Complex event relation causal tracing method based on subgraph evaluation
By combining the improved Viterbi algorithm with pruning strategies, a multidimensional evaluation system was constructed, which solved the problems of low efficiency and unreliable results in causal attribution in complex event relationship networks, and achieved efficient and interpretable causal attribution results.
Patent Information
- Application Number
- CN202511578627.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-31
- Publication Date
- 2026-02-10
AI Technical Summary
Existing methods for tracing the origins of events are inefficient in complex networks and struggle to provide reliable explanations of causal paths when faced with noisy relationships and idealized model assumptions.
A subgraph-based evaluation method is adopted. By improving the Viterbi algorithm and combining it with pruning strategies, multi-path backtracking search is performed in the event relationship network to construct a multi-dimensional evaluation system, including node-level, path-level and graph structure-level evaluation, and extract the subgraphs corresponding to candidate source nodes.
It improves the efficiency of causal attribution, achieves accurate screening of optimal source nodes and causal paths, provides interpretable attribution results, and enhances the accuracy and reliability of attribution results in noisy environments.
Smart Images
Figure CN121503598A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of causal attribution technology, specifically relating to a causal attribution method for complex event relationships based on subgraph evaluation. Background Technology
[0002] With the development of the internet and information technology, events in society, economy, science and technology are increasingly recorded and disseminated in digital form, forming massive event data. Constructing event relationship networks from this data and extracting causal evolutionary paths for subsequent understanding and analysis is of great significance in data mining and analysis, and also has considerable value in tasks such as public opinion analysis and risk warning. However, in reality, event relationship networks are usually large in scale and complex in structure. A single outcome event is often caused by multiple causal events acting together through complex and diverse paths in terms of length, structure, and distribution. In such challenging event relationship networks, accurately and efficiently locating the root cause of a given outcome event—that is, causal tracing—becomes a highly challenging yet valuable problem.
[0003] Existing methods for causal attribution problems in complex networks can be mainly divided into two categories: one is the structured method based on graph theory and rule reasoning, which treats the event relationship network as a deterministic graph structure, and the attribution problem is defined as a search and matching task based on graph topology; the other is the statistical association method based on probabilistic models, which treats causal relationships as a kind of statistical association, discovers correlation patterns by constructing probabilistic models or using data-driven learning, and infers source node events accordingly.
[0004] Structured methods based on graph theory and rule-based reasoning rely on predefined graph traversal algorithms (such as BFS, DFS, and shortest path algorithms) to explore paths in the network, supplemented by simple static metrics such as weighted sums to filter and rank candidate paths. It can be seen that these methods are extremely sensitive to data noise (such as incorrectly extracted causal edges) and lack mechanisms for quantifying and tolerating uncertainty. Furthermore, their evaluation system is limited to a single-dimensional metric, resulting in poor reliability. In addition, they face the problem of combinatorial explosion in complex networks; exhaustive search is inefficient, while heuristic rules may miss critical paths, making it difficult to achieve a good balance between efficiency and accuracy. Therefore, such methods have low practicality when dealing with complex event-relationship networks.
[0005] Statistical association methods based on probabilistic models treat causal relationships as statistical associations. They discover correlation patterns by constructing probabilistic models or utilizing data-driven learning, and then infer source events from these patterns. This involves building abstract mathematical models (such as independent cascade models or various graph neural networks, GNNs) for event propagation, and learning and inference by maximizing likelihood probability or minimizing prediction error. This approach focuses on mining statistical association features between nodes from data, rather than explicit causal logic. The performance of probabilistic model-based statistical association methods heavily depends on the degree to which the selected model matches the complex causal mechanisms of the real world. Faced with complex and ever-changing realities, the model assumptions made are often quite idealistic. Furthermore, these methods typically have poor interpretability; their "black box" operation makes it difficult to provide clear and credible explanations of causal paths, and verifying the reliability of the results during application is challenging.
[0006] In summary, existing methods for tracing the origins of events suffer from significant shortcomings when faced with an explosive volume of news data. Noisy relationships and invalid events within the data greatly impact the tracing performance of structured methods based on graph theory and rule-based reasoning. Furthermore, these methods suffer from combinatorial explosion problems in complex networks, resulting in low efficiency. Statistical association methods based on probabilistic models heavily rely on the degree to which the selected model matches the complex causal mechanisms of the real world; however, the model assumptions are often overly idealistic, and the tracing results lack interpretability. Summary of the Invention
[0007] To address the aforementioned technical problems, this invention provides a causal tracing method for complex event relationships based on subgraph evaluation. By improving the Viterbi algorithm and combining it with a pruning strategy, multi-path backtracking search is achieved in the event relationship network, thereby extracting the subgraphs corresponding to candidate source nodes. The optimal tracing result is obtained by employing a multi-dimensional evaluation system that includes node-level, path-level, and graph structure-level evaluation.
[0008] The technical solution adopted in this invention is as follows:
[0009] A causal tracing method for complex event relationships based on subgraph evaluation, the specific steps of which are as follows:
[0010] S1. Perform data preprocessing on the event relationship data, and construct a directed graph using events as nodes and event relationships as directed edges;
[0011] S2. Based on the target tail event to be traced, perform a multi-path hierarchical backtracking search in the directed graph according to the causal rationality and temporal continuity to obtain a set of candidate paths, including a set of candidate source events and the optimal path to the target tail event corresponding to each source event.
[0012] S3. Extract subgraphs for each candidate path, construct connected subgraphs, and calculate the comprehensive score of multi-dimensional indicators in each connected subgraph; among which, the subgraph evaluation indicators include: path quality and event relationship network diagram structure;
[0013] S4. Sort the candidate source events according to the comprehensive score of the corresponding subgraph, and select the event with the highest comprehensive score as the optimal source event; and use the node identification method based on the degree centrality calculation to identify the key nodes in the subgraph by calculating the frequency of occurrence of nodes in the shortest path and the in-degree of nodes; output the optimal source event and the corresponding event development subgraph and key nodes in the causal evolution path.
[0014] Furthermore, S1 is specifically as follows:
[0015] S11. Standardize the format of the time field in the input data for event relationship data, remove invalid, negative confidence, or noisy relationships with missing head and tail nodes, and sort the events according to the time field to establish a time index, transforming the event relationship data into a processable and computable directed graph structure; wherein, the event relationship data includes: event description, time attribute, event relationship type, and relationship confidence;
[0016] S12. Construct a weighted directed graph with comprehensive information, using events as nodes and event relationships as directed edges. The node set represents a specific set of events, and the edge set represents the directed causal or sequential relationship between events. Each edge carries multiple attributes, including: relationship type and relationship confidence.
[0017] S13. A depth-based search method is used to identify the cyclic structure in the weighted directed graph, and selective pruning is performed based on the comprehensive confidence of the edges. The comprehensive confidence of the edges is affected by three factors: relation type, relation confidence score, and time consistency. By calculating the confidence gradient of each edge, the edges with the lowest gradient are deleted first. After pruning, connectivity is verified, and isolated nodes are filtered out by checking the in-degree and out-degree of nodes through graph traversal, finally obtaining a directed acyclic graph.
[0018] Furthermore, S2 is specifically as follows:
[0019] S21, establish a hierarchical structure starting from the target tail event, where each node represents a possible upstream cause of the event in the previous level; this is set in a directed graph. middle, Represents a set of nodes. Represents an edge set, nodes Represents an event, edge Indicates an event and events There is a causal or sequential relationship between the events. As the subject of the relationship between events, the event As the object, the right of the side The overall confidence score represents the relationship edge; a state value is defined for each node. , representing the optimal path score for reaching the target tail event;
[0020] S22: Starting with the target tail event, expand the directed graph layer by layer in reverse order until all reachable nodes are traversed. Obtain the optimal path score from each node to the target tail event based on the state transition recursive relation expression, and record the specific path. At the same time, prune the graph according to the optimal path score.
[0021] The recursive expression for the state transition is as follows:
[0022]
[0023] In the formula For nodes All predecessor nodes, The path length reward function, The overall confidence score for the relationship edge between the next level and the current level;
[0024] The pruning strategy expression is as follows:
[0025] like Then discard the node. Path to the target tail event
[0026] In the formula, This is the set of optimal path scores for all nodes at the same level. This is a threshold for the score difference; it also sets a limit to retain only the top K highest-scoring paths in each search layer, forming a Top-K set. K is a set value;
[0027] S23, define the total path score as a weighted combination of multidimensional features; calculate the scores of each path after pruning. Total score The expression is as follows:
[0028]
[0029] In the formula Score the structural features. Score semantic features For time-series feature scoring, Here are the weight coefficients for the three types of features, and n is the path. The number of nodes, structural feature score, semantic feature score, and temporal feature score are determined based on the path. The optimal path score for each node is obtained;
[0030] The final set of alternative paths is obtained, which includes a set of alternative source events and the optimal path to the target tail event for each source event.
[0031] Furthermore, S3 is specifically as follows:
[0032] S31. Set the target tail event as t, and the alternative source event as... ;by Starting from t, a set of k reachable paths to t is extracted using a directed random walk method. And form a connected subgraph from the nodes and edges involved in the path. The candidate source nodes and the target tail events are compared in a subgraph; and during the extraction process, only those with a confidence level higher than the threshold are retained. The path;
[0033] S32. Establish a complete causal subgraph quality assessment system to evaluate the reliability of tracing results from two aspects: path quality and event relationship network graph structure.
[0034] In terms of path quality, the rationality of semantic and temporal consistency of each causal evolution path in the subgraph is considered. Quantitative indicators include the average path confidence of the subgraph, the proportion of extremely high and extremely low-scoring paths, and the proportion of edges with correct temporal direction in the links. The specific definitions are as follows:
[0035]
[0036] In the formula, The average path confidence. Let P represent the subgraph where the source event is event s and the tail event is event t. |P| and |E| represent the number of paths and relation edges within the subgraph, respectively. This indicates the proportion of the side in the diagram that is aligned with the correct time direction. This represents the directed edge connecting event nodes i and j. Indicates the time when event i occurred. The proportion of extremely high-scoring paths, This indicates the high confidence threshold. The percentage of paths with extremely low scores Score(P) represents the low confidence threshold. i () represents the total score of a path in the subgraph;
[0037] The structural index of the event relationship network diagram is expressed through graph density. Clustering coefficients measure structural compactness, and the specific quantification method is expressed as follows:
[0038]
[0039] In the formula, To calculate the average clustering coefficient, represents the local clustering coefficient of the node. This indicates the number of nodes within the subgraph. It is the number of actual edges between the neighbors of node i. It is the degree of node i;
[0040] Taking into account both path quality and event relationship network diagram structure metrics, a subgraph scoring function is defined. The expression is as follows:
[0041]
[0042] In the formula, F(G) s,t () is a comprehensive score based on multiple indicators. This represents the mean link confidence, derived from the average path confidence. The ratio of the side with the correct time direction We get the weighted sum; The structural complexity index is represented by graph density D and graph average clustering coefficient. The weighted average of the graph structure indicators is obtained; Indicates the weighting parameter;
[0043] The weight parameter calculation expression is as follows:
[0044]
[0045] In the formula, This indicates the original weight parameter to be replaced. This indicates the value to be changed for the corresponding weight parameter, where δ is the set adjustment coefficient, and k=1,2,3,4.
[0046] The beneficial effects of this invention are:
[0047] The method of this invention first preprocesses the input event relationship data, constructs a directed graph using the purified data, and then performs a multi-path hierarchical backtracking search in the directed graph based on the received target tail events that need to be traced, obtaining a set of candidate source events and their corresponding optimal paths. Next, it extracts causal relationship subgraphs corresponding to different source events, evaluates these subgraphs, and analyzes the corresponding event development context subgraphs and key nodes in the causal evolution path, thus completing causal tracing in complex event relationship networks. This method improves the Viterbi algorithm and combines it with pruning strategies to achieve multi-path backtracking search in event relationship networks, extracts subgraphs corresponding to candidate source nodes, and uses a multi-dimensional evaluation system including node-level, path-level, and graph structure-level evaluation to obtain the optimal tracing result. This improves the efficiency of causal tracing in complex event relationship networks and enables the comparison and selection of optimal source nodes and their corresponding optimal development paths. Attached Figure Description
[0048] Figure 1 This is a flowchart of a causal tracing method for complex event relationships based on subgraph evaluation according to the present invention.
[0049] Figure 2 This is a schematic diagram of the backtracking search process in an embodiment of the present invention. Detailed Implementation
[0050] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0051] like Figure 1 The flowchart shown is a method for causal tracing of complex event relationships based on subgraph evaluation according to the present invention. The specific steps are as follows:
[0052] S1. Preprocess the event relationship data, and construct a directed graph using events as nodes and event relationships as directed edges; the specific process is as follows:
[0053] S11. Transform the event relationship data into a directed graph structure that can be processed and computed;
[0054] After receiving event relationship data, the system formats the input data's time field, removes invalid relationships, those with negative confidence, or those missing head or tail nodes, and sorts the events according to the time field to create a time index, transforming the event relationship data into a processable and computationally comprehensible directed graph structure. Through this processing flow, all input event relationship information is transformed into valid data that meets the structural and quality requirements needed for subsequent graph construction.
[0055] The event relationship data includes: event description, time attribute, event relationship type, and relationship confidence level.
[0056] S12. Construct a weighted directed graph with comprehensive information, using events as nodes and event relationships as directed edges. The node set represents a specific set of events, and the edge set represents the directed causal or sequential relationships between events. Each edge carries multiple attributes, including relationship type and relationship confidence. This design improves the efficiency of graph traversal and provides abundant and easily accessible attribute information for subsequent event relationship data analysis.
[0057] S13. Preprocess the weighted directed graph constructed in step S12 to obtain a directed acyclic graph;
[0058] In the process of constructing an event graph, the existence of loop structures violates the temporal consistency of causality or sequence, indicating the presence of unreasonable noisy relationships. Therefore, this embodiment employs a depth-first search method to identify loop structures in the weighted directed graph and performs selective pruning based on the overall confidence of the edges. The overall confidence of an edge is influenced by three factors: relationship type, relationship confidence score, and temporal consistency. By calculating the confidence gradient of each edge, edges with the lowest gradients are prioritized for deletion, breaking loop structures with minimal cost. After pruning, the system performs connectivity verification to ensure that the overall structure of the graph is not excessively damaged. This process is designed to maintain the connectivity of the constructed event relationship directed graph while filtering out potentially unreasonable loop structures. Furthermore, isolated nodes in the graph are useless data for tracing the development of events. Therefore, this embodiment filters out isolated nodes by checking the in-degree and out-degree of nodes through graph traversal, ultimately obtaining a directed acyclic graph.
[0059] After this series of preprocessing steps, the resulting directed acyclic graph not only has good event propagation properties, but also ensures the rationality of event relationships in the time dimension, providing a stable and reliable structural foundation for subsequent source tracing calculations.
[0060] By incorporating indicators such as relationship confidence and temporal consistency, the method of this invention can effectively screen high-quality paths in the event development network. Therefore, the method of this invention has a certain degree of anti-interference capability against noisy data that appears during the event extraction process, and to a certain extent overcomes the problem of overall path deviation caused by individual noise relationships in existing methods, thereby improving the accuracy of source tracing in noisy environments.
[0061] S2. Based on the target tail event to be traced, perform a multi-path hierarchical backtracking search in the directed graph according to the causal rationality and temporal continuity to obtain a set of candidate paths, including a set of candidate source events and the optimal path to the target tail event corresponding to each source event.
[0062] Upon receiving the target tail event to be traced, a reverse source tracing search is performed in the graph data. The Viterbi algorithm is employed, recursively calculating the maximum probability path to a given state at each time step using dynamic programming. Backtracking is then used to obtain the globally optimal sequence. A pruning strategy for graph search is introduced to ultimately identify a high-quality set of candidate source events and the optimal causal evolution path corresponding to each source event, i.e., the candidate path set. The main innovation of this process lies in modifying the sequential dynamic programming algorithm of the Viterbi algorithm to adapt it for use in graph structure environments.
[0063] In the method of this invention, the system defines a state vector for each event node, recording the optimal path score and corresponding path information to reach the target event. Through a redesigned recursive formula, the system can comprehensively consider multi-dimensional event relationship information such as relationship confidence, semantic relevance, temporal consistency, and relationship type weight, to find high-quality source event sets and corresponding causal evolution paths in complex event relationship networks. The specific implementation process is as follows: Figure 2 As shown.
[0064] In this embodiment, S2 is specifically as follows:
[0065] S21, establish a hierarchical structure starting from the target tail event, where each node represents a possible upstream cause of the event in the previous level; this is set in a directed graph. middle, Represents a set of nodes. Represents an edge set, nodes Represents an event, edge Indicates an event and events There is a causal or sequential relationship between the events. As the subject of the relationship between events, the event As the object, the right of the side The overall confidence score represents the relationship edge; a state value is defined for each node. , representing the optimal path score for reaching the target tail event.
[0066] S22: Starting with the target tail event, expand the directed graph layer by layer in reverse order until all reachable nodes are traversed. Obtain the optimal path score from each node to the target tail event based on the state transition recursive relation expression, and record the specific path. At the same time, prune the graph according to the optimal path score.
[0067] The recursive expression for the state transition is as follows:
[0068]
[0069] In the formula For nodes All predecessor nodes, The path length reward function, The overall confidence score for the relationship edge between the next level and the current level;
[0070] The pruning strategy expression is as follows:
[0071] like Then discard the node. Path to the target tail event
[0072] In the formula, This is the set of optimal path scores for all nodes at the same level. This is a threshold for the score difference; it also sets a limit to retain only the top K highest-scoring paths in each search layer, forming a Top-K set. K is a set value; thus, the amount of calculation is further reduced while maintaining accuracy.
[0073] S23, define the total path score as a weighted combination of multidimensional features; calculate the scores of each path after pruning. Total score The expression is as follows:
[0074]
[0075] In the formula For structural feature scoring, path length is the primary consideration; The semantic feature score is calculated by combining the relation type weight and the event relation confidence score. The scoring of temporal features primarily considers whether events within a causal or sequential relationship exhibit temporal disorder. Here are the weight coefficients for the three types of features, and n is the path. The number of nodes, structural feature score, semantic feature score, and temporal feature score are determined based on the path. The optimal path score for each node is obtained;
[0076] The final set of alternative paths is obtained, which includes a set of alternative source events and the optimal path to the target tail event for each source event.
[0077] The core causal tracing process of this invention is based on the Viterbi algorithm. By constructing a reverse multi-path optimization recursive mechanism, the state space is reconstructed from a temporal sequence into a graph topology, and the recursive mechanism is transformed from a single-predecessor optimization process to a multi-predecessor optimization process. Furthermore, a graph-search-oriented pruning strategy is introduced. This successfully yields the causal tracing method, which can efficiently discover the optimal development path between events and identify the most reliable source event leading to the tail event, providing a new solution for causal tracing in complex event relationship networks.
[0078] S3. Extract subgraphs for each candidate path, construct connected subgraphs, and calculate the comprehensive score of multi-dimensional indicators in each connected subgraph; among which, the subgraph evaluation indicators include: path quality and event relationship network diagram structure.
[0079] The specific process for S3 is as follows:
[0080] S31. Set the target tail event as t, and the alternative source event as... ;by Starting from t, a set of k reachable paths to t is extracted using a directed random walk method. And form a connected subgraph from the nodes and edges involved in the path. The candidate source nodes and the target tail events are compared in a subgraph; and during the extraction process, only those with a confidence level higher than the threshold are retained. The path is designed to avoid the damage caused by noisy data to the subgraph evaluation.
[0081] S32. Establish a complete causal subgraph quality assessment system to evaluate the reliability of tracing results from two aspects: path quality and event relationship network graph structure.
[0082] In terms of path quality, the rationality of semantic and temporal consistency of each causal evolution path in the subgraph is considered. Quantitative indicators include the average path confidence of the subgraph, the proportion of extremely high and extremely low-scoring paths, and the proportion of edges with correct temporal direction in the links. The specific definitions are as follows:
[0083]
[0084] In the formula, The average path confidence. Let P represent the subgraph where the source event is event s and the tail event is event t. |P| and |E| represent the number of paths and relation edges within the subgraph, respectively. This indicates the proportion of the side in the diagram that is aligned with the correct time direction. This represents the directed edge connecting event nodes i and j. Indicates the time when event i occurred. The proportion of extremely high-scoring paths, This indicates the high confidence threshold. The percentage of paths with extremely low scores Score(P) represents the low confidence threshold. i () represents the total score of a path in the subgraph;
[0085] The structural index of the event relationship network diagram is expressed through graph density. Clustering coefficients measure structural compactness, and the specific quantification method is expressed as follows:
[0086]
[0087] In the formula, To calculate the average clustering coefficient, represents the local clustering coefficient of the node. This indicates the number of nodes within the subgraph. It is the number of actual edges between the neighbors of node i. It is the degree of node i;
[0088] Taking into account both path quality and event relationship network diagram structure metrics, a subgraph scoring function is defined. The expression is as follows:
[0089]
[0090] In the formula, F(G) s,t () is a comprehensive score based on multiple indicators. This represents the mean link confidence, derived from the average path confidence. The ratio of the side with the correct time direction We get the weighted sum; The structural complexity index is represented by graph density D and graph average clustering coefficient. The weighted average of the graph structure indicators is obtained; Indicates the weighting parameter;
[0091] To account for the different sizes of the subgraphs generated under the random walk strategy, a special weighting mechanism is introduced, and the weight parameters are calculated as follows:
[0092]
[0093] In the formula, This indicates the original weight parameter to be replaced. This indicates the value to be changed for the corresponding weight parameter, where δ is the adjustment coefficient, and k = 1, 2, 3, 4. The purpose of this design is to reward subgraphs with greater information content based on their size, thus avoiding unfairness between subgraphs of different sizes during subgraph evaluation.
[0094] Based on these multi-dimensional evaluation methods, the system uses a flexible weighting mechanism to conduct a comprehensive quality score, which can provide a more reasonable quantitative basis for comparing the credibility of traceability results.
[0095] S4. Based on the subgraph evaluation results obtained in step S3, determine the optimal source event and the corresponding optimal causal evolution path, and analyze the key nodes in the corresponding event development subgraph and causal evolution path to complete the causal tracing in the complex event relationship network.
[0096] After evaluating the subgraph, the system ranks the candidate source events based on their comprehensive scores and selects the event with the highest score as the optimal source event. To identify key influencing nodes in the event's development, the system uses a node identification method based on median centrality. By calculating the frequency of a node's occurrence in the shortest path and its in-degree and out-degree, the system identifies key nodes in the subgraph. It then outputs the optimal source event, along with the corresponding event development subgraph and key nodes in the causal evolution path. These nodes often represent turning points or intensifying points in the event's development, greatly aiding in analyzing the overall evolution of the event.
[0097] The system's final output includes the optimal source event and its corresponding high-confidence subgraph and key node sequence. The output clearly demonstrates the causal development of the event, and also includes some of the impact on possible side events beyond the main event. Therefore, this system has high practical value in analyzing the development of events. Furthermore, the method of this invention provides interpretable discriminant records containing quantitative data for verifying the causal relationships of the tracing results.
[0098] In summary, the method of this invention is based on the recursive search framework of the dynamic programming method of the improved Viterbi algorithm, combined with a scientific pruning strategy, to avoid the combinatorial explosion problem in exhaustive search in large-scale event relationship networks. It concentrates computational resources on causal paths with high reliability, thus improving search efficiency. To evaluate the quality of the event development network graph, this invention introduces indicators such as average path confidence, proportion of temporally consistent relationships, graph density, and average clustering coefficient, constructing a multi-dimensional causal subgraph quality evaluation system. This system quantitatively evaluates the tracing results from multiple perspectives, including path quality, graph structure quality, and temporal rationality. Addressing the contingency of single-path evaluation, it provides credibility criteria for the results from multiple dimensions, and the output results include a comprehensive quality score, enhancing the practical value and credibility of the tracing results.
[0099] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.
Claims
1. A causal attribution method for complex event relationships based on subgraph evaluation, characterized in that, The specific steps are as follows: S1. Perform data preprocessing on the event relationship data, and construct a directed graph using events as nodes and event relationships as directed edges; S2. Based on the target tail event to be traced, perform a multi-path hierarchical backtracking search in the directed graph according to the causal rationality and temporal continuity to obtain a set of candidate paths, including a set of candidate source events and the optimal path to the target tail event corresponding to each source event. S3. Extract subgraphs for each candidate path, construct connected subgraphs, and calculate the comprehensive score of multi-dimensional indicators in each connected subgraph; among which, the subgraph evaluation indicators include: path quality and event relationship network diagram structure; S4. Sort the candidate source events according to the comprehensive score of the corresponding subgraph, and select the event with the highest comprehensive score as the optimal source event; and use the node identification method based on the degree centrality calculation to identify the key nodes in the subgraph by calculating the frequency of occurrence of nodes in the shortest path and the in-degree of nodes; output the optimal source event and the corresponding event development subgraph and key nodes in the causal evolution path.
2. The causal tracing method for complex event relationships based on subgraph evaluation according to claim 1, characterized in that, S1 is specifically as follows: S11. Standardize the format of the time field in the input data for event relationship data, remove invalid, negative confidence, or noisy relationships with missing head and tail nodes, and sort the events according to the time field to establish a time index, transforming the event relationship data into a processable and computable directed graph structure; wherein, the event relationship data includes: event description, time attribute, event relationship type, and relationship confidence; S12. Construct a weighted directed graph with comprehensive information, using events as nodes and event relationships as directed edges. The node set represents a specific set of events, and the edge set represents the directed causal or sequential relationship between events. Each edge carries multiple attributes, including: relationship type and relationship confidence. S13. A depth-based search method is used to identify the cyclic structure in the weighted directed graph, and selective pruning is performed based on the comprehensive confidence of the edges. The comprehensive confidence of the edges is affected by three factors: relation type, relation confidence score, and time consistency. By calculating the confidence gradient of each edge, the edges with the lowest gradient are deleted first. After pruning, connectivity is verified, and isolated nodes are filtered out by checking the in-degree and out-degree of nodes through graph traversal, finally obtaining a directed acyclic graph.
3. The method for causal tracing of complex event relationships based on subgraph evaluation according to claim 1, characterized in that, S2 is specifically as follows: S21, establish a hierarchical structure starting from the target tail event, where each node represents a possible upstream cause of the event in the previous level; this is set in a directed graph. middle, Represents a set of nodes. Represents an edge set, nodes Represents an event, edge Indicates an event and events There is a causal or sequential relationship between the events. As the subject of the relationship between events, the event As the object, the right of the side The overall confidence score representing the relationship edge; Define state values for each node , representing the optimal path score for reaching the target tail event; S22: Starting with the target tail event, expand the directed graph layer by layer in reverse order until all reachable nodes are traversed. Obtain the optimal path score from each node to the target tail event based on the state transition recursive relation expression, and record the specific path. At the same time, prune the graph according to the optimal path score. The recursive expression for the state transition is as follows: In the formula For nodes All predecessor nodes, The path length reward function, The overall confidence score for the relationship edge between the next level and the current level; The pruning strategy expression is as follows: like Then discard the node. Path to the target tail event In the formula, This is the set of optimal path scores for all nodes at the same level. This is a threshold for the score difference; it also sets a limit to retain only the top K highest-scoring paths in each search layer, forming a Top-K set. K is a set value; S23, define the total path score as a weighted combination of multidimensional features; calculate the scores of each path after pruning. Total score The expression is as follows: In the formula Score the structural features. Score semantic features For time-series feature scoring, Here are the weight coefficients for the three types of features, and n is the path. The number of nodes, structural feature score, semantic feature score, and temporal feature score are determined based on the path. The optimal path score for each node is obtained; The final set of alternative paths is obtained, which includes a set of alternative source events and the optimal path to the target tail event for each source event.
4. The method for causal tracing of complex event relationships based on subgraph evaluation according to claim 1, characterized in that, S3 is specifically as follows: S31. Set the target tail event as t, and the alternative source event as... ;by Starting from t, a set of k reachable paths to t is extracted using a directed random walk method. And form a connected subgraph from the nodes and edges involved in the path. The candidate source nodes and the target tail events are compared in a subgraph; and during the extraction process, only those with a confidence level higher than the threshold are retained. The path; S32. Establish a complete causal subgraph quality assessment system to evaluate the reliability of tracing results from two aspects: path quality and event relationship network graph structure. In terms of path quality, the rationality of semantic and temporal consistency of each causal evolution path in the subgraph is considered. Quantitative indicators include the average path confidence of the subgraph, the proportion of extremely high and extremely low-scoring paths, and the proportion of edges with correct temporal direction in the links. The specific definitions are as follows: In the formula, The average path confidence. Let P represent the subgraph where the source event is event s and the tail event is event t. |P| and |E| represent the number of paths and relation edges within the subgraph, respectively. This indicates the proportion of the side in the diagram that is aligned with the correct time direction. This represents the directed edge connecting event nodes i and j. Indicates the time when event i occurred. The proportion of extremely high-scoring paths, This indicates the high confidence threshold. The percentage of paths with extremely low scores Score(P) represents the low confidence threshold. i () represents the total score of a path in the subgraph; The structural index of the event relationship network diagram is expressed through graph density. Clustering coefficients measure structural compactness, and the specific quantification method is expressed as follows: In the formula, To calculate the average clustering coefficient, represents the local clustering coefficient of the node. This indicates the number of nodes within the subgraph. It is the number of actual edges between the neighbors of node i. It is the degree of node i; Taking into account both path quality and event relationship network diagram structure metrics, a subgraph scoring function is defined. The expression is as follows: In the formula, F(G) s,t This is a comprehensive score based on multiple indicators. This represents the mean link confidence, derived from the average path confidence. The ratio of the side with the correct time direction We get the weighted sum; The structural complexity index is represented by graph density D and graph average clustering coefficient. The weighted average of the graph structure indicators is obtained; Indicates the weighting parameter; The weight parameter calculation expression is as follows: In the formula, This indicates the original weight parameter to be replaced. This indicates the value to be changed for the corresponding weight parameter, where δ is the set adjustment coefficient, and k=1,2,3,4.