Multi-source data-oriented intelligent dynamic character event atlas construction method

Abstract

The invention discloses a multi-source data-oriented intelligent dynamic character event atlas construction method, and aims to solve the core problems of causality deficiency, structure static stiffness and uninterpretable prediction in the traditional technology. The method comprises the following steps: firstly, unifying heterogeneous data into a'character-behavior-time 'three-dimensional event tensor through multi-source analysis; further, semantic similarity and causal strength calculation are creatively fused, and a dynamic map with an explicit causal relationship is constructed; driving adaptive growth and updating of the atlas through a structural uncertainty index; an individual behavior latent variable is learned by adopting a variational auto-encoder, so that personalized behavior prediction is realized; and finally, generating a prediction result with natural language interpretation in combination with causal path consistency constraint and a large language model. According to the method, the crossing from static association to dynamic causal reasoning is realized, and the accuracy, the self-adaptability and the interpretability of character behavior modeling under the scenes of public security criminal investigation, financial risk control and the like are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to a method and system for constructing a knowledge graph. In particular, this invention relates to a method for constructing an intelligent person event graph that can integrate multi-source heterogeneous data, achieve dynamic modeling and growth of event chains, and support causal reasoning and interpretable behavior prediction. Background Technology

[0002] In today's deeply information-driven society, data from the internet, government systems, and financial transactions is exploding. The behavioral and relational information contained within this data is of paramount value to national security, financial risk control, and social governance. Achieving accurate modeling, dynamic correlation analysis, and behavioral trend prediction of key individuals has become a core requirement for enhancing the intelligence of decision-making in various institutions. For example, in criminal investigations, it is necessary to reconstruct the action chain and command relationships of criminal organizations from fragmented information; in financial regulation, it is hoped that the past behavior of corporate executives can predict their risk of violations; and in public safety, it is hoped that the evolution of the social relationships of specific individuals can be tracked to provide early warnings of their potential movements. However, the realization of these high-value applications first encounters a fundamental challenge at the data level: the sources of data related to individuals are extremely broad and highly heterogeneous, encompassing structured database records, unstructured text reports, and semi-structured log documents. Traditional data processing methods are usually designed for single, well-organized data sources, lacking a unified framework for multi-source information fusion and representation, resulting in severe information silos and making it difficult to form a comprehensive and consistent understanding of the target individuals.

[0003] To address these challenges, knowledge graph technology, due to its powerful entity relation representation capabilities, has been widely applied in the field of interpersonal relationship analysis in recent years. Traditional knowledge graphs, centered on the "entity-relationship-entity" triple, combined with graph neural networks and embedding techniques, have made significant progress in static relation reasoning and link prediction tasks. However, when this paradigm is directly applied to dynamic, evolving interpersonal event analysis, its inherent limitations become apparent. First, its modeling granularity is too coarse; the triple structure struggles to accommodate multi-dimensional attributes such as time, location, and motivation in complex events, and cannot directly describe behavioral chains or workflows composed of a series of sub-events. Second, reasoning capabilities are limited; existing methods mostly rely on entity co-occurrence or temporal proximity for association, lacking the ability to explicitly model and deduce causal logic between events, preventing the system from answering the crucial question of "why it happened." Furthermore, existing knowledge graph structures are inherently static; once constructed, they cannot automatically evolve based on newly incoming data, failing to achieve intelligent growth and iterative updates of nodes, relationships, and event paths, thus limiting their practicality and timeliness in real-world dynamic environments.

[0004] In summary, the current state of technology, in its evolution from "static character relationship graphs" to "dynamic character event graphs," suffers from a series of systemic shortcomings. The core issues can be summarized as follows: at the data level, it cannot efficiently integrate multi-source heterogeneous information; at the representation level, it lacks fine-grained modeling of temporal event chains; at the reasoning level, it is weak in causal explanation and path deduction; and at the system level, it lacks the ability for continuous learning and adaptive growth. These deficiencies collectively lead to the poor interpretability of existing system behavior predictions and their inability to adapt to complex dynamic scenarios. Therefore, there is an urgent need in this field for a novel technical solution that can connect the entire chain from multi-source data fusion to dynamic event chain construction, and then to causal reasoning and explanation generation, thereby providing core technical support for achieving truly intelligent dynamic character behavior analysis. Summary of the Invention

[0005] This invention aims to provide an intelligent dynamic character event graph construction method for multi-source data, in order to solve the problems of causal loss, static and rigid structure and uninterpretable prediction in the existing technology. Its core lies in constructing a causal event tensor graph, introducing structural uncertainty to drive graph growth, and integrating latent variable generation and large language model, so as to finally achieve interpretable dynamic character behavior modeling and prediction.

[0006] The present invention provides a method for constructing an intelligent dynamic person event graph based on multi-source data, comprising the following steps: Step 1) Multi-source data structured parsing: receiving structured, semi-structured, and unstructured raw data containing person-related information, and extracting event units from the raw data using a large language model and a tabular semantic parsing model, wherein each event unit is represented as... ,in Indicates the first A person, Indicates the first One action, Indicates the first Each time segment; all event units constitute a three-dimensional event tensor space. ,in Indicates the total number of people. Indicates the total number of action types. Indicates the total number of time periods within the time span; Step 2) Event graph construction and link reasoning: The event units obtained in Step 1 Mapped to a high-dimensional embedding vector Based on event embedding vectors, the connection weights between events are calculated. To construct a causal event graph ,in The set representing the directed causal relationships between events; Step 3) Calculation and update of the structural uncertainty of the event graph: For each event node e in the event graph, calculate its structural uncertainty. ;when When the preset threshold is exceeded, a graph update operation is triggered, including adding new event nodes or new causal edges; Step 4) Individual potential behavior modeling and generation: targeting the individual Based on its historical event sequence The latent factor representation of behavior is learned using a variational autoencoder. ; Using the decoder according to Predict the character's behavior at the next time step. Meanwhile, based on latent variables Step 5) Tracing back its causal path in the causal event graph and calling the large language model to generate a natural language explanation of the predicted behavior; Step 6) Comprehensive training on causal path consistency: Constructing the total loss function. The model is trained such that the total loss function includes at least the behavior prediction cross-entropy loss. Latent variable regularization loss Consistency loss of causal path Step 6) Visualization output: Visualize the constructed causal event graph, event timeline, predicted behavior, and natural language interpretation.

[0007] Optimally, in step 2, the event unit Mapped to a high-dimensional embedding vector The specific method is as follows: calculate entity embedding ,in It is an entity embedding matrix; it calculates action embeddings. ,in It is the action embedding matrix; the temporal embedding is calculated. The entity is generated using a sine / cosine position encoding method; the entity is then embedded... Action embedding and time embedding The data is concatenated and projected onto a unified dimension using a linear transformation to obtain the event embedding vector. ,in It is a fused projection matrix. It is a bias term.

[0008] Optimally, in step 2, the connection weights between events are calculated. The specific formula is: ,in, and It is the linear transformation matrix between the query and the key in the attention mechanism. It is the vector dimension. It is a hyperparameter that adjusts the weights of causal terms. It is used for judgment A causal scoring function to determine whether a causal relationship exists; when At that time, then in the event and Create directed causal edges between them ,in This is a preset threshold.

[0009] Ideally, in step 3, the structural uncertainty of event node e The calculation formula is: ,in, It is the information entropy of the predicted distribution for the next row, used to measure the uncertainty of the prediction; It is a variational generator With spectral structure model The KL divergence between the predicted behavior distributions is used to measure the inconsistency between the map and the simulation; α is a hyperparameter that adjusts the relative weights of the two terms.

[0010] Optimally, in step 4, the behavioral latent factor representation is learned using a variational autoencoder. The specific method is as follows: Encoding stage: ,in It is an encoder network that outputs the mean value. and variance Through reparameterization techniques Sampling; Decoding stage: via decoder network Output a probability distribution of actions and sample the specific actions to be predicted. .

[0011] Ideally, in step 5, the causal path consistency loss... The calculation formula is: , where f(e i ) and g(e' i ) are events e i Its target node e' in the real event path i causal path embedding vector; It is the actual path distribution Distribution of generated paths The KL divergence between them; λ is the balance coefficient.

[0012] Optimally, the true path distribution It is estimated by statistically analyzing the frequency of path π in historical trajectory data, and the calculation formula is as follows: ,in It represents the number of times the trajectory π appears in the real dataset D.

[0013] The present invention also describes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described schemes.

[0014] The present invention also describes a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] 1. In terms of system modeling capabilities and adaptability, this invention achieves a leap from static correlation analysis to dynamic causal reasoning. Traditional character graphs are mostly limited to static triple relationships, making it difficult to characterize the temporal and causal nature of behavior. This invention, by introducing a causal weight mechanism and a three-dimensional event tensor, not only unifies and integrates multi-source heterogeneous data but also accurately identifies implicit causal chains in high-noise environments. In particular, through a graph adaptive growth mechanism driven by the structural uncertainty index U(e), the system can automatically update and correct its structure in real-world scenarios with incomplete or continuously evolving data, significantly improving the system's modeling accuracy and robustness for complex and dynamic character behavior patterns.

[0017] 2. Regarding the depth and generalization of behavior prediction, this invention achieves a deeper understanding, moving from surface behavior matching to inferring underlying intentions. Addressing the shortcomings of traditional methods in modeling individual decision-making motivations, this invention utilizes a variational autoencoder (VAE) to decouple explicit event features from implicit causal intentions within the latent space, thereby learning individualized behavioral representations that incorporate personality preferences, strategies, and risk appetite. This latently causally driven embedding method enables the model to maintain good generalization ability even when faced with low-label data or requiring cross-domain transfer, significantly improving the accuracy and reliability of tasks such as abnormal transaction detection and high-risk behavior warning in small-sample scenarios.

[0018] 3. Regarding the interpretability and decision support value of the results, this invention solves the long-standing "black box" problem of intelligent analysis systems. By introducing a causal path consistency loss function, it ensures that the model's predicted behavior is highly consistent with its inference chain in the graph. Furthermore, it innovatively utilizes a Large Language Model (LLM) to transform abstract causal paths into clear and coherent natural language explanations, forming an integrated output of "prediction-reasoning-explanation." This mechanism enables analysts not only to know "what will happen," but also to understand "why it will happen," enhancing the auditability and credibility of the results and providing direct and powerful support for high-risk decisions such as national security and financial risk control. Attached Figure Description

[0019] Figure 1 This is a flowchart of the intelligent dynamic human event graph construction method for multi-source data according to the present invention. Detailed Implementation

[0020] The following detailed description, in conjunction with the accompanying drawings and embodiments, illustrates a method for constructing an intelligent dynamic person-event graph based on multi-source data, according to the present invention. Those skilled in the art will understand that the embodiments described are merely illustrative of the invention and not intended to limit it.

[0021] The overall process of implementing this invention is as follows: Figure 1 As shown, the process mainly includes six core steps: multi-source data structured parsing, event graph construction and link reasoning, graph structure uncertainty calculation and updating, individual potential behavior modeling and generation, causal path consistency comprehensive training, and visualization output. Each step is described in detail below.

[0022] Step 1: Multi-source data structured parsing: The goal of this step is to transform the original multi-source heterogeneous data into structured event units.

[0023] Input data: The raw data received by the system includes, but is not limited to:

[0024] Structured data: such as bank statement tables and communication record tables in a database.

[0025] Semi-structured data: such as meeting minutes and survey logs in JSON or XML format.

[0026] Unstructured data: such as news reports, social media posts, and legal documents.

[0027] Event extraction: A pre-trained large language model is used to process unstructured text, and a specialized tabular semantic parsing model is used to process tabular data to extract core semantic information. The key to extraction is identifying the three core elements required to constitute an event: people ( ),action( ) and time ( ).

[0028] Tensor construction: Each successfully extracted event is represented as a triple. For example, from a news article "Company A Chairman Zhang San announced a major asset restructuring on May 20, 2024," the event (Zhang San, announced asset restructuring, 2024-05-20) can be extracted. All event units are systematically organized into a three-dimensional tensor. In this tensor, P is the set of all characters, A is the set of all action types, and T is the set of time segments divided in chronological order. This tensor forms the data foundation for all subsequent calculations.

[0029] Step 2: Event Graph Construction and Linking Reasoning: This step aims to connect discrete event units into a graph structure with causal relationships.

[0030] Event embedding: To perform computation, each symbolized event must first be embedded... Mapped to a low-dimensional, dense vector representation The specific implementation is as follows:

[0031] Entity embedding: from a learnable embedding matrix Search for the corresponding person vector .

[0032] Action embedding: from a learnable embedding matrix Search for the corresponding action vector .

[0033] Time embedding: Timestamps are encoded using the sine and cosine positional encoding formulas from the Transformer model. Encode to obtain a vector .

[0034] Fusion: After concatenating the three vectors above, a fully connected layer is used for fusion and dimensionality reduction. .

[0035] Causal edge construction: calculating any two events and Causal connection weights between This weight consists of two parts:

[0036] Semantic similarity terms: It is a standard attention mechanism that measures the semantic relevance of events on the surface.

[0037] Causal strength term: The key innovation is the use of intervention techniques in causal inference to calculate... In practice, this score can be estimated from the data using PC algorithms based on conditional independent tests or LiNGAM.

[0038] Ultimately, if the overall weighting If the value exceeds a preset threshold δ (e.g., 0.5), then in the map... Create a path from i points to A directed edge represents a causal relationship.

[0039] Step 3: Calculation and update of event graph structure uncertainty: This step enables the graph to have the ability to self-evaluate and dynamically evolve.

[0040] Uncertainty calculation: For each event node e, its structural uncertainty. Calculate using the following formula:

[0041] Predicting entropy Using a Transformer predictor, the probability distribution p of the next possible action of node e is predicted based on the context, and its information entropy is calculated. The larger the entropy value, the more uncertain the prediction.

[0042] KL divergence Calculate the difference between the distribution q output by the Individual Behavior Generator (VAE decoder) and the distribution p output by the Graph Predictor. The larger the difference, the further the individual behavior deviates from the Graph Consensus, and the more "abnormal" the node is.

[0043] α is a hyperparameter, usually set to 0.5 to balance the contributions of the two terms.

[0044] Triggered update: The system continuously monitors all nodes. When a certain node When the value exceeds the threshold ϵ, the following update operation is triggered:

[0045] Add new events: Based on the current high uncertainty context, propose and validate new candidate event nodes to be added to tensor T using a generative model.

[0046] Adding / removing edges: Recalculate the relationship between the node and its neighbors. j, update the adjacency matrix, which may add causal edges or remove incorrect edges.

[0047] Step 4: Individual Potential Behavior Modeling and Generation: This step builds a personalized behavioral model for each individual and generates an explanation.

[0048] Latent variable learning: Using variational autoencoders (VAEs) to learn about human variables. Historical event sequence Modeling is performed.

[0049] encoder It is a multilayer perceptron (MLP) or recurrent neural network (RNN), with input... Output hidden variables The mean μ and variance σ of the Gaussian distribution are given. This is determined using reparameterization techniques. Sampling is performed to obtain the latent vector representing the character's behavioral intention. .

[0050] decoder It is another MLP, with Given the input, output a probability distribution over all actions A, and based on this distribution, predict the most likely action at the next moment. .

[0051] Interpretable generation:

[0052] Path backtracking: in causal event graphs In, from the current latent variables Starting from the most relevant event node, trace back its upstream causal path to form an event chain. .

[0053] LLM interpretation: This involves combining the explicit causal path sequence with the predicted behavior. Together, they are used as prompt words and input into the large language model. LLM leverages its powerful semantic understanding and generation capabilities to transform this path into a fluent natural language explanation, such as: "Because Zhang (… ) received multiple transfers from high-risk accounts in the past week. Furthermore, the company issued an abnormal financial announcement yesterday. Therefore, the model predicts that its next action may be a "large-scale fund transfer". .

[0054] Step 5: Comprehensive Training for Causal Path Consistency: This step ensures the accuracy and causal plausibility of the model's predictions by designing a multi-task loss function. The model's total loss function is:

[0055] Behavioral prediction loss Standard cross-entropy loss measures predictive behavior. The difference between actual behavior and real behavior.

[0056] Latent variable regularization loss KL divergence loss according to VAE standard, This constrains the latent variable space to conform to the prior distribution, preventing overfitting.

[0057] Causal path consistency loss The core loss term of this invention is calculated using the following formula:

[0058] Path embedding alignment items: and These are the embedding vectors of the real path and the corresponding events in the generated path, respectively, encoded by the graph neural network. This term requires that they be as close as possible in the vector space.

[0059] Path distribution alignment items: It is the distribution of true causal paths obtained statistically from the training data. ). It is the path distribution generated by the model, by The calculations show that the KL divergence term forces the causal paths generated by the model to maintain an overall distribution consistent with those observed in the real world.

[0060] Step 6: Visualization Output: This step presents the analysis results from the previous steps to the end user in a graphical format.

[0061] Timeline view: The system displays events in chronological order. Arrange the data to generate an interactive timeline that clearly shows the evolution of events involving key figures.

[0062] Causal Graph View: The system displays a causal event graph. Rendered as a node-link graph, where nodes represent events and directed edges represent causal relationships, with connection weights visualized using color, thickness, and other methods. and uncertainty .

[0063] Prediction Report: The system generates a comprehensive report that includes: the model's predicted next action, key causal paths extracted from the graph, and a natural language explanation generated by LLM. This enables analysts to quickly understand the basis of the predictions and make final decisions.

[0064] In summary, this invention provides an intelligent dynamic person-event graph construction method for multi-source data, successfully solving the core bottlenecks of traditional technologies in terms of causality, dynamism, and interpretability. The fundamental breakthrough of this method lies in elevating discrete person behavior data into a continuous event evolution chain with causal logic. By constructing a three-dimensional event tensor of "person-behavior-time" and innovatively introducing a causal weighting mechanism and a structural uncertainty-driven model, the graph achieves a leap from static association to dynamic causal reasoning, and from a fixed structure to adaptive growth. Simultaneously, the method deeply integrates variational autoencoders and large language models, enabling not only accurate prediction of individual behavior but also the generation of clear natural language explanations, revealing the "causal logic behind the behavior." Ultimately, this forms an intelligent analysis system capable of unifying and integrating multi-source information, supporting causal reasoning, possessing self-evolution and growth capabilities, and providing highly interpretable prediction results. This provides reliable technical support for in-depth analysis and decision-making in key areas such as national security and financial risk control.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an intelligent dynamic person-event graph for multi-source data, characterized in that, Includes the following steps: Step 1) Multi-source data structured parsing: Receive structured, semi-structured, and unstructured raw data containing information about individuals. Use a large language model and a tabular semantic parsing model to extract event units from the raw data, where each event unit is represented as... ,in Indicates the first A person, Indicates the first One action, Indicates the first Each time segment; all event units constitute a three-dimensional event tensor space. ,in Indicates the total number of people. Indicates the total number of action types. Indicates the total number of time periods within the time span; Step 2) Event graph construction and link reasoning: The event units obtained in Step 1 Mapped to a high-dimensional embedding vector ; Based on event embedding vectors, the connection weights between events are calculated. To construct a causal event graph ,in The set representing the directed causal relationships between events; Step 3) Calculation and update of the structural uncertainty of the event graph: For each event node e in the event graph, calculate its structural uncertainty. ;when When the preset threshold is exceeded, a graph update operation is triggered, including adding new event nodes or new causal edges; Step 4) Individual potential behavior modeling and generation: targeting the individual Based on its historical event sequence The latent factor representation of behavior is learned using a variational autoencoder. ; Using the decoder according to Predict the character's behavior at the next time step. Meanwhile, based on latent variables Step 5) Tracing back its causal path in the causal event graph and calling the large language model to generate a natural language explanation of the predicted behavior; Step 6) Comprehensive training on causal path consistency: Constructing the total loss function. The model is trained such that the total loss function includes at least the behavior prediction cross-entropy loss. Latent variable regularization loss Consistency loss of causal path Step 6) Visualization output: Visualize the constructed causal event graph, event timeline, predicted behavior, and natural language interpretation.

2. The intelligent dynamic person event graph construction method for multi-source data according to claim 1, characterized in that, In step 2, the event unit Mapped to a high-dimensional embedding vector The specific method is as follows: calculate entity embedding ,in It is an entity embedding matrix; it calculates action embeddings. ,in It is the action embedding matrix; the temporal embedding is calculated. The entity is generated using a sine / cosine position encoding method; the entity is then embedded... Action embedding and time embedding The data is concatenated and projected onto a unified dimension using a linear transformation to obtain the event embedding vector. ,in It is a fused projection matrix. It is a bias term.

3. The intelligent dynamic person-event graph construction method for multi-source data according to claim 1, characterized in that, In step 2, the connection weights between events are calculated. The specific formula is: ,in, and It is the linear transformation matrix between the query and the key in the attention mechanism. It is the vector dimension. It is a hyperparameter that adjusts the weights of causal terms. It is used for judgment A causal scoring function to determine whether a causal relationship exists; when At that time, then in the event and Create directed causal edges between them ,in This is a preset threshold.

4. The intelligent dynamic person event graph construction method for multi-source data according to claim 1, characterized in that, In step 3, the structural uncertainty of event node e The calculation formula is: ,in, It is the information entropy of the predicted distribution for the next row, used to measure the uncertainty of the prediction; It is a variational generator With spectral structure model The KL divergence between the predicted behavior distributions is used to measure the inconsistency between the map and the simulation; α is a hyperparameter that adjusts the relative weights of the two terms.

5. The intelligent dynamic person-event graph construction method for multi-source data according to claim 1, characterized in that, In step 4, the behavioral latent factor representation is learned using a variational autoencoder. The specific method is as follows: Encoding stage: ,in It is an encoder network that outputs the mean value. and variance Through reparameterization techniques Sampling; Decoding stage: via decoder network Output a probability distribution of actions and sample the specific actions to be predicted. .

6. The intelligent dynamic person-event graph construction method for multi-source data according to claim 1, characterized in that, In step 5, the causal path consistency loss The calculation formula is: , where f(e i ) and g(e' i ) are events e i Its target node e' in the real event path i causal path embedding vector; It is the actual path distribution Distribution of generated paths The KL divergence between them; λ is the balance coefficient.

7. The intelligent dynamic person event graph construction method for multi-source data according to claim 6, characterized in that, The true path distribution It is estimated by statistically analyzing the frequency of path π in historical trajectory data, and the calculation formula is as follows: ,in It represents the number of times the trajectory π appears in the real dataset D.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images