping relationships.

[0059] Real-time monitoring of incremental information in multi-source data streams. When the following types of incremental data events are detected, the update process of the corresponding enterprise node is triggered: the enterprise's business scope changes; the enterprise has a new patent publication announcement; the enterprise completes a new round of financing; the daily average number of news reports related to the enterprise exceeds three times the daily average of the past 30 days and lasts for more than 3 days; the enterprise node has been updated for more than 30 days.

[0060] For companies that trigger the above conditions, step 9 may specifically include: re-collecting the company's latest data from various data sources and regenerating the comprehensive company description text using the methods from steps 1 to 3. and enterprise semantic vectors Using the updated enterprise semantic vector as the initial embedding, perform forward inference using a graph attention network on the current enterprise knowledge graph to obtain the updated enhanced embedding. Based on the updated enterprise semantic vector and enhanced embedding Recalculate the final mapping score between the company and the candidate track set. Write the updated enterprise semantic vectors into the enterprise node attributes of the graph database and update the corresponding vector records in the vector database; update the score attributes of enterprise track mapping edges whose mapping score changes exceed the threshold; create new MAPPED_TO edges for newly added mapping relationships that meet the threshold conditions; set the is_active attribute of the corresponding MAPPED_TO edge to False for mapping relationships that no longer meet the threshold conditions, and retain historical records; write the detailed information of this update operation into the update log database.

[0061] In this embodiment of the application, to address the emergence of new tracks during industrial development, in addition to step 9, the following adaptive evolution steps for the track system are also included: periodically enhancing the embedding vectors of all enterprises. Perform HDBSCAN density clustering analysis, where the input is the enhanced embedding vector set of all firms, and the output is several clusters and the set of firms within each cluster; for each cluster found, calculate the centroid vector of that cluster. (The mean of the enhanced embedding vectors of all enterprises within the cluster) In the track semantic vector space, the track node with the highest cosine similarity to the centroid vector is retrieved. If the highest cosine similarity is lower than a preset threshold, the cluster is marked as a potential emerging track candidate. For each potential emerging track candidate, a candidate track description report is generated, which includes the number of enterprises in the cluster, a list of representative enterprises, a list of high-frequency keywords extracted from the text of enterprises in the cluster using the TFIDF algorithm, the nearest neighboring existing track nodes and their similarity scores, and is pushed to the domain expert review queue. After confirmation by the domain experts, a new track node is created in the graph database, and the mapping relationship of enterprises in the cluster is updated to the new track.

[0062] Furthermore, it also includes the following degradation detection steps: periodically perform degradation detection on all track nodes; if the track... The following conditions must be met: ; in, The track to which the data belongs during the statistical period The change in the number of enterprises, This represents the number of enterprises at the start of the statistical period. This represents the change in the track popularity index within the statistical period. and If the threshold for the percentage decrease in the number of enterprises and the threshold for the decrease in the popularity index are respectively set, the track will be marked as a shrinking track, and a shrinking track handling suggestion report will be generated. After review and confirmation by domain experts, the track merging or archiving operation will be performed. After the track is archived, the is_active attribute of the original track node will be set to False, the historical mapping relationship data will be retained, and it will no longer participate in the new mapping calculation.

[0063] This implementation solves the following technical problems in determining enterprise industry track affiliation by constructing enterprise knowledge graphs and industry track knowledge graphs, and by adopting a mapping score calculation method that integrates semantic similarity, graph attention network neighborhood aggregation enhancement, and rule constraints.

[0064] To address the issue that a single data source cannot fully reflect a company's technological direction, this implementation method combines four types of heterogeneous data—business scope text, patent abstracts, recruitment descriptions, and product introductions—into a comprehensive company description text. This text is then encoded into a unified semantic vector representation using a pre-trained language model. This allows the company's semantic vector to comprehensively reflect its technological direction information across multiple dimensions, enabling a more comprehensive similarity comparison with the semantic vectors of different tracks at the semantic level. This overcomes the problem of biased track attribution determination caused by incomplete information from a single data source.

[0065] To address the issue that relying solely on textual semantic similarity cannot capture implicit connections within the industrial ecosystem, this implementation introduces a graph attention network to perform neighborhood aggregation on the enterprise knowledge graph. By using attention weights, information from different types of neighboring nodes, such as competitors, partners, supply chain companies, and investee companies, is aggregated with differentiated weights. This allows the enhanced embedding vector of an enterprise to be integrated into its relational structure information within the industrial ecosystem. The graph neural network-enhanced mapping score can reflect the indicative role of an enterprise's associated position within the industrial ecosystem in terms of track affiliation. This also provides track affiliation identification capabilities for SMEs with insufficient public disclosure of business information.

[0066] To address the issue that purely data-driven methods may overlook explicit technical feature identifiers, this implementation introduces rule-constrained scoring based on IPC classification numbers and keyword dictionaries. Expert knowledge is integrated into the mapping score calculation in the form of structured rules, which effectively strengthens enterprises with explicit technical feature identifiers in the final fusion score, avoiding the problem of insufficient accuracy of pure vector similarity calculation in scenarios with explicit technical features.

[0067] To address the problem that static track systems cannot adapt to changes in industrial development, this implementation method uses an event-driven dynamic update method for enterprise nodes and an automatic discovery method for emerging tracks based on HDBSCAN clustering. This allows the enterprise track mapping relationship to be continuously updated as enterprise business changes and industrial landscape evolves. The track system itself can also be dynamically expanded as new industrial directions emerge. At the same time, the effectiveness of the track system is maintained through shrinking track detection and archiving, thereby maintaining the consistency between the mapping results and the actual industrial status.

[0068] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for dynamic mapping of enterprise industry tracks based on knowledge graphs, characterized in that, Includes the following steps: The system collects original enterprise data from multiple databases, performs standardized preprocessing and deduplication on the collected text data, and combines the business scope text, patent abstract text, recruitment description text, and product introduction text into a comprehensive enterprise description text. Named entity recognition and relation extraction are performed on the comprehensive description text of enterprises to generate structured knowledge triples, and multi-strategy fusion entity alignment is performed on the identified entities. Enterprise knowledge graphs are constructed based on structured knowledge triples. The comprehensive description text of the enterprise is input into the pre-trained language model, and the hidden state vector of the output classification position is taken and normalized by L2 to obtain the semantic vector of the enterprise. A three-level industry track knowledge graph is constructed. The track description text and key technology tags are concatenated and input into a pre-trained language model. The hidden state vector of the classification position is taken and normalized by L2 to obtain the track semantic vector. Calculate the cosine similarity between the semantic vectors of enterprises and the semantic vectors of tracks to obtain semantic similarity scores, and generate a set of candidate tracks; A graph attention network is used to perform neighborhood aggregation on enterprise nodes in the enterprise knowledge graph to generate enhanced embedding vectors. The enhanced mapping score of the graph neural network is calculated based on the enhanced embedding vectors. The scores for the two rule constraints are calculated separately based on the number of IPC classification number hits and the weighted number of keyword hits, and then merged into a single rule constraint score. The semantic similarity score, graph neural network enhanced mapping score, and rule constraint score are weighted and fused to obtain the final mapping score. Tracks with a final mapping score not lower than the mapping decision threshold are determined as the track to which the enterprise belongs. The mapping relationship between the enterprise and the track is generated and written into the graph database. Real-time monitoring of incremental data events in multi-source data streams triggers data re-collection, semantic vector updates, graph attention network forward inference, and mapping score recalculation for corresponding enterprise nodes, updating the mapping relationships in the graph database.

2. The method according to claim 1, characterized in that: The generation of the comprehensive enterprise description text also includes: performing deduplication using the combination of the enterprise's unified social credit code and the data release timestamp as the deduplication key, marking and filtering records with missing key fields, inserting special delimiters between text fragments; and setting the corresponding text fragment to an empty string when a certain type of data source cannot collect valid data.

3. The method according to claim 1, characterized in that: The multi-strategy fusion entity alignment includes: performing exact matching with the unified social credit code as the primary key; generating a candidate alignment set by using a weighted combination of edit distance and Jaccard similarity for entities that cannot be matched by the primary key; inputting the candidate aligned entity pairs into a pre-trained language model to obtain semantic vectors and calculating cosine similarity; and determining that they are the same entity when both the rule matching result and the semantic similarity result meet the threshold condition.

4. The method according to claim 1, characterized in that: The graph attention network adopts a two-layer stacked structure, with each layer using multiple independent attention heads. The outputs of each attention head are concatenated and then subjected to linear transformation for dimensionality reduction. When aggregating the neighborhood, only the relationship edges between enterprise nodes are considered, including competitive, cooperative, supply, and investment relationship edges. The attention weights are obtained by concatenating the transformation vectors of adjacent nodes, applying a linear rectified activation function with a negative slope, and then normalizing them using Softmax.

5. The method according to claim 4, characterized in that: The graph attention network uses known enterprise and track affiliation labeling data as supervision signals. The training objective is to minimize the distance between the enterprise's enhanced embedding vector and its corresponding track semantic vector, while maximizing the distance with semantic vectors of enterprises not belonging to their tracks. The loss function is triplet loss, and the training uses the Adam optimization algorithm.

6. The method according to claim 1, characterized in that: The IPC classification number constraint score is calculated using an exponential saturation function with the number of patents hit by the IPC classification number as the independent variable; the keyword constraint score is calculated using an exponential saturation function with the sum of the weighted number of hits of each keyword in the track keyword dictionary as the independent variable; the two scores are weighted and summed using a fusion weight coefficient that satisfies the condition that the sum of the weights equals 1 to obtain the rule constraint score.

7. The method according to claim 1, characterized in that: In the three-tiered industry track knowledge graph, the three-tiered tracks expand dynamically with industry development; the track nodes include a track popularity index attribute, which is obtained by weighted summation of four sub-scores: financing popularity score, patent popularity score, news popularity score, and recruitment popularity score, followed by MinMax normalization. Each sub-score is obtained by summing the relevant quantitative indicators of the enterprises within the statistical time window and then taking the logarithmic transformation.

8. The method according to claim 1, characterized in that: The mapping relationship generation also includes: setting upper limit constraints on the number of tracks belonging to each enterprise at each level; after the three-level track mapping is completed, the mapping relationship is propagated upward along the hierarchical subordinate relationship edge of the track body, and the corresponding second-level tracks and first-level tracks are included in the enterprise's track belonging set, and the highest final mapping score in each level is used as the score of the corresponding level mapping edge.

9. The method according to claim 1, characterized in that: The dynamic update also includes the track system adaptive evolution step: periodically perform density clustering analysis on the enhanced embedding vectors of all enterprises, calculate the centroid vector of each cluster, and search the track node with the highest cosine similarity to the centroid vector in the track semantic vector space. If the highest cosine similarity is lower than the preset threshold, the cluster is marked as a potential emerging track candidate and pushed to the domain expert review queue. After confirmation by domain experts, a new track node is created in the graph database and the mapping relationship is updated.

10. The method according to claim 9, characterized in that: The adaptive evolution steps of the track system also include the detection of declining tracks: periodically detecting the decline rate of the number of companies belonging to each track and the decline rate of the track popularity index within the statistical period. When both indicators exceed the corresponding thresholds, the track is marked as a shrinking track, a processing suggestion report is generated, and after review and confirmation by domain experts, track merging or archiving operations are performed. The archived track nodes will no longer participate in new mapping calculations, and historical mapping relationship data will be retained.

Images