An enterprise data intelligent analysis system based on a knowledge graph

Abstract

This invention discloses a knowledge graph-based enterprise data intelligent analysis system, belonging to the field of enterprise data intelligent analysis technology. The system includes acquiring internal business transaction logs, human resource files, and external market sentiment text data; generating an initial set of fact triples through unified timestamp alignment and entity identifier extraction; constructing a foundational knowledge graph skeleton based on entity relationships; extracting product-related sentiment events through natural language processing and transforming them into event triples with sentiment polarity labels, integrating them into the foundational graph, and updating edge weights to generate an enhanced knowledge graph; and identifying potential collaborative links and isolated business nodes based on the density of graph nodes and the distribution of edge weights, generating a graph insight report. This system achieves multi-source data integration and dynamic graph updates, improving the comprehensiveness and timeliness of enterprise data analysis.

Description

Technical Field

[0001] This invention belongs to the field of enterprise data intelligent analysis technology, specifically an enterprise data intelligent analysis system based on knowledge graphs. Background Technology

[0002] In current enterprise data intelligent analysis, knowledge graph technology is often used to model and analyze enterprise data. Existing technologies mostly process single types of internal or external data. Some systems simply organize internal business transaction data or human resource file data to generate basic entity relationships and build static knowledge graphs. Other systems analyze market sentiment text data separately, extracting sentiment information but without integrating it with internal enterprise data.

[0003] Current technologies cannot achieve unified integration of various types of internal enterprise data with external public opinion data. The lack of unified timestamp alignment and entity identification standards for data from different sources and of different types leads to severe data fragmentation, making it difficult to form cross-domain entity relationships and comprehensively reflect the overall business picture. Furthermore, existing knowledge graphs are mostly static structures, only reflecting fixed business interaction relationships between entities. They cannot dynamically adjust the graph's association strength based on sentiment trends in external public opinion, making it difficult to capture the dynamic business situation of the enterprise. Consequently, they cannot accurately identify cross-departmental collaboration potential and isolated business nodes, affecting the accuracy and relevance of data analysis. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, the present invention proposes an enterprise data intelligent analysis system based on knowledge graphs, comprising: The preprocessing module acquires internal business transaction data, human resource file data, and external market sentiment text data. It then performs unified timestamp alignment and entity identifier extraction on the business transaction data, human resource file data, and market sentiment text data to generate an initial set of fact triples containing employee entities, customer entities, and product entities. The graph construction module constructs a lower-level knowledge graph skeleton based on the entity association relationships in the initial set of fact triples. The lower-level knowledge graph skeleton consists of a node set and an edge set, where nodes represent entities and edges represent business interaction relationships between entities. The event fusion module uses natural language processing technology to perform semantic role labeling on the market sentiment text data, extracts negative and positive evaluation events related to the product entity, and transforms them into event triples with sentiment polarity labels. The event triples are then fused with the underlying knowledge graph skeleton, and the edge weights in the underlying knowledge graph skeleton are updated through an event triggering mechanism to generate an enhanced knowledge graph containing dynamic business situations. The graph analysis module identifies potential cross-departmental collaboration links and isolated business nodes based on the node connection density and edge weight distribution in the enhanced knowledge graph, and generates a graph insight report to guide the data analysis perspective.

[0005] Furthermore, based on the entity relationships in the initial set of fact triples, a lower-level knowledge graph skeleton is constructed, including: Each entity in the initial set of fact triples is encoded with a unique identifier, and a mapping dictionary between entity names and identifiers is established. Traverse all triples in the initial set of fact triples, extract the direct relationships between entities, and classify them according to the type of relationship to generate various types of edge sets; For each of the various types of edge sets, count the number of common neighbors of the node pairs connected by the edge of the corresponding type, and calculate the confidence score of the edge of the corresponding type. The node set, the classified edge set, and the corresponding confidence scores are imported into the graph database to construct the underlying knowledge graph skeleton that supports efficient adjacency queries. After the underlying knowledge graph skeleton is constructed, a background thread is started to verify the structural integrity of the underlying knowledge graph skeleton and remove redundant edges with confidence scores lower than a preset threshold.

[0006] Furthermore, the process involves using natural language processing technology to perform semantic role labeling on the market sentiment text data, extracting negative and positive evaluation events related to the product entity, and transforming them into event triples with sentiment polarity labels, including: Load a pre-trained domain-adaptive language model, and use the domain-adaptive language model to perform word segmentation and dependency parsing on the market sentiment text data; Based on the results of the dependency parsing, the predicate core word of the sentence and the argument components associated with the predicate core word are identified. The argument components include the agent, the patient, and the adverbial of time and place. Determine whether the subject refers to the product entity in the underlying knowledge graph skeleton; if so, mark the current sentence as the target evaluation sentence. A sentiment discrimination algorithm combining a sentiment dictionary and contextual information is used to calculate the sentiment tendency score of the predicate core word in the target evaluation sentence, and the sentiment polarity label is determined based on the positive or negative value of the sentiment tendency score. The agent, the predicate core word, and the patient are combined into a structured event triple, and the sentiment polarity label is attached as an attribute to the edge of the event triple.

[0007] Furthermore, the event triples are fused with the underlying knowledge graph skeleton, and the edge weights in the underlying knowledge graph skeleton are updated through an event triggering mechanism, including: Retrieve node pairs in the underlying knowledge graph skeleton that match the subject entity and object entity of the event triple; When a matching pair of nodes is found, the weight values ​​of the original edges connecting the two nodes in the underlying knowledge graph skeleton are read. The weight value is adjusted based on the emotional polarity label of the event triple. If it is a negative evaluation event, the weight value is reduced, and if it is a positive evaluation event, the weight value is increased. The adjusted weight values ​​are written into the edge attribute field of the underlying knowledge graph skeleton to complete the fusion update of a single event; If multiple events targeting the same node pair occur within a preset time window, the weighted average of the sentiment polarity scores of these events is calculated, and the edge weights are updated in a one-time batch based on the weighted average as the final adjustment basis.

[0008] Furthermore, based on the node connectivity and edge weight distribution in the enhanced knowledge graph, potential cross-departmental collaboration links and isolated business nodes are identified, including: In the enhanced knowledge graph, nodes with departmental attributes are selected, and nodes with the same departmental attributes are divided into different community subsets; Calculate the number of node connection paths between different community subsets, and combine the sum of the weights of all edges in the path to calculate the cross-departmental interaction activity index; The interaction activity index is sorted, and the interaction paths that rank high and meet the preset frequency threshold are selected as the potential collaboration links. In the enhanced knowledge graph, the degree centrality and betweenness centrality indices of each node are calculated, and nodes whose degree centrality and betweenness centrality are both below a set lower threshold are identified and marked as the isolated business nodes. The analysis results of the potential collaborative links and the isolated business nodes are encapsulated into structured data objects for subsequent graph insight reports.

[0009] Furthermore, it also includes a risk warning module, used for: Extract supply chain relationship edges involving supplier entities and material entities from the enhanced knowledge graph to form a dynamic supply chain topology view; Retrieve the delivery timestamps for material procurement from the business transaction log data and calculate the time interval volatility between two adjacent delivery cycles; The time interval volatility is correlated and mapped with the credit rating of the supplier nodes in the supply chain topology view to construct a time series feature vector; The time series feature vector is input into a preset pattern matching library to identify abnormal fluctuation sequences that match the characteristics of historical supply disruption crises. Once the abnormal fluctuation sequence is identified, the node of the supplier entity is marked as high-risk in the enhanced knowledge graph, and a traceability query link pointing to the node of the supplier entity is automatically generated.

[0010] Furthermore, the time series feature vector is input into a preset pattern matching library to identify abnormal fluctuation sequences that match the characteristics of historical supply disruption crises, including: The pattern matching library pre-stores several sets of multi-dimensional feature templates for historical crisis scenarios. Each set of multi-dimensional feature templates includes a combination of time interval volatility, price volatility, and the number of logistics delays. The time series feature vector is processed by a sliding window to generate feature sub-vectors for multiple consecutive time slices; Calculate the Euclidean distance metric between each of the said feature sub-vectors and all of the said multidimensional feature templates; When the Euclidean distance between a certain feature subvector and a certain specific multidimensional feature template is less than the similarity threshold, the corresponding time slice is determined to be an abnormal time slice; The consecutively occurring abnormal time slices are merged to form the abnormal fluctuation sequence, and the start and end times corresponding to the abnormal fluctuation sequence are recorded.

[0011] Furthermore, it also includes multi-hop reasoning and anomaly detection steps for corporate financial data: Extract fund transfer records from the business transaction flow data and convert the fund transfer records into financial triples representing fund flows; The financial triple is injected into the enhanced knowledge graph, and a graph traversal algorithm is used to perform a multi-hop path search from the expenditure account node to the income account node. The number of hops in the multi-hop path is limited to three. During each multi-hop path search, the monetary attribute values ​​of all edges on the path are accumulated and compared with the preset fund pool size. If the cumulative amount of a certain multi-hop path is found to exceed the size of the fund pool, the multi-hop path is determined to be an abnormal fund loop. Extract the path node sequence and amount data of the abnormal funding loop to generate abnormal transaction cases for audit verification.

[0012] Furthermore, it also includes an anomaly detection module, used for: Monitor new data changes in the enterprise data interface, and when changes are detected in the business transaction data or human resource file data, extract the incremental facts from the changed data; Perform a logical consistency check between the incremental facts and the current version of the enhanced knowledge graph to identify whether there are attribute value conflicts or relationship contradictions. For conflicts where only attribute values ​​differ, the conflict is resolved by overwriting the older data with the newest timestamp. For conflicts involving contradictory relationships, suspend the entry of incremental facts into the database and initiate a secondary data verification request targeting the points of contradiction. After the secondary data verification request is approved, the disambiguated incremental facts are added as new triples to the enhanced knowledge graph, and the relevant index structure is updated.

[0013] Furthermore, it also includes a visual configuration module for: Define a set of analysis primitives for describing analysis tasks, including data extraction primitives, graph query primitives, and result rendering primitives; The system receives user input by dragging and dropping the analysis primitives through a graphical interface, and generates a directed acyclic graph composed of the analysis primitives, which serves as the executable workflow. The input and output parameters of each node in the executable workflow are analyzed to establish data dependencies between nodes; The query interface of the enhanced knowledge graph is encapsulated as a standard operator in the analysis primitives, and the calling permissions of the standard operator are configured. When a user triggers the execution of the executable workflow, each analysis primitive is scheduled sequentially according to the topological order of the directed acyclic graph, and the final execution result is sent back to the graph insight report.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This solution unifies timestamp alignment and entity identifier extraction for three types of heterogeneous data: internal business transaction logs, human resource records, and external market sentiment text data. This directly generates an initial set of fact triples containing employee, customer, and product entities. This approach breaks down the barriers between internal and external data, solves the problem of data fragmentation from different sources and types, and achieves efficient integration of multi-source heterogeneous data. It allows various data types to form organic connections, comprehensively presenting the basic business relationships between employees, customers, and products within the enterprise. This provides a complete and unified data foundation for subsequent data analysis, avoiding the problems of broken relationships and data disconnects caused by domain-based data processing in conventional technologies.

[0015] This solution utilizes natural language processing (NLP) technology to semantically tag market sentiment text data, extracting negative and positive evaluation events related to product entities. These events are then transformed into event triples with sentiment polarity labels. These event triples are integrated with the underlying knowledge graph skeleton, and the edge weights in the skeleton are updated through an event-triggered mechanism, generating an enhanced knowledge graph that reflects dynamic business dynamics. This approach transforms the knowledge graph from a static structure to a dynamic one that reflects evolving business dynamics. It can adjust the strength of relationships between entities in real time based on the sentiment of external public opinion, accurately capturing the impact of changes in product reputation on business relationships. Compared to conventional static knowledge graphs, it can more realistically and timely reflect enterprise business dynamics, accurately identify potential cross-departmental collaboration links and isolated business nodes, and improve the relevance and timeliness of data analysis. Attached Figure Description

[0016] Figure 1 This is a timeline diagram of an enterprise data intelligent analysis system based on knowledge graphs as described in this invention. Figure 2 Flowchart for constructing the underlying knowledge graph skeleton; Figure 3 A flowchart for the fusion and updating of event triples and knowledge graphs; Figure 4 The trend of delivery time interval and volatility for supplier S1-material M1; Figure 5 This provides a performance analysis diagram for the execution of primitive workflows during the visualization configuration phase. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 This paper describes the implementation of an enterprise data intelligent analysis system based on knowledge graphs, comprising a preprocessing module, a graph construction module, an event fusion module, and a graph analysis module. The preprocessing module acquires internal business transaction data, human resource file data, and external market sentiment text data, cleaning and standardizing the multi-source heterogeneous data. Subsequently, the preprocessing module performs unified timestamp alignment on the business transaction data, human resource file data, and market sentiment text data, and extracts entity identifiers from the text to generate an initial set of fact triples containing employee entities, customer entities, and product entities. Based on the entity relationships in the initial set of fact triples, the graph construction module constructs the underlying knowledge graph skeleton, which consists of a node set and an edge set, where nodes represent various entities and edges represent the business interaction relationships between entities. The event fusion module uses natural language processing technology to perform semantic role labeling on the market sentiment text data, extracting negative and positive evaluation events involving product entities, and transforming them into event triples with sentiment polarity labels. The event fusion module integrates event triples with the underlying knowledge graph skeleton, updating the edge weights in the skeleton through an event-triggered mechanism to generate an enhanced knowledge graph containing dynamic business dynamics. The graph analysis module calculates relevant graph metrics based on the node connectivity and edge weight distribution in the enhanced knowledge graph, identifying potential cross-departmental collaboration links and isolated business nodes, ultimately generating a graph insight report to guide data analysis.

[0019] In one embodiment of the present invention, see [reference] Figure 2 Each entity in the initial set of fact triples is uniquely identified, and a mapping dictionary between entity names and identifiers is established. All triples in the initial set of fact triples are traversed to extract direct relationships between entities, and these relationships are categorized according to their type, generating multiple types of edge sets. For each type of edge set, the number of common neighbors of the node pairs connected by the corresponding edge is counted, and the confidence score of the corresponding edge type is calculated. The node set, the categorized edge set, and the corresponding confidence scores are imported into a graph database to construct the underlying knowledge graph skeleton that supports efficient adjacency queries. After the underlying knowledge graph skeleton is constructed, a background thread is started to verify the structural integrity of the underlying knowledge graph skeleton, removing redundant edges with confidence scores below a preset threshold.

[0020] In practical implementation, the initial set of fact triples generated by the preprocessing module is the starting point for the construction work. This initial set of fact triples contains fact records such as "Employee A, signed contract, Customer X", "Employee A, responsible for project, Project P", and "Customer X, purchase, Product Z". Each triple consists of a subject, a relation, and an object. The graph construction module assigns a unique identifier to each entity in the initial set of fact triples, assigning the identifier "EMP_001" to "Employee A", "CUS_100" to "Customer X", and "PROD_Z01" to "Product Z", and establishing a one-to-one mapping dictionary between entity names and identifiers. In practical implementation, the graph construction module traverses all triples in the initial set of fact triples, extracting direct relationships between entities. For example, from "Employee A, signed contract, Customer X", it extracts the "signed contract" relationship between "Employee A" and "Customer X". The graph construction module categorizes relationships based on their type, generating edge sets of various types such as "employment relationship", "contract relationship", "responsibility relationship", and "purchase relationship".

[0021] In some embodiments, for each of the multiple types of edge sets, the graph construction module counts the number of common neighbors among the node pairs connected by the corresponding type of edge. For example, for the edge set of the "contractual relationship" type, it counts the number of "project" nodes that the two ends of each edge connecting the "employee" node and the "customer" node have jointly participated in. The graph construction module calculates the confidence score of the corresponding type of edge, taking into account the number of common neighbors, the frequency of relationship occurrence, and the authority of the data source. An optional confidence score calculation formula is as follows: Where: characters Represents the calculated confidence score, character Represents the number of common neighbors, characters Represents the normalized frequency of occurrence of that specific relation in the data, character The character represents the authority rating of the data source that generated this relationship. , , The preset weighting coefficients are used to adjust the contribution of different factors. In practice, the graph construction module imports the encoded node set, the classified edge set, and the corresponding confidence scores calculated using the above formula into graph databases such as Neo4j or JanusGraph to construct the underlying knowledge graph skeleton that supports efficient adjacency queries.

[0022] Understandably, after the underlying knowledge graph skeleton is constructed, the system initiates a background thread to verify the structural integrity of the skeleton. This verification process includes checking the integrity of node attributes, edge attributes, and the logical consistency of connections between nodes and edges. The graph construction module performs a removal operation based on preset thresholds, such as edges with confidence scores below 0.3, removing redundant edges with confidence scores below the preset threshold from the underlying knowledge graph skeleton. After removing redundant edges, the structure of the underlying knowledge graph skeleton is more streamlined, and the relationship network is more reliable, providing a higher-quality foundation graph structure for subsequent event fusion and graph analysis. The entire implementation process starts from the original triples, and through systematic encoding, classification, quantitative evaluation, and filtering, ultimately forms a structured, queryable, and relationally reliable knowledge graph skeleton.

[0023] In one embodiment of the present invention, a pre-trained domain-adaptive language model is loaded, and this model is used to perform word segmentation and dependency parsing on market sentiment text data. Based on the results of dependency parsing, the predicate core word of the sentence and the argument components associated with the predicate core word are identified. The argument components include the agent, the patient, and adverbs of time and place. It is determined whether the patient points to a product entity in the underlying knowledge graph skeleton; if so, the current sentence is marked as the target evaluation sentence. A sentiment discrimination algorithm combining a sentiment lexicon and contextual information is used to calculate the sentiment tendency score of the predicate core word in the target evaluation sentence, and the sentiment polarity label is determined based on the positive or negative value of the sentiment tendency score. The agent, predicate core word, and patient are combined into structured event triples, and the sentiment polarity label is attached as an attribute to the edges of the event triples. See also... Figure 3 The process involves retrieving node pairs from the underlying knowledge graph skeleton that match the subject and object entities of the event triple. When a matching node pair is found, the weight values ​​of the existing edges connecting these two nodes are read from the underlying knowledge graph skeleton. The weight values ​​are adjusted based on the sentiment polarity label of the event triple: decreasing the weight value for a negative event and increasing it for a positive event. The adjusted weight values ​​are then written into the edge attribute fields of the underlying knowledge graph skeleton, completing the fusion update for a single event. If multiple events targeting the same node pair occur within a preset time window, the weighted average of the sentiment polarity scores of these events is calculated, and this weighted average is used as the final adjustment basis for a one-time batch update of the edge weights.

[0024] In its implementation, the event fusion module processes market sentiment text data containing sentences such as "Users generally report that the battery life of the Alpha smartphone is severely insufficient" or "Technology media reviews believe that the location function of the Beta smartwatch is very accurate." The event fusion module loads a pre-trained domain-adaptive language model, such as a BERT model further trained on consumer electronics industry commentary corpora, and uses this model to perform word segmentation and dependency parsing on the market sentiment text data. Based on the results of dependency parsing, the system identifies the predicate core word of the sentence and the argument components associated with it. These argument components include the agent, the patient, and adverbs of time and place. For example, for the sentence "Users generally report that the battery life of the Alpha smartphone is severely insufficient," dependency parsing will identify the predicate core word "insufficient," the agent as "battery life," the patient as implicit, and "Alpha smartphone" as a modifier of "battery life."

[0025] In some embodiments, the system determines whether the recipient or its associated entity points to a product entity in the underlying knowledge graph skeleton. If so, the current sentence is marked as the target evaluation sentence. In the example, "Smartphone Alpha" matches a product entity node in the graph, therefore the sentence is marked as the target evaluation sentence. The event fusion module uses a sentiment discrimination algorithm that combines a sentiment dictionary with contextual information to calculate the sentiment tendency score of the predicate core words in the target evaluation sentence. Optionally, the calculation of the sentiment tendency score can be expressed as: in: This represents the calculated sentiment score. As the core word of the predicate, Representative words Polarity scores in the basic sentiment lexicon Represents based on context Contextual sentiment scores calculated using an attention mechanism model It is a balancing factor between 0 and 1. Based on the sentiment tendency score... The positive or negative value determines the sentiment polarity label; for example, a score greater than 0 is labeled as positive, and a score less than 0 is labeled as negative. In specific implementation, the event fusion module combines the identified agent, predicate core word, and the product entity to which the recipient belongs into a structured event triple, such as battery life, insufficient, smartphone Alpha, and attaches the calculated sentiment polarity label as an attribute to the edges of the event triple.

[0026] Understandably, after event extraction, the event fusion module merges the event triples with the underlying knowledge graph skeleton. It retrieves node pairs in the underlying knowledge graph skeleton that match the subject and object entities of the event triples; for example, it retrieves the edge between "battery life" (which might map to an attribute node of the product "Smartphone Alpha") and the "Smartphone Alpha" product node. When a matching node pair is found, the event fusion module reads the weight values ​​of the original edges connecting these two nodes in the underlying knowledge graph skeleton. It adjusts the weight values ​​based on the sentiment polarity label of the event triples: decreasing the weight for negative events and increasing it for positive events. For example, if the original edge weight is 0.8, a negative event reduces the weight by 0.1, resulting in a new weight of 0.7; a positive event increases the weight by 0.05, resulting in a new weight of 0.85. The adjusted weight values ​​are then written into the edge attribute field of the underlying knowledge graph skeleton, completing the fusion update for a single event.

[0027] In some embodiments, if multiple events concerning the same node pair occur within a preset time window, a weighted average of the sentiment polarity scores of these events is calculated. For example, if three event records are generated within 24 hours regarding the "screen display" attribute of "Smartphone Alpha," with sentiment polarity scores of 0.5, -0.8, and -0.3 respectively, and their weights can be set to 0.7, 1.0, and 0.5 based on the authority of the data source, then the weighted average is... The event fusion module uses this weighted average as the final adjustment basis, performing a one-time batch update of edge weights to avoid frequent incremental operations. Through the above process, external unstructured public opinion texts are transformed into structured events, dynamically affecting the quantitative weights of relationships within the knowledge graph, enabling the enhanced knowledge graph to reflect changes in external market attitudes.

[0028] In one embodiment of the present invention, nodes with departmental attributes are selected from the enhanced knowledge graph, and nodes with the same departmental attributes are divided into different community subsets. The number of node connection paths between different community subsets is calculated, and the cross-departmental interaction activity index is calculated by combining the sum of the weights of all edges in the path. The interaction activity index is sorted, and the interaction paths with the highest ranking and meeting a preset frequency threshold are selected as potential collaboration links. The degree centrality and betweenness centrality indices of each node are calculated in the enhanced knowledge graph, and nodes whose degree centrality and betweenness centrality are both below the set lower limit threshold are identified and marked as isolated business nodes. The analysis results of potential collaboration links and isolated business nodes are encapsulated into structured data objects for subsequent graph insight reports.

[0029] In practical implementation, the graph analysis module filters nodes with departmental attributes from the enhanced knowledge graph, such as "R&D Department," "Marketing Department," and "Sales Department." Departmental attributes exist in the form of node labels or attribute fields. For example, the enhanced knowledge graph contains a node named "Employee Zhang San" with the departmental attribute "R&D Department," and a node named "Employee Li Si" with the departmental attribute "Marketing Department." The graph analysis module divides nodes with the same departmental attribute into different community subsets. For instance, all nodes with the departmental attribute "R&D Department" are grouped into community subset C1, all nodes with the departmental attribute "Marketing Department" into community subset C2, and all nodes with the departmental attribute "Sales Department" into community subset C3. In practical implementation, the graph analysis module calculates the number of node connection paths between different community subsets. A connection path refers to a path whose starting point and ending point belong to two different community subsets. Paths can pass through intermediate nodes, but the departmental attributes of intermediate nodes are not limited. For example, there is a path from the "Project P" node in community subset C1 to the "Customer K" node in community subset C2, consisting of an edge labeled "Responsible" and an edge labeled "Consulting." The graph analysis module calculates the cross-departmental interaction activity index by combining the sum of the weights of all edges in the path. The interaction activity index quantifies the frequency and strength of the interaction between two departmental subsets.

[0030] In some embodiments, the cross-departmental interaction activity index can be calculated using the following formula: in: Representatives from community subsets To community subset Interaction activity index Represents connecting a subset of communities and community subset The set of all pathways, Representative set One of the specific pathways in Representative pathway Path length, Representative pathway The set of all edges contained therein Representative set One of the edges in, Representing the side The weight value is calculated by penalizing long paths with the reciprocal of the path length and summing the weights of all edges on the path, thus favoring connections with higher edge weights and shorter paths. For example, if there are three paths between community subsets C1 and C2, with calculated interaction activity indices of 2.1, 1.5, and 0.8 respectively, the total interaction activity index is the sum of the three, which is 4.4. The graph analysis module sorts the interaction activity indices between all the calculated community subset pairs. The graph analysis module selects the top-ranked interaction paths that meet a preset frequency threshold as potential collaboration links. For example, if the preset frequency threshold requires an interaction activity index greater than 3.0, then the specific paths corresponding to the C1-C2 path with an interaction activity index of 4.4 and the C2-C3 path with an interaction activity index of 3.7 are selected as potential collaboration links.

[0031] It is understandable that in an enhanced knowledge graph, the degree centrality and betweenness centrality metrics for each node are calculated. Degree centrality refers to the number of edges directly connected to a node; for a weighted graph, it is calculated as the sum of edge weights. Betweenness centrality refers to the proportion of times a node appears on the shortest path between other nodes. The graph analysis module identifies nodes whose degree centrality and betweenness centrality are both below a set lower threshold. For example, if the lower threshold for degree centrality is set to 2 and the lower threshold for betweenness centrality is set to 0.01, and the node "Server S1" has a degree centrality of 1 (connected to only one administrator node) and a betweenness centrality of 0, then this node is marked as an isolated business node. In some embodiments, the graph analysis module encapsulates the analysis results of potential collaboration links and isolated business nodes into a structured data object. For example, it uses JSON format to record the starting department, ending department, path details, and interaction activity index of potential collaboration links, as well as the identifier, department, degree centrality value, and betweenness centrality value of isolated business nodes. This structured data object is output for subsequent graph insight reports.

[0032] Optionally, when calculating the interaction activity index, paths exceeding a certain length can be ignored; for example, only paths with a length of no more than 3 can be considered to focus on direct or near-direct inter-departmental interactions. Optionally, when marking isolated business nodes, filtering can be performed based on node attribute information; for example, this judgment can be applied only to "personnel" and "project" type nodes, but not to "document" type nodes. The graph analysis module is executed entirely based on the topological structure and edge weight data of the enhanced knowledge graph. Through quantified indices and indicators, it identifies high-value collaboration opportunities and potentially overlooked nodes, thereby providing graph-based insights for management decisions.

[0033] In one embodiment of the present invention, supply chain relationship edges involving supplier entities and material entities are extracted from an enhanced knowledge graph to form a dynamic supply chain topology view. Delivery timestamps for material procurement are retrieved from business transaction log data, and the time interval volatility between two adjacent delivery cycles is calculated. The time interval volatility is correlated and mapped with the credit ratings of supplier nodes in the supply chain topology view to construct a time series feature vector. Several sets of multi-dimensional feature templates for historical crisis scenarios are pre-stored in a pattern matching library. Each set of multi-dimensional feature templates includes a combination of time interval volatility, price volatility, and the number of logistics delays. A sliding window process is applied to the time series feature vector to generate feature sub-vectors for multiple consecutive time slices. The Euclidean distance metric between each feature sub-vector and all multi-dimensional feature templates is calculated. When the Euclidean distance metric between a feature sub-vector and a multi-dimensional feature template is less than a similarity threshold, the corresponding time slice is determined to be an abnormal time slice. Consecutive abnormal time slices are merged to form an abnormal volatility sequence, and the start and end times corresponding to the abnormal volatility sequence are recorded. Once an abnormal fluctuation sequence is identified, the supplier entity's node is marked as high-risk in the enhanced knowledge graph, and a traceability query link pointing to the supplier entity's node is automatically generated. Fund transfer records are extracted from business transaction log data and converted into financial triples representing fund flows. These financial triples are injected into the enhanced knowledge graph, and a graph traversal algorithm is used to perform multi-hop path searches from expenditure account nodes to revenue account nodes, with the number of hops limited to three. During each multi-hop path search, the monetary attribute values ​​of all edges on the path are accumulated and compared with a preset fund pool size. If the accumulated value of a multi-hop path exceeds the fund pool size, the multi-hop path is determined to be an abnormal fund loop. The path node sequence and monetary data of the abnormal fund loop are extracted to generate abnormal transaction cases for audit verification.

[0034] In its implementation, the risk warning module extracts supply chain relationship edges involving supplier entities and material entities from the enhanced knowledge graph, such as relationships like "Supplier S1, supplies, material M1" and "Supplier S2, supplies, material M2," forming a dynamic supply chain topology view. The risk warning module retrieves delivery timestamps related to material procurement from business transaction log data, such as the most recent delivery timestamp data for records of supplier S1 supplying material M1. The risk warning module calculates the volatility of the time interval between two adjacent delivery cycles, for example, by calculating the ratio of the standard deviation to the average of the six most recent delivery time intervals. See Table 1 for an example set of delivery records and calculation processes.

[0035] Table 1: Supplier S1 Material M1 Delivery Records and Time Interval Volatility Calculation Table In some embodiments, time interval volatility is correlated with the credit rating of supplier nodes in the supply chain topology view. For example, if supplier S1 has a credit rating of "A", a time series feature vector is constructed that includes time interval volatility, credit rating quantification score, and the number of delivery delays in the last three deliveries. Several sets of multi-dimensional feature templates for historical crisis scenarios are pre-stored in the pattern matching library. For example, "Historical Supply Disruption Crisis Feature Template A" includes a combination of [time interval volatility: above 0.6, price volatility: above 0.15, number of logistics delays: 2]. The risk warning module performs sliding window processing on the time series feature vector to generate feature sub-vectors for multiple consecutive time slices. For example, a feature sub-vector is generated using the last four deliveries as a time window. The Euclidean distance metric between each feature sub-vector and all multi-dimensional feature templates is calculated. The formula for calculating the Euclidean distance metric is expressed as: Where: characters Represents the calculated Euclidean distance metric, character These represent the values ​​of the feature subvector in three dimensions: time interval volatility, price volatility, and number of logistics delays, respectively. This represents the value of a multidimensional feature template in its corresponding dimension. When the Euclidean distance metric between a feature subvector and a multidimensional feature template is less than a similarity threshold, the corresponding time slice is determined to be an anomalous time slice. Consecutive anomalous time slices are merged to form an anomalous fluctuation sequence, and the start and end times of this sequence are recorded. For example, deliveries from the 4th to the 6th are considered an anomalous fluctuation sequence. Once an anomalous fluctuation sequence is identified, the node of supplier entity S1 is marked as high-risk in the enhanced knowledge graph, and a traceability query link pointing to the node of supplier entity S1 is automatically generated. This link displays all orders, contracts, and responsible personnel nodes related to that supplier.

[0036] It's understandable that the multi-hop reasoning and anomaly detection steps for enterprise financial data are independent of the risk warning process, but they are still based on an enhanced knowledge graph. Fund transfer records are extracted from business transaction data, such as "Account A, transfer 500,000 to Account B" and "Account B, transfer 300,000 to Account C," and these records are converted into financial triples representing fund flows. These financial triples are then injected into the enhanced knowledge graph, and a graph traversal algorithm is used to perform multi-hop path searches from expenditure account nodes to revenue account nodes. The number of hops in a multi-hop path is limited to three; for example, searching whether it's possible to return to account A from account A within three hops. During each multi-hop path search, the monetary attribute values ​​of all edges on the path are accumulated. For example, the accumulated value for the path "Account A -> Account B -> Account C -> Account A" is 500,000 + 300,000 + 200,000 = 1,000,000. If the accumulated value of a multi-hop path exceeds a preset fund pool size (e.g., a preset fund pool size of 800,000), this multi-hop path is determined to be an abnormal fund loop. Extract the path node sequence and amount data of abnormal funding loops to generate abnormal transaction cases for audit verification.

[0037] Optionally, more dimensions can be introduced when constructing time-series feature vectors, such as the order amount change rate or supplier concentration index. Optionally, more complex filtering conditions can be set when performing multi-hop path finding, such as ignoring edges with amounts below a specific threshold, or requiring nodes in the path to belong to different entity types. The risk warning module and the multi-hop inference steps share an enhanced knowledge graph as their data foundation, but they are analyzed from two different dimensions: supply chain stability and cash flow security. Together, they expand the depth of knowledge graph applications in risk insight.

[0038] See Figure 4This chart presents the core risk characteristics of the supply chain data preprocessing stage. The bar chart represents the interval between deliveries (days), and the line chart represents the time interval volatility (σ), corresponding to the left and right vertical axes respectively, with the horizontal axis representing the delivery order. From a data perspective, the delivery interval exhibits significant non-stationary fluctuations from the 2nd to the 6th deliveries: the 2nd and 4th delivery intervals are both 26 days, the 3rd is 38 days, the 5th reaches a peak of 45 days, and the 6th drops sharply to 11 days. This drastic interval fluctuation directly drives the increase in time interval volatility, which reaches 0.632 in the 6th delivery, far exceeding the previous stable level of 0.316. This trend highly matches the multi-dimensional feature template in the risk warning module, which states that "a time interval volatility exceeding 0.6 triggers a high-risk judgment," indicating that supplier S1 has entered an abnormal fluctuation range in the delivery of material M1. From an analytical perspective, this chart, through a dual-axis linked visualization method, correlates the absolute value of the delivery interval with the relative degree of volatility, providing an intuitive basis for subsequently constructing time series feature vectors. In the pattern matching stage, the consecutive abnormal time slices formed by the 4th to 6th deliveries can be directly compared with historical supply disruption crisis feature templates using Euclidean distance calculation. This allows supplier S1 to be marked as a high-risk node in the enhanced knowledge graph, generating a traceability query link. This deep coupling between visualization and algorithmic logic demonstrates the application value of knowledge graphs in supply chain risk insight.

[0039] In one embodiment of the invention, new data changes in the enterprise data interface are monitored. When changes in business transaction log data or human resource file data are detected, incremental facts are extracted from the changed data. The incremental facts are then logically consistent with the current version of the enhanced knowledge graph to identify any attribute value conflicts or relationship contradictions. For conflicts involving only attribute value differences, the conflict is resolved by overwriting older data with the latest timestamp. For conflicts involving relationship contradictions, the incremental fact insertion process is paused, and a secondary data verification request is initiated for the point of conflict. After passing the secondary data verification request, the disambiguated incremental facts are appended as new triples to the enhanced knowledge graph, and the relevant index structure is updated. A set of analytical primitives is defined to describe the analytical tasks, including data extraction primitives, graph query primitives, and result rendering primitives. User drag-and-drop operations of analytical primitives through a graphical interface are received, generating a directed acyclic graph (DAG) composed of these primitives as an executable workflow. The input and output parameters of each node in the executable workflow are parsed to establish data dependencies between nodes. The query interface of the enhanced knowledge graph is encapsulated as a standard operator in the analysis primitives, and the calling permissions of the standard operator are configured. When a user triggers the execution of an executable workflow, each analysis primitive is scheduled sequentially according to the topological order of the directed acyclic graph, and the final execution result is sent back to the graph insight report.

[0040] In practice, the anomaly detection module monitors new data changes in enterprise data interfaces, such as the data update interface of the human resources system or the transaction log change log of the enterprise resource planning system. When a change is detected in business transaction data or human resources file data, such as the human resources system adding a record "Employee Zhao Liu, position changed to Project Manager", the anomaly detection module extracts the incremental facts from the changed data. The anomaly detection module performs a logical consistency check on the incremental facts and the current version of the enhanced knowledge graph to identify whether there are attribute value conflicts or relationship contradictions. For example, it checks whether the current position attribute of the "Employee Zhao Liu" node in the enhanced knowledge graph is "Senior Engineer", while the incremental facts show that its position has changed to "Project Manager", which constitutes an attribute value conflict; or, for example, the enhanced knowledge graph shows that "Employee Zhao Liu" belongs to the "R&D Department", while the incremental facts attempt to establish the relationship "Employee Zhao Liu, belongs to the Sales Department", which constitutes a relationship contradiction.

[0041] In practice, for conflicts where only attribute values ​​differ, the system resolves them by overwriting older data with the most recent timestamp. For example, in the enhanced knowledge graph, the timestamp for the job attribute "Senior Engineer" of "Employee Zhao Liu" is 2025-12-01, while the timestamp for the newly extracted incremental fact "Employee Zhao Liu, Position, Project Manager" is 2026-01-15. The system will overwrite "Senior Engineer" with "Project Manager". For conflicts involving contradictory relationships, the system pauses the addition of incremental facts and initiates a secondary data verification request for the point of conflict. For example, it sends a query request to the authoritative data source in the human resources system to confirm the department to which "Employee Zhao Liu" currently belongs. After passing the secondary data verification request, confirmation information from the authoritative data source is obtained. The anomaly detection module then appends the disambiguated incremental fact as a new triple to the enhanced knowledge graph and updates the relevant index structure, such as updating the inverted index based on employee name.

[0042] The visualization configuration module defines a set of analytical primitives to describe analytical tasks. These primitives include data extraction primitives, graph query primitives, and result rendering primitives. Data extraction primitives are used to retrieve raw data from specified data tables or interfaces; graph query primitives are used to send query statements to the enhanced knowledge graph, such as Cypher or Gremlin queries; and result rendering primitives are used to visualize the query results in the form of charts or tables. The visualization configuration module receives user drag-and-drop operations of analytical primitives through the graphical interface. For example, a user can drag three analytical primitives—"Extract Department Employee List," "Query Employee Collaboration Projects," and "Render Project Network Diagram"—from the left component library to the canvas and specify the execution order through connecting lines, generating a directed acyclic graph (DAG) composed of these analytical primitives as an executable workflow. The visualization configuration module parses the input and output parameters of each node in the executable workflow, establishing data dependencies between nodes. For example, the input of the "Query Employee Collaboration Projects" node is the output of the "Extract Department Employee List" node.

[0043] In some embodiments, the query interface of the enhanced knowledge graph is encapsulated as a standard operator in the analysis primitives, and the calling permissions of the standard operator are configured, such as restricting certain user roles to only using the graph query template. When a user triggers the execution of an executable workflow, the visualization configuration module schedules each analysis primitive sequentially according to the topological order of the directed acyclic graph. First, it executes "Extract Department Employee List" to obtain the employee list, then passes the list as a parameter to "Query Employee Collaboration Projects" to perform a graph query, and finally passes the queried project network data to "Render Project Network Graph" to generate a visualization graph. Specifically, for conflicts with contradictory relationships, the resolution strategy can be formally expressed as a decision function. An optional decision function is used to determine whether to initiate secondary verification: Where: characters A boolean result indicating whether to initiate a secondary data verification request; character. Represents the number of detected relational conflicts, characters Represents the time delay from the detection of the contradiction to the current moment, character It is a preset time tolerance threshold. When When True, the system initiates a secondary data verification request. Optionally, the visualization configuration module can perform static syntax checks when parsing the executable workflow, such as checking whether the data types of the connections match. The visualization configuration module then sends the final execution results, such as the rendered project network diagram and its associated data, back to the graph insight report, forming an analysis section that users can view and interact with.

[0044] See Figure 5In the performance analysis of the analysis primitives workflow during the visualization configuration phase, the system quantitatively evaluated the execution efficiency and reliability of three types of core analysis primitives. From the perspective of execution time (bar chart), different primitive types exhibited significant performance differences: data extraction primitives had the highest execution time, reaching approximately 1.9 seconds, mainly due to the need for batch data retrieval and format conversion from multiple heterogeneous data sources; graph query primitives had relatively low execution time, approximately 0.85 seconds, thanks to the efficient optimization of adjacency queries by the enhanced knowledge graph's underlying skeleton; result rendering primitives had a moderate execution time, approximately 1.27 seconds, with the time primarily consumed in the layout calculation and rendering pipeline for converting query results into visual graphics. From the perspective of execution success rate (line graph), all three types of primitives maintained high reliability: graph query primitives had the highest success rate, reaching approximately 99.2%, which is closely related to their encapsulation as standardized operators and relatively fixed calling logic; data extraction primitives had the second highest success rate, at approximately 98.5%, with fluctuations mainly stemming from the availability of external data sources and network latency; result rendering primitives had a success rate of 97.8%, with their main risks concentrated in compatibility adaptation scenarios for complex graph layouts. Overall, the performance at this stage reflects the design trade-off of the analysis primitive workflow: data extraction primitives bore the heaviest IO load, while graph query primitives achieved efficient and stable execution thanks to the indexing advantages of graph databases, and result rendering primitives achieved a balance between visual richness and execution efficiency.

[0045] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A knowledge graph-based enterprise data intelligent analysis system, characterized in that, include: The preprocessing module acquires internal business transaction data, human resource file data, and external market sentiment text data. It then performs unified timestamp alignment and entity identifier extraction on the business transaction data, human resource file data, and market sentiment text data to generate an initial set of fact triples containing employee entities, customer entities, and product entities. The graph construction module constructs a lower-level knowledge graph skeleton based on the entity association relationships in the initial set of fact triples. The lower-level knowledge graph skeleton consists of a node set and an edge set, where nodes represent entities and edges represent business interaction relationships between entities. The event fusion module uses natural language processing technology to perform semantic role labeling on the market sentiment text data, extracts negative and positive evaluation events related to the product entity, and transforms them into event triples with sentiment polarity labels. The event triples are then fused with the underlying knowledge graph skeleton, and the edge weights in the underlying knowledge graph skeleton are updated through an event triggering mechanism to generate an enhanced knowledge graph containing dynamic business situations. The graph analysis module identifies potential cross-departmental collaboration links and isolated business nodes based on the node connection density and edge weight distribution in the enhanced knowledge graph, and generates a graph insight report to guide the data analysis perspective. 2.The enterprise data intelligent analysis system based on knowledge graph of claim 1, wherein, Based on the entity relationships in the initial set of fact triples, a basic knowledge graph skeleton is constructed, including: Each entity in the initial set of fact triples is encoded with a unique identifier, and a mapping dictionary between entity names and identifiers is established. Traverse all triples in the initial set of fact triples, extract the direct relationships between entities, and classify them according to the type of relationship to generate various types of edge sets; For each of the various types of edge sets, count the number of common neighbors of the node pairs connected by the edge of the corresponding type, and calculate the confidence score of the edge of the corresponding type. The node set, the classified edge set, and the corresponding confidence scores are imported into the graph database to construct the underlying knowledge graph skeleton that supports efficient adjacency queries. After the underlying knowledge graph skeleton is constructed, a background thread is started to verify the structural integrity of the underlying knowledge graph skeleton and remove redundant edges with confidence scores lower than a preset threshold. 3.The enterprise data intelligent analysis system based on knowledge graph of claim 2, characterized in that, The process involves using natural language processing technology to perform semantic role labeling on the market sentiment text data, extracting negative and positive evaluation events related to the product entity, and transforming them into event triplets with sentiment polarity labels, including: Load a pre-trained domain-adaptive language model, and use the domain-adaptive language model to perform word segmentation and dependency parsing on the market sentiment text data; Based on the results of the dependency parsing, the predicate core word of the sentence and the argument components associated with the predicate core word are identified. The argument components include the agent, the patient, and the adverbial of time and place. Determine whether the subject refers to the product entity in the underlying knowledge graph skeleton; if so, mark the current sentence as the target evaluation sentence. A sentiment discrimination algorithm combining a sentiment dictionary and contextual information is used to calculate the sentiment tendency score of the predicate core word in the target evaluation sentence, and the sentiment polarity label is determined based on the positive or negative value of the sentiment tendency score. The agent, the predicate core word, and the patient are combined into a structured event triple, and the sentiment polarity label is attached as an attribute to the edge of the event triple. 4.The enterprise data intelligent analysis system based on knowledge graph of claim 3, characterized in that, The event triples are fused with the underlying knowledge graph skeleton, and the edge weights in the underlying knowledge graph skeleton are updated through an event triggering mechanism, including: Retrieve node pairs in the underlying knowledge graph skeleton that match the subject entity and object entity of the event triple; When a matching pair of nodes is found, the weight values ​​of the original edges connecting the two nodes in the underlying knowledge graph skeleton are read. The weight value is adjusted based on the emotional polarity label of the event triple. If it is a negative evaluation event, the weight value is reduced, and if it is a positive evaluation event, the weight value is increased. The adjusted weight values ​​are written into the edge attribute field of the underlying knowledge graph skeleton to complete the fusion update of a single event; If multiple events targeting the same node pair occur within a preset time window, the weighted average of the sentiment polarity scores of these events is calculated, and the edge weights are updated in a one-time batch based on the weighted average as the final adjustment basis. 5.The enterprise data intelligent analysis system based on knowledge graph of claim 4, characterized in that, Based on the node connectivity and edge weight distribution in the enhanced knowledge graph, potential cross-departmental collaboration links and isolated business nodes are identified, including: In the enhanced knowledge graph, nodes with departmental attributes are selected, and nodes with the same departmental attributes are divided into different community subsets; Calculate the number of node connection paths between different community subsets, and combine the sum of the weights of all edges in the path to calculate the cross-departmental interaction activity index; The interaction activity index is sorted, and the interaction paths that rank high and meet the preset frequency threshold are selected as the potential collaboration links. In the enhanced knowledge graph, the degree centrality and betweenness centrality indices of each node are calculated, and nodes whose degree centrality and betweenness centrality are both below a set lower threshold are identified and marked as the isolated business nodes. The analysis results of the potential collaborative links and the isolated business nodes are encapsulated into structured data objects for subsequent graph insight reports. 6.The enterprise data intelligent analysis system based on knowledge graph according to claim 5, characterized in that, It also includes a risk warning module, used for: Extract supply chain relationship edges involving supplier entities and material entities from the enhanced knowledge graph to form a dynamic supply chain topology view; Retrieve the delivery timestamps for material procurement from the business transaction log data and calculate the time interval volatility between two adjacent delivery cycles; The time interval volatility is correlated and mapped with the credit rating of the supplier nodes in the supply chain topology view to construct a time series feature vector; The time series feature vector is input into a preset pattern matching library to identify abnormal fluctuation sequences that match the characteristics of historical supply disruption crises. Once the abnormal fluctuation sequence is identified, the node of the supplier entity is marked as high-risk in the enhanced knowledge graph, and a traceability query link pointing to the node of the supplier entity is automatically generated. 7.The enterprise data intelligent analysis system based on knowledge graph of claim 6, wherein, The time series feature vector is input into a preset pattern matching library to identify abnormal fluctuation sequences that match the characteristics of historical supply disruption crises, including: The pattern matching library pre-stores several sets of multi-dimensional feature templates for historical crisis scenarios. Each set of multi-dimensional feature templates includes a combination of time interval volatility, price volatility, and the number of logistics delays. The time series feature vector is processed by a sliding window to generate feature sub-vectors for multiple consecutive time slices; Calculate the Euclidean distance metric between each of the said feature sub-vectors and all of the said multidimensional feature templates; When the Euclidean distance between a certain feature subvector and a certain specific multidimensional feature template is less than the similarity threshold, the corresponding time slice is determined to be an abnormal time slice; The consecutively occurring abnormal time slices are merged to form the abnormal fluctuation sequence, and the start and end times corresponding to the abnormal fluctuation sequence are recorded.

8. The enterprise data intelligent analysis system based on knowledge graphs according to claim 7, characterized in that, It also includes multi-hop reasoning and anomaly detection steps for corporate financial data: Extract fund transfer records from the business transaction flow data and convert the fund transfer records into financial triples representing fund flows; The financial triples are injected into the enhanced knowledge graph, and a graph traversal algorithm is used to perform multi-hop path lookup from the expenditure account node to the income account node. The number of hops in the multi-hop path is limited to three. During each multi-hop path search, the monetary attribute values ​​of all edges on the path are accumulated and compared with the preset fund pool size. If the cumulative amount of a certain multi-hop path is found to exceed the size of the fund pool, the multi-hop path is determined to be an abnormal fund loop. Extract the path node sequence and amount data of the abnormal funding loop to generate abnormal transaction cases for audit verification.

9. The enterprise data intelligent analysis system based on knowledge graphs according to claim 8, characterized in that, It also includes an anomaly detection module, used for: Monitor new data changes in the enterprise data interface, and when changes are detected in the business transaction data or human resource file data, extract the incremental facts from the changed data; Perform a logical consistency check between the incremental facts and the current version of the enhanced knowledge graph to identify whether there are attribute value conflicts or relationship contradictions. For conflicts where only attribute values ​​differ, the conflict is resolved by overwriting the older data with the newest timestamp. For conflicts involving contradictory relationships, suspend the entry of incremental facts into the database and initiate a secondary data verification request targeting the points of contradiction. After the secondary data verification request is approved, the disambiguated incremental facts are added as new triples to the enhanced knowledge graph, and the relevant index structure is updated.

10. The enterprise data intelligent analysis system based on knowledge graphs according to claim 9, characterized in that, It also includes a visual configuration module for: Define a set of analysis primitives for describing analysis tasks, including data extraction primitives, graph query primitives, and result rendering primitives; The system receives user input by dragging and dropping the analysis primitives through a graphical interface, and generates a directed acyclic graph composed of the analysis primitives, which serves as the executable workflow. The input and output parameters of each node in the executable workflow are analyzed to establish data dependencies between nodes; The query interface of the enhanced knowledge graph is encapsulated as a standard operator in the analysis primitives, and the calling permissions of the standard operator are configured. When a user triggers the execution of the executable workflow, each analysis primitive is scheduled sequentially according to the topological order of the directed acyclic graph, and the final execution result is sent back to the graph insight report.

Images