A knowledge graph-based method and system for visual analysis of specific targets

Abstract

This invention proposes a specific target visualization analysis method and system based on knowledge graphs, comprising: acquiring a basic database from multiple data sources, including entity data and dynamic event data; classifying topics from each data source in the basic database using a streaming text single-pass clustering method; extracting key information, entities, and inter-entity relationships from each data source; and constructing a knowledge graph by fusing the extraction results from each data source; retrieving the knowledge graph based on the specific target to be analyzed to obtain the target entity and related events; performing personnel and organizational analysis on the target entity to form an entity profile; and displaying the relationships between the target entity and its related entities or between the target entity and its related events in the form of nodes and edges. This invention provides the ability to comprehensively analyze multi-domain data, offering temporal clue reasoning and implicit clue reasoning functions, and can uncover clue data that does not exist in the database.

Description

Technical Field

[0001] This invention belongs to the field of data analysis technology, and in particular relates to a specific target visualization analysis system based on knowledge graphs. Background Technology

[0002] The rise of the Internet age has accelerated the generation and transmission of data, and society's demand for data analysis is growing stronger. How to extract high-value information from massive amounts of data has become an important means of survival in this era.

[0003] Data analytics, as a crucial branch of big data applications, involves collecting, cleaning, denoising, disambiguating, and aligning data from various fields to create structured data. It then uses model algorithms to fuse and correlate multi-source, heterogeneous data, constructing multi-domain knowledge graphs. Knowledge graphs, as a key tool for relationship analysis, display the relationships between entities in the form of nodes and edges. By retrieving a specific target from the data, extracting multi-dimensional entity profiles or event timelines, and presenting them in visual formats such as charts and relationship graphs, knowledge graphs ultimately provide efficient analytical capabilities and deep relationship mining capabilities, assisting users in strategic decision-making.

[0004] Currently, most data analysis technologies are applied to fields such as public opinion, technology, and finance. These different fields have significantly different data and analytical dimensions, resulting in existing technologies only being able to solve data analysis within a single field. They lack the ability to perform cross-domain comprehensive analysis and have poor adaptability, which manifests in two ways: 1. They cannot effectively integrate knowledge from multiple fields, leading to a single-dimensional analysis of specific targets; 2. The analytical operators have poor adaptability, only catering to specific business needs in a particular field, and cannot flexibly adapt to changing analytical business requirements. When encountering new analytical business needs, operators need to be redesigned. This limits the application capabilities of data analysis technologies and fails to meet the needs of users in different fields for efficient data analysis. Existing data analysis systems typically present the analysis results directly, greatly reducing the user's time for visual analysis and failing to support interactive clue mining, which does not align with users' thinking patterns.

[0005] Existing technologies primarily focus on data analysis within specific domains, lacking the capability for comprehensive multi-domain analysis. This results in incomplete data, limited analytical dimensions, and constraints in social relationship analysis when targeting specific objectives. In terms of analytical methods, most products rely heavily on public opinion data, neglecting the analysis of relationships between entities within events. Entities are categorized into various types, such as individuals, schools, businesses, and government agencies. Current technologies largely fail to integrate data from different entity types, leading to limited information content and a singular analytical dimension. Regarding visualization, analysis results are mostly presented as charts or relationship graphs. However, the workflow is inefficient for in-depth relationship mining and granular analysis, failing to fully showcase the exploration process and hindering time-series event clue analysis for specific targets. Furthermore, existing technologies do not support custom nodes and edges when displaying different node types, resulting in unclear presentation of analysis results and a poor user experience. Summary of the Invention

[0006] The purpose of this invention is to solve the problems of existing technologies that do not support multi-domain comprehensive analysis when performing specific target analysis, and whose visualization analysis process does not conform to the thinking logic of users when mining clues. This invention proposes a specific target visualization analysis system based on knowledge graphs.

[0007] Specifically, this invention proposes a specific target visualization analysis method based on knowledge graphs, including:

[0008] Step S1: Obtain a basic database from multiple data sources, including entity data and dynamic event data. Use the streaming text single-pass clustering method to classify the topics of each data source in the basic database, and extract key information, entity and inter-entity relationship information from each data source. Combine the extraction results from each data source to construct a knowledge graph.

[0009] Step S2: Retrieve the knowledge graph for the specific target to be analyzed to obtain the target entity and related events of the specific target; perform personnel analysis and organizational analysis on the target entity to form an entity profile of the target entity; display the relationship between the target entity and its related entities or between the target entity and the related events in the form of nodes and edges.

[0010] The knowledge graph-based specific target visualization analysis method, in step 2, the associated event includes a time-series event clue reasoning event; based on the specific target and the knowledge graph, the trend of the specific target's participation in the event is predicted by a time-series event reasoning model based on an event graph neural attention network, thus obtaining the time-series event clue reasoning event;

[0011] The temporal event reasoning model includes a clue collection module and a clue reasoning module. The clue collection module is used to recall event clues that are helpful for reasoning query q from an event knowledge graph F constructed from known dynamic events. The clue collection module consists of a random walk collector and a path encoder. The former retrieves several N path clues of length l starting from s in the event knowledge graph based on the given query information and assembles them into a clue set. The path encoder is implemented by a unidirectional LSTM encoder, which encodes the paths collected by the random walk collector into vector information and calculates the weight coefficient of each path in the clue set so that the clue reasoning module can perform the next step of reasoning. The goal of the clue reasoning module is to infer the set of hidden events from the event clue information that is helpful for reasoning query q recalled by the clue collection module. The clue reasoning module consists of a subgraph encoding module and a temporal reasoning module. The subgraph encoding module is composed of a graph neural network, which is used to understand the topological features of the clues at each time step. The temporal reasoning module is composed of a unidirectional recurrent neural network, which is used to learn the law of unidirectional evolution of clues over time and predict the future development of the temporal event to obtain the inference event of the temporal event clue.

[0012] The aforementioned knowledge graph-based target visualization analysis method includes the following training process for the time-series event reasoning model:

[0013] Initialize the entity vector, relation vector, and time vector, setting them to trainable mode; for the query q = s, r, t and the correct entity o, initialize the query representation vector as follows:

[0014] Starting from a specific target entity s, N path clues of length l are sampled through the random walk collector and assembled into a clue set P = {p0, p1, ..., p...} N}, where N and l are user-defined parameters;

[0015] The paths in the clue set are encoded by a path encoder. For a path p = [s,r0,t0,o0,r1,t1,o1,...,r] l-1 ,t l-1 ,o l-1 (where s0 = s), the encoding formula is (0 ≤ k ≤ l-1):

[0016] t k =σ(W r Δt k +U r h k-1

[0017]

[0018]

[0019] Δt in the above formula k =t k -t k-1 (When k = 0, Δt) k =0); t k Represents timing threshold control. W represents the semantic information encoded in this unit. r W h with U r U h Representing 4 mapping matrices, The k-th step in the path represents the correspondence and entity vector; the output of the LSTM at the k-th step is h. k ;

[0020] The path p corresponds to the output h of step l-1. l-1 This becomes the encoding vector p for the path, from which the encoding vectors {p0, p1, ..., p} for the path clue set are obtained. N};

[0021] Softmax is used to calculate the attention score {α0, α1, ..., α} for each encoded vector with respect to the query. N};

[0022] The global thread representation is obtained by using the encoded vectors of the path thread set.

[0023] Using the clue set P, a query subgraph sequence {SG0,SG1,...,SG} is constructed based on the event time information in the path. t-1};

[0024] The graph neural network uses a subgraph encoding module to encode each subgraph in the subgraph sequence, obtaining the node representation {SG0,SG1,...,SG} in the subgraph. t-1};

[0025] The unidirectional recurrent neural network using this time-series inference module encodes the node representations of the aforementioned subgraph sequence, outputting the subgraph node representation SG at time t. t ;

[0026] Combined with global clues Calculate SG t The matching score for each node in the dataset;

[0027] Select the entity o′ with the highest matching score and the correct entity o, calculate the loss for this training using the binary classification loss function, and train and update the time-series event reasoning model using the backpropagation algorithm.

[0028] The aforementioned knowledge graph-based target visualization analysis method includes step S1, which comprises:

[0029] A multi-source heterogeneous data fusion model is constructed to complete the mapping relationship between entities from different data sources based on the extraction results from different data sources, thereby achieving the fusion of extraction results from various data sources and constructing the knowledge graph.

[0030] The multi-source heterogeneous data fusion model includes an entity representation module and an entity mapping relationship reasoning module;

[0031] The entity representation module encodes entity features based on the entity's textual description, temporal information, and the structural information of its neighbors, serving as the basis for mapping relationship mining. This module consists of an entity textual description encoding module, a temporal information encoding module, and a structural information encoding module. The entity textual description encoding is achieved using a pre-trained language model BERT combined with a fully connected layer of a neural network. Temporal information is encoded using the Time2vec algorithm. For structural information, the entity description and temporal information representations are concatenated as input and encoded using a graph neural network with an attention mechanism. The resulting vector is then used as the final representation of the entity.

[0032] The entity mapping relationship reasoning module is used to infer the mapping relationship between entities based on the final representation of the entities. The entity mapping relationship reasoning module consists of two parts: an entity representation translation module and an entity mapping matching module. The entity representation translation module is composed of a multilayer perceptron (MLP) and is used to map the representations of entities from different data sources to a unified semantic space. The entity mapping matching module first calculates the matching degree of entity representations from different sources using cosine similarity, and then performs entity mapping matching based on the numerical value of the matching degree to finally obtain the mapping relationship between entities, thereby realizing the fusion of multi-source heterogeneous data and obtaining the knowledge graph.

[0033] This invention also proposes a knowledge graph-based visual analysis system for specific targets, comprising:

[0034] The knowledge graph construction module is used to obtain a basic database from multiple data sources, including entity data and dynamic event data. It uses a streaming text single-pass clustering method to classify the topics of each data source in the basic database, and extracts key information, entities and relationships between entities from each data source. The extraction results from each data source are then integrated to construct a knowledge graph.

[0035] The target visualization module is used to retrieve the knowledge graph for a specific target to be analyzed, obtain the target entity and related events of the specific target; perform personnel analysis and organizational analysis on the target entity to form an entity profile of the target entity; and display the relationship between the target entity and its related entities or between the target entity and the related event in the form of nodes and edges.

[0036] The aforementioned knowledge graph-based specific target visualization analysis system includes a target visualization module where the associated events include time-series event clue reasoning events. Based on the specific target and the knowledge graph, the system uses a time-series event reasoning model based on an event graph neural attention network to predict the trend of the specific target's participation in events, thereby obtaining the time-series event clue reasoning events.

[0037] The temporal event reasoning model includes a clue collection module and a clue reasoning module. The clue collection module is used to recall event clues that are helpful for reasoning query q from an event knowledge graph F constructed from known dynamic events. The clue collection module consists of a random walk collector and a path encoder. The former retrieves several N path clues of length l starting from s in the event knowledge graph based on the given query information and assembles them into a clue set. The path encoder is implemented by a unidirectional LSTM encoder, which encodes the paths collected by the random walk collector into vector information and calculates the weight coefficient of each path in the clue set so that the clue reasoning module can perform the next step of reasoning. The goal of the clue reasoning module is to infer the set of hidden events from the event clue information that is helpful for reasoning query q recalled by the clue collection module. The clue reasoning module consists of a subgraph encoding module and a temporal reasoning module. The subgraph encoding module is composed of a graph neural network, which is used to understand the topological features of the clues at each time step. The temporal reasoning module is composed of a unidirectional recurrent neural network, which is used to learn the law of unidirectional evolution of clues over time and predict the future development of the temporal event to obtain the inference event of the temporal event clue.

[0038] The training process of the time-series event reasoning model in the knowledge graph-based target visualization analysis system includes:

[0039] Initialize the entity vector, relation vector, and time vector, setting them to trainable mode; for the query q = (s, r, t) and the correct entity o, initialize the query representation vector as follows:

[0040] Starting from a specific target entity s, N path clues of length l are sampled through the random walk collector and assembled into a clue set P = {p0, p1, ..., P}. N}, where N and l are user-defined parameters;

[0041] The paths in the clue set are encoded by a path encoder. For a path p = [s,r0,t0,o0,r1,t1,o1,...,r] -1 ,t l-1 ,o l-1 (where s0 = s), the encoding formula is (0 ≤ k ≤ l-1):

[0042] t k =σ(Wr Δt k +U r h k-1 )

[0043]

[0044]

[0045] Δt in the above formula k =t k -t k-1 (When k = 0, Δt) k =0); t k Represents timing threshold control. W represents the semantic information encoded in this unit. r W h with U r U h Representing 4 mapping matrices, The k-th step in the path represents the correspondence and entity vector; the output of the LSTM at the k-th step is h. k ;

[0046] The path p corresponds to the output h of step l-1. l-1 This becomes the encoding vector p for the path, from which the encoding vectors {p0, p1, ..., p} for the path clue set are obtained. N};

[0047] Softmax is used to calculate the attention score {α0, α1, ..., α} for each encoded vector with respect to the query. N};

[0048] The global thread representation is obtained by using the encoded vectors of the path thread set.

[0049] Using the clue set P, a query subgraph sequence {SG0,SG1,...,SG} is constructed based on the event time information in the path. t-1};

[0050] The graph neural network uses a subgraph encoding module to encode each subgraph in the subgraph sequence, obtaining the node representation {SG0,SG1,...,SGt} in the subgraph. -1};

[0051] The unidirectional recurrent neural network using this time-series inference module encodes the node representations of the aforementioned subgraph sequence, outputting the subgraph node representation SG at time t. t ;

[0052] Combined with global clues Calculate SG tThe matching score for each node in the dataset;

[0053] Select the entity o′ with the highest matching score and the correct entity o, calculate the loss for this training using the binary classification loss function, and train and update the time-series event reasoning model using the backpropagation algorithm.

[0054] The aforementioned knowledge graph-based target visualization and analysis system includes a knowledge graph construction module comprising:

[0055] A multi-source heterogeneous data fusion model is constructed to complete the mapping relationship between entities from different data sources based on the extraction results from different data sources, thereby achieving the fusion of extraction results from various data sources and constructing the knowledge graph.

[0056] The multi-source heterogeneous data fusion model includes an entity representation module and an entity mapping relationship reasoning module;

[0057] The entity representation module encodes entity features based on the entity's textual description, temporal information, and the structural information of its neighbors, serving as the basis for mapping relationship mining. This module consists of an entity textual description encoding module, a temporal information encoding module, and a structural information encoding module. The entity textual description encoding is achieved using a pre-trained language model BERT combined with a fully connected layer of a neural network. Temporal information is encoded using the Time2vec algorithm. For structural information, the entity description and temporal information representations are concatenated as input and encoded using a graph neural network with an attention mechanism. The resulting vector is then used as the final representation of the entity.

[0058] The entity mapping relationship reasoning module is used to infer the mapping relationship between entities based on the final representation of the entities. The entity mapping relationship reasoning module consists of two parts: an entity representation translation module and an entity mapping matching module. The entity representation translation module is composed of a multilayer perceptron (MLP) and is used to map the representations of entities from different data sources to a unified semantic space. The entity mapping matching module first calculates the matching degree of entity representations from different sources using cosine similarity, and then performs entity mapping matching based on the numerical value of the matching degree to finally obtain the mapping relationship between entities, thereby realizing the fusion of multi-source heterogeneous data and obtaining the knowledge graph.

[0059] The present invention also proposes a storage medium for storing programs that execute any of the knowledge graph-based target visualization analysis methods.

[0060] The present invention also proposes a client for any of the knowledge graph-based target visualization and analysis systems.

[0061] This invention integrates heterogeneous data from multiple sources to form a foundational library. This library is not limited to any particular domain and includes, but is not limited to, data on individuals, schools, enterprises, government agencies, and democratic alliances. It analyzes the relationships between entities using model algorithms and constructs a knowledge graph by combining entity information related to dynamic events. Following the user's thought process, this invention supports specific target retrieval by entity and event, providing various analytical tools such as entity profiling, event profiling, relationship analysis, entity-event association analysis, and implicit clue mining. When performing specific target association analysis, users can freely define the search scope (e.g., search for entities or events), select operators suitable for the current business from the operator library, and independently switch view layouts, customizing node styles by type or selected scope to achieve a higher level of visualization analysis. During the analysis process, this invention can perform granular searches of the search target, eliminating redundant targets and helping users obtain more accurate analysis results. Because the visualization analysis process of this invention originates from the user's thought process, it helps improve analysis efficiency and uncover more high-value implicit clues.

[0062] As can be seen from the above solutions, the advantages of the present invention are:

[0063] Compared with existing technologies, this invention provides the ability to comprehensively analyze data from multiple fields. Its designed visual analysis process is more in line with the user's thinking pattern, satisfying the exploration steps from retrieving specific targets to discovering related clues. This invention provides temporal clue reasoning and implicit clue reasoning functions, which can uncover clue data not found in the database. This improves analysis efficiency and user experience, facilitating the discovery and reasoning of high-value implicit clues. Attached Figure Description

[0064] Figure 1 Constructing graphs for knowledge graphs;

[0065] Figure 2 This is a flowchart for knowledge graph-based visualization analysis. Detailed Implementation

[0066] In the research on the comprehensive clue mining and visualization of specific targets in multi-domain data, it was found that if the massive amount of collected entity information is used as the basic data, and multi-source heterogeneous data is integrated and correlated, combined with different domain scenarios and dynamic events, the problem of the single analysis domain and analysis dimension when the existing technology is used to analyze specific targets can be solved.

[0067] In conducting research on the visualization of specific target clue mining processes, this invention is based on the integration of multi-source heterogeneous data to construct a multi-domain knowledge graph. Taking a task-oriented approach, it mines specific target clues, starting from the functional point of target retrieval. Through the design of analysis prototype engines for entity profiling, relationship analysis, opinion analysis, multi-hop related entity analysis, temporal clue mining, and implicit clue mining, the business logic is more aligned with users' thinking patterns and usage processes. Furthermore, when using the analysis prototype engine, users can prioritize relationship analysis as a key module for clue mining, employing more reasonable layout algorithms, style layering, custom nodes, and operator libraries to improve analysis efficiency and result output. Designing according to users' thinking logic is beneficial for improving analysis efficiency and uncovering implicit clues.

[0068] To make the above features and effects of the present invention clearer and easier to understand, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings.

[0069] This invention comprises two parts: constructing a multi-domain knowledge graph that includes entity data and dynamic event data, and visual analysis of specific target clues discovered based on the aforementioned multi-domain knowledge graph.

[0070] The flowchart for constructing the multi-domain knowledge graph designed in this invention is attached. Figure 1 As shown, the process steps are described below:

[0071] Step 1: Entity Data and Dynamic Event Acquisition

[0072] This invention requires the collection of two types of data: entities and dynamic events. The collected entity data mainly includes think tank scholars, corporate executives, enterprises, schools, government organizations, democratic parties, opinion leaders, and social media accounts from various fields. Entities are classified and labeled using model algorithms to construct a multi-domain knowledge graph, serving as a visual analysis tool for discovering specific target clues. The collected dynamic events mainly include news updates, social media posts, and official websites. A single-pass clustering method using streaming text processing is employed to classify article topics. Key information is extracted from the text through keyword matching and other methods, identifying entities such as people, events, and things mentioned in the text. Relationships between entities and argument information such as the time and location of events are extracted to initially construct a multi-domain knowledge graph.

[0073] Step 2: Multi-source heterogeneous data fusion

[0074] Business requirements necessitate the deep integration of heterogeneous data from different data sources for subsequent clue analysis. Therefore, the model needs to supplement the mapping relationships between entities from different data sources based on existing descriptive information to achieve fusion. Accordingly, this invention designs a multi-source heterogeneous data fusion model based on graph neural networks.

[0075] The input to this model is:

[0076] The knowledge graphs KG1 = (E1, R1, Q1, T1) and KG2 = (E2, R2, Q2, T2) from different data sources need to be merged, where E represents the entity set, R represents the relation set, and Q represents the fact set. A fact q ∈ Q is represented as (e... head ,r,e tail ,t), where e head ,e tail r and t represent the head entity, tail entity, relation, and the time of occurrence of the dynamic event, respectively, and e head ,e tail ∈E, r∈R, t∈T.

[0077] The output of this model is:

[0078] The mapping relationship between entities in KG1 and KG2 is {(e′0,e0),(e′1,e1),…(e′…} n ,e n )}, where e′0,e′1,…,e′ n ∈E1,e0,e1,…,e n ∈E2.

[0079] Model structure functions:

[0080] The model is divided into two modules: entity representation and entity mapping relationship reasoning.

[0081] The entity representation module aims to encode entity features based on their textual descriptions, temporal information, and the structural information of their neighbors, serving as the basis for mapping relationship mining. The entity representation module consists of three modules: an entity textual description encoding module, a temporal information encoding module, and a structural information encoding module. The textual description encoding is achieved using a pre-trained language model BERT combined with a fully connected layer of a neural network; temporal information is encoded using the Time2vec algorithm; and structural information is encoded by concatenating the entity description and temporal information representations as input and then encoding them using a graph neural network with an attention mechanism. Finally, the encoded vector is used as the final representation of the entity and is then used for the next step of mapping relationship inference.

[0082] The goal of the entity mapping relationship reasoning module is to infer the mapping relationships between entities based on their representations. This module consists of two parts: an entity representation translation module and an entity mapping matching module. The entity representation translation module, constructed using a multilayer perceptron (MLP), maps entity representations from different data sources to a unified semantic space. The entity mapping matching module first calculates the matching degree of entity representations from different sources using cosine similarity, and then uses this matching degree as a basis, combined with the Sinkhorn algorithm, to perform entity mapping matching, ultimately obtaining the mapping relationships between entities and achieving multi-source heterogeneous data fusion.

[0083] Model training methods:

[0084] The training process for this model is as follows:

[0085] Step (1): Input the textual description information, time information and structural information of entities from knowledge graphs from different data sources.

[0086] Step (2): Encode the textual description information of entities using the pre-trained language model BERT to obtain a high-dimensional representation of the entity textual description information, and then reduce the dimensionality through a fully connected layer of a neural network to obtain a representation of the entity name information. Where n is the number of entities in the knowledge graph.

[0087] Step (3): Represent the time information using the Time2Vec algorithm, with the following formula:

[0088]

[0089] In the above formula, t2v(.) is the encoder, k is the number of dimensions representing the time information, and ω i and These are learnable parameters; given a time t, the corresponding representation can be obtained through the above encoding method. Where n is the number of entities in the knowledge graph.

[0090] Step (4): Represent the entity name information Representation of time information The concatenation is used as input, and the Graph Attention Neural Network (GAT) is used to calculate the representation of entity structure information. Specifically, it includes an information transfer module, an attention weighting module, and an information aggregation module, as shown in the following formula:

[0091] The information transmission part calculates the interaction message representation transmitted from the target node o to the source node s, and transforms it according to the node type φ and the edge type ψ, as formally represented as follows:

[0092]

[0093] Where l represents the number of network layers, W M and W MSG The information transfer mapping matrix is ​​selected based on the node and edge types.

[0094] The attention weight calculation part is used to measure the importance of the interaction messages passed from the target node o to the source node s, and is formally represented as follows:

[0095]

[0096] W Q and W ATT Calculate the mapping matrix for the attention weights selected based on node and edge types.

[0097] Finally, information is aggregated based on information delivery and attention weights, as shown below:

[0098]

[0099] The entity representation is updated by using a multi-layered graph neural network. The final learned entity representation includes the entity's textual description information, time information, and structural information about the entity's surroundings captured by the graph neural network.

[0100] Step (5): For two knowledge graphs that need to be merged, compare the entities in each pair and calculate the similarity of the entity representations by using cosine similarity.

[0101] Step (6): Perform entity mapping matching using the Sinkhorn algorithm to obtain the mapping relationship between entities in KG1 and KG2 {(e'0,e0),(e'1,e1),…(e' n ,e n )}.

[0102] Step (7): Compare the predicted mapping relationship with the correct mapping relationship, calculate the training loss using the cross-entropy loss function, and correct the model parameters using the backpropagation algorithm.

[0103] Model inference steps:

[0104] Same as steps (1) to (6) above.

[0105] Model results:

[0106] The model's entity text description encoding module, temporal information encoding module, and structural information encoding module can comprehensively utilize entity text description information, temporal information, and graph structure information from multi-source heterogeneous data to help infer the mapping relationships between entities from different sources. Furthermore, in real-world application scenarios where some data noise or missing data exists, it avoids the problem of being unable to obtain entity representations due to the absence of certain types of data. The final set of mapping relationships between entity pairs obtained by the model can help business personnel analyze specific targets more comprehensively, integrate clues from different sources, and serve subsequent clue reasoning and other business needs.

[0107] Step 3: Create a new analysis task

[0108] Users can create new analysis tasks based on their business needs. Analysis tasks can include fields such as task name, task tag, task description, analysis results, and creation time, and are used to record clue discovery records and analysis results for a specific target.

[0109] Step 4: Target-Specific Retrieval

[0110] After creating a new analysis task, proceed to the specific target retrieval step. Target retrieval is divided into two categories: entity retrieval and event retrieval. When retrieval by entity, you can search according to three patterns: "one person," "a type of person," and "a group of people." The goal of the search is to select target entities from the search results for analysis.

[0111] "A person" is a uniquely identified entity. This person's actual identity must be unique, but they can have several virtual identities. By searching through the "person search box" and entering a person's name / ID / nickname, you can retrieve that person's actual identity and the virtual identities that may correspond to them on different platforms.

[0112] "A category of people" refers to individuals with one or more common attributes, which can be searched using the "Category Search Box". The "Category Search Box" has a variable number of input fields (one or any number) for searching with different numbers and combinations of category attributes. Attributes can be numeric, string, or category attributes, and can also be searched using fuzzy matching or natural language descriptions. After entering the category characteristics to be searched, the operation area will then display individuals who match the search criteria.

[0113] "A group of people" refers to multiple individuals indirectly related through other entities / events, which can be searched using the "Group Search Box". First, you need to find the associated entities / events (one or more) in the "Group Search Box" and click on the corresponding associated entity / event to search. After that, the operation area will display the individuals (i.e., "a group of people") that meet the search criteria.

[0114] When searching by event, users can search for events by event keywords, time range, tags, etc. From the candidate events returned by the system, they can select the target event for profile analysis (step 5). At the same time, it supports displaying all entities involved in multiple suspected events in the form of a list for entity analysis.

[0115] Step 5: Event Profiling Analysis

[0116] After selecting a specific target event in step 4's event retrieval, the event profiling analysis includes modules such as event details, timeline of major events, word cloud of keywords, locations involved, people involved, organizations involved, and similar event recommendations. This comprehensive display of details about the specific target event helps users trace the causes and consequences of the event. Editing of event information is also supported, improving the accuracy of event-related information.

[0117] Step 6: Obtain the target entity

[0118] Through target retrieval, users can select entities to be analyzed from the entity results returned in step 4 "Specific Target Retrieval" according to their own needs, in order to conduct opinion analysis (step 7), entity profile analysis (step 8), and relationship analysis (step 9).

[0119] Step 7: Viewpoint Analysis

[0120] After selecting a specific target entity in step 4 (entity retrieval), users can perform opinion analysis on that target entity. The purpose of opinion analysis is to analyze the target entity's sentiment towards related entities or events. This invention uses a text sentiment polarity analysis model as the basic module for opinion analysis. By analyzing the text information corresponding to entities and events related to the target entity retrieved from the dynamic event database, the sentiment polarity value of the target entity towards these entities and events is determined. Considering that the sentiment of a target entity may change over time, this invention uses a line graph to express the development trend of the target entity's sentiment tendency. The horizontal axis of the line graph corresponds to time, and the vertical axis ranges from [-1, 1], representing the polarity value of the opinion, where -1 represents completely negative and 1 represents completely positive. The points in the line graph represent the sentiment value of the target entity. This invention can trace the origins of event opinions, helping users understand the reasons for sentiment trends. Through the sentiment analysis model algorithm, the future sentiment trend of a specific target towards that entity or event is predicted.

[0121] Step 8: Entity Profile Analysis

[0122] After selecting a specific target entity in step 4 (Entity Search), users can perform entity profiling analysis on that target. Depending on the specific target, entity profiling analysis includes two categories: personnel analysis and organizational analysis. Personnel analysis focuses on an "individual" model, including modules such as basic personnel information, personal resume, personal abilities, tags, social keywords, participation in important activities, core viewpoints, social relationship graph, latest developments, and social activity. Organizational analysis analyzes "a type of person" and "a group of people" models, including modules such as basic organizational information, organizational evolution information, core member structure, tags, keywords, core viewpoints, member analysis, and social relationship graph. Various charts, such as pie charts, line charts, radar charts, relationship diagrams, word clouds, and bubble charts, are used to present the multi-dimensional analysis results of the specific target, forming an entity profile.

[0123] Step 9: Relationship Analysis

[0124] After selecting a specific target in step 4, the user can perform relationship analysis on that target. Relationship analysis can directly reflect the association between multiple entities or between an entity and an event, displayed in the form of nodes and edges. Nodes are labeled with the entity's name and category, and relationships are labeled with the relationship type.

[0125] Because entities in multi-domain knowledge graphs exist in various types, this invention supports customizing node icons, colors, sizes, shapes, and other styles to more intuitively display node types, with default style differentiation based on type. It also supports adjusting the color, shape, and display attributes of edges of different types. Regarding relationship graph layout, in addition to providing common force-guided layouts, grid layouts, ring layouts, concentric circle layouts, and radial layouts, a new clustered hierarchical layout has been added to address the diverse entity types in knowledge graphs. This allows for the clustering and hierarchical display of different types of nodes, facilitating the viewing of entities of the same type and improving the effectiveness of diverse analysis.

[0126] For common business scenarios, this invention provides an algorithm tool library function in step 9, which provides algorithm tools including path analysis, node filtering by attribute, weight analysis, etc., to help users quickly obtain the results they want to analyze with one click.

[0127] Step 10: Multi-hop related entity analysis

[0128] Multi-hop relational entity analysis refers to displaying other entities related to a specific target entity, as well as multi-hop relationship paths, through methods such as node expansion. In the relationship analysis described in step 9, only entities with a single association with the specific target entity are displayed, resulting in limited content. In this step, the user can define the number of hops H (H>1, and must be an integer). Based on the output of step 9, the system retrieves entities within H hops of the specific target entity in a multi-domain knowledge graph. As the relationships expand, more multi-level social relationships and background information about the target entity are discovered. Right-clicking a node provides quick access to common analysis tools, listing frequently used functions in a menu format to increase analysis efficiency and ease of use. For convenient display, this invention provides a community discovery function in step 10, allowing users to understand the closeness or distance between entities after dividing the discovered multi-hop relational entities into different communities.

[0129] Step 11: Reasoning based on chronological event clues

[0130] Business requirements necessitate predicting the future development trend of a specific target entity based on historical event clues. Therefore, the model needs to predict the future evolution trend of the entity based on collected time-series event (hereinafter referred to as "event") clue data. Based on this, this invention designs a time-series event clue reasoning model based on an event graph neural attention network.

[0131] The input to this model is:

[0132] (1) Query q = s, r, t, which contains a specific target entity s, a specific relation type r, and a prediction time t;

[0133] (2) Event knowledge graph F = {f0, f1, ..., f i}, where the clue event f i Represented as (s i ,r i ,o i ,t i In the form of ), where s i The header entity representing the event, r i Indicates the relationship type of the event, o i Represented as the tail entity of the event, t i This represents the time information of the event.

[0134] The output of this model is:

[0135] The inferred set of j implicit events related to query q is F' = {f'0, f'1, ..., f'}. j}, where the clue event f' j Represented as (s,r,o) j ,t).

[0136] Model structure functions:

[0137] The model is divided into two modules: clue collection and clue reasoning.

[0138] The goal of the clue collection module is to recall event clues that are helpful for reasoning query q from the event knowledge graph F constructed from known dynamic events. The clue collection module consists of a random walk collector and a path encoder. The former retrieves several N path clues of length l starting from s in the event knowledge graph based on the given query information and assembles them into a clue set. The path encoder is implemented by a unidirectional LSTM encoder, which encodes the paths collected by the random walk collector into vector information and calculates the weight coefficient of each path in the clue set so that the clue reasoning module can perform the next step of reasoning.

[0139] The goal of the clue reasoning module is to deduce the set of implicit events from the event clues retrieved by the clue collection module that are helpful in reasoning query q. The clue reasoning module consists of a subgraph encoding module and a temporal reasoning module. The subgraph encoding module is composed of a graph neural network, whose function is to understand the topological features of the clues at each time step. The temporal reasoning module is composed of a unidirectional recurrent neural network, whose function is to learn the unidirectional evolution of clues over time and predict the future development trend of temporal event clues.

[0140] Model training methods:

[0141] The training process for this model is as follows:

[0142] Step (1): Initialize the entity vector, relation vector, and time vector, setting them to trainable mode; for the query q = s, r, t and the correct entity o, initialize the query representation vector as follows:

[0143] Step (2): Starting from a specific target entity s, sample N paths of length l using a random walk collector, and assemble them into a path thread set P = {p0, p1, ..., p...} N}, where N and l are user-defined parameters;

[0144] Step (3): Encode the path in the thread set using a path encoder. For the path p = [s, r0, t0, o0, r1, t1, o1, ..., r l-1 ,t l-1 ,o l-1 The encoding formula is (0≤k≤l-1):

[0145] t k =σ(W r Δt k +U r h k-1

[0146]

[0147]

[0148] Δt in the above formula k =t k -t k-1 (When k = 0, Δt) k =0), t k Represents timing threshold control. W represents the semantic information encoded in this unit. r W h with U r U h Representing 4 mapping matrices, The k-th step in the path represents the correspondence and entity vector; the output of the LSTM at the k-th step is h. k .

[0149] Finally, path p corresponds to the output h of step l-1. 1-1 This becomes the encoding vector p for the path. Based on this, the present invention obtains the encoding vector {p0, p1, ..., p} of the path clue set. N}

[0150] Step (4): Use Softmax to calculate the attention score {α0, α1, ..., α} for each encoded vector in relation to the query. N}

[0151] Step (5): Calculate the global thread representation using the encoded vectors of the path thread set.

[0152] Step (6): Using the clue set P, construct the query subgraph sequence {SG0,SG1,...,SG} based on the event time information in the path. t-1}

[0153] Step (7): Encode each subgraph in the subgraph sequence using the graph neural network of the subgraph encoding module to obtain the node representation {SG0,SG1,...,SG} in the subgraph. t-1}

[0154] Step (8): Encode the node representation of the above subgraph sequence using the unidirectional recurrent neural network (RNN) of the temporal reasoning module, and output the subgraph node representation SG at time t. t .

[0155] Step (9): Combine global clues Calculate SG tThe matching score of each node is calculated and the matching scores of the nodes are sorted.

[0156] Step (10): Select the entity o' with a predicted rank of 1 and the correct entity o, calculate the loss for this training using the binary classification loss function, and correct the model parameters and node, relation, and time representation vectors using the backpropagation algorithm.

[0157] Model inference steps:

[0158] Same as steps (1) to (8) above.

[0159] In step (9), the set of entities with the top j matching scores (j can be specified by the user) is output, and a set of implicit events is constructed as the output.

[0160] Model results:

[0161] Temporal event clue reasoning can predict the future participation trends of specific target individuals in events, such as the establishment of social connections between individuals and the conduct of group activities. In the clue collection module, this model filters event clues that are helpful for reasoning query q for the next stage of clue reasoning and calculates the weight values ​​of the clues. This helps avoid selecting weakly related or irrelevant event clues during the reasoning process, thus reducing the efficiency of the reasoning model. In the clue reasoning module, a graph neural network is used to encode subgraph features, and a one-way recurrent neural network is used to learn the temporal features between subgraphs, enhancing the reasoning ability and achieving better results in temporal clue reasoning.

[0162] Step 12: Reasoning based on hidden clues

[0163] Considering the limitations of data sources and the errors inherent in information extraction models, it cannot be guaranteed that the data collection step in step 1 contains all the knowledge. Therefore, it is necessary to use the implicit clue reasoning step to uncover implicit clues. Implicit clue reasoning refers to using the multi-domain knowledge graph constructed and integrated in steps 1 and 2, along with the multi-hop related entities and multi-hop relationship paths returned in step 10, to discover implicit relationships between entities through rule mining algorithms such as PRA. These implicit relationships do not exist in the multi-domain knowledge graph. The principle of the implicit clue reasoning step is to infer the implicit direct association between two entities A and B that do not have a direct and obvious relationship through the indirect multi-hop relationship path between them. For example, if entity A and entity B have a multi-hop relationship path "A-birthplace-Honolulu-located-Hawaii-located-B", the implicit clue "(A, nationality, B)" can be inferred from this. Through the implicit clue reasoning step, the knowledge related to specific targets can be supplemented and improved, allowing users to comprehensively analyze specific target entities.

[0164] Step 13: Save the relationship analysis task

[0165] This invention can save the relationship analysis results formed in steps 10 to 12 to the selected analysis task so that the analysis results can be viewed at any time, and analysis and judgment can be carried out. At the same time, phased exploration can be carried out based on this.

[0166] Step 14: Discovering Clues

[0167] End users analyze and judge the search target based on the analysis results of steps 7, 8, and 13, and obtain high-value specific target-related clues.

[0168] The following are system embodiments corresponding to the above method embodiments. This embodiment can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0169] This invention also proposes a knowledge graph-based visual analysis system for specific targets, comprising:

[0170] The knowledge graph construction module is used to obtain a basic database from multiple data sources, including entity data and dynamic event data. It uses a streaming text single-pass clustering method to classify the topics of each data source in the basic database, and extracts key information, entities and relationships between entities from each data source. The extraction results from each data source are then integrated to construct a knowledge graph.

[0171] The target visualization module is used to retrieve the knowledge graph for a specific target to be analyzed, obtain the target entity and related events of the specific target; perform personnel analysis and organizational analysis on the target entity to form an entity profile of the target entity; and display the relationship between the target entity and its related entities or between the target entity and the related event in the form of nodes and edges.

[0172] The aforementioned knowledge graph-based specific target visualization analysis system includes a target visualization module where the associated events include time-series event clue reasoning events. Based on the specific target and the knowledge graph, the system uses a time-series event reasoning model based on an event graph neural attention network to predict the trend of the specific target's participation in events, thereby obtaining the time-series event clue reasoning events.

[0173] The temporal event reasoning model includes a clue collection module and a clue reasoning module. The clue collection module is used to recall event clues that are helpful for reasoning query q from an event knowledge graph F constructed from known dynamic events. The clue collection module consists of a random walk collector and a path encoder. The former retrieves several N path clues of length l starting from s in the event knowledge graph based on the given query information and assembles them into a clue set. The path encoder is implemented by a unidirectional LSTM encoder, which encodes the paths collected by the random walk collector into vector information and calculates the weight coefficient of each path in the clue set so that the clue reasoning module can perform the next step of reasoning. The goal of the clue reasoning module is to infer the set of hidden events from the event clue information that is helpful for reasoning query q recalled by the clue collection module. The clue reasoning module consists of a subgraph encoding module and a temporal reasoning module. The subgraph encoding module is composed of a graph neural network, which is used to understand the topological features of the clues at each time step. The temporal reasoning module is composed of a unidirectional recurrent neural network, which is used to learn the law of unidirectional evolution of clues over time and predict the future development of the temporal event to obtain the inference event of the temporal event clue.

[0174] The training process of the time-series event reasoning model in the knowledge graph-based target visualization analysis system includes:

[0175] Initialize the entity vector, relation vector, and time vector, setting them to trainable mode; for the query q = (s, r, t) and the correct entity o, initialize the query representation vector as follows:

[0176] Starting from a specific target entity s, N path clues of length l are sampled through the random walk collector and assembled into a clue set P = {p0, p1, ..., p...} N}, where N and l are user-defined parameters;

[0177] The paths in the clue set are encoded by a path encoder. For a path p = [s,r0,t0,o0,r1,t1,o1,...,r] l-1 ,t l-1 ,o l-1 (where s0 = s), the encoding formula is (0 ≤ k ≤ l-1):

[0178] t k =σ(W r Δt k +U r h k-1 )

[0179]

[0180]

[0181] Δt in the above formula k =t k -t k-1 (When k = 0, Δt) k =0); t k Represents timing threshold control. w represents the semantic information encoded in this unit r W h with U r U h Representing 4 mapping matrices, The k-th step in the path represents the correspondence and entity vector; the output of the LSTM at the k-th step is h. k ;

[0182] The path p corresponds to the output h of step l-1. l-1 This becomes the encoding vector p for the path, from which the encoding vectors {p0, p1, ..., p} for the path clue set are obtained. N};

[0183] Softmax is used to calculate the attention score {α0, α1, ..., α} for each encoded vector with respect to the query. N};

[0184] The global thread representation is obtained by using the encoded vectors of the path thread set.

[0185] Using the clue set P, a query subgraph sequence {SG0,SG1,...,SG} is constructed based on the event time information in the path. t-1};

[0186] The graph neural network uses a subgraph encoding module to encode each subgraph in the subgraph sequence, obtaining the node representation {SG0,SG1,...,SG} in the subgraph. t-1};

[0187] The unidirectional recurrent neural network using this time-series inference module encodes the node representations of the aforementioned subgraph sequence, outputting the subgraph node representation SG at time t. t ;

[0188] Combined with global clues Calculate SG t The matching score for each node in the dataset;

[0189] Select the entity o′ with the highest matching score and the correct entity o, calculate the loss for this training using the binary classification loss function, and train and update the time-series event reasoning model using the backpropagation algorithm.

[0190] The aforementioned knowledge graph-based target visualization and analysis system includes a knowledge graph construction module comprising:

[0191] A multi-source heterogeneous data fusion model is constructed to complete the mapping relationship between entities from different data sources based on the extraction results from different data sources, thereby achieving the fusion of extraction results from various data sources and constructing the knowledge graph.

[0192] The multi-source heterogeneous data fusion model includes an entity representation module and an entity mapping relationship reasoning module;

[0193] The entity representation module encodes entity features based on the entity's textual description, temporal information, and the structural information of its neighbors, serving as the basis for mapping relationship mining. This module consists of an entity textual description encoding module, a temporal information encoding module, and a structural information encoding module. The entity textual description encoding is achieved using a pre-trained language model BERT combined with a fully connected layer of a neural network. Temporal information is encoded using the Time2vec algorithm. For structural information, the entity description and temporal information representations are concatenated as input and encoded using a graph neural network with an attention mechanism. The resulting vector is then used as the final representation of the entity.

[0194] The entity mapping relationship reasoning module is used to infer the mapping relationship between entities based on the final representation of the entities. The entity mapping relationship reasoning module consists of two parts: an entity representation translation module and an entity mapping matching module. The entity representation translation module is composed of a multilayer perceptron (MLP) and is used to map the representations of entities from different data sources to a unified semantic space. The entity mapping matching module first calculates the matching degree of entity representations from different sources using cosine similarity, and then performs entity mapping matching based on the numerical value of the matching degree to finally obtain the mapping relationship between entities, thereby realizing the fusion of multi-source heterogeneous data and obtaining the knowledge graph.

[0195] The present invention also proposes a storage medium for storing programs that execute any of the knowledge graph-based target visualization analysis methods.

[0196] The present invention also proposes a client for any of the knowledge graph-based target visualization and analysis systems.

Claims

1. A specific target visualization analysis method based on knowledge graphs, characterized in that, include: Step S1: Obtain a basic database from multiple data sources, including entity data and dynamic event data. Use the streaming text single-pass clustering method to classify the topics of each data source in the basic database, and extract key information, entity and inter-entity relationship information from each data source. Combine the extraction results from each data source to construct a knowledge graph. Step S2: Retrieve the knowledge graph for the specific target to be analyzed to obtain the target entity and related events of the specific target; perform personnel analysis and organizational analysis on the target entity to form an entity profile of the target entity; display the relationship between the target entity and its related entities or between the target entity and the related events in the form of nodes and edges; In step 2, the associated event includes a time-series event clue reasoning event; based on the specific target and the knowledge graph, the trend of the specific target's participation in the event is predicted by a time-series event reasoning model based on an event graph neural attention network, thus obtaining the time-series event clue reasoning event; The training process for this time-series event reasoning model includes: Initialize the entity vector, relation vector, and time vector, setting them to trainable mode; for the query q = (s, r, t) and the correct entity o, initialize the query representation vector as follows: Starting from a specific target entity s, N path clues of length l are sampled using a random walk collector and assembled into a clue set P = {p0, p1, ..., p...} N }, where N and l are user-defined parameters; The paths in the clue set are encoded by a path encoder. For a path p = [s,r0,t0,o0,r1,t1,o1,...,r] l-1 ,t l-1 ,o l-1 (where s0 = s), the encoding formula is (0 ≤ k ≤ l-1): t k =σ(W r Δt k +U r h k-1 ) Δt in the above formula k =t k -t k-1 (When k = 0, Δt) k =0); t k Represents timing threshold control. W represents the semantic information encoded in this unit. r W h with U r U h Representing 4 mapping matrices, s⊕r k ⊕o k The k-th step in the path represents the correspondence and entity vector; the output of the LSTM at the k-th step is h. k ; The path p corresponds to the output h of step l-1. l-1 This becomes the encoding vector p for the path, from which the encoding vectors {p0, p1, ..., p} for the path clue set are obtained. N }; Softmax is used to calculate the attention score {α0, α1, ..., α} for each encoded vector with respect to the query. N }; The global thread representation is obtained by using the encoded vectors of the path thread set. Using the clue set P, a query subgraph sequence {SG0,SG1,...,SG} is constructed based on the event time information in the path. t-1 }; The graph neural network uses a subgraph encoding module to encode each subgraph in the subgraph sequence, obtaining the node representation {SG0,SG1,...,SG} in the subgraph. t-1 }; The node representations of the above subgraph sequence are encoded using a unidirectional recurrent neural network in the temporal inference module, and the subgraph node representation SG at time t is output. t ; Combined with global clues Calculate SG t The matching score for each node in the dataset; Select the entity o′ with the highest matching score and the correct entity o, calculate the loss for this training using the binary classification loss function, and train and update the time-series event reasoning model using the backpropagation algorithm.

2. The knowledge graph-based specific target visualization analysis method as described in claim 1, characterized in that, The temporal event reasoning model includes a clue collection module and a clue reasoning module. The clue collection module is used to recall event clues that are helpful for reasoning query q from an event knowledge graph F constructed from known dynamic events. The clue collection module consists of a random walk collector and a path encoder. The former retrieves several N path clues of length l starting from s in the event knowledge graph based on the given query information and assembles them into a clue set. The path encoder is implemented by a unidirectional LSTM encoder, which encodes the paths collected by the random walk collector into vector information and calculates the weight coefficient of each path in the clue set so that the clue reasoning module can perform the next step of reasoning. The goal of the clue reasoning module is to infer the set of hidden events from the event clue information that is helpful for reasoning query q recalled by the clue collection module. The clue reasoning module consists of a subgraph encoding module and a temporal reasoning module. The subgraph encoding module is composed of a graph neural network, which is used to understand the topological features of the clues at each time step. The temporal reasoning module is composed of a unidirectional recurrent neural network, which is used to learn the law of unidirectional evolution of clues over time and predict the future development of the temporal event to obtain the inference event of the temporal event clue.

3. The knowledge graph-based specific target visualization analysis method as described in claim 1, characterized in that, Step S1 includes: A multi-source heterogeneous data fusion model is constructed to complete the mapping relationship between entities from different data sources based on the extraction results from different data sources, thereby achieving the fusion of extraction results from various data sources and constructing the knowledge graph. The multi-source heterogeneous data fusion model includes an entity representation module and an entity mapping relationship reasoning module; The entity representation module encodes entity features based on the entity's textual description, temporal information, and the structural information of its neighbors, serving as the basis for mapping relationship mining. This module consists of an entity textual description encoding module, a temporal information encoding module, and a structural information encoding module. The entity textual description encoding is achieved using a pre-trained language model BERT combined with a fully connected layer of a neural network. Temporal information is encoded using the Time2vec algorithm. For structural information, the entity description and temporal information representations are concatenated as input and encoded using a graph neural network with an attention mechanism. The resulting vector is then used as the final representation of the entity. The entity mapping relationship reasoning module is used to infer the mapping relationship between entities based on the final representation of the entities. The entity mapping relationship reasoning module consists of two parts: an entity representation translation module and an entity mapping matching module. The entity representation translation module is composed of a multilayer perceptron and is used to map the representations of entities from different data sources to a unified semantic space. The entity mapping matching module first calculates the matching degree of entity representations from different sources using cosine similarity, and then performs entity mapping matching based on the numerical value of the matching degree to finally obtain the mapping relationship between entities, thereby realizing the fusion of multi-source heterogeneous data and obtaining the knowledge graph.

4. A knowledge graph-based visualization and analysis system for specific targets, characterized in that, include: The knowledge graph construction module is used to obtain a basic database from multiple data sources, including entity data and dynamic event data. It uses a streaming text single-pass clustering method to classify the topics of each data source in the basic database, and extracts key information, entities and relationships between entities from each data source. The extraction results from each data source are then integrated to construct a knowledge graph. The target visualization module is used to retrieve the knowledge graph for a specific target to be analyzed, and obtain the target entities and related events of that specific target. Perform personnel and organizational analysis on the target entity to form an entity profile; display the relationship between the target entity and its related entities or between the target entity and the related event in the form of nodes and edges; In the target visualization module, the associated events include time-series event clue reasoning events. Based on the specific target and the knowledge graph, the trend of the specific target's participation in events is predicted through a time-series event reasoning model based on an event graph neural attention network, thus obtaining the time-series event clue reasoning events. The training process for this time-series event reasoning model includes: Initialize the entity vector, relation vector, and time vector, setting them to trainable mode; for the query q = (s, r, t) and the correct entity o, initialize the query representation vector as follows: Starting from a specific target entity s, N path clues of length l are sampled using a random walk collector and assembled into a clue set P = {p0, p1, ..., p...} N }, where N and l are user-defined parameters; The paths in the clue set are encoded by a path encoder. For a path p = [s,r0,t0,o0,r1,t1,o1,...,r] l-1 ,t l-1 ,o l-1 (where s0 = s), the encoding formula is (0 ≤ k ≤ l-1): t k =σ(W r Δt k +U r h k-1 ) Δt in the above formula k =t k -t k-1 (When k = 0, Δt) k =0); t k Represents timing threshold control. W represents the semantic information encoded in this unit. r W h with U r U h Representing 4 mapping matrices, The k-th step in the path represents the correspondence and entity vector; the output of the LSTM at the k-th step is h. k ; The path p corresponds to the output h of step l-1. l-1 This becomes the encoding vector p for the path, from which the encoding vectors {p0, p1, ..., p} for the path clue set are obtained. N }; Softmax is used to calculate the attention score {α0, α1, ..., α} for each encoded vector with respect to the query. N }; The global thread representation is obtained by using the encoded vectors of the path thread set. Using the clue set P, a query subgraph sequence {SG0,SG1,...,SG} is constructed based on the event time information in the path. t-1 }; The graph neural network uses a subgraph encoding module to encode each subgraph in the subgraph sequence, obtaining the node representation {SG0,SG1,...,SG} in the subgraph. t-1 }; The node representations of the above subgraph sequence are encoded using a unidirectional recurrent neural network in the temporal inference module, and the subgraph node representation SG at time t is output. t ; Combined with global clues Calculate SG t The matching score for each node in the dataset; Select the entity o′ with the highest matching score and the correct entity o, calculate the loss for this training using the binary classification loss function, and train and update the time-series event reasoning model using the backpropagation algorithm.

5. The knowledge graph-based target visualization analysis system as described in claim 4, characterized in that, The temporal event reasoning model includes a clue collection module and a clue reasoning module. The clue collection module is used to recall event clues that are helpful for reasoning query q from an event knowledge graph F constructed from known dynamic events. The clue collection module consists of a random walk collector and a path encoder. The former retrieves several N path clues of length l starting from s in the event knowledge graph based on the given query information and assembles them into a clue set. The path encoder is implemented by a unidirectional LSTM encoder, which encodes the paths collected by the random walk collector into vector information and calculates the weight coefficient of each path in the clue set so that the clue reasoning module can perform the next step of reasoning. The goal of the clue reasoning module is to infer the set of hidden events from the event clue information that is helpful for reasoning query q recalled by the clue collection module. The clue reasoning module consists of a subgraph encoding module and a temporal reasoning module. The subgraph encoding module is composed of a graph neural network, which is used to understand the topological features of the clues at each time step. The temporal reasoning module is composed of a unidirectional recurrent neural network, which is used to learn the law of unidirectional evolution of clues over time and predict the future development of the temporal event to obtain the inference event of the temporal event clue.

6. The knowledge graph-based target visualization analysis system as described in claim 4, characterized in that, The knowledge graph construction module includes: A multi-source heterogeneous data fusion model is constructed to complete the mapping relationship between entities from different data sources based on the extraction results from different data sources, thereby achieving the fusion of extraction results from various data sources and constructing the knowledge graph. The multi-source heterogeneous data fusion model includes an entity representation module and an entity mapping relationship reasoning module; The entity representation module encodes entity features based on the entity's textual description, temporal information, and the structural information of its neighbors, serving as the basis for mapping relationship mining. This module consists of an entity textual description encoding module, a temporal information encoding module, and a structural information encoding module. The entity textual description encoding is achieved using a pre-trained language model BERT combined with a fully connected layer of a neural network. Temporal information is encoded using the Time2vec algorithm. For structural information, the entity description and temporal information representations are concatenated as input and encoded using a graph neural network with an attention mechanism. The resulting vector is then used as the final representation of the entity. The entity mapping relationship reasoning module is used to infer the mapping relationship between entities based on the final representation of the entities. The entity mapping relationship reasoning module consists of two parts: an entity representation translation module and an entity mapping matching module. The entity representation translation module is composed of a multilayer perceptron and is used to map the representations of entities from different data sources to a unified semantic space. The entity mapping matching module first calculates the matching degree of entity representations from different sources using cosine similarity, and then performs entity mapping matching based on the numerical value of the matching degree to finally obtain the mapping relationship between entities, thereby realizing the fusion of multi-source heterogeneous data and obtaining the knowledge graph.

7. A storage medium for storing a program that executes any one of the knowledge graph-based target visualization analysis methods as described in claims 1 to 3.

8. A client for use in any one of the knowledge graph-based target visualization and analysis systems of claims 4 to 6.

Images