Knowledge graph driven inference center system

Abstract

The invention relates to the technical field of knowledge reasoning, in particular to a knowledge graph driven reasoning center system which comprises a label entity screening module, a graph edge state recognition module, a path activation construction module, a logic jump matching module and a path structure state summarization module. According to the method, consistency judgment is carried out on scene type marks of entity nodes, and a number set matched with task requirements is screened, so that uniform binding of accurate constraint and task semantics of entities in the atlas is realized, structure screening and state classification are carried out based on available states of connecting edges, and the structure effectiveness of edges participating in path construction is ensured; the continuity and logic definition of a path structure are improved by adopting a grouping mode of an edge sequence and a connection direction, and a structure type attribute comparison mechanism is introduced into a path interruption part, so that the path communication accuracy and the reasoning extension capability are improved; and high-integrating-degree screening of reasoning paths, dynamic complementation of path links and systematic control of a graph structure state are realized.

Description

Technical Field

[0001] This invention relates to the field of knowledge reasoning technology, and in particular to a reasoning center system driven by knowledge graphs. Background Technology

[0002] The field of knowledge reasoning technology involves the process of analyzing, judging, and deriving information through formalized knowledge representations and logical rules. It includes knowledge representation, semantic modeling, logical reasoning, rule computation, and ontology construction. Based on semantic networks, knowledge graphs, and reasoning engines, it is widely applied in AI-related scenarios such as intelligent question answering, decision support, and natural language understanding. Specifically, a reasoning center system refers to a system that, after constructing ontology concept relationships and rule sets and uniformly modeling input data, utilizes propositional logic reasoning, descriptive logic reasoning, or rule-chain-based reasoning to complete the knowledge deduction or induction process for specific application scenarios. It typically uses the RDF triplet storage model, SPARQL query language, or an embedded rule interpreter to achieve knowledge acquisition and reasoning process control.

[0003] In current task reasoning processes, the unified modeling relies mainly on static rule sets and pre-defined ontology relationship models. This makes it difficult to dynamically select matching graph nodes in the context of diverse semantic changes in tasks, resulting in a wide range of entity selections with insufficient accuracy. In terms of edge connection state processing, there is a lack of clear division of participation in path construction, which can easily introduce semantically deviated connection paths. In the process of generating reasoning paths, the connection order and directional relationship are not combined to form a standardized path sequence, making it difficult to ensure the integrity and logical consistency of the path structure. When the original path is interrupted, no jump mechanism based on structural attributes is provided, resulting in jump connections that lack a structural semantic basis, affecting the overall connectivity effectiveness of the path and the reliability of the reasoning conclusions. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a knowledge graph-driven reasoning center system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a knowledge graph-driven reasoning center system includes:

[0006] The tag entity filtering module obtains the inference task scenario type information, collects the scenario type tags of all entity nodes in the graph to perform type consistency judgment, filters the entity numbers that match the current task scenario, establishes task-related node mapping, and generates a list of task-matching entity numbers.

[0007] The graph edge state recognition module extracts edge connection information based on the task matching entity number list, determines whether the starting node is in the list, marks edges that do not meet the conditions as unparticipable states and removes them, summarizes the connection state classification of participating edges, and generates a graph edge connection available state record table.

[0008] The path activation construction module performs path grouping processing based on the graph edge connection available status record table, constructs the attribute path activation sequence in sequence, and generates a task applicable path sequence structure table.

[0009] The logical jump matching module locates the starting node number of the path interruption position according to the task applicable path sequence structure table, screens the connection information in the graph that starts from the corresponding node but belongs to other paths, compares the attribute of the end node structure type with the target structure type of the interrupted path, counts the number of attribute hits and calculates the comparison score, and generates a graph path connection jump mapping detail.

[0010] As a further embodiment of the present invention, the task matching entity number list includes an entity number list, entity type labels, and a task scenario mapping table; the graph edge connection available status record table includes participating edge numbers, connection start and end node numbers, and edge type status identifiers; the task applicable path sequence structure table includes path numbers, edge connection order, and attribute path syntax sequence; and the graph path connectivity jump mapping details include interrupted path start number, candidate jump path number, and attribute comparison score.

[0011] As a further aspect of the present invention, the tag entity filtering module includes:

[0012] The entity scene extraction submodule collects the scene type labels of all entity nodes in the graph, records the scene label number and scene classification attribute of each entity node, organizes the correspondence between scene number and node number in a set manner, and obtains the node scene mapping set.

[0013] The type consistency judgment submodule performs unified and standardized processing on the scene classification attributes of each entity node based on the node scene mapping set, judges the correspondence between the scene type to which the node belongs and the inference task scene type, identifies entity nodes that are consistent with the task scene type, filters the set of entity numbers whose scene classification attributes are consistent with the current task scene type, and obtains the scene consistent entity number set.

[0014] The task entity matching submodule performs task number association processing based on the scenario-consistent entity number set, establishes a one-to-one mapping between task numbers and entity numbers, extracts entity numbers that meet the requirements based on the matching of the tag attributes of entity nodes with the task scenario, and obtains a list of task-matched entity numbers.

[0015] As a further aspect of the present invention, the graph edge state recognition module includes:

[0016] The edge connection information extraction submodule extracts all edge information that has a connection relationship with the entity number in the task matching list based on the task matching entity number list. It collects the starting node number, ending node number and corresponding relationship type identifier of each edge, organizes and cleans the format according to the structure field, establishes a combined record of starting node, ending node and relationship type, and obtains the edge node structure table.

[0017] The edge state filtering submodule judges the starting node number of each edge record one by one according to the edge node structure table, and verifies whether there is a matching relationship by comparing it with the task matching entity number list. If the starting node number does not appear in the list, the corresponding edge is marked as unparticipable; otherwise, it is marked as participateable. After marking, all edge records in the unparticipable state are removed to obtain the set of available edge connections.

[0018] The edge state record construction submodule extracts the relationship types of all edges based on the available edge connection set, categorizes and organizes edges with the same relationship type, records the start and end node numbers associated with each type of edge, integrates them into a standard record format, and establishes a graph edge connection available state record table.

[0019] As a further aspect of the present invention, the path activation building module includes:

[0020] The edge path grouping submodule, based on the available state record table of the graph edge connection, detects the edge connection structure one by one according to the continuous connection relationship between the starting node and the ending node, constructs the edge path chain with the first and last connection features, establishes the edge set of the connection path based on the continuity of the first and last nodes, summarizes it into a structured dataset according to the path number, and obtains the edge path grouping set.

[0021] The path direction numbering submodule extracts the connection order and direction information of the edges in each group of paths based on the edge path group set. The path direction is marked as positive by the direction from the start point of the edge to the end point. If there is a reverse connection in the path, the direction is marked with a negative sign. The direction order comparison record is established by combining the position and direction of each edge in the path to obtain the path edge direction order table.

[0022] The path sequence generation submodule processes the edge direction number sequence recorded in each path according to the path edge direction order table, performs format standardization processing, concatenates the edge direction expressions in sequence, establishes a one-to-one correspondence between the path number and the string result, binds it to the task number for output, and establishes a task applicable path sequence structure table.

[0023] As a further aspect of the present invention, the logical jump matching module includes:

[0024] The path interruption location submodule extracts the connection structure expression in each path according to the task applicable path sequence structure table, judges the position of the syntax structure interruption point one by one, obtains the end node number of the tail edge of the interrupted path as the starting point of the jump node, compares the end symbols of the path structure one by one according to the path number, judges the connection continuity indicator, extracts the starting node number corresponding to all interruption positions, and establishes a path interruption starting node set.

[0025] The jump path screening submodule, based on the set of path interruption starting point nodes, screens all connection edge records in the graph that start from the corresponding node, excludes edge items that are the same as the original path number, retains only the connection records that belong to different path numbers, extracts the corresponding end node number and connection type, and obtains the jump path candidate information table.

[0026] The structural comparison and evaluation submodule extracts the structural attributes of the endpoint nodes in the candidate paths based on the jump path candidate information table, compares them with the attributes of the target nodes of the interrupted paths, and uses attribute-level similarity weighted calculation to obtain the jump structure comparison score. The candidate paths are sorted and summarized to establish a graph path connectivity jump mapping detail.

[0027] As a further aspect of the present invention, the system further includes:

[0028] The path structure state summary module identifies the current state marker of each edge in each path based on the graph path connection jump mapping details, classifies, counts and archives all edges according to triplet state identifiers, forms a structure state summary record that runs through all jump segments and original paths, and generates a knowledge graph-driven reasoning center path control structure table.

[0029] The knowledge graph-driven reasoning center path control structure table includes triplet state classification statistics, edge state archive list, and path segment structure markers.

[0030] As a further aspect of the present invention, the path structure status summary module includes:

[0031] The triplet state recognition submodule, based on the graph path connectivity jump mapping details and combined with the structural record of each edge in the mapping path, extracts the starting node, ending node and connection predicate code corresponding to each edge to form a standard triplet field combination, identifies the predicate state in the triplet, and obtains the edge state identification dataset.

[0032] The edge state archiving submodule classifies all triplet structure records according to their state identifiers based on the edge state identifier dataset. It counts the edges after grouping them by state category and establishes a classification index number. It collects the edge structures corresponding to each state category and establishes a state archiving directory with the grouped state as the primary key to obtain the structure edge state classification table.

[0033] The connected path structure generation submodule summarizes the state of each edge in each path according to the structure edge state classification table, gathers all edge state identifiers by path number, counts the proportion of states in each path, merges them to establish a structure control path number set, and generates a knowledge graph-driven reasoning center path control structure table.

[0034] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0035] In this invention, by performing consistency judgment on the scene type tags of entity nodes and filtering the set of numbers that match the task requirements, precise constraints on entities in the graph and unified binding of task semantics are achieved. Based on the available state of the connecting edges between entities, structural screening and state classification are performed to ensure that all edges participating in path construction have semantic relevance and structural validity. The grouping method of edge order and connection direction is adopted to improve the continuity and logical clarity of the path structure. A structural type attribute comparison mechanism is introduced for the path interruption part, so that the jump path has structural similarity support, improving the accuracy of path connectivity and reasoning extension ability, and realizing high-fitness screening of reasoning paths, dynamic completion of path links, and systematic control of graph structure state. Attached Figure Description

[0036] Figure 1 This is a system flowchart of the present invention;

[0037] Figure 2 This is a flowchart of the label entity filtering module of the present invention;

[0038] Figure 3 This is a flowchart of the graph edge state recognition module of the present invention;

[0039] Figure 4 This is a flowchart of the path activation construction module of the present invention;

[0040] Figure 5 This is a flowchart of the logical jump matching module of the present invention;

[0041] Figure 6 This is a flowchart of the path structure status summary module of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0044] Please see Figure 1 Knowledge graph-driven reasoning center systems include:

[0045] The tag entity filtering module obtains the inference task scenario type information, collects the scenario type tags of all entity nodes in the graph, performs type consistency judgment on all nodes, filters the set of entity numbers that match the current task scenario, establishes a task-related node mapping, and generates a list of task-matching entity numbers.

[0046] The graph edge state recognition module extracts edge connection information associated with task-related nodes based on the task matching entity number list, determines whether the starting node of each edge is in the list, marks edges that do not meet the conditions as unparticipating states and removes them, summarizes the connection state classification of participating edges (using the relation type state identifier of the Neo4j graph database), and generates a graph edge connection available state record table.

[0047] The path activation construction module is based on the graph edge connection available state record table, counts the set of edges that can participate in the state and performs path grouping processing (executes path traversal optimization algorithm), extracts the occurrence order and connection direction number of the edges in each path, constructs the attribute path activation sequence in order (refer to SPARQL1.1 PropertyPath syntax), and generates a task applicable path sequence structure table.

[0048] The logical jump matching module locates the starting node number corresponding to the path interruption position based on the task applicable path sequence structure table, screens all connection information in the graph that starts from the corresponding node but belongs to other paths, compares the structure type of the endpoint node with the target structure type of the interrupted path, counts the number of attribute hits and calculates the comparison score (using ontology reasoning similarity algorithm), and generates a graph path connectivity jump mapping detail.

[0049] The path structure state summary module identifies the current state marker of each edge in each path based on the graph path connectivity jump mapping details. It classifies, counts, and archives all edges according to triplet state identifiers (using predicate state encoding of the graph database), forming a structure state summary record that runs through all jump segments and original paths, and generating a knowledge graph-driven reasoning center path control structure table.

[0050] The task matching entity number list includes an entity number list, entity type labels, and a task scenario mapping table. The graph edge connection available status record table includes participating edge numbers, connection start and end node numbers, and edge type status identifiers. The task applicable path sequence structure table includes path numbers, edge connection order, and attribute path syntax sequence. The graph path connectivity jump mapping details include interrupted path start number, candidate jump path number, and attribute comparison score. The knowledge graph-driven reasoning center path control structure table includes triple state classification statistics, edge state archive list, and path segment structure markers.

[0051] Please see Figure 2 The tag entity filtering module includes:

[0052] The entity scene extraction submodule collects the scene type labels of all entity nodes in the graph, records the scene label number and scene classification attribute of each entity node, organizes the correspondence between scene number and node number in a set manner, and obtains the node scene mapping set.

[0053] To collect scene type labels for all entity nodes in the knowledge graph, the node number must first be clearly defined as a unique identifier. In the business knowledge graph, each node corresponds to a specific entity. For example, entity "E001" represents a customer account, "E002" represents a medical case, and "E003" represents a device component. For each node, the pre-defined scene type labels in the system need to be extracted. These labels are generally derived from manual definitions or industry standards in the early stages of data construction. For example, "financial risk control" is set as "T1", "medical auxiliary diagnosis" is set as "T2", and "intelligent manufacturing testing" is set as "T3". By traversing the entity list, the scene labels corresponding to each node are collected and combined with their tag classification attributes, such as category A representing rule-based data, category B representing perception-based data, and category C representing control-based data, to form a structured mapping relationship. Table 1 lists the scene labels and tag classification information of some nodes.

[0054] Table 1 Example of Node Scene Type Mapping

[0055] Node number Scene type marker Tag category attributes E001 T1 Category A E002 T2 Category B E003 T1 Category A E004 T3 Category C E005 T1 Category A

[0056] As shown in Table 1, nodes E001, E003 and E005 all correspond to scenario type T1 and belong to type A label. This indicates that they may belong to different objects in the same business domain, such as accounts, transactions and risk control records in the financial field. After the data collection is completed, a mapping set data structure is established to form a one-to-one correspondence between node numbers and scenario labels, and the node scenario mapping set is obtained.

[0057] The type consistency judgment submodule performs unified and standardized processing on the scene classification attributes of each entity node based on the node scene mapping set, judges the correspondence between the scene type to which the node belongs and the inference task scene type, identifies entity nodes that are consistent with the task scene type, filters the set of entity numbers whose scene classification attributes are consistent with the current task scene type, and obtains the scene consistent entity number set.

[0058] Based on the node scenario mapping set, the label classification attributes of all nodes in the record are standardized. This process converts character-type categories into numeric classification codes for easier subsequent operations. For example, categories A, B, and C are uniformly converted into codes 1, 2, and 3, respectively, resulting in a standardized field list. Then, a matching judgment is made based on the target scenario type value specified by the inference task. Assuming the current inference task requires processing the "financial risk control" scenario (scenario type labeled T1), the system filters out all node numbers corresponding to T1 from the mapping set, namely E001, E003, and E005. Next, it determines whether their label attributes are consistent with the task requirements. For example, if the task is labeled as type A, then all three tags are type A and meet the requirements, so the system includes them in the consistency set. If there is a node E006 that is T1 but labeled as type B, then that node is excluded and not included in the set. The filtering process does not require complex calculations; it only requires direct comparison of string fields or corresponding codes. If there is a requirement for fuzzy matching, then a classification consistency judgment threshold needs to be set. For example, if it is set to complete consistency, that is, the proportion of matching fields is 100%. Otherwise, the node is not matched. After the above processing, a set of entity numbers containing all scene types and label attributes that are consistent with the task is obtained, and finally, a scene consistency entity number set is obtained.

[0059] The task entity matching submodule performs task number association processing based on the scenario-consistent entity number set, establishes a one-to-one mapping between task numbers and entity numbers, extracts the entity numbers that meet the requirements based on the matching of the tag attributes of the entity nodes and the task scenario, and obtains a list of task-matched entity numbers.

[0060] Based on the scenario-consistent entity ID set, the system further establishes a mapping relationship between task IDs and entity IDs. Each task is assigned a unique ID; for example, T20240601 represents a financial task generated on June 1, 2024. When establishing the mapping, the task ID is used as the key, and the three entity IDs E001, E003, and E005 are grouped into an array as the value, forming a key-value pair for storage. This facilitates fast retrieval and indexing in the database or memory structure. If the task has additional requirements, such as limiting the selection to only recently created or recently active entities, the node timestamp or activity field can be further called for additional filtering. For example, from E001, E003, and E005, E003 was last updated to June 1, 2024, E005 to May 30, and E001 to May 28. After sorting by time, only the first two items can be selected to form the final task entity list. This list serves as the core input in the inference task execution process and runs through the subsequent task processing flow, ultimately obtaining the task matching entity ID list.

[0061] Please see Figure 3 The graph edge state recognition module includes:

[0062] The edge connection information extraction submodule extracts all edge information that has a connection relationship with the entity number in the task matching list. It collects the starting node number, ending node number and corresponding relationship type identifier of each edge, organizes and cleans the format according to the structure field, establishes a combined record of starting node, ending node and relationship type, and obtains the edge node structure table.

[0063] When operating based on a task-matched entity ID list, it is necessary to first retrieve all entity IDs recorded in the list and identify their valid node identities in the graph database. For example, if the task ID is T20240620, and the corresponding entity ID list is E001, E005, E009, etc., then it is necessary to query all records in the Neo4j graph database that have direct edge connections with E001, E005, and E009. The extracted content includes the starting node ID, ending node ID, and relation type fields of the edge. This field is generally represented in English form, such as "ASSOCIATED_WITH", "LINKED_TO", "LOCATED_IN", etc., and then the extracted raw data is processed. Edge records are converted into structured data rows, with the field order arranged as "starting point number → relation type → ending point number". A table structure or dataset is then created for subsequent processing. During the structuring process, missing or redundant fields in the original data need to be processed. For example, if an edge is missing ending point information, it should be removed. At the same time, the relation type identifier should be uniformly set to lowercase to avoid duplicate classifications caused by field branches. For example, if node E001 has an edge connecting to E010 with the relation "connected_to", the record should be "E001-connected_to-E010". If the same node has multiple edges, they should be recorded repeatedly. Table 2 lists some data examples from the edge connection information extraction stage.

[0064] Table 2 Examples of graph edge connection extraction

[0065] Starting point number Relationship type End point number E001 connected_to E010 E001 associated_with E011 E005 linked_to E015 E009 located_in E012 E009 connected_to E018

[0066] As shown in Table 2, node E001 has two outgoing edges that connect to E010 and E011 respectively. The relationship types are different, so each one needs to be recorded. After the data collection is completed, all records are integrated into a structured data table, which is used as the input basis for subsequent edge state judgment and classification, thus obtaining the edge node structure table.

[0067] The edge state filtering submodule judges the starting node number of each edge record one by one according to the edge node structure table, and verifies whether there is a matching relationship by comparing it with the task matching entity number list. If the starting node number does not appear in the list, the corresponding edge is marked as an unparticipating state; otherwise, it is marked as a participating state. After marking, all edge records in the unparticipating state are removed to obtain the set of available edge connections.

[0068] Based on the edge node structure table, the consistency of the starting node number field of each record needs to be checked against the task matching entity number list from the previous step. This check is performed by comparing the starting node number of each edge with each entity number in the list. If a match is found, the edge is considered a valid participating edge; otherwise, it is considered invalid and removed. This check needs to be performed row by row on all edges in the structure table. For example, if the list contains three nodes: E001, E005, and E009, and the starting node number of an edge is E010, it will not appear in the list, and thus the edge will be considered unparticipating and marked. If a record is marked as "invalid", it is removed from the structure table. If it starts with E005, it is marked as "valid" and retained for the next processing step. To enhance the consistency of judgment, an auxiliary identification field "status flag" can be set. A value of "1" indicates participation and "0" indicates non-participation. In actual operation, if 15 edge records are extracted from the structure table, and only 9 starting nodes appear in the task matching list after filtering, 6 non-participation edges are removed to form a new set as the set of valid edge connections. This set will only contain connection behaviors initiated by task matching nodes, and finally, the set of usable edge connections is obtained.

[0069] The edge state record construction submodule extracts the relationship types of all edges based on the available edge connection set, categorizes and organizes edges with the same relationship type, records the start and end node numbers associated with each type of edge, integrates them into a standard record format, and establishes a graph edge connection available state record table.

[0070] Based on the available edge connection set, each edge needs to be categorized and organized according to its "relationship type" field. Edges with the same relationship type identifier are merged and summarized to form a mapping structure between relationship categories and edge sets. After categorization, the basic information of each type of edge is statistically analyzed. The statistical items include the number of edges under each relationship category, a list of starting node numbers, and a list of ending node numbers. The statistical structure is uniformly output in the form of "relationship identifier + number of edges + starting set + ending set" and organized in a table structure. For example, if the "connected_to" category contains 4 edges, involving starting node numbers E001, E005, and ending node numbers... For edge numbers E010, E011, and E018, the record would be "connected_to, 4, E001|E005, E010|E011|E018". Additionally, a unique identifier value needs to be set as a field key for each type of relation identifier to facilitate indexing in the database and generating a structured record table. Simultaneously, the task numbers are bound to form a state mapping record between tasks and relation categories. If the task number is T20240620, the record table under this task contains three edge types. After integrating all edge information, a unified record data table is output. The completed record table is the graph edge connection available state record table.

[0071] Please see Figure 4 The path activation building block includes:

[0072] The edge path grouping submodule is based on the graph edge connection available state record table. According to the continuous connection relationship between the start node and the end node, it detects the edge connection structure one by one, constructs the edge path chain with the first and last connection features, establishes the edge set of the connection path based on the continuity of the first and last nodes, and summarizes it into a structured dataset according to the path number to obtain the edge path grouping set.

[0073] Based on the available state record table of graph edge connections, we extract all edge entries with a state field marked "1" from all edge records. Each edge contains a start node number, an end node number, and a relationship type identifier. Before grouping paths, we need to perform a continuity analysis on the edge set based on the node connection situation, that is, determine whether there is a start-end connection relationship between adjacent edges. If the end node number of edge A is the same as the start node number of edge B, then they are grouped into the same path group. During the processing, the edge set needs to be sorted according to the start node order, and then the structure record is traversed from top to bottom, recording the path number, and then... The path content is expanded sequentially according to the connection order. If a broken edge is encountered (without a continuous endpoint connection), the current path is ended and the next path is started. For example, the edges “E001→E003” and “E003→E006” can belong to the same path group, while “E001→E004” is independently identified as a new path if it has no connection with the previous one. All paths need to record their path number and the composition of the edge items they contain, and output them in a list in the structure table. The path number is represented by an increasing sequence, such as P001, P002, etc. The table below shows the edge status markers and path connections.

[0074] Table 3 Example of Edge State Path Connection

[0075]

[0076]

[0077] As shown in Table 3, path P001 is a path consisting of two connected edges, and path P002 is an independent single-sided path. Each path in this table is composed of continuous connected edges. After grouping, a basic structure record that can be directly used for direction extraction and structure generation is formed, resulting in the edge path grouping set.

[0078] The path direction numbering submodule groups the edge paths into sets, extracts the connection order and direction information of the edges in each set of paths, and identifies the path direction as positive by the direction from the start point of the edge to the end point. If there is a reverse connection in the path, the direction is marked with a negative sign. Combine the position and direction of each edge in the path to establish a direction order comparison record and obtain the path edge direction order table.

[0079] Based on the edge path grouping set, extract the connection order and direction of all edges in the path. The path direction is marked as "+1" if the starting point points to the ending point. If the ending point points back to the previous node in the actual connection, it is marked as "-1". Each path record needs to be expanded according to the path number, and a sequence number is set for each edge in the path, starting from 1 and incrementing. This forms a three-field record structure of path number + sequence number + direction value. At the same time, the mapping relationship between the direction value and the edge number is recorded. If the path P001 contains 3 edges with directions of "+1, +1, -1", it means that the first two edges continue normally according to the connection direction, and the third edge is a backtracking connection. During the recording process, the relationship between the starting point and the ending point number needs to be judged for each edge. If the starting point is earlier than the ending point on the path, the direction is marked as "+1"; otherwise, it is marked as "-1". In this way, a direction order structure record table is constructed. The following is the direction order style data.

[0080] Table 4 Example of Path Direction Sequence

[0081] Path number edge sequence number Direction value P001 1 +1 P001 2 +1 P001 3 -1 P002 1 +1

[0082] As shown in Table 4, the direction value in each path is determined by the edge order and the node connection relationship. If there is no path connection break, the direction value is continuous and consistent; otherwise, a complex path reverse jump structure can be formed. After the table is completed, it can be directly referenced in the subsequent path syntax construction stage to finally obtain the path edge direction order table.

[0083] The path sequence generation submodule processes the format of the edge direction number sequence recorded in each path according to the path edge direction order table, concatenates the edge direction expressions in sequence, establishes a one-to-one correspondence between the path number and the string result, binds it to the task number for output, and establishes a task applicable path sequence structure table.

[0084] Based on the path edge direction order table, each path needs to map its sequential direction data into a SPARQL1.1 PropertyPath format syntax string. Each direction marker in the path corresponds to a path operator: the positive direction marker "+1" is converted to " / ", and the negative direction marker "-1" is converted to "^". Combining the path number and edge type order, each edge is concatenated into a path expression. For example, if path P001 is "connected_to / linked_with / ^located_in", it is expressed as "connected_to / linked_with / ^located_in". All paths... The path expression is organized uniformly with the path number as the index, and the output structure is "task number + path number + syntax path string". If the task number is T20240701, then a record in the path structure table is "task number: T20240701, path number: P001, path string: connected_to / linked_with / ^located_in". The final output data structure adopts the standard table structure, with fields "task number, path number, path expression" respectively. All path records are aggregated by task number to form a task-level path structure list and establish a task-applicable path sequence structure table.

[0085] Please see Figure 5 The logical jump matching module includes:

[0086] The path interruption location submodule extracts the connection structure expression in each path according to the applicable path sequence structure table of the task, judges the position of the break point of the syntax structure one by one, obtains the end node number of the tail edge of the interrupted path as the starting point of the jump node, compares the end symbols of the path structure one by one according to the path number, judges the connection continuity indicator, extracts the starting node number corresponding to all interruption positions, and establishes the path interruption starting node set.

[0087] Based on the task's applicable path sequence structure table, each path needs to be parsed sequentially to extract the final structural form of the path expression. The presence of connection interruptions is identified through symbolic structure. If a path terminates at a reverse structure marker or the symbol combination is incomplete, the path is marked as interrupted, and the interruption location is located. The endpoint node number connected by the last edge of this path is then extracted as the starting node of a potential jump path. All task paths are traversed, and the terminating edges of interrupted paths are recorded sequentially, along with their endpoint node numbers, forming a set of numbers. In a real-world scenario, if the path expression is " / typeX / typeY / ^typeZ", it indicates that the connection path is interrupted at the typeZ node. The program needs to extract the node number corresponding to typeZ. The path termination structure is identified by a combination of path number and termination node number, such as N_005, and recorded as a broken link node. In the system implementation, this combination is used to identify the termination structure of the path, such as "P002→N_005". This combination facilitates the subsequent screening of the starting position of the jump connection. In addition, for some paths with incomplete syntax but actual broken connections, the endpoint node field of the terminating edge should be directly extracted from its edge combination structure. That is, if the last edge in the path record is "E_012: N_004→N_007", then N_007 should be extracted as the jump starting node. All broken node numbers are indexed in the form of "node number + path number" to facilitate structural comparison and mapping with subsequent paths. Some broken node structure records are shown in the table below:

[0088] Table 5. Example of Termination Node for Interrupted Path

[0089] Path number End number Starting point node End point P002 E_012 N_004 N_007 P005 E_024 N_011 N_019 P007 E_038 N_021 N_028

[0090] As shown in Table 5, paths P002, P005, and P007 are interrupted at nodes N_007, N_019, and N_028, respectively, and need to be used as the starting nodes for jump path screening to obtain the set of path interruption starting nodes.

[0091] The jump path screening submodule is based on the set of starting nodes of the path interruption. It screens all connection edge records in the graph that start from the corresponding node, excludes edge items with the same path number as the original path number, retains only the connection records belonging to different path numbers, extracts the corresponding end node number and connection type, and obtains the jump path candidate information table.

[0092] Based on the set of starting nodes of the path break, the edge connection information in the graph structure is screened to extract all connection edges originating from these nodes. The target endpoint node, connection type, and path affiliation number are recorded. In practical applications, if N_007 has a connection edge "E_048: N_007→N_030" belonging to path P009, and the broken path number is P002, then this edge can constitute a candidate jump path record. Otherwise, it should be excluded from the candidate set. Connections with the same path number or different directions should be excluded. Furthermore, the connection type field should be identified using a standard dictionary classification. Examples of acceptable field names include "linked_with", "controlled_by", and "triggered_to". Field lengths should be standardized to no more than 20 characters. During screening, the maximum number of candidate paths for the same node should be limited to 20. If the number of candidate edges exceeds this threshold, the paths should be sorted in descending order by connection timestamp, retaining the 10 most recent connection path records. During field extraction, connection direction markers should also be included. Direction values ​​use "+1" to indicate a connection between the start and end points, and "-1" to indicate that the end point points to the start point. Some screened path information records are shown below:

[0093] Table 6. Jump Candidate Path Structure Information Table

[0094] Starting point node Jump edge number End point Path number Connection type Direction markers N_007 E_048 N_030 P009 linked_with +1 N_007 E_055 N_033 P011 triggers -1 N_019 E_065 N_045 P012 controlled_by +1

[0095] As shown in Table 6, nodes N_007 and N_019 both have jump candidate edges that connect to target nodes under other path numbers in the graph, providing a data basis for subsequent structural similarity comparison and obtaining the jump path candidate information table.

[0096] The structural comparison and evaluation submodule extracts the structural attributes of the endpoint nodes in the candidate paths from the jump path candidate information table, compares them with the attributes of the target nodes in the interrupted paths, and uses attribute-level similarity weighting calculation, employing the formula:

[0097]

[0098] The calculation yields the jump structure comparison score, sorts and summarizes the comparison scores of candidate paths, and establishes a detailed map of path connectivity and jump mapping; where A i B represents the value of the i-th structural attribute field of the candidate path endpoint node. i W represents the value of the i-th structural attribute field of the target node in the broken chain path. i D represents the importance weight of attribute field i. iThe standard number of the dimension category where attribute field i is located, n represents the total number of comparable attribute fields between the candidate path endpoint structure and the target structure, m represents the number of all available structural attribute fields, and S′ represents the structural comparison normalization score. The closer this value is to 0, the more similar the path structures are.

[0099] Based on the jump path candidate information table, the structural attribute values ​​carried by the endpoint nodes of the candidate paths need to be extracted and compared with the attribute values ​​of the target nodes of the interrupted paths. A unified attribute numbering rule is set, and the attribute fields of the candidate nodes and target nodes must completely match the numbering before entering the comparison process. In the comparison, the normalized difference term is calculated first. Then extract the weight value W of this field. i With dimension number D i The weight coefficient term is obtained by multiplying the two and taking the square root. The results of each attribute field are calculated and then summed in a weighted manner to form the numerator. The sum of the weighted coefficients of all attribute fields is the denominator. The denominator is then substituted into the formula to calculate the structural comparison score S′.

[0100] Taking candidate node N_030 and target node N_019 as examples, their structural attributes are compared in the following table:

[0101] Table 7 Example Table of Jump Structure Attribute Comparison

[0102] Attribute Number <![CDATA[Candidate structure attribute value A i > <![CDATA[Target structure attribute value B i > <![CDATA[Weight value W i > <![CDATA[Dimension number D i > Attr_1 12.50 10.0 0.4 1 Attr_2 0.76 0.8 0.2 2 Attr_3 240.00 300.0 0.3 3 Attr_4 1.00 1.0 0.1 1

[0103] Among them, the weight value W i The settings are determined based on the influence of each structural attribute field in path semantic matching, with values ​​limited to between 0.05 and 0.5. Taking Attr_1 as an example, its attribute type is "entity label matching frequency," which is the primary basis for judging whether a jump node has structural consistency, and therefore it is given a high weight of 0.4. Attr_2 is "control domain coverage ratio," and its influence fluctuates significantly with the domain span. When the path span is less than 3 hops, the weight is usually lower than 0.25, so it is set to 0.2. Attr_3 is "node intra-connection edge density," which reflects the aggregation ability of the jump structure in the local network. The weight setting is proportional to the average number of outgoing edges of the node in the graph, and it is set to 0.3 when the current average number of outgoing edges is 15. Attr_4 is a Boolean flag field indicating whether it is a terminal node. It has little impact on path continuation, and its weight is set to 0.1, which remains constant across all paths. The weight values ​​mentioned above are all fixed and do not change with the task number, but can be adjusted according to changes in path type and node metric. An automatic weight optimization mechanism can be introduced for iterative fine-tuning in subsequent expansions.

[0104] The calculation process is as follows:

[0105]

[0106] Sum of numerators:

[0107] 0.158 + 0.0316 + 0.1898 = 0.3794;

[0108] Sum of denominators:

[0109]

[0110] The final comparison score is:

[0111]

[0112] The results indicate that the structural properties have relatively small errors and can be used as a basis for judging the validity of structural jump paths, generating detailed mapping information for graph path connectivity jumps.

[0113] The jump structure comparison score is used to measure the similarity between the endpoint node of the candidate jump path and the target node of the broken link path at the structural attribute level. This score is calculated by weighted aggregation of normalized differences between multiple structural attribute fields. The closer the value is to 0, the closer the candidate node and the target node are in each key attribute, and the stronger their overall structural consistency. Conversely, the closer the score is to 1, the greater the attribute difference between the two and the higher the degree of structural deviation. This score not only considers the similarity of the numerical values ​​of each attribute field, but also comprehensively incorporates attribute weights and dimensional complexity factors. Through normalization and weighting mechanisms, a unified measurement of different types of attributes is achieved in the comparison process. Therefore, this comparison score can serve as an important criterion for judging whether the path connectivity jump relationship is valid, and provides a clear quantitative standard for subsequent jump path screening, structural reconstruction and path repair.

[0114] The formula first examines the candidate structural attribute value A. i With the target structural attribute value B i The difference is normalized, that is... The differences in attributes with different units and numerical levels can be compressed to the same order of magnitude to ensure horizontal comparability; then, the weight W corresponding to the attribute field is introduced. i With dimension number D i The product of the two and the square root result creates a non-linear amplification mechanism, meaning that the higher the attribute weight or the more complex the attribute dimension, the greater its influence on the overall comparison. The introduction of the square root function is used to balance the sensitivity difference between attribute importance and dimensional complexity, avoiding calculation bias caused by an excessive proportion of a certain type of factor. All weighted normalized differences are accumulated in the numerator to form the weighted sum of the overall differences. At the same time, all attribute items are... The summation is used as the normalization benchmark value and placed in the denominator to form a normalized ratio structure, ensuring that the comparison score S′ is always in a unified range and has context independence. The entire formula logic maintains mathematical rigor and compatibility with engineering universality, and has logical support in three aspects: dimensional consistency, numerical normalization and weight adjustment.

[0115] Please see Figure 6 The path structure status summary module includes:

[0116] The triplet state recognition submodule is based on the graph path connectivity jump mapping details. It combines the structural record of each edge in the mapping path to extract the start node, end node and connection predicate code corresponding to each edge, forming a standard triplet field combination. It identifies the predicate state in the triplet and obtains the edge state identification dataset.

[0117] Based on the graph path connectivity jump mapping details, the edge structure data recorded in the mapping details is called one by one. The triplet data field corresponding to each edge is obtained and split into three elements: the starting node, the ending node, and the connection predicate. The state identifier value in the connection predicate field is extracted as the judgment criterion. By matching the state value with the state encoding table set in the graph database, the state category judgment operation is performed. When the state value is S1, it is directly marked as 1 to indicate an active state; when the state value is S2, it is marked as 0 to indicate a frozen state; and when the state value is S3, it is marked as -1 to indicate an abandoned state. The state judgment of all edges must be based on the predicate state mapping already set in the current structure version of the graph database. The rules ensure that the actual meaning of different state codes remains consistent in the node control logic. During implementation, the state value needs to be obtained through the "predicate_status" field in the predicate field. This field is usually embedded in the edge attribute and represented in JSON structure. In actual application scenarios, if the value of the predicate_status field in edge record E_014 is "S2", it should be mapped to 0 as a frozen state to participate in subsequent processing. Furthermore, the triplet structure of all edges and the state determination results are recombined into a quadruple structure and bound and collected according to the path number to finally obtain the edge state identifier dataset.

[0118] The edge state archiving submodule classifies all triplet structure records according to their state identifiers based on the edge state identifier dataset. It counts the edges after grouping them by state category and establishes a classification index number. It collects the edge structures corresponding to each state category and establishes a state archiving directory with the grouped state as the primary key to obtain the structure edge state classification table.

[0119] Based on the edge state identifier dataset, each edge structure record in the dataset is grouped according to the state identifier field. Edges with a state value of 1 are classified as active edges, edges with a state value of 0 as frozen edges, and edges with a state value of -1 as abandoned edges. An edge number index structure is established for each edge set, and the context path attribution information of the record set is grouped using the path number as the primary key. In implementation, the edge number uniformly adopts the format "E_xx", such as E_001, E_002, etc., where "xx" is the sequential number. By counting the number of edge structures in each edge set, the number of edge structures in each state can be obtained. In the edge archiving operation, the path number, edge number, and state identifier fields of each edge record are used as core fields to construct a state classification mapping table. A tree index structure is established based on the path number to group the states, facilitating the state summary operation at the path level.

[0120] In practical applications, if edge records E_003, E_004, and E_005 all belong to path P002, and their state values ​​are -1, 1, and 0 respectively, then they should be recorded as 1 abandoned edge, 1 active edge, and 1 frozen edge under path P002. This summary structure will be stored as a structural edge state classification table, as shown in the following examples:

[0121] Table 8 Examples of Side Structure State Classification

[0122] Path number Side numbering Status indicator P001 E_01 1 P001 E_02 0 P002 E_03 -1 P002 E_04 1 P002 E_05 0 P003 E_06 1

[0123] As shown in Table 8, the state distribution under different path numbers is clearly marked. The state under path P001 is 1 and 0, path P002 contains three types of state structures, and path P003 only contains active state edges. Finally, the structure edge state classification table is obtained.

[0124] The connected path structure generation submodule summarizes the state of the edges in each path according to the structure edge state classification table, gathers all edge state identifiers according to path number, counts the proportion of state in each path, merges and establishes a structure control path number set, and generates a knowledge graph-driven reasoning center path control structure table.

[0125] Based on the edge state classification table, a traversal extraction operation is performed on the edge state records under each path number. First, the edge set belonging to the path number is retrieved, and the state value corresponding to each edge is recorded. Then, three indicators are calculated: the number of active edges, the number of frozen edges, and the number of abandoned edges in the path. Each of the three indicators is divided by the total number of edges in the path to obtain the proportion of the three types of states for the path. In the identification of path state structure differences, the range of the three proportions is used as the basic judgment parameter for the fluctuation of path structure state. For example, if path P002 has 1 active edge, 1 frozen edge, and 1 abandoned edge, and a total of 3 edges, then the proportion of the three types of states is... The values ​​are 0.33, 0.33, and 0.33, respectively, and the state range is 0. During the judgment process, the critical value for structural state fluctuation is set to 0.6. That is, when the difference between any two of the three state ratios is greater than 0.6, it is marked as an abnormal structural state fluctuation. The abnormal structural state fluctuation path number will be set to 1 in the state monitoring identifier field "ctrl_flag" in the final path control table. All paths that do not exceed this critical value are marked as 0. Finally, all path numbers, state ratios, fluctuation judgment values, and state flags are integrated into the path control table field set to establish a knowledge graph-driven reasoning center path control structure table.

[0126] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A knowledge graph-driven reasoning center system, characterized in that, The system includes: The tag entity filtering module obtains the inference task scenario type information, collects the scenario type tags of all entity nodes in the graph to perform type consistency judgment, filters the entity numbers that match the current task scenario, establishes task-related node mapping, and generates a list of task-matching entity numbers. The graph edge state recognition module extracts edge connection information based on the task matching entity number list, determines whether the starting node is in the list, marks edges that do not meet the conditions as unparticipable states and removes them, summarizes the connection state classification of participating edges, and generates a graph edge connection available state record table. The path activation construction module performs path grouping processing based on the graph edge connection available status record table, constructs the attribute path activation sequence in sequence, and generates a task applicable path sequence structure table. The logical jump matching module locates the starting node number of the path interruption position according to the task applicable path sequence structure table, screens the connection information in the graph that starts from the corresponding node but belongs to other paths, compares the attribute of the end node structure type with the target structure type of the interrupted path, counts the number of attribute hits and calculates the comparison score, and generates a graph path connection jump mapping detail.

2. The knowledge graph-driven reasoning center system according to claim 1, characterized in that, The task matching entity number list includes an entity number list, entity type labels, and a task scenario mapping table. The graph edge connection available status record table includes participating edge numbers, connection start and end node numbers, and edge type status identifiers. The task applicable path sequence structure table includes path numbers, edge connection order, and attribute path syntax sequence. The graph path connectivity jump mapping details include interrupted path start number, candidate jump path number, and attribute comparison score.

3. The knowledge graph-driven reasoning center system according to claim 1, characterized in that, The tag entity filtering module includes: The entity scene extraction submodule collects the scene type labels of all entity nodes in the graph, records the scene label number and scene classification attribute of each entity node, organizes the correspondence between scene number and node number in a set manner, and obtains the node scene mapping set. The type consistency judgment submodule performs unified and standardized processing on the scene classification attributes of each entity node based on the node scene mapping set, judges the correspondence between the scene type to which the node belongs and the inference task scene type, identifies entity nodes that are consistent with the task scene type, filters the set of entity numbers whose scene classification attributes are consistent with the current task scene type, and obtains the scene consistent entity number set. The task entity matching submodule performs task number association processing based on the scenario-consistent entity number set, establishes a one-to-one mapping between task numbers and entity numbers, extracts entity numbers that meet the requirements based on the matching of the tag attributes of entity nodes with the task scenario, and obtains a list of task-matched entity numbers.

4. The knowledge graph-driven reasoning center system according to claim 1, characterized in that, The graph edge state recognition module includes: The edge connection information extraction submodule extracts all edge information that has a connection relationship with the entity number in the task matching list based on the task matching entity number list. It collects the starting node number, ending node number and corresponding relationship type identifier of each edge, organizes and cleans the format according to the structure field, establishes a combined record of starting node, ending node and relationship type, and obtains the edge node structure table. The edge state filtering submodule judges the starting node number of each edge record one by one according to the edge node structure table, and verifies whether there is a matching relationship by comparing it with the task matching entity number list. If the starting node number does not appear in the list, the corresponding edge is marked as unparticipable; otherwise, it is marked as participateable. After marking, all edge records in unparticipable states are removed to obtain the set of available edge connections. The edge state record construction submodule extracts the relationship types of all edges based on the available edge connection set, categorizes and organizes edges with the same relationship type, records the start and end node numbers associated with each type of edge, integrates them into a standard record format, and establishes a graph edge connection available state record table.

5. The knowledge graph-driven reasoning center system according to claim 1, characterized in that, The path activation building module includes: The edge path grouping submodule, based on the available state record table of the graph edge connection, detects the edge connection structure one by one according to the continuous connection relationship between the starting node and the ending node, constructs the edge path chain with the first and last connection features, establishes the edge set of the connection path based on the continuity of the first and last nodes, summarizes it into a structured dataset according to the path number, and obtains the edge path grouping set. The path direction numbering submodule extracts the connection order and direction information of the edges in each group of paths based on the edge path group set. The path direction is marked as positive by the direction from the start point of the edge to the end point. If there is a reverse connection in the path, the direction is marked with a negative sign. The direction order comparison record is established by combining the position and direction of each edge in the path to obtain the path edge direction order table. The path sequence generation submodule processes the edge direction number sequence recorded in each path according to the path edge direction order table, performs format standardization processing, concatenates the edge direction expressions in sequence, establishes a one-to-one correspondence between the path number and the string result, binds it to the task number for output, and establishes a task applicable path sequence structure table.

6. The knowledge graph-driven reasoning center system according to claim 1, characterized in that, The logical jump matching module includes: The path interruption location submodule extracts the connection structure expression in each path according to the task applicable path sequence structure table, judges the position of the syntax structure interruption point one by one, obtains the end node number of the tail edge of the interrupted path as the starting point of the jump node, compares the end symbols of the path structure one by one according to the path number, judges the connection continuity indicator, extracts the starting node number corresponding to all interruption positions, and establishes a path interruption starting node set. The jump path screening submodule, based on the set of path interruption starting point nodes, screens all connection edge records in the graph that start from the corresponding node, excludes edge items that are the same as the original path number, retains only the connection records that belong to different path numbers, extracts the corresponding end node number and connection type, and obtains the jump path candidate information table. The structural comparison and evaluation submodule extracts the structural attributes of the endpoint nodes in the candidate paths based on the jump path candidate information table, compares them with the attributes of the target nodes of the interrupted paths, and uses attribute-level similarity weighted calculation to obtain the jump structure comparison score. The candidate paths are sorted and summarized to establish a graph path connectivity jump mapping detail.

7. The knowledge graph-driven reasoning center system according to claim 1, characterized in that, The system also includes: The path structure state summary module identifies the current state marker of each edge in each path based on the graph path connection jump mapping details, classifies, counts and archives all edges according to triplet state identifiers, forms a structure state summary record that runs through all jump segments and original paths, and generates a knowledge graph-driven reasoning center path control structure table. The knowledge graph-driven reasoning center path control structure table includes triplet state classification statistics, edge state archive list, and path segment structure markers.

8. The knowledge graph-driven reasoning center system according to claim 7, characterized in that, The path structure status summary module includes: The triplet state recognition submodule, based on the graph path connectivity jump mapping details and combined with the structural record of each edge in the mapping path, extracts the starting node, ending node and connection predicate code corresponding to each edge to form a standard triplet field combination, identifies the predicate state in the triplet, and obtains the edge state identification dataset. The edge state archiving submodule classifies all triplet structure records according to their state identifiers based on the edge state identifier dataset. It counts the edges after grouping them by state category and establishes a classification index number. It collects the edge structures corresponding to each state category and establishes a state archiving directory with the grouped state as the primary key to obtain the structure edge state classification table. The connected path structure generation submodule summarizes the states of edges in each path according to the structure edge state classification table, gathers all edge state identifiers by path number, counts the proportion of states in each path, merges them to establish a structure control path number set, and generates a knowledge graph-driven reasoning center path control structure table.

Images