A multi-format enterprise-oriented global data knowledge fusion management method

Abstract

This invention relates to the fields of data processing and knowledge graph technology, and discloses a method for comprehensive data knowledge fusion and governance across multiple business sectors within an enterprise. The invention monitors data change event streams to extract the identifiers and attribute values ​​of changed entities, formalizes cross-business linkage rules into a logical constraint formula network, uses the changed entity identifier as the trigger variable, and derives the set of affected related entities through a constraint propagation algorithm. It then extracts the minimum connected subgraph containing the changed entity and related entities from the comprehensive knowledge graph; freezes read / write permissions for the remaining areas; performs knowledge embedding calculations and constraint consistency checks within the minimum connected subgraph; and synchronously updates features and releases the frozen lock after successful verification. This invention introduces business rules into the incremental update scope definition stage, overcoming the shortcomings of conventional local updates that cause cross-business nodes to violate business linkage constraints, maintaining the logical consistency of the knowledge graph, and reducing computational consumption.

Description

Technical Field

[0001] This invention relates to the field of data processing and knowledge graph technology, and discloses a method for the comprehensive data knowledge fusion and governance of enterprises across multiple business formats. Background Technology

[0002] For enterprise-wide knowledge graph updates across multiple business sectors, a common approach is to employ local reconstruction techniques based on timestamps or change quantity thresholds. When the system receives a data write request from a multi-business front-end, it records a timestamp for the write operation. The system continuously accumulates timestamped data change events. When the accumulated number of change events reaches a preset threshold, or when the time interval since the last update operation reaches a preset time window, a local update process is triggered. In this process, the system extracts the knowledge graph nodes corresponding to these change events, re-runs the knowledge embedding computation model within a fixed local scope or directly on the entire graph, and after computation, overwrites the original node features with the generated new feature vectors, completing the physical update of the knowledge graph.

[0003] The aforementioned triggering mechanisms based on timestamps or change quantity thresholds suffer from a core technical problem: the failure of cross-business linkage constraints. This conventional approach determines the timing and scope of updates solely based on the time or quantity of data changes, neglecting the complex business logic coupling relationships between multiple business sectors. In multi-business scenarios, data changes in a single sector often require business linkage rules to constrain related data in other sectors. Existing technologies, when triggering partial updates, fail to identify and incorporate related nodes from other sectors constrained by business linkage rules; they only reconstruct the features of physically changed nodes. This leads to logical contradictions between the updated node features and the unupdated cross-business related node features, causing the business constraints of the entire knowledge graph to fail. Summary of the Invention

[0004] The purpose of this invention is to provide a comprehensive data knowledge fusion and governance method for enterprises with multiple business formats, which can solve the problems mentioned in the background technology.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A comprehensive data knowledge fusion governance method for enterprises with multiple business formats includes: monitoring the data change event stream of multiple business formats and extracting the change entity identifier and change attribute value; The cross-business linkage rules are formalized into a logical constraint formula network. The changed entity identifier is used as a trigger variable and input into the logical constraint formula network. The set of related entities affected by the changed event is derived through the constraint propagation algorithm. In the global knowledge graph, starting from the changed entity and ending with the entities in the set of associated entities, a minimum connected subgraph containing the starting point and the ending point is extracted using breadth-first traversal. Freeze read and write permissions in the global knowledge graph except for the minimum connected subgraph, and re-execute knowledge embedding calculation and constraint consistency verification within the minimum connected subgraph; After the verification is passed, the updated minimum connected subgraph nodes and edge features are synchronized to the main copy of the global knowledge graph, and the freeze lock is released.

[0006] Preferably, the step of monitoring the multi-business data change event stream and extracting the change entity identifier and change attribute value includes: deploying a multi-level cache queue between the multi-business front-end data source and the global knowledge graph, wherein the multi-level cache queue is divided into independent cache shards according to the business category; When the multi-level cache queue receives a data write request, it calculates the first hash value of the current attribute value of the entity corresponding to the write request, and compares the first hash value with the second hash value of the corresponding entity in the global knowledge graph. If the first hash value does not match the second hash value, a data change event is determined to have occurred. The change entity identifier and the change attribute value with hash difference are parsed from the write request. The change entity identifier and the change attribute value are encapsulated into a change event message and stored in the cache fragment of the corresponding business type.

[0007] Preferably, the step of formalizing the cross-business linkage rules into a logical constraint formula network, inputting the changed entity identifier as a trigger variable into the logical constraint formula network, and deriving the set of related entities affected by the change event through a constraint propagation algorithm includes: parsing the entity dependency relationship in the business linkage rules, and constructing a directed acyclic graph with entity identifiers as nodes and dependency relationships as directed edges as the logical constraint formula network; The changed entity identifier is injected into the directed acyclic graph as the initial activation node. The arc consistency algorithm is executed, and the changed attribute value domain corresponding to the changed entity identifier is used as a constraint condition to propagate to the downstream nodes along the directed edge. When the constraint received by a downstream node in the directed acyclic graph causes its own attribute value range to be reduced, the downstream node is marked as an active node, and the entity identifiers corresponding to all the active nodes after traversal are aggregated into the associated entity set.

[0008] Preferably, the step of extracting the minimum connected subgraph containing the starting point and the ending point in the global knowledge graph by using breadth-first traversal with the changed entity as the starting point and the entities in the associated entity set as the ending point includes: locating the first starting node corresponding to the changed entity and the multiple target nodes corresponding to the associated entity set in the global knowledge graph. A breadth-first traversal is initiated with the first starting node as the root node. During the traversal, the edge relationship type weight between the current traversed node and its adjacent nodes is calculated. If the edge relationship type weight is lower than the preset business coupling threshold, the branch is pruned and the traversal of the branch is blocked. Continue traversing until all target nodes have been visited, extract all nodes traversed along the traversal path and all unpruned connecting edges, and combine all extracted nodes and connecting edges to generate the minimum connected subgraph.

[0009] Preferably, freezing the read and write permissions of the global knowledge graph except for the minimum connected subgraph, and re-performing the knowledge embedding calculation and constraint consistency verification within the minimum connected subgraph, includes: using a multi-version concurrency control mechanism to allocate a globally incrementing timestamp to the main copy of the global knowledge graph, and marking the nodes and edges except for the minimum connected subgraph as read-only; Within the minimum connected subgraph, the feature vectors of the nodes corresponding to the changed entity are initialized, and the node features within the minimum connected subgraph are updated by message passing using a local graph attention network. After completing a preset number of message transmissions, the updated feature vectors of all nodes in the minimum connected subgraph are extracted. The updated feature vectors are then input into a pre-trained constraint judgment model to verify whether the relative distance between the updated feature vectors satisfies the business constraint conditions defined by the logical constraint formula network.

[0010] Preferably, after the verification passes, synchronizing the updated nodes and edge features of the minimum connected subgraph to the main copy of the global knowledge graph and releasing the freeze lock includes: when the constraint consistency verification passes, constructing a shadow copy of the minimum connected subgraph in memory and writing the verified nodes and edge features into the shadow copy; Establish a pointer mapping relationship between the shadow copy and the main copy of the global knowledge graph, and redirect the old pointer of the corresponding minimum connected subgraph in the main copy to the memory address of the shadow copy through atomic operations; After the pointer redirection is completed, the read-only status markers of nodes and edges in the main replica, except for the minimum connected subgraph, are cleared, triggering the garbage collection process of the multi-version concurrency control mechanism to release the storage space occupied by the old version data, and completing the release of the frozen lock.

[0011] Preferably, the step of encapsulating the changed entity identifier and the changed attribute value into a change event message and storing it in the cache shard of the corresponding business type includes: configuring an independent time sliding window and an event aggregation counter for each cache shard; When the change event message is written to the cache shard, the time sliding window corresponding to the cache shard is reset, and the value of the event aggregation counter is incremented by one; If the value of the event aggregation counter reaches the preset aggregation threshold within the time sliding window period, the batch output of all the change event messages in the cache fragment is triggered in advance. If the time sliding window times out and the value of the event aggregation counter does not reach the aggregation threshold, then all the current change event messages in the cache fragment will be forcibly output to the next processing stage, and the event aggregation counter will be cleared.

[0012] Preferably, the execution arc consistency algorithm, which propagates the changed attribute value range corresponding to the changed entity identifier as a constraint condition to downstream nodes along the directed edge, includes: extracting the constraint condition set corresponding to all incoming edges of the currently active node in the directed acyclic graph, wherein the constraint condition set contains attribute value range intervals connected by logical operators; For each attribute dimension of the currently active node, traverse all constraints related to that attribute dimension in the constraint set and calculate the intersection interval of all related constraints; If the intersection interval is an empty set, then it is determined that a constraint propagation conflict has occurred and the current process is terminated; If the intersection interval is a proper subset of the original attribute value range of the current active node, then the original attribute value range of the current active node is replaced by the intersection interval, and the outgoing edges of the current active node are added to the propagation queue for further iterative processing.

[0013] Preferably, the initialization of the feature vector of the node corresponding to the changed entity and the message passing update of the node features in the minimum connected subgraph using a local graph attention network include: obtaining the changed attribute value of the node corresponding to the changed entity and mapping the changed attribute value to a high-dimensional feature vector through a multilayer perceptron to replace the initial feature vector of the node corresponding to the changed entity. In the single-layer message passing of the local graph attention network, the central node and its first-order neighbor nodes in the minimum connected subgraph are used as computational units. The feature vector of the central node and the feature vector of the first-order neighbor node are concatenated and then input into the shared linear transformation layer. The attention weight coefficients of the first-order neighbor nodes relative to the center node are calculated using the softmax normalization function. The feature vectors of the first-order neighbor nodes are then weighted and summed using these attention weight coefficients to generate the updated feature vector of the center node.

[0014] Preferably, if the intersection interval is an empty set, then determining that a constraint propagation conflict has occurred and terminating the current process includes: recording the target attribute dimension that generates the empty set intersection interval and the target constraint condition that triggers the conflict, and generating an abnormal alarm containing the conflict source node identifier and the conflict constraint identifier. The abnormal alarm is sent to a preset conflict degradation processing queue. The preset fault tolerance strategy corresponding to the target constraint is queried in the conflict degradation processing queue. The preset fault tolerance strategy includes relaxing the value range boundary or skipping the target constraint. Modify the constraint parameters of the target constraint in the directed acyclic graph according to the preset fault tolerance strategy found in the query, and re-trigger the constraint propagation process starting from the conflict source node. If the number of retries exceeds the preset retry limit and there is still an empty set intersection interval, then remove the change event corresponding to the changed entity identifier from the multi-business data change event stream.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention formalizes cross-business linkage rules into a logical constraint formula network and uses a constraint propagation algorithm to deduce the set of related entities affected by change events along directed edges, changing the triggering logic of existing technologies that rely on time or quantity dimensions. This approach directly introduces business rules into the triggering and scope definition stages of incremental updates, ensuring that cross-business related nodes affected by changes are fully included in the update process. It overcomes the defect that cross-business knowledge node attributes violate business linkage constraints after partial updates and maintains the logical consistency of the entire knowledge graph during incremental updates.

[0016] 2. This invention filters invalid change events through multi-level cache queues and hash comparisons, and performs low-coupling branch pruning by calculating the edge relationship type weights during breadth-first traversal, thereby reducing the size of the minimum connected subgraph participating in knowledge embedding computation. It also employs a multi-version concurrency control mechanism to freeze non-updating regions and performs data synchronization by combining shadow copy construction and pointer mapping atomic operations, avoiding the computational overhead of full graph reconstruction, reducing system resource consumption during incremental fusion, and ensuring the service availability of the full-domain knowledge graph during continuous updates. Attached Figure Description

[0017] Figure 1 This is the main flowchart of the full-domain data knowledge fusion governance method of the present invention; Figure 2This is a flowchart of the data change event monitoring and extraction sub-process of the present invention; Figure 3 This is a flowchart illustrating the constraint propagation and associated entity derivation sub-flowchart of the present invention; Figure 4 This is a flowchart of the minimum connected subgraph extraction process of the present invention; Figure 5 This is a flowchart of the knowledge embedding calculation and consistency verification sub-process of the present invention; Figure 6 This is a flowchart of the synchronization and lock release sub-process of the present invention. Detailed Implementation

[0018] Please refer to the attached document. Figure 1 This embodiment provides a method for the fusion and governance of full-domain data knowledge for multiple business formats of an enterprise, which is deployed on an enterprise-level distributed data governance platform. This platform connects to the front-end business systems of the enterprise's retail, supply chain, finance, and human resources business formats. The full-domain knowledge graph is stored in a distributed graph database. The graph database supports the attribute graph model. Each node corresponds to a business entity, and each edge corresponds to the business relationship between entities. Both nodes and edges store corresponding feature vectors and attribute key-value pairs.

[0019] In this embodiment, the front-end business system of the retail sector generates business data related to products, stores, members, and sales orders; the front-end business system of the supply chain sector generates business data related to suppliers, inventory, purchase orders, and logistics orders; the front-end business system of the finance sector generates business data related to accounting vouchers, accounting subjects, accounts payable, and accounts receivable; and the front-end business system of the human resources sector generates business data related to employees, departments, positions, and payroll calculation. There are cross-sector business linkage rules and data dependencies among the business data of each sector.

[0020] The platform utilizes a data change monitoring component deployed between the front-end business systems of various business sectors and the distributed graph database to monitor the data change event stream across multiple business sectors in real time, extracting the identifiers of changed entities and the values ​​of changed attributes. The data change monitoring component connects to the databases of the front-end business systems of each business sector. The data acquisition module and the business interface write proxy module establish a long connection to receive event streams corresponding to data write, update, and delete operations generated by various business systems in real time. During the reception of the event stream, the data change monitoring component parses each received event message and extracts the entity's unique identifier, the business category corresponding to the entity, and all key-value pairs of the entity's attributes contained in the message.

[0021] Furthermore, the data change monitoring component generates a query request for each extracted entity unique identifier and sends it to the distributed graph database corresponding to the global knowledge graph. It obtains the current attribute key-value pairs stored in the graph database for the node corresponding to the entity unique identifier. It compares the attribute key-value pairs in the received event message with the attribute key-value pairs stored in the graph database field by field, identifies the attribute fields with numerical differences, marks the entity unique identifier as the changed entity identifier, and encapsulates the field name and field value corresponding to the attribute fields with numerical differences into the changed attribute value.

[0022] In this embodiment, the basic specifications of each business entity are uniformly defined through the table below, providing a standardized comparison benchmark for the identification and extraction of change events.

[0023] Table 1 Definition of Enterprise Multi-Business Entities and Core Attributes

[0024] The table above defines the basic specifications of the business entities corresponding to the front-end data sources of each business type in this embodiment. The entity unique identifier coding rule provides a unified standard for the unique identification of cross-business type entities. The core attribute fields provide a comparison benchmark for the identification and extraction of change events, ensuring that change events of different business types can be uniformly parsed and processed, and avoiding change identification errors caused by cross-business type entity identifier conflicts.

[0025] Furthermore, the platform formalizes cross-business linkage rules into a logical constraint formula network, inputting the changed entity identifier as a trigger variable into the logical constraint formula network, and deriving the set of related entities affected by the change event through a constraint propagation algorithm. The platform pre-collects cross-business linkage rules within the enterprise, including cross-business data verification rules, business process dependency rules, and data standardization rules defined by the enterprise's internal management system. Semantic parsing is performed on each business linkage rule to extract the entity type, entity attributes, logical constraint relationships between attributes, and dependency directions between entities.

[0026] Based on the entity dependencies obtained from the parsing, a directed acyclic graph is constructed with entity identifiers as nodes and dependencies as directed edges. This directed acyclic graph is the logical constraint formula network, where each directed edge corresponds to a set of logical constraint formulas. The logical constraint formulas define the constraint relationship between the attribute value domain of the starting node and the attribute value domain of the ending node of the edge.

[0027] In this embodiment, the extracted change entity identifier is used as a trigger variable and mapped to an initial activation node in the logical constraint formula network. The value range of the change attribute corresponding to the initial activation node is used as the initial constraint condition. The constraint condition is propagated layer by layer along the directed edges in the directed acyclic graph through the constraint propagation algorithm. During the propagation process, for each node that receives the constraint condition, it is determined whether the constraint condition affects the attribute value range of the node. If it does, the node is marked as an affected node. After the constraint propagation algorithm has completed its traversal, the entity identifiers corresponding to all marked affected nodes are aggregated to generate a set of associated entities affected by the change event.

[0028] Furthermore, within the global knowledge graph, the platform uses breadth-first traversal to extract the minimum connected subgraph containing the starting and ending points, starting with the changed entity and ending with entities in the associated entity set. In the distributed graph database corresponding to the global knowledge graph, the platform locates the starting node corresponding to the changed entity and the target node corresponding to each entity identifier in the associated entity set using the entity's unique identifier, establishing a node address mapping table between the starting and target nodes.

[0029] Starting with the root node, a breadth-first traversal is initiated across the entire knowledge graph. During the traversal, the parent node pointer and access path of each visited node are recorded. Simultaneously, for each edge traversed, the relationship type corresponding to that edge is obtained. Based on a preset mapping rule between relationship types and business coupling, the business coupling weight corresponding to that edge is calculated. If the calculated weight is lower than a preset business coupling threshold, the traversal of that branch is terminated, and the adjacent nodes corresponding to that edge are not visited. This breadth-first traversal continues until all target nodes corresponding to the associated entity set have been visited. After the traversal terminates, the set of nodes traversed in all access paths and the set of unpruned connecting edges between nodes are extracted. The node set and the connecting edge set are combined to generate a minimum connected subgraph containing the starting node and all target nodes.

[0030] Furthermore, the platform freezes read and write permissions in the entire knowledge graph except for the least connected subgraph, and re-executes knowledge embedding computation and constraint consistency verification within the least connected subgraph. A multi-version concurrency control mechanism is initiated for the distributed graph database corresponding to the entire knowledge graph, assigning a globally incrementing transaction timestamp to the current update operation. Based on this timestamp, all nodes and edges in the entire knowledge graph except those contained in the least connected subgraph are marked as read-only. Nodes and edges in read-only state only support data read operations and do not support any data write, update, or delete operations, thus completing the freeze of read and write permissions for non-update areas.

[0031] In the memory computation space corresponding to the minimum connected subgraph, the initial feature vectors of all nodes and edges within the minimum connected subgraph are loaded. Based on the changed attribute values ​​corresponding to the changed entities, the attribute key-value pairs of the nodes corresponding to the changed entities are updated. Based on the updated attribute key-value pairs, the knowledge embedding computation is re-executed to generate the updated feature vectors of all nodes and edges within the minimum connected subgraph. After the knowledge embedding computation is completed, based on the business constraints defined in the logical constraint formula network, constraint consistency verification is performed on the updated feature vectors and attribute values ​​of all nodes within the minimum connected subgraph. During the verification process, each constraint formula in the logical constraint formula network is verified one by one to determine whether the node attribute values ​​and feature vectors within the minimum connected subgraph meet the requirements of the constraint formula. If all constraint formulas pass the verification, the constraint consistency verification is considered successful.

[0032] Furthermore, after successful verification, the platform synchronizes the updated node and edge features of the minimum connected subgraph to the main copy of the global knowledge graph and releases the freeze lock. Once the constraint consistency verification passes, a shadow copy of the minimum connected subgraph is constructed in the shared memory region of the distributed graph database. The shadow copy contains complete attribute key-value pairs for all nodes and edges within the minimum connected subgraph, the updated feature vectors, and the mapping of relationships between nodes and edges.

[0033] A pointer mapping relationship is established between each node and edge in the shadow copy and the corresponding node and edge in the main copy of the global knowledge graph. Using atomic write operations provided by the distributed graph database, the old data pointers for the corresponding minimum connected subgraph regions in the main copy are batch-redirected to the corresponding memory addresses in the shadow copy. This pointer redirection operation is completed within a single atomic transaction, ensuring that there are no intermediate states during the data switching of the main copy. After the pointer redirection operation is completed, the read-only status markers of all nodes and edges in the main copy of the global knowledge graph, except for the minimum connected subgraph, are cleared. This triggers the garbage collection process of the multi-version concurrency control mechanism, releasing the memory and disk storage space occupied by the replaced old version minimum connected subgraph data in the main copy. This completes the full release of the frozen locks, concluding this global data knowledge fusion governance process.

[0034] In this embodiment, by using a network of logical constraint formulas formalized from business linkage rules and a constraint propagation algorithm, it is ensured that cross-business related entities affected by change events are fully included in the update process. By extracting the minimum connected subgraph and freezing non-updating regions, the scope of knowledge embedding calculation and constraint verification is limited to the minimum business-related region. This ensures the consistency of business constraints across the entire knowledge graph while avoiding the resource overhead caused by full graph reconstruction.

[0035] In a preferred embodiment, please refer to the appendix. Figure 2The monitoring and processing of multi-business data change event streams are implemented through a multi-level cache queue. Constraint propagation of cross-business linkage rules is executed using an arc consistency algorithm, and conflicts during constraint propagation are handled using a pre-defined fault-tolerance strategy. In this embodiment, the multi-level cache queue is deployed between the data change monitoring component and the subsequent constraint propagation processing component. The cache queue is implemented using a distributed message queue, and a corresponding number of independent cache shards are divided according to the number of business categories in the enterprise. Each cache shard is bound to a single business category and only receives change event messages from that corresponding business category. Data processing between different cache shards is isolated, and there is no resource contention or data interference. Each cache shard is configured with independent write permission control, data persistence strategy, and consumption progress management mechanism to ensure that sudden traffic in a single business category does not affect the normal operation of cache shards in other business categories.

[0036] After receiving a data write request from the front-end business system, the data change monitoring component first parses the write request, extracts the entity's unique identifier and all attribute key-value pairs contained in the request, sorts all attribute key-value pairs according to the preset attribute field sorting rules, concatenates the sorted attribute names and attribute values ​​to generate an attribute string, and performs a hash calculation on the attribute string to generate the first hash value of the current attribute value of the entity corresponding to the write request. The calculation formula is as follows:

[0037] in, The first hash value, For the preset cryptographic hash function, This is a string concatenation function. This is a sorting function that sorts attributes in ascending order (lexicographical order). This is the collection of all entity attribute key-value pairs extracted from the write request. For the first The field name of each attribute. For the first The field values ​​of each attribute, This represents the total number of attribute fields.

[0038] Furthermore, based on the extracted unique entity identifier, a query request is initiated to the distributed graph database corresponding to the global knowledge graph to obtain all attribute key-value pairs stored in the graph database for the corresponding node of the entity. Using the same calculation process as the first hash value, the full set of attribute key-value pairs stored in the graph database is sorted, concatenated, and hashed to generate the second hash value. Compare the first hash value. With the second hash value The value, if and If they are completely identical, it is determined that the write request has not resulted in any data change, and the write request is discarded without further processing; if... and If there is a mismatch, a data change event is determined. From the full set of attribute key-value pairs obtained from the write request parsing, the single-field hash value corresponding to each attribute field is calculated field by field. The single-field hash value is compared with the single-field hash value of the corresponding attribute field in the graph database to identify the attribute fields with hash differences. The field name and field value of the field are extracted as the changed attribute value. At the same time, the unique identifier of the entity corresponding to the write request is extracted as the changed entity identifier.

[0039] According to the preset message format specifications, the changed entity identifier, changed attribute value, business type, event generation timestamp, and business serial number corresponding to the write request are encapsulated to generate a standardized change event message. Based on the business type field in the message, the change event message is routed to the cache shard of the corresponding business type for storage.

[0040] In this embodiment, the configuration parameters of each shard of the multi-level cache queue and the standardized format of change event messages are uniformly defined through the following table, providing differentiated aggregation processing rules for change events of different business types.

[0041] Table 2 Multi-level cache queue fragmentation configuration and event message format definition table

[0042] The table above defines the configuration parameters of each shard of the multi-level cache queue and the standardized format of change event messages in this embodiment. It provides differentiated aggregation processing rules for change events of different business types, ensuring that the reception, caching and output of change events can adapt to the business traffic characteristics of different business types, avoiding invalid change events from entering the subsequent processing flow, and balancing the requirements of data reliability and processing throughput through differentiated configuration of persistence strategies.

[0043] Furthermore, each cache shard is configured with an independent time sliding window and an event aggregation counter. The event aggregation counter is used to count the total number of change event packets written to the corresponding cache shard within the current time sliding window period. The time sliding window adopts a sliding timer mode, and the window's timing start is updated in real time with the latest event write operation. When a change event packet is written to the corresponding cache shard, the timing end of the time sliding window corresponding to that cache shard is reset, and the timing end is updated to the current system time plus the preset window period duration. At the same time, the value of the event aggregation counter corresponding to that cache shard is incremented by one.

[0044] Within the time-sliding window's timing period, the event aggregation counter is continuously monitored. If the event aggregation counter reaches a preset aggregation threshold, a batch output operation of all stored change event messages in the cache shard is immediately triggered, sending all change event messages in the cache shard to the subsequent constraint propagation processing component. Simultaneously, the event aggregation counter for the cache shard is cleared, and the time-sliding window is reset. If the time-sliding window's timing period ends (i.e., the current system time reaches the time-sliding window's timing end) and the event aggregation counter has not reached the preset aggregation threshold, then all currently stored change event messages in the cache shard are forcibly output to the subsequent constraint propagation processing component. Simultaneously, the event aggregation counter for the cache shard is cleared, and the time-sliding window is reset.

[0045] For further details, please refer to the appendix. Figure 3 Each cross-business linkage rule is structured and parsed to extract the premise entity, result entity, and logical constraints between the premise entity attributes and result entity attributes. The premise entity is the trigger of the constraint, and the result entity is the constrained entity. Directed edges from the premise entity to the result entity represent the dependency direction between entities. Each directed edge corresponds to a set of attribute value domain constraints connected by logical operators, including AND, OR, NOT, greater than, less than, equal to, contain, and belong. Loop detection is performed on the constructed directed acyclic graph. If a loop is detected, a rule anomaly alarm is generated, prompting business personnel to correct the business linkage rules containing loops, ensuring that the final generated logical constraint formula network is a strictly directed acyclic graph structure.

[0046] Furthermore, the changed entity identifier is injected into the directed acyclic graph as the initial active node. The changed attribute value corresponding to the initial active node is extracted, generating the attribute value range corresponding to the changed attribute. This attribute value range is used as the initial constraint and added to the propagation queue for execution of the arc consensus algorithm. During the iteration of the arc consensus algorithm, the currently active node is retrieved from the propagation queue, and all outgoing edges of the currently active node and the constraint set corresponding to each outgoing edge are extracted. The constraint set contains the attribute value range connected by logical operators. For each downstream node corresponding to an outgoing edge, the constraint of the currently active node is propagated to the downstream node along the directed edge. For each attribute dimension of the downstream node, all constraints related to that attribute dimension in the constraint set are traversed, and the intersection interval of all related constraints is calculated using the following formula:

[0047] in, For downstream node attribute dimensions The updated attribute value range The set intersection operator. For the first Individual and attribute dimensions The range of attribute values ​​corresponding to the relevant constraints. For attribute dimensions The total number of relevant constraints.

[0048] Furthermore, the calculated updated attribute value range is... Make a judgment, if If it is an empty set, then a constraint propagation conflict is determined, and the current constraint propagation process is terminated; if For the downstream node, this is the original attribute value range of this attribute dimension. proper subsets of, i.e. Then use Replace the original attribute value range of this attribute dimension. Mark the downstream node as an active node, add all outgoing edges of the downstream node to the propagation queue, and continue iterative processing; if and If all nodes are completely equal, no processing is performed on that downstream node, and it is not added to the propagation queue. When the propagation queue is empty, the arc consensus algorithm terminates its iteration, aggregating the entity identifiers of all entities marked as active nodes to generate a set of associated entities affected by the change event.

[0049] Furthermore, if the calculated updated attribute value range is... If the set is empty, the target attribute dimension that generates the intersection interval of the empty set, the node identifier corresponding to the target attribute dimension, and the target constraint identifier that triggers the conflict are recorded. An anomaly alarm containing the conflict source node identifier, conflict attribute dimension, and conflict constraint identifier is generated. The generated anomaly alarm is sent to a preset conflict degradation processing queue. The conflict degradation processing queue adopts a first-in-first-out scheduling mode to process the anomaly alarms serially. For each anomaly alarm, the preset fault tolerance strategy corresponding to the target constraint is queried in the preset constraint fault tolerance strategy library. The preset fault tolerance strategies include relaxing the value range boundary, skipping the target constraint, and manual review and intervention.

[0050] If the preset fault tolerance strategy is to relax the value range boundary, the original value range interval corresponding to the target constraint is extracted, and the upper and lower boundaries of the value range interval are expanded according to the preset relaxation ratio to generate the modified constraint. The modified constraint replaces the target constraint in the directed acyclic graph, and the constraint propagation process starting from the conflict source node is retried. If the preset fault tolerance strategy is to skip the target constraint, the target constraint is temporarily removed from the constraint set of the directed acyclic graph, and the constraint propagation process starting from the conflict source node is retried. If an empty set intersection interval still exists after retrieding the constraint propagation process, the conflict degradation processing process is repeated. If the number of retries exceeds the preset retries limit and an empty set intersection interval still exists, the change event corresponding to the change entity identifier of this change event is removed from the multi-business data change event stream, the subsequent processing process of this change event is terminated, and a manual processing work order is generated and sent to the enterprise data governance management personnel.

[0051] In this embodiment, the business type isolation and hash comparison mechanism of the multi-level cache queue reduces the computing power overhead of subsequent processing. The time sliding window and event aggregation mechanism adapt to the business traffic characteristics of different business types and improve the throughput of change event processing. The arc consistency algorithm realizes the accurate propagation of constraints, ensuring the complete identification of related entities affected by changes. The preset fault tolerance strategy realizes the automatic degradation processing of constraint propagation conflicts, improving the stability of system operation.

[0052] In a preferred embodiment, please refer to the appendix. Figure 4 The extraction of the minimum connected subgraph is achieved through breadth-first traversal with weighted pruning; knowledge embedding computation is performed through a local graph attention network; constraint consistency verification is achieved through a pre-trained constraint decision model; and updated data synchronization is completed through shadow copy and atomic pointer redirection operations.

[0053] In this embodiment, based on the changed entity identifier and the entity identifier in the associated entity set, a precise query is performed in the node index of the distributed graph database to obtain the unique internal ID of the node corresponding to each entity identifier, the physical shard address of the node storage, and the list of the first-order neighbor nodes of the node. A mapping relationship between the entity identifier and the node internal ID is established, the node corresponding to the changed entity identifier is marked as the first starting node, and the node corresponding to the entity identifier in the associated entity set is marked as the target node, generating a target node ID set.

[0054] Initialize the traversal queue by adding the starting node, and simultaneously initialize the set of visited nodes by adding the starting node to the set of visited nodes. Also initialize the traversal path mapping table, recording the parent node information of each node. In each iteration of the breadth-first traversal, retrieve the currently traversed node from the head of the traversal queue, obtain all its adjacent nodes, and the relationship type of the connecting edges between the currently traversed node and each adjacent node. Based on the preset mapping rules between edge relationship types and weights, calculate the edge relationship type weight corresponding to the current connecting edge using the following formula:

[0055] in, For connecting edges The corresponding edge relationship type weight, For the first Preset weighting coefficients corresponding to each business coupling dimension As an indicator function, when connecting edges Satisfy the first When conditions are met for each business coupling dimension, The value is 1 otherwise the value is 0. This represents the total number of business coupling dimensions. In this embodiment, business coupling dimensions include cross-business process dependencies, data scope associations, organizational structure affiliation, and financial accounting entities. The weight coefficients for each dimension are pre-configured based on the importance of the business.

[0056] Furthermore, the calculated edge relation type weights are... Compare with the preset business coupling threshold, if If the business coupling degree is lower than the preset threshold, it is determined that the business coupling degree between the adjacent node corresponding to the connection edge and the currently traversed node is insufficient. A pruning operation is then performed, the adjacent node is not added to the traversal queue, and the traversal of that branch is terminated. If the value is greater than or equal to the preset business coupling threshold, then it is further determined whether the adjacent node exists in the set of visited nodes. If it does not exist in the set of visited nodes, then the adjacent node is added to the traversal queue and the set of visited nodes, and the parent node of the adjacent node is recorded as the current traversal node in the traversal path mapping table.

[0057] The breadth-first traversal iteration continues until all nodes in the target node ID set have been added to the visited node set, at which point the traversal terminates. After the traversal terminates, based on the traversal path mapping table, starting from each target node, backtracking along the parent node pointer to the first starting node, extracting all nodes traversed in the backtracking path to generate the node set of the minimum connected subgraph. At the same time, extracting all unpruned connecting edges between nodes in the node set to generate the edge set of the minimum connected subgraph, and combining the node set and the edge set to generate the minimum connected subgraph containing the first starting node and all target nodes.

[0058] In this embodiment, the weight calculation rules and business coupling thresholds corresponding to different edge relationship types in the global knowledge graph are uniformly defined through the following table, providing a quantitative judgment standard for pruning operations in the breadth-first traversal process.

[0059] Table 3. Weight Mapping of Edge Relationship Types and Business Coupling Threshold Configuration Table

[0060] The table above defines the weight calculation rules and business coupling thresholds for different edge relationship types in the global knowledge graph in this embodiment. It provides a quantitative judgment standard for pruning operations during the breadth-first traversal process, ensuring that only related branches with sufficient business coupling are retained during the traversal process, while removing irrelevant branches with low coupling, thereby further narrowing the processing scope of subsequent knowledge embedding calculations.

[0061] Furthermore, a multi-version concurrency control mechanism is employed to assign globally incrementing timestamps to the primary replica of the global knowledge graph, and nodes and edges, except for the least connected subgraph, are marked as read-only. A distributed transaction is initiated for this update operation, and a globally incrementing transaction timestamp is assigned to this transaction through the global transaction manager of the distributed graph database. This timestamp serves as a unique identifier for this transaction, used to distinguish different versions of the data. Based on the transaction timestamp... Version tags are generated for all nodes and edges in the main replica of the global knowledge graph. Nodes and edges contained in the minimum connected subgraph are marked as writable versions, which support write and update operations within the current transaction. All nodes and edges except those in the minimum connected subgraph are marked as read-only versions, which only support timestamps less than [timestamp value missing]. The read operation of the transaction does not support any write or update operation of the transaction, that is, the read and write permissions of the non-updating area are frozen. In this embodiment, each node and edge corresponds to a version chain. Each node in the version chain contains data content, transaction timestamp, and pointer to the next version. The multi-version concurrency control mechanism realizes multi-version management of data through the version chain, avoiding mutual blocking of read and write operations.

[0062] For further details, please refer to the appendix. Figure 5 Within the minimum connected subgraph, the feature vectors of the nodes corresponding to the changed entity are initialized. A local graph attention network is used to update the node features within the minimum connected subgraph through message passing. The changed attribute values ​​of the nodes corresponding to the changed entity are obtained. Discrete attributes in the changed attribute values ​​are encoded using one-hot encoding, and continuous attributes are normalized. The encoded and processed attribute values ​​are concatenated into a one-dimensional attribute vector. This one-dimensional attribute vector is input into a pre-trained multilayer perceptron. Through linear transformation and nonlinear activation of the multilayer perceptron, it is mapped to a fixed-dimensional high-dimensional feature vector. This high-dimensional feature vector replaces the initial feature vector of the nodes corresponding to the changed entity, completing the feature initialization of the changed entity nodes. In this embodiment, the multilayer perceptron contains two fully connected layers. The output dimension of the first fully connected layer is 256, and the output dimension of the second fully connected layer is 128. The activation function is... The function ensures that attribute values ​​can be mapped to high-dimensional feature vectors that meet the requirements of knowledge embedding.

[0063] Furthermore, a local graph attention network is constructed, with the number of layers matching the number of message passing iterations, pre-set to 3 layers. The network input consists of the initial feature vectors of all nodes within the minimum connected subgraph, and the output consists of the updated feature vectors of all nodes. In the single-layer message passing of the local graph attention network, the central node and its first-order neighbors in the minimum connected subgraph are used as computational units. The feature vector of the central node is concatenated with the feature vectors of its first-order neighbors and then input into a shared linear transformation layer. For any central node in the minimum connected subgraph... Get the set of all its first-order neighbors. The feature vector of the central node i With each neighbor node eigenvectors The concatenation process generates a concatenated feature vector, which is then input into a shared linear transformation layer to obtain intermediate values ​​of the attention coefficients. The attention weight coefficients of the first-order neighbor nodes relative to the center node are calculated using the softmax normalization function. These attention weight coefficients are then used to perform a weighted summation of the feature vectors of the first-order neighbor nodes to generate the updated feature vector of the center node. The calculation formula is as follows:

[0064] in, For neighboring nodes Relative to the central node The unnormalized attention coefficient, This is the weight vector for the attention mechanism. For the shared linear transformation weight matrix, This is a vector concatenation operator. For a leaky linear rectified activation function, These are the normalized attention weight coefficients. For the set of neighbor nodes Normalization function, As the central node Updated feature vector, It is a non-linear activation function. As the central node The set of first-order neighbor nodes.

[0065] Furthermore, after message passing is completed at each layer of the local graph attention network, layer normalization is performed on the feature vectors of all nodes to ensure that the numerical distribution of the feature vectors is stable. After message passing is completed at all layers according to the preset number of iterations, the updated feature vectors of all nodes in the minimum connected subgraph and the updated feature vectors of the edges are extracted to complete the knowledge embedding calculation.

[0066] Furthermore, after completing a preset number of message passes, the updated feature vectors of all nodes within the minimum connected subgraph are extracted. These updated feature vectors are then input into a pre-trained constraint judgment model to verify whether the relative distance between the updated feature vectors satisfies the business constraints defined by the logical constraint formula network. The pre-trained constraint judgment model is a binary classification deep neural network model. The model's input is the difference vector between the updated feature vectors of two nodes, and its output is the binary classification result indicating whether the two nodes satisfy the corresponding business constraints, with 0 representing non-satisfaction and 1 representing satisfaction.

[0067] During the training process of the constraint decision model, node feature vector pairs that meet the business constraints are used as positive samples, and node feature vector pairs that do not meet the business constraints are used as negative samples. The model is trained using the cross-entropy loss function until the classification accuracy reaches a preset threshold. In this embodiment, the constraint decision model contains three fully connected layers. The first fully connected layer has an input dimension of 128 and an output dimension of 64; the second fully connected layer has an output dimension of 32; and the third fully connected layer has an output dimension of 2. The activation function is... Function, output layer adopts The function ensures that the determination of constraints has sufficient accuracy.

[0068] During the constraint consistency verification process, based on the constraints in the logical constraint formula network, node pairs with constraint relationships are extracted from the minimum connected subgraph. For each node pair, the updated feature vectors of the two nodes are extracted, and the difference vector between the two feature vectors is calculated using the following formula:

[0069] in, Let be the difference vector between the updated feature vectors of node i and node j. Let be the updated feature vector of node i. Let be the updated feature vector of node j. Input the difference vector into the constraint judgment model and obtain the model's output. If the model output of all node pairs is 1, that is, all node pairs satisfy the corresponding business constraints, then the constraint consistency check is passed; if the model output of any node pair is 0, then the constraint consistency check is failed, the update process is terminated, a constraint check anomaly alarm is generated, and sent to the data governance administrator.

[0070] For further details, please refer to the appendix. Figure 6 When the constraint consistency check passes, a shadow copy of the minimum connected subgraph is constructed in memory, and the node and edge features that passed the check are written to the shadow copy. In the shared memory area of ​​the distributed graph database, contiguous memory space is allocated to construct the shadow copy of the minimum connected subgraph. The structure of the shadow copy is completely consistent with the structure of the main copy of the global knowledge graph, including a node table, edge table, attribute index, and feature vector storage area. All attribute key-value pairs, updated feature vectors, and node relationships of all nodes in the minimum connected subgraph after the constraint consistency check passes, as well as all attribute key-value pairs, updated feature vectors, and edge relationships of all edges, are written to the corresponding storage area of ​​the shadow copy, completing the construction of the shadow copy. After writing, the integrity of all nodes and edges in the shadow copy is checked, the checksum of the shadow copy is calculated, and compared with the checksum of the minimum connected subgraph to ensure that the data in the shadow copy is complete and undamaged.

[0071] Furthermore, a pointer mapping relationship is established between the shadow copy and the main copy of the global knowledge graph. Through atomic operations, the old pointers of the corresponding minimum connected subgraphs in the main copy are redirected to the memory addresses of the shadow copy. For each node and edge in the shadow copy, a corresponding memory address pointer is generated, and a one-to-one pointer mapping table is established between the old nodes and edges in the main copy and the new nodes and edges in the shadow copy.

[0072] By leveraging the atomic transaction operations provided by the distributed graph database, within a single indivisible transaction, the old data pointers of all corresponding minimum connected subgraph regions in the primary replica are batch redirected to the memory addresses corresponding to the shadow replica according to the pointer mapping table. After the pointer redirection operation is completed, all read requests for that region in the primary replica will be routed to the memory space corresponding to the shadow replica, ensuring the atomicity of data switching, eliminating intermediate states, and avoiding inconsistencies in data read and write operations.

[0073] Furthermore, after pointer redirection, the read-only status markers of all nodes and edges in the primary replica, except for the minimum connected subgraph, are cleared. This triggers the garbage collection process of the multi-version concurrency control mechanism to release the storage space occupied by the old version data, thus releasing the frozen lock. After the pointer redirection operation is completed, the distributed transaction is committed, and the transaction commit timestamp is updated to the current global timestamp. Based on the commit timestamp, the read-only status markers of all nodes and edges in the primary replica of the global knowledge graph are cleared, restoring normal read and write permissions for all nodes and edges, thus releasing the frozen lock.

[0074] The garbage collection process, which triggers the multi-version concurrency control mechanism, scans all data versions in the main replica of the global knowledge graph, identifies the smallest connected subgraph data of the old version that was replaced in this transaction, that is, all old data whose timestamps are less than the commit timestamp of this transaction and have been replaced by pointer redirection. The memory space and disk storage space occupied by these old data are released, the garbage collection operation is completed, and the entire global data knowledge fusion governance process is completed.

[0075] In this embodiment, the scope of knowledge embedding computation is further narrowed and the computational cost is reduced by using a breadth-first traversal with weighted pruning; it ensures that the updated node features can fully capture the business relationships between entities; the constraint judgment model realizes automated verification of constraint consistency, ensuring that the updated knowledge graph conforms to cross-business linkage rules; and the shadow copy and atomic pointer redirection operations avoid the impact on the availability of the global knowledge graph service during the update process.

Claims

1. A comprehensive data knowledge fusion governance method for enterprises with multiple business formats, characterized in that, include: Monitor data change event streams across multiple business formats and extract the entity identifier and attribute value of the changed entity. The cross-business linkage rules are formalized into a logical constraint formula network. The changed entity identifier is used as a trigger variable and input into the logical constraint formula network. The set of related entities affected by the changed event is derived through the constraint propagation algorithm. In the global knowledge graph, the changed entity is the starting point and the entities in the set of associated entities are the ending points. Breadth-first traversal is used to extract the minimum connected subgraph containing the starting point and the ending point. Freeze read and write permissions in the global knowledge graph except for the minimum connected subgraph, and re-execute knowledge embedding calculation and constraint consistency verification within the minimum connected subgraph; After the verification is passed, the updated minimum connected subgraph nodes and edge features are synchronized to the main copy of the global knowledge graph, and the freeze lock is released.

2. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises as described in claim 1, characterized in that, The process of monitoring multi-business data change event streams and extracting change entity identifiers and change attribute values ​​includes: deploying multi-level cache queues between multi-business front-end data sources and the global knowledge graph, wherein the multi-level cache queues are divided into independent cache shards according to business categories; When the multi-level cache queue receives a data write request, it calculates the first hash value of the current attribute value of the entity corresponding to the write request, and compares the first hash value with the second hash value of the corresponding entity in the global knowledge graph. If the first hash value does not match the second hash value, a data change event is determined to have occurred. The change entity identifier and the change attribute value with hash difference are parsed from the write request. The change entity identifier and the change attribute value are encapsulated into a change event message and stored in the cache fragment of the corresponding business type.

3. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises according to claim 1, characterized in that, The process of formalizing cross-business linkage rules into a logical constraint formula network, inputting the changed entity identifier as a trigger variable into the logical constraint formula network, and deriving the set of associated entities affected by the change event through a constraint propagation algorithm includes: parsing the entity dependency relationships in the business linkage rules, and constructing a directed acyclic graph with entity identifiers as nodes and dependency relationships as directed edges as the logical constraint formula network. The changed entity identifier is injected into the directed acyclic graph as the initial activation node. The arc consistency algorithm is executed, and the changed attribute value domain corresponding to the changed entity identifier is used as a constraint condition to propagate to the downstream nodes along the directed edge. When the constraint received by a downstream node in the directed acyclic graph causes its own attribute value range to be reduced, the downstream node is marked as an active node, and the entity identifiers corresponding to all the active nodes after traversal are aggregated into the associated entity set.

4. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises as described in claim 1, characterized in that, In the global knowledge graph, the changed entity is the starting point and the entities in the associated entity set are the ending points. The minimum connected subgraph containing the starting point and the ending point is extracted using breadth-first traversal, including: locating the first starting node corresponding to the changed entity and multiple target nodes corresponding to the associated entity set in the global knowledge graph. The first starting node initiates a breadth-first traversal as the root node. During the traversal, the edge relationship type weight between the current traversed node and its neighboring nodes is calculated. If the edge relationship type weight is lower than a preset business coupling threshold, the branch is pruned to block the traversal of the branch. Continue traversing until all target nodes have been visited, extract all nodes traversed along the traversal path and all unpruned connecting edges, and combine all extracted nodes and connecting edges to generate the minimum connected subgraph.

5. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises according to claim 1, characterized in that, The step of freezing read and write permissions in the global knowledge graph except for the minimum connected subgraph, and re-executing knowledge embedding calculation and constraint consistency verification in the minimum connected subgraph, includes: using a multi-version concurrency control mechanism to allocate a globally incrementing timestamp to the main copy of the global knowledge graph, and marking nodes and edges except for the minimum connected subgraph as read-only; Within the minimum connected subgraph, the feature vectors of the nodes corresponding to the changed entity are initialized, and the node features within the minimum connected subgraph are updated by message passing using a local graph attention network. After completing a preset number of message transmissions, the updated feature vectors of all nodes in the minimum connected subgraph are extracted. The updated feature vectors are then input into a pre-trained constraint judgment model to verify whether the relative distance between the updated feature vectors satisfies the business constraint conditions defined by the logical constraint formula network.

6. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises according to claim 1, characterized in that, After the verification passes, the updated nodes and edge features of the minimum connected subgraph are synchronized to the main copy of the global knowledge graph, and the freeze lock is released. This includes: when the constraint consistency verification passes, constructing a shadow copy of the minimum connected subgraph in memory, and writing the verified nodes and edge features into the shadow copy. Establish a pointer mapping relationship between the shadow copy and the main copy of the global knowledge graph, and redirect the old pointer of the corresponding minimum connected subgraph in the main copy to the memory address of the shadow copy through atomic operations; After the pointer redirection is completed, the read-only status markers of nodes and edges in the main replica, except for the minimum connected subgraph, are cleared, triggering the garbage collection process of the multi-version concurrency control mechanism to release the storage space occupied by the old version data, and completing the release of the frozen lock.

7. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises according to claim 2, characterized in that, The step of encapsulating the changed entity identifier and the changed attribute value into a change event message and storing it in the cache segment of the corresponding business type includes: configuring an independent time sliding window and an event aggregation counter for each cache segment; When the change event message is written to the cache shard, the time sliding window corresponding to the cache shard is reset, and the value of the event aggregation counter is incremented by one; If the value of the event aggregation counter reaches the preset aggregation threshold within the time sliding window period, the batch output of all the change event messages in the cache fragment is triggered in advance. If the time sliding window times out and the value of the event aggregation counter does not reach the aggregation threshold, then all the current change event messages in the cache fragment will be forcibly output to the next processing stage, and the event aggregation counter will be cleared.

8. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises according to claim 3, characterized in that, The execution arc consistency algorithm propagates the changed attribute value range corresponding to the changed entity identifier as a constraint condition to downstream nodes along the directed edge, including: extracting the constraint condition set corresponding to all incoming edges of the currently active node in the directed acyclic graph, wherein the constraint condition set contains attribute value range intervals connected by logical operators; For each attribute dimension of the currently active node, traverse all constraints related to that attribute dimension in the constraint set and calculate the intersection interval of all related constraints; If the intersection interval is an empty set, then it is determined that a constraint propagation conflict has occurred and the current process is terminated; If the intersection interval is a proper subset of the original attribute value range of the current active node, then the original attribute value range of the current active node is replaced by the intersection interval, and the outgoing edges of the current active node are added to the propagation queue for further iterative processing.

9. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises according to claim 5, characterized in that, The initialization of the feature vector of the node corresponding to the changed entity and the message passing update of the node features in the minimum connected subgraph using a local graph attention network include: obtaining the changed attribute value of the node corresponding to the changed entity and mapping the changed attribute value to a high-dimensional feature vector through a multilayer perceptron to replace the initial feature vector of the node corresponding to the changed entity. In the single-layer message passing of the local graph attention network, the central node and its first-order neighbor nodes in the minimum connected subgraph are used as computational units. The feature vector of the central node and the feature vector of the first-order neighbor nodes are concatenated and then input into the shared linear transformation layer. The attention weight coefficients of the first-order neighbor nodes relative to the center node are calculated using the softmax normalization function. The feature vectors of the first-order neighbor nodes are then weighted and summed using these attention weight coefficients to generate the updated feature vector of the center node.

10. The method for comprehensive data knowledge fusion and governance for multi-business operations of enterprises according to claim 8, characterized in that, If the intersection interval is an empty set, then it is determined that a constraint propagation conflict has occurred and the current process is terminated, including: recording the target attribute dimension that generates the empty set intersection interval and the target constraint condition that triggers the conflict, and generating an abnormal alarm containing the conflict source node identifier and the conflict constraint identifier. The abnormal alarm is sent to a preset conflict degradation processing queue. The preset fault tolerance strategy corresponding to the target constraint is queried in the conflict degradation processing queue. The preset fault tolerance strategy includes relaxing the value range boundary or skipping the target constraint. Modify the constraint parameters of the target constraint in the directed acyclic graph according to the preset fault tolerance strategy found in the query, and re-trigger the constraint propagation process starting from the conflict source node. If the number of retries exceeds the preset retry limit and there is still an empty set intersection interval, then remove the change event corresponding to the changed entity identifier from the multi-business data change event stream.

Images