Event clue multi-dimensional knowledge base construction method and system based on knowledge graph

Abstract

The application discloses a knowledge graph-based event clue multi-dimensional knowledge base construction method and system, which comprises the following steps: collecting and preprocessing multi-source event clues, constructing a standardized knowledge graph, and generating a multi-dimensional knowledge base; performing graph layering based on improved K-core decomposition, identifying high-cohesion event groups, forming K-core subgraphs and boundary structures; setting walk parameters of subgraphs with different K values, formulating in-and-out community walk strategies, and updating knowledge structures; performing multi-round layering walks, generating event node path sequences covering in-and-out community structures; aggregating walk sequence features, generating clue vectors, completing clustering, reasoning and completion, and optimizing the knowledge base; triggering graph layering and walks when new clues are added, and periodically or real-timely updating the multi-dimensional knowledge base structure. The application realizes the structured expression and efficient aggregation of multi-source event clues by introducing K-core decomposition and layering random walk methods, and constructs a multi-dimensional knowledge base which can dynamically evolve to support intelligent analysis and reasoning of complex events.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence reasoning technology, and in particular to a method and system for constructing a multidimensional knowledge base of event clues based on knowledge graphs. Background Technology

[0002] With the rapid development of information technology and big data technology, knowledge graphs, as structured semantic networks, have been widely applied in various fields such as event analysis, intelligent recommendation, financial risk control, and public opinion monitoring. Existing knowledge graph construction methods typically rely on extracting entities and relationships from multi-source heterogeneous data to initially form a network structure of nodes and edges. However, in the actual process of multi-dimensional event clue fusion and intelligent knowledge base optimization, traditional techniques generally suffer from shortcomings such as insufficient structural layering capabilities, difficulty in finely identifying complex event communities, limited multi-attribute expression of nodes and edges, and insufficient support for dynamic evolution. Especially when facing massive heterogeneous event clues, existing methods often only use simple node degrees or static features for structural division, making it difficult to mine multi-level, highly cohesive event subgroups, thus limiting the performance of knowledge graphs in scenarios such as event reasoning, anomaly detection, and source tracing analysis.

[0003] Traditional random walk sampling methods, while possessing some local sampling capabilities in knowledge graph structure analysis, lack mechanisms for dynamic adjustment based on community structure, hierarchical attributes, and node importance. This often results in uneven structural coverage, insufficient sampling of key event nodes, or information redundancy, limiting the diversity and representativeness of event clue feature representation. In practical applications, knowledge bases often need to continuously integrate newly added event data to achieve real-time structural optimization and adaptive knowledge evolution. However, most current methods offer limited support for dynamic incremental updates and structural self-optimization of knowledge bases, failing to meet the continuous management and intelligent optimization needs of high-dimensional, high-frequency dynamic knowledge in complex scenarios.

[0004] Therefore, how to provide a method and system for constructing a multidimensional knowledge base based on knowledge graphs for event clues is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for constructing a multi-dimensional knowledge base based on knowledge graphs for event clues. This invention integrates an improved K-kernel decomposition algorithm with a structured hierarchical adaptive random walk mechanism, and describes in detail the entire process of standardization processing of multi-source event clues, hierarchical modeling of knowledge graphs, adaptive parameter sampling, and feature aggregation reasoning. This invention can accurately identify complex event community structures, dynamically extract multi-granularity key clues, support real-time optimization and evolution of the knowledge base, and has the advantages of strong structural expression ability, high event reasoning accuracy, comprehensive sampling coverage, and strong knowledge adaptive evolution ability.

[0006] The method for constructing a multi-dimensional knowledge base for event clues based on knowledge graphs according to embodiments of the present invention includes:

[0007] Collect multi-source event clue data, preprocess the multi-source event clue data, construct a knowledge graph, standardize the attribute information of nodes and edges in the knowledge graph, and form a multi-dimensional knowledge base;

[0008] Based on the improved K-kernel decomposition, the knowledge graph in the multidimensional knowledge base is processed in layers to identify high cohesive event subgroups at each level, determine the number of kernels for each node, and form multiple K-kernel subgraphs and boundary structures.

[0009] Based on multiple K-core subgraphs and boundary structures, hierarchical adaptive random walk parameters, including walk path length and step probability, are set for subgraphs with different K values. Random walk strategies within and between communities are designed to update the structural parameters of the multidimensional knowledge base.

[0010] Based on the random walk parameters, high-frequency long-path random walks are implemented in the internal region of the community in the K-core graph, and low-frequency short-path random walks are implemented at the community boundary and in the low K-value region to obtain multiple event node walk sequences.

[0011] Based on multiple event node walking sequences, the node, edge, kernel number and attribute information are aggregated to generate multi-dimensional event clue feature vectors. Event clue clustering, relation reasoning and knowledge completion are performed on the knowledge graph in the multi-dimensional knowledge base to optimize the structure of the multi-dimensional knowledge base.

[0012] When new event clues are added, K-core decomposition and hierarchical random walk are dynamically triggered, and the structure evolution and continuous optimization of the multidimensional knowledge base are executed periodically or in real time.

[0013] Optionally, the multi-source event clue data specifically includes event-related text, images, logs, and sensor data from structured data, semi-structured data, and unstructured data.

[0014] Optionally, the preprocessing of multi-source event clue data specifically includes data cleaning, format conversion, noise removal, and information standardization.

[0015] Optionally, the construction of the knowledge graph and the standardization of the attribute information of the nodes and edges in the knowledge graph to form a multidimensional knowledge base refers to extracting entities, relationships and attributes from the preprocessed multi-source event clue data, establishing a graph structure with entities as nodes and relationships as edges, and performing unified format conversion, encoding standardization and numerical normalization processing on the node attributes and edge attributes, and storing them in the multidimensional knowledge base.

[0016] Optionally, the improved K-kernel decomposition-based hierarchical processing of the knowledge graph in the multidimensional knowledge base, identifying high-cohesion event subgroups at each level, determining the kernel number of each node, and forming multiple K-kernel subgraphs and boundary structures includes:

[0017] Knowledge graphs are extracted from multidimensional knowledge bases and represented as undirected graphs consisting of sets of nodes and edges. Attribute information is extracted from each node and node attribute weights are assigned. Edge weights are assigned to each type of edge. The effective degree of each node is calculated based on the node attribute weights and the edge weights of the connected edges, forming a weighted node degree dataset of the knowledge graph.

[0018] Based on the weighted node degree dataset, the knowledge graph is stripped in ascending order of degree. Nodes and related edges with a weighted effective degree lower than the current K value are recursively deleted. After completing one round of weighted stripping, the weights of the remaining nodes and edges are updated to obtain the subgraph structure and weighted node degree data of the current round.

[0019] For the obtained subgraph structure and weighted node degree data, the node degree distribution and attribute distribution are statistically analyzed. Based on the changes in the structure and attributes within the subgraph, the K value used in this round of stripping operation is dynamically adjusted to achieve adaptive layering threshold setting and output a new K value and subgraph.

[0020] Using the new K value and subgraph as input, a new round of weighted stripping is performed, recursively deleting nodes and edges in a loop. After each round of stripping, the core nodes and attribute features stripped from the previous layer are passed to the current subgraph, and the layering judgment conditions of the current node are updated in real time, forming a layered stripping process based on context feature interaction. After each round of stripping, the nodes that have a connection relationship between the current subgraph and the newly stripped nodes are marked as the boundary nodes of this layer.

[0021] During the layered stripping process, based on the weighting coefficients and stripping priorities of different edge types, retention thresholds and deletion thresholds are set for each edge type. Edges with weighting coefficients greater than or equal to the retention threshold are retained, and edges with weighting coefficients less than the deletion threshold are deleted. This ensures that the stripping results simultaneously reflect node attributes, edge weights, and relationship diversity, generating a layered subgraph structure that integrates multiple relationship types.

[0022] When processing large-scale knowledge graphs, the layering and stripping process is applied in parallel to multiple subgraphs. Each subgraph is synchronized through the boundary nodes of the intersection area to maintain the consistency of the overall structure. After each round of batch processing, the global subgraph structure and node core allocation results are merged and updated.

[0023] When there are changes in the knowledge graph structure due to the addition or deletion of nodes or edges, for the changed area, based on the latest subgraph structure and weighted node degree data, incremental K-kernel decomposition is adopted, and only the affected subgraph and its neighborhood are locally stripped and the kernel number allocation is adjusted.

[0024] The generated subgraphs, boundary nodes, and boundary structures are aggregated, and the final output is a hierarchical result containing multiple K-core subgraphs and boundary structures.

[0025] Optionally, based on multiple K-core subgraphs and boundary structures, hierarchical adaptive random walk parameters are set for subgraphs with different K values, including walk path length and step probability; random walk strategies within and between communities are designed; and the structural parameters of the multidimensional knowledge base are updated, including:

[0026] The obtained multiple K-kernel subgraphs and their boundary structures are used as input. For each K-kernel subgraph, the node set, edge set, and weighted effective degree information of the nodes are obtained.

[0027] For each K-core subgraph, the random walk path length parameter of the subgraph is set according to the size of the node set, the connectivity of the edge set, and the weighted effective degree of the nodes, and the step probability parameter is set. The path length parameter is used to limit the maximum number of steps in each walk, and the step probability parameter is used to control the probability of transition between nodes during the walk.

[0028] Based on the hierarchical structure of the K-core graph, different parameters are set for random walks within communities and random walks between communities. For walks within communities, the path length parameter is set to be greater than or equal to a first threshold, and the step probability parameter is set to be greater than or equal to a second threshold. For walks between communities, the path length parameter is set to be less than the first threshold, and the step probability parameter is set to be less than the second threshold.

[0029] By combining edge weights and node attributes, a transition probability matrix is ​​constructed, where the matrix element m uv This represents the transition probability from node u to node v in the k-th subgraph. The transition probability is jointly determined by the node attribute weight, edge weight, and step probability parameter.

[0030] For different K-kernel subgraphs and boundary structures, the random walk path length parameter and step probability parameter of each subgraph are written into the multidimensional knowledge base.

[0031] Optionally, based on the random walk parameters, high-frequency long-path random walks are performed within the community region of the K-core graph, and low-frequency short-path random walks are performed at the community boundary and in low-K value regions to obtain multiple event node walk sequences, including:

[0032] The system includes a strategy switching and start determination module, which comprises a walk start point selection unit and a walk strategy switching unit. The walk start point selection unit dynamically determines the walk start node within the community, the community boundary, and the low K value region based on the hierarchical structure, node attributes, and boundary structure of the K-core subgraph. During the walk process, the walk strategy switching unit dynamically determines the strategy of normal walk, jump walk, or directional walk based on the region to which the current node belongs, the node attribute weight, and the historical sampling state, and adjusts the path generation method accordingly.

[0033] The system includes an information feedback and self-regulation module, which comprises a walk sequence statistics unit and a parameter self-regulation unit. The walk sequence statistics unit records in real time the nodes, edges, path length, number of steps, node coverage frequency, and sampling area distribution of each walk sequence. The parameter self-regulation unit dynamically adjusts the walk path length parameter, step probability parameter, starting point selection, and termination condition based on the statistical results.

[0034] A local parallel and interactive fusion module is set up, which includes a subgraph parallel walking unit and a walking information interaction unit. The subgraph parallel walking unit divides multiple K-core subgraphs and boundary structures into several subgraphs, and performs random walks in parallel within each subgraph to collect the walking sequence of event nodes. The walking information interaction unit periodically summarizes the walking paths and node coverage of each subgraph and feeds them back to the parameter self-regulation unit to achieve cross-subgraph sampling coordination and global sampling balance.

[0035] Under the coordinated action of the strategy switching and initial determination module, the information feedback and self-regulation module, and the local parallel and interactive fusion module, a random walk is implemented, specifically as follows:

[0036] The starting point selection unit dynamically selects the starting node for this round of random walks based on the hierarchical adaptive parameters, within the community, at the community boundary, or in a low K-value region.

[0037] During each round of walking, the walking strategy switching unit determines in real time whether to use a normal walking, jump walking, or directional walking strategy for the current step based on the current node attributes, layer position, historical path, and sampling status, and dynamically adjusts the path generation method.

[0038] Under the normal traversal strategy, based on the path length parameter and the step probability parameter, the transition probability is calculated among all the adjacent nodes of the current node by combining the node attribute weight and the edge weight, the next traversal node is randomly selected, and the path is recorded.

[0039] Under the jump-walking or directional walking strategy, the walking strategy switching unit selects the node that is farther away or has specific attributes as the next walking node based on node attributes, structural features or walking history, and generates a long-distance or directional path.

[0040] The event node traversal sequence of a long path is completed until the path length threshold, sampling number threshold, or termination condition set within the community area is reached. In the community boundary and low K value area, multiple rounds of short path event node traversal are performed based on the criteria that the path length parameter is less than the first threshold and the step probability parameter is less than the second threshold, forming a sampling sequence covering the community boundary and low K area.

[0041] During the traversal, all collected path, node, edge, and parameter information is fed back to the information feedback and self-regulation module and the traversal information interaction unit in real time, dynamically optimizing the sampling distribution and realizing diversified structural sampling of different regions of the knowledge graph structure.

[0042] After each walk, record all nodes, edges, path parameters, sampling areas, and node coverage information involved in the walk path, and accumulate multiple event node walk sequences covering different regions of the knowledge graph;

[0043] All event node walk sequences, related parameters, and statistical information collected through random walks are written into a multidimensional knowledge base to complete the collection and storage of walk sequences.

[0044] Optionally, the step of aggregating node, edge, core, and attribute information based on multiple event node walk sequences to generate a multi-dimensional event clue feature vector, performing event clue clustering, relation reasoning, and knowledge completion on the knowledge graph in the multi-dimensional knowledge base, and optimizing the multi-dimensional knowledge base structure includes:

[0045] Using all event node walk sequences as input data, for each event node walk sequence, extract the set of nodes involved, the set of edges, node attributes, edge weights, path lengths and sampling area information to form a basic feature set;

[0046] Feature aggregation is performed on the basic feature set of all walk sequences. The frequency of occurrence of the same node in different walk sequences, the cumulative weight of the connected edges, the statistical characteristics of node attributes, the path length distribution, and the number of node cores are comprehensively statistically analyzed to form a multi-dimensional feature matrix at the node level and the path level.

[0047] Based on the multidimensional feature matrix, a clustering algorithm is used to cluster event nodes into event clues, grouping nodes with similar features into the same event clue cluster, and outputting the node set and clustering label of each cluster.

[0048] By utilizing the node connection relationships and feature clustering results in the walking sequence, a graph reasoning method is used to mine the potential relationships between nodes and fill in the missing node attributes or edge relationships. The completion criteria are based on the similarity of node feature vectors and structural proximity.

[0049] Based on clustering results, relational reasoning, and completion operations, the knowledge graph structure in the multidimensional knowledge base is updated in real time, including node attributes, edge relationships, number of node cores, and hierarchical structure.

[0050] The event clue multidimensional knowledge base construction system based on knowledge graph according to an embodiment of the present invention includes the following modules:

[0051] The data acquisition and preprocessing module is used to collect multi-source event clue data and preprocess it to build and standardize the knowledge graph, forming a multi-dimensional knowledge base.

[0052] The K-kernel decomposition and hierarchical module is used to perform K-kernel decomposition on the knowledge graph in the multidimensional knowledge base, identify high-cohesion event subgroups in layers, determine the number of node kernels, and form multiple K-kernel subgraphs and boundary structures.

[0053] The hierarchical walk parameter setting module is used to set hierarchical adaptive random walk parameters for K-core graphs and boundary structures with different K values, and to design walk strategies within and between communities;

[0054] The hierarchical random walk module is used to perform multiple rounds of high-frequency long-path or low-frequency short-path random walks in each K-core subgraph region based on the random walk parameters, and collect the walk sequence of event nodes.

[0055] The feature aggregation and analysis module is used to aggregate features from the event node walking sequence, generate multi-dimensional event clue feature vectors, and realize event clue clustering, relation reasoning and knowledge completion to optimize the knowledge base structure.

[0056] The dynamic evolution module is used to dynamically trigger K-core decomposition and hierarchical random walk when new event clues are added, and to periodically or in real time optimize the knowledge base structure.

[0057] The beneficial effects of this invention are:

[0058] This invention, based on structured modeling of knowledge graphs, innovatively introduces an improved K-kernel decomposition hierarchical method and a hierarchical adaptive random walk mechanism, achieving efficient fusion and multi-dimensional representation of multi-source event clues. Compared with existing technologies, this invention can not only accurately segment and identify highly cohesive, multi-level event subgroups in the knowledge graph, but also adaptively adjust the walk parameters for different community structures and node attributes, making event clue sampling more diverse and representative, and improving the ability of event feature aggregation and intelligent reasoning. This invention supports the dynamic evolution and incremental updates of the knowledge base structure, and can cope with the needs of frequently accessing new event data and knowledge structure changes in practical applications, ensuring that the knowledge graph always has good structural rationality and information integrity. This invention improves the expressive power, reasoning depth, and real-time response capability of the event knowledge base, meeting the needs of efficient management and intelligent processing of multi-dimensional heterogeneous event clues in complex scenarios, and has broad practical application prospects and technical promotion value. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of the method for constructing a multi-dimensional knowledge base based on knowledge graphs for event clues proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the structure of the event clue multidimensional knowledge base construction system based on knowledge graph proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1 A method for constructing a multi-dimensional knowledge base for event clues based on knowledge graphs, including:

[0064] Collect multi-source event clue data, preprocess the multi-source event clue data, construct a knowledge graph, standardize the attribute information of nodes and edges in the knowledge graph, and form a multi-dimensional knowledge base;

[0065] Based on the improved K-kernel decomposition, the knowledge graph in the multidimensional knowledge base is processed in layers to identify high cohesive event subgroups at each level, determine the number of kernels for each node, and form multiple K-kernel subgraphs and boundary structures.

[0066] Based on multiple K-core subgraphs and boundary structures, hierarchical adaptive random walk parameters, including walk path length and step probability, are set for subgraphs with different K values. Random walk strategies within and between communities are designed to update the structural parameters of the multidimensional knowledge base.

[0067] Based on the random walk parameters, high-frequency long-path random walks are implemented in the internal region of the community in the K-core graph, and low-frequency short-path random walks are implemented at the community boundary and in the low K-value region to obtain multiple event node walk sequences.

[0068] Based on multiple event node walking sequences, the node, edge, kernel number and attribute information are aggregated to generate multi-dimensional event clue feature vectors. Event clue clustering, relation reasoning and knowledge completion are performed on the knowledge graph in the multi-dimensional knowledge base to optimize the structure of the multi-dimensional knowledge base.

[0069] When new event clues are added, K-core decomposition and hierarchical random walk are dynamically triggered, and the structure evolution and continuous optimization of the multidimensional knowledge base are executed periodically or in real time.

[0070] In this embodiment, the multi-source event clue data specifically includes event-related text, images, logs, and sensor data from structured data, semi-structured data, and unstructured data.

[0071] In this embodiment, the preprocessing of multi-source event clue data specifically includes data cleaning, format conversion, noise removal, and information standardization.

[0072] In this embodiment, constructing a knowledge graph and standardizing the attribute information of nodes and edges in the knowledge graph to form a multidimensional knowledge base refers to extracting entities, relationships, and attributes from preprocessed multi-source event clue data, establishing a graph structure with entities as nodes and relationships as edges, and performing unified format conversion, encoding standardization, and numerical normalization on node attributes and edge attributes, and storing them in the multidimensional knowledge base.

[0073] In this embodiment, the step of performing hierarchical processing on the knowledge graph in the multidimensional knowledge base based on improved K-kernel decomposition, identifying high-cohesion event subgroups at each level, determining the number of kernels for each node, and forming multiple K-kernel subgraphs and boundary structures includes:

[0074] Knowledge graphs are extracted from multidimensional knowledge bases and represented as undirected graphs consisting of sets of nodes and edges. Attribute information is extracted from each node and node attribute weights are assigned. Edge weights are assigned to each type of edge. The effective degree of each node is calculated based on the node attribute weights and the edge weights of the connected edges, forming a weighted node degree dataset of the knowledge graph.

[0075] Based on the weighted node degree dataset, the knowledge graph is stripped in ascending order of degree. Nodes and related edges with a weighted effective degree lower than the current K value are recursively deleted. After completing one round of weighted stripping, the weights of the remaining nodes and edges are updated to obtain the subgraph structure and weighted node degree data of the current round.

[0076] For the obtained subgraph structure and weighted node degree data, the node degree distribution and attribute distribution are statistically analyzed. Based on the changes in the structure and attributes within the subgraph, the K value used in this round of stripping operation is dynamically adjusted to achieve adaptive layering threshold setting and output a new K value and subgraph.

[0077] Using the new K value and subgraph as input, a new round of weighted stripping is performed, recursively deleting nodes and edges in a loop. After each round of stripping, the core nodes and attribute features stripped from the previous layer are passed to the current subgraph, and the layering judgment conditions of the current node are updated in real time, forming a layered stripping process based on context feature interaction. After each round of stripping, the nodes that have a connection relationship between the current subgraph and the newly stripped nodes are marked as the boundary nodes of this layer.

[0078] During the layered stripping process, based on the weighting coefficients and stripping priorities of different edge types, retention thresholds and deletion thresholds are set for each edge type. Edges with weighting coefficients greater than or equal to the retention threshold are retained, and edges with weighting coefficients less than the deletion threshold are deleted. This ensures that the stripping results simultaneously reflect node attributes, edge weights, and relationship diversity, generating a layered subgraph structure that integrates multiple relationship types.

[0079] When processing large-scale knowledge graphs, the layering and stripping process is applied in parallel to multiple subgraphs. Each subgraph is synchronized through the boundary nodes of the intersection area to maintain the consistency of the overall structure. After each round of batch processing, the global subgraph structure and node core allocation results are merged and updated.

[0080] When there are changes in the knowledge graph structure due to the addition or deletion of nodes or edges, for the changed area, based on the latest subgraph structure and weighted node degree data, incremental K-kernel decomposition is adopted, and only the affected subgraph and its neighborhood are locally stripped and the kernel number allocation is adjusted.

[0081] The generated subgraphs, boundary nodes, and boundary structures are aggregated, and the final output is a hierarchical result containing multiple K-core subgraphs and boundary structures.

[0082] In this embodiment, the step of setting hierarchical adaptive random walk parameters for subgraphs with different K values ​​based on multiple K-core subgraphs and boundary structures, including walk path length and step probability, designing random walk strategies within and between communities, and updating the structural parameters of the multidimensional knowledge base includes:

[0083] The obtained multiple K-kernel subgraphs and their boundary structures are used as input. For each K-kernel subgraph, the node set, edge set, and weighted effective degree information of the nodes are obtained.

[0084] For each K-core subgraph, the random walk path length parameter of the subgraph is set according to the size of the node set, the connectivity of the edge set, and the weighted effective degree of the nodes, and the step probability parameter is set. The path length parameter is used to limit the maximum number of steps in each walk, and the step probability parameter is used to control the probability of transition between nodes during the walk.

[0085] Based on the hierarchical structure of the K-core graph, different parameters are set for random walks within communities and random walks between communities. For walks within communities, the path length parameter is set to be greater than or equal to a first threshold, and the step probability parameter is set to be greater than or equal to a second threshold. For walks between communities, the path length parameter is set to be less than the first threshold, and the step probability parameter is set to be less than the second threshold.

[0086] By combining edge weights and node attributes, a transition probability matrix is ​​constructed, where the matrix element m uv This represents the transition probability from node u to node v in the k-th subgraph. The transition probability is jointly determined by the node attribute weight, edge weight, and step probability parameter.

[0087] For different K-kernel subgraphs and boundary structures, the random walk path length parameter and step probability parameter of each subgraph are written into the multidimensional knowledge base.

[0088] In this embodiment, based on the random walk parameters, high-frequency long-path random walks are performed within the community region of the K-core graph, while low-frequency short-path random walks are performed at the community boundary and in low-K value regions to obtain multiple event node walk sequences, including:

[0089] The system includes a strategy switching and start determination module, which comprises a walk start point selection unit and a walk strategy switching unit. The walk start point selection unit dynamically determines the walk start node within the community, the community boundary, and the low K value region based on the hierarchical structure, node attributes, and boundary structure of the K-core subgraph. During the walk process, the walk strategy switching unit dynamically determines the strategy of normal walk, jump walk, or directional walk based on the region to which the current node belongs, the node attribute weight, and the historical sampling state, and adjusts the path generation method accordingly.

[0090] The system includes an information feedback and self-regulation module, which comprises a walk sequence statistics unit and a parameter self-regulation unit. The walk sequence statistics unit records in real time the nodes, edges, path length, number of steps, node coverage frequency, and sampling area distribution of each walk sequence. The parameter self-regulation unit dynamically adjusts the walk path length parameter, step probability parameter, starting point selection, and termination condition based on the statistical results.

[0091] A local parallel and interactive fusion module is set up, which includes a subgraph parallel walking unit and a walking information interaction unit. The subgraph parallel walking unit divides multiple K-core subgraphs and boundary structures into several subgraphs, and performs random walks in parallel within each subgraph to collect the walking sequence of event nodes. The walking information interaction unit periodically summarizes the walking paths and node coverage of each subgraph and feeds them back to the parameter self-regulation unit to achieve cross-subgraph sampling coordination and global sampling balance.

[0092] Under the coordinated action of the strategy switching and initial determination module, the information feedback and self-regulation module, and the local parallel and interactive fusion module, a random walk is implemented, specifically as follows:

[0093] The starting point selection unit dynamically selects the starting node for this round of random walks based on the hierarchical adaptive parameters, within the community, at the community boundary, or in a low K-value region.

[0094] During each round of walking, the walking strategy switching unit determines in real time whether to use a normal walking, jump walking, or directional walking strategy for the current step based on the current node attributes, layer position, historical path, and sampling status, and dynamically adjusts the path generation method.

[0095] Under the normal traversal strategy, based on the path length parameter and the step probability parameter, the transition probability is calculated among all the adjacent nodes of the current node by combining the node attribute weight and the edge weight, the next traversal node is randomly selected, and the path is recorded.

[0096] Under the jump-walking or directional walking strategy, the walking strategy switching unit selects the node that is farther away or has specific attributes as the next walking node based on node attributes, structural features or walking history, and generates a long-distance or directional path.

[0097] The event node traversal sequence of a long path is completed until the path length threshold, sampling number threshold, or termination condition set within the community area is reached. In the community boundary and low K value area, multiple rounds of short path event node traversal are performed based on the criteria that the path length parameter is less than the first threshold and the step probability parameter is less than the second threshold, forming a sampling sequence covering the community boundary and low K area.

[0098] During the traversal, all collected path, node, edge, and parameter information is fed back to the information feedback and self-regulation module and the traversal information interaction unit in real time, dynamically optimizing the sampling distribution and realizing diversified structural sampling of different regions of the knowledge graph structure.

[0099] After each walk, record all nodes, edges, path parameters, sampling areas, and node coverage information involved in the walk path, and accumulate multiple event node walk sequences covering different regions of the knowledge graph;

[0100] All event node walk sequences, related parameters, and statistical information collected through random walks are written into a multidimensional knowledge base to complete the collection and storage of walk sequences.

[0101] In this embodiment, the step of aggregating node, edge, core, and attribute information based on multiple event node walk sequences to generate a multi-dimensional event clue feature vector, performing event clue clustering, relation reasoning, and knowledge completion on the knowledge graph in the multi-dimensional knowledge base, and optimizing the multi-dimensional knowledge base structure includes:

[0102] Using all event node walk sequences as input data, for each event node walk sequence, extract the set of nodes involved, the set of edges, node attributes, edge weights, path lengths and sampling area information to form a basic feature set;

[0103] Feature aggregation is performed on the basic feature set of all walk sequences. The frequency of occurrence of the same node in different walk sequences, the cumulative weight of the connected edges, the statistical characteristics of node attributes, the path length distribution, and the number of node cores are comprehensively statistically analyzed to form a multi-dimensional feature matrix at the node level and the path level.

[0104] Based on the multidimensional feature matrix, a clustering algorithm is used to cluster event nodes into event clues, grouping nodes with similar features into the same event clue cluster, and outputting the node set and clustering label of each cluster.

[0105] By utilizing the node connection relationships and feature clustering results in the walking sequence, a graph reasoning method is used to mine the potential relationships between nodes and fill in the missing node attributes or edge relationships. The completion criteria are based on the similarity of node feature vectors and structural proximity.

[0106] Based on clustering results, relational reasoning, and completion operations, the knowledge graph structure in the multidimensional knowledge base is updated in real time, including node attributes, edge relationships, number of node cores, and hierarchical structure.

[0107] refer to Figure 2 The knowledge graph-based multidimensional knowledge base construction system for event clues includes the following modules:

[0108] The data acquisition and preprocessing module is used to collect multi-source event clue data and preprocess it to build and standardize the knowledge graph, forming a multi-dimensional knowledge base.

[0109] The K-kernel decomposition and hierarchical module is used to perform K-kernel decomposition on the knowledge graph in the multidimensional knowledge base, identify high-cohesion event subgroups in layers, determine the number of node kernels, and form multiple K-kernel subgraphs and boundary structures.

[0110] The hierarchical walk parameter setting module is used to set hierarchical adaptive random walk parameters for K-core graphs and boundary structures with different K values, and to design walk strategies within and between communities;

[0111] The hierarchical random walk module is used to perform multiple rounds of high-frequency long-path or low-frequency short-path random walks in each K-core subgraph region based on the random walk parameters, and collect the walk sequence of event nodes.

[0112] The feature aggregation and analysis module is used to aggregate features from the event node walking sequence, generate multi-dimensional event clue feature vectors, and realize event clue clustering, relation reasoning and knowledge completion to optimize the knowledge base structure.

[0113] The dynamic evolution module is used to dynamically trigger K-core decomposition and hierarchical random walk when new event clues are added, and to periodically or in real time optimize the knowledge base structure.

[0114] Example 1:

[0115] To verify the feasibility of this invention in practice, it was applied to an emergency command center. The invention's knowledge graph-based multi-dimensional knowledge base construction method for event clues was fully implemented to enhance intelligent response and data analysis capabilities for public safety emergencies. The project focused on the daily operation and emergency management of a landmark shopping mall, with five major data sources distributed on-site: fire alarms, security video surveillance, environmental sensors, social media sentiment interfaces, and patrol check-in records. During the "localized fire inside the building" emergency drill on March 10th, the system collected 10,215 pieces of raw data, including 1,423 alarm texts, 2,107 surveillance video summaries, 2,315 environmental sensor logs, 2,184 social media sentiment records, and 2,186 patrol check-in records. These data were originally distributed across different systems and in various formats, exhibiting duplication, missing data, and false alarms, making it difficult to support rapid and comprehensive event clue analysis.

[0116] Using the method of this invention, all raw data is first processed in a unified structure through a data acquisition and preprocessing module, and the attributes of nodes and edges are standardized. A knowledge graph with a total of 26,710 nodes and 73,850 edges is generated, covering personnel, locations, equipment, risk points, and related events. Using a K-core decomposition and hierarchical processing module, the system automatically divides the knowledge graph into 12 K-core subgraphs, significantly improving the ability to identify core areas and peripheral risk points. A hierarchical adaptive random walk parameter setting module sets the walk path and step probability for each area according to different K-core levels. In actual execution, the system collects an average long-path walk sequence length of 30 steps and a step probability of 0.88 in the core fire area, and an average short-path walk length of 10 steps and a step probability of 0.44 in the peripheral low-K areas, automatically collecting and analyzing 1,400 walk sequences every 10 minutes. Compared with the passive data aggregation and manual event sorting of traditional systems, this invention can automatically capture the propagation paths of multi-source high-frequency events and identify suspected abnormal nodes and potential risk links in real time.

[0117] In real-world scenarios, traditional methods rely on manual retrieval during fire drills, averaging 32 minutes. This often fails to promptly integrate social media sentiment and video anomalies, leading to delays in risk assessment due to the omission of some crucial nodes. After the system of this invention went live, in a drill test in March 2025, the average response time of the knowledge base was reduced to 14 minutes, the automatic clustering and knowledge completion rate of event clues increased to 23.1%, and the accuracy rate of critical path early warning reached 93%, significantly exceeding previous levels. During the actual drill, the command center detected congestion in the core area's evacuation routes 5 minutes in advance three times through the system's intelligent early warning, preventing the risk from spreading as with traditional methods. Post-event surveys showed that over 80% of on-site management personnel believed the new system significantly reduced the burden of manual analysis and markedly improved the anomaly detection rate. User satisfaction reached 92%.

[0118] Table 1 Comparison of Emergency Response Effects between the Method of the Invention and Traditional Solutions

[0119]

[0120]

[0121] The data in Table 1 above fully demonstrates the significant advantages of the method of this invention in actual emergency management scenarios. Regarding data fusion and knowledge structure construction, this invention comprehensively processes the original data, resulting in a substantial increase in the number of knowledge graph nodes and edges. The number of nodes increases from 17,893 in the traditional scheme to 26,710, and the number of edges increases from 47,350 to 73,850, representing increases of 49.3% and 55.9% respectively. This greatly enhances the ability to integrate information from multiple event clues.

[0122] In terms of event response efficiency, this invention, through structured layering and intelligent walkthrough mechanisms, reduces the average event response time from 32 minutes in traditional solutions to 14 minutes, improving response speed by 56.3% and enhancing the timeliness of event handling. The automatic clustering capability for event clues represents a qualitative leap from "manual-based" to "system-automated," making event analysis more intelligent and precise. The accuracy rate of critical path warnings has increased from 71% to 93%, a 22 percentage point improvement, indicating that this invention is more reliable in risk node identification and critical clue reasoning. The knowledge completion rate has increased to 23.1%, effectively compensating for the shortcomings of traditional solutions in identifying implicit relationships and missing information.

[0123] The system significantly increases the number of walk sequences that can be automatically sampled and analyzed, processing 1,400 walk sequences every 10 minutes, more than six times that of traditional methods, thus expanding the depth of structural exploration and diversity analysis. The automatic identification capability of suspected anomalous nodes has improved from "low" to "high," highlighting the value of intelligent algorithms in detecting and warning of anomalies. On-site user satisfaction has increased from 65% to 92%, with the vast majority of users acknowledging the efficiency, intelligence, and workload reduction benefits of this invention.

[0124] The method of this invention has core advantages in many aspects such as data fusion, knowledge extraction, emergency response, intelligent analysis and user experience in complex scenarios, and provides strong support for urban public safety and large-scale heterogeneous event clue management.

[0125] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a multi-dimensional knowledge base for event clues based on knowledge graphs, characterized in that, include: Collect multi-source event clue data, preprocess the multi-source event clue data, construct a knowledge graph, standardize the attribute information of nodes and edges in the knowledge graph, and form a multi-dimensional knowledge base; Based on the improved K-kernel decomposition, the knowledge graph in the multidimensional knowledge base is processed in layers to identify high cohesive event subgroups at each level, determine the number of kernels for each node, and form multiple K-kernel subgraphs and boundary structures. Based on multiple K-core subgraphs and boundary structures, hierarchical adaptive random walk parameters, including walk path length and step probability, are set for subgraphs with different K values. Random walk strategies within and between communities are designed to update the structural parameters of the multidimensional knowledge base. Based on the random walk parameters, high-frequency long-path random walks are implemented in the internal region of the community in the K-core graph, and low-frequency short-path random walks are implemented at the community boundary and in the low K-value region to obtain multiple event node walk sequences. Based on multiple event node walking sequences, the node, edge, kernel number and attribute information are aggregated to generate multi-dimensional event clue feature vectors. Event clue clustering, relation reasoning and knowledge completion are performed on the knowledge graph in the multi-dimensional knowledge base to optimize the structure of the multi-dimensional knowledge base. When new event clue data is integrated, K-core decomposition and hierarchical random walk are dynamically triggered, and the structure evolution and continuous optimization of the multidimensional knowledge base are executed periodically or in real time. Specifically, the multi-source event clue data includes event-related text, images, logs, and sensor data from structured data, semi-structured data, and unstructured data; The improved K-kernel decomposition method is used to perform hierarchical processing on the knowledge graph in the multidimensional knowledge base, identify high-cohesion event subgroups at each level, determine the kernel number of each node, and form multiple K-kernel subgraphs and boundary structures, including: Knowledge graphs are extracted from multidimensional knowledge bases and represented as undirected graphs consisting of sets of nodes and edges. Attribute information is extracted from each node and node attribute weights are assigned. Edge weights are assigned to each type of edge. The effective degree of each node is calculated based on the node attribute weights and the edge weights of the connected edges, forming a weighted node degree dataset of the knowledge graph. Based on the weighted node degree dataset, the knowledge graph is stripped in ascending order of degree. Nodes and related edges with a weighted effective degree lower than the current K value are recursively deleted. After completing one round of weighted stripping, the weights of the remaining nodes and edges are updated to obtain the subgraph structure and weighted node degree data of the current round. For the obtained subgraph structure and weighted node degree data, the node degree distribution and attribute distribution are statistically analyzed. Based on the changes in the structure and attributes within the subgraph, the K value used in this round of stripping operation is dynamically adjusted to achieve adaptive layering threshold setting and output a new K value and subgraph. Using the new K value and subgraph as input, a new round of weighted stripping is performed, recursively deleting nodes and edges in a loop. After each round of stripping, the core nodes and attribute features stripped from the previous layer are passed to the current subgraph, and the layering judgment conditions of the current node are updated in real time, forming a layered stripping process based on context feature interaction. After each round of stripping, the nodes that have a connection relationship between the current subgraph and the newly stripped nodes are marked as the boundary nodes of this layer. During the layered stripping process, based on the weighting coefficients and stripping priorities of different edge types, retention thresholds and deletion thresholds are set for each edge type. Edges with weighting coefficients greater than or equal to the retention threshold are retained, and edges with weighting coefficients less than the deletion threshold are deleted. This ensures that the stripping results simultaneously reflect node attributes, edge weights, and relationship diversity, generating a layered subgraph structure that integrates multiple relationship types. When processing large-scale knowledge graphs, the layering and stripping process is applied in parallel to multiple subgraphs. Each subgraph is synchronized through the boundary nodes of the intersection area to maintain the consistency of the overall structure. After each round of batch processing, the global subgraph structure and node core allocation results are merged and updated. When there are changes in the knowledge graph structure due to the addition or deletion of nodes or edges, for the changed area, based on the latest subgraph structure and weighted node degree data, incremental K-kernel decomposition is adopted, and only the affected subgraph and its neighborhood are locally stripped and the kernel number allocation is adjusted. The generated subgraphs, boundary nodes, and boundary structures are aggregated, and the final output is a hierarchical result containing multiple K-core subgraphs and boundary structures.

2. The method for constructing a multi-dimensional knowledge base for event clues based on knowledge graphs according to claim 1, characterized in that, The preprocessing of multi-source event clue data specifically includes data cleaning, format conversion, noise removal, and information standardization.

3. The method for constructing a multi-dimensional knowledge base for event clues based on knowledge graphs according to claim 1, characterized in that, The construction of the knowledge graph, which standardizes the attribute information of nodes and edges in the knowledge graph to form a multidimensional knowledge base, refers to extracting entities, relationships and attributes from preprocessed multi-source event clue data, establishing a graph structure with entities as nodes and relationships as edges, and performing unified format conversion, encoding standardization and numerical normalization on node attributes and edge attributes, and storing them in the multidimensional knowledge base.

4. The method for constructing a multi-dimensional knowledge base for event clues based on knowledge graphs according to claim 1, characterized in that, The process involves setting hierarchical adaptive random walk parameters, including walk path length and step probability, for subgraphs with different K values ​​based on multiple K-core subgraphs and boundary structures. It also involves designing random walk strategies within and between communities, and updating the structural parameters of the multidimensional knowledge base, including: The obtained multiple K-kernel subgraphs and their boundary structures are used as input. For each K-kernel subgraph, the node set, edge set, and weighted effective degree information of the nodes are obtained. For each K-core subgraph, the random walk path length parameter of the subgraph is set according to the size of the node set, the connectivity of the edge set, and the weighted effective degree of the nodes, and the step probability parameter is set. The path length parameter is used to limit the maximum number of steps in each walk, and the step probability parameter is used to control the probability of transition between nodes during the walk. Based on the hierarchical structure of the K-core graph, different parameters are set for random walks within communities and random walks between communities. For walks within communities, the path length parameter is set to be greater than or equal to a first threshold, and the step probability parameter is set to be greater than or equal to a second threshold. For walks between communities, the path length parameter is set to be less than the first threshold, and the step probability parameter is set to be less than the second threshold. By combining edge weights and node attributes, a transition probability matrix is ​​constructed, where the matrix elements are... Indicates the node in the k-th kernel graph To the node The transition probability is determined jointly based on the node attribute weight, edge weight, and step probability parameter. For different K-kernel subgraphs and boundary structures, the random walk path length parameter and step probability parameter of each subgraph are written into the multidimensional knowledge base.

5. The method for constructing a multi-dimensional knowledge base for event clues based on knowledge graphs according to claim 1, characterized in that, Based on the random walk parameters, high-frequency long-path random walks are performed within the community region of the K-core graph, while low-frequency short-path random walks are performed at the community boundary and in low-K value regions, to obtain multiple event node walk sequences, including: The system includes a strategy switching and start determination module, which comprises a walk start point selection unit and a walk strategy switching unit. The walk start point selection unit dynamically determines the walk start node within the community, the community boundary, and the low K value region based on the hierarchical structure, node attributes, and boundary structure of the K-core subgraph. During the walk process, the walk strategy switching unit dynamically determines the strategy of normal walk, jump walk, or directional walk based on the region to which the current node belongs, the node attribute weight, and the historical sampling state, and adjusts the path generation method accordingly. The system includes an information feedback and self-regulation module, which comprises a walk sequence statistics unit and a parameter self-regulation unit. The walk sequence statistics unit records in real time the nodes, edges, path length, number of steps, node coverage frequency, and sampling area distribution of each walk sequence. The parameter self-regulation unit dynamically adjusts the walk path length parameter, step probability parameter, starting point selection, and termination condition based on the statistical results. A local parallel and interactive fusion module is set up, which includes a subgraph parallel walking unit and a walking information interaction unit. The subgraph parallel walking unit divides multiple K-core subgraphs and boundary structures into several subgraphs, and performs random walks in parallel within each subgraph to collect the walking sequence of event nodes. The walking information interaction unit periodically summarizes the walking paths and node coverage of each subgraph and feeds them back to the parameter self-regulation unit to achieve cross-subgraph sampling coordination and global sampling balance. Under the synergistic effect of the strategy switching and initial determination module, the information feedback and self-regulation module, and the local parallel and interactive fusion module, a random walk is implemented, specifically as follows: The starting point selection unit dynamically selects the starting node for this round of random walks based on the hierarchical adaptive parameters, within the community, at the community boundary, or in a low K-value region. During each round of walking, the walking strategy switching unit determines in real time whether to use a normal walking, jump walking, or directional walking strategy for the current step based on the current node attributes, layer position, historical path, and sampling status, and dynamically adjusts the path generation method. Under the normal traversal strategy, based on the path length parameter and the step probability parameter, the transition probability is calculated among all the adjacent nodes of the current node by combining the node attribute weight and the edge weight, the next traversal node is randomly selected, and the path is recorded. Under the jump-walking or directional walking strategy, the walking strategy switching unit selects the node that is farther away or has specific attributes as the next walking node based on node attributes, structural features or walking history, and generates a long-distance or directional path. The event node traversal sequence of a long path is completed until the path length threshold, sampling number threshold, or termination condition set within the community area is reached. In the community boundary and low K value area, multiple rounds of short path event node traversal are performed based on the criteria that the path length parameter is less than the first threshold and the step probability parameter is less than the second threshold, forming a sampling sequence covering the community boundary and low K area. During the traversal, all collected path, node, edge, and parameter information is fed back to the information feedback and self-regulation module and the traversal information interaction unit in real time, dynamically optimizing the sampling distribution and realizing diversified structural sampling of different regions of the knowledge graph structure. After each walk, record all nodes, edges, path parameters, sampling areas and node coverage information involved in the walk path, and accumulate multiple event node walk sequences covering different areas of the knowledge graph; All event node walk sequences, related parameters, and statistical information collected through random walks are written into a multidimensional knowledge base to complete the collection and storage of walk sequences.

6. The method for constructing a multi-dimensional knowledge base for event clues based on knowledge graphs according to claim 1, characterized in that, The method involves aggregating node, edge, core, and attribute information based on multiple event node walk sequences to generate a multi-dimensional event clue feature vector. This vector is then used to perform event clue clustering, relation reasoning, and knowledge completion on the knowledge graph in the multi-dimensional knowledge base, thereby optimizing the multi-dimensional knowledge base structure. This includes: Using all event node walk sequences as input data, for each event node walk sequence, extract the set of nodes involved, the set of edges, node attributes, edge weights, path lengths and sampling area information to form a basic feature set; Feature aggregation is performed on the basic feature set of all walk sequences. The frequency of occurrence of the same node in different walk sequences, the cumulative weight of the connected edges, the statistical characteristics of node attributes, the path length distribution, and the number of node cores are comprehensively statistically analyzed to form a multi-dimensional feature matrix at the node level and the path level. Based on the multidimensional feature matrix, a clustering algorithm is used to cluster event nodes into event clues, grouping nodes with similar features into the same event clue cluster, and outputting the node set and clustering label of each cluster. By utilizing the node connection relationships and feature clustering results in the walking sequence, a graph reasoning method is used to mine the potential relationships between nodes and fill in the missing node attributes or edge relationships. The completion criteria are based on the similarity of node feature vectors and structural proximity. Based on clustering results, relational reasoning, and completion operations, the knowledge graph structure in the multidimensional knowledge base is updated in real time, including node attributes, edge relationships, number of node cores, and hierarchical structure.

7. A knowledge graph-based event clue multidimensional knowledge base construction system, comprising the knowledge graph-based event clue multidimensional knowledge base construction method according to any one of claims 1 to 6, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect multi-source event clue data and preprocess it to build and standardize the knowledge graph, forming a multi-dimensional knowledge base. The K-kernel decomposition and hierarchical module is used to perform K-kernel decomposition on the knowledge graph in the multidimensional knowledge base, identify high-cohesion event subgroups in layers, determine the number of node kernels, and form multiple K-kernel subgraphs and boundary structures. The hierarchical walk parameter setting module is used to set hierarchical adaptive random walk parameters for K-core graphs and boundary structures with different K values, and to design walk strategies within and between communities; The hierarchical random walk module is used to perform multiple rounds of high-frequency long-path or low-frequency short-path random walks in each K-core subgraph region based on the random walk parameters, and collect the walk sequence of event nodes. The feature aggregation and analysis module is used to aggregate features from the event node walking sequence, generate multi-dimensional event clue feature vectors, and realize event clue clustering, relation reasoning and knowledge completion to optimize the knowledge base structure. The dynamic evolution module is used to dynamically trigger K-core decomposition and hierarchical random walk when new event clues are added, and to periodically or in real time optimize the knowledge base structure.

Images