Method and system for evaluating holographic portraits and dynamic risks of concerned persons

Abstract

The invention provides a method and system for evaluating holographic portraits and dynamic risks of concerned persons, and belongs to the technical field of intelligent information processing, and the method comprises the steps: carrying out the data fusion and entity alignment of to-be-analyzed data through an intelligence analysis model based on a large language model in response to an analysis request, and constructing a knowledge graph; the intelligence analysis model generates a standard holographic data packet according to the first preset prompt word and the knowledge graph so as to generate a holographic portrait of the concerned person; performing multi-factor analysis on the knowledge graph to determine a dynamic risk index; and continuously monitoring the dynamic risk index, and when the dynamic risk index exceeds a preset risk threshold, generating and pushing early warning information. The method has the beneficial effects that deep fusion and entity alignment of multi-source heterogeneous data are realized through the semantic understanding ability of a large language model, automatic generation of holographic archives of concerned personnel is realized, and dynamic, objective and interpretable risk assessment is provided, so that the intelligent level of public safety intelligence work is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent information processing technology, and in particular to a method and system for creating holographic profiles and dynamic risk assessments of individuals. Background Technology

[0002] Public safety is the cornerstone of social stability and development. The integrated operation of public safety intelligence, command, and action (hereinafter referred to as intelligence, command, and action) plays a crucial role in maintaining public safety. Among these operations, the effective management of individuals of interest is a core component of this integrated operation. Accurate and efficient processing and analysis of information on individuals of interest are of critical importance for the timely detection of potential security threats, the prevention of criminal activity, and the maintenance of social order.

[0003] However, existing technologies face a series of serious technical challenges when processing information about individuals of interest:

[0004] First, data integration faces technical barriers, leading to fragmented information and the formation of data silos. Information about individuals of interest, such as past criminal records, accommodation registrations, travel history, communications, interrogation transcripts, and online statements, is typically stored across multiple independent, heterogeneous business databases. For law enforcement officers to gain a comprehensive understanding of an individual of interest, they must manually log into multiple systems to query, compare, and summarize information. This process is not only time-consuming and inefficient but also highly susceptible to human error, resulting in omissions or errors and hindering a complete and unified understanding of the target. Furthermore, current technologies for multi-source data fusion often rely on simple rule matching, such as matching only ID numbers, which fails to effectively address the entity alignment issues of "same name, different person" or "different name, same person" (such as nicknames or former names).

[0005] Secondly, existing analytical methods struggle to uncover deep-level relationships. Traditional analytical systems largely rely on clearly documented direct connections, such as known family members or accomplices. For deeper information hidden within massive datasets, such as strong potential connections, unusual social circles, and behavioral patterns, current technologies fall short. For instance, while some systems can perform "fellowship analysis," they often rely on simple spatiotemporal co-occurrence rules, lacking a deep semantic understanding of relationship strength and nature, making it difficult to effectively identify concealed, indirect gang structures or key contacts. Furthermore, simple graph database queries can only discover direct connections (first-degree connections), failing to uncover hidden networks spanning multiple intermediate nodes, such as indirect relationships like money laundering through non-immediate family members.

[0006] Finally, risk assessment models are subjective and static, lacking scientific rigor and dynamism. Currently, the assessment of the actual risks to individuals of concern largely relies on the personal experience and subjective judgment of frontline investigators. This model leads to inconsistent assessment standards, difficulty in reproducing and quantifying assessment results, and an inability to continuously and dynamically track risks. When the behavioral patterns or social relationships of target individuals change, static assessment conclusions cannot be updated in a timely manner, potentially leading to missed early warning opportunities or unnecessary waste of resources. Although some existing automated systems have proposed the concept of "data analysis models," they are essentially more inclined towards the automated execution of preset rules rather than dynamic and adaptive risk reasoning based on real-time data streams. Other systems primarily focus on processing police voice data, a solution for single-modal data, failing to address the comprehensive risk assessment problem of cross-modal, multi-source, heterogeneous data.

[0007] In conclusion, these challenges significantly restrict the efficiency, depth, and accuracy of intelligence analysis, thereby affecting the implementation of public safety work. Summary of the Invention

[0008] To address the above technical problems, this invention provides a method for creating a holographic profile and dynamic risk assessment of a person under observation; furthermore, it also provides a system for creating a holographic profile and dynamic risk assessment of a person under observation.

[0009] The technical problem solved by this invention can be achieved by the following technical solutions:

[0010] A method for holographic profiling and dynamic risk assessment of personnel includes:

[0011] Step S1: In response to the analysis request, use the intelligence analysis model based on the large language model to perform data fusion and entity alignment on the data to be analyzed, and construct a knowledge graph;

[0012] Step S2: The intelligence analysis model generates a standard holographic data package based on the first preset prompt word and the knowledge graph, so as to generate a holographic portrait of the person of interest based on the standard holographic data package;

[0013] Step S3: Perform multi-factor analysis on the knowledge graph and determine the dynamic risk index based on the results of the multi-factor analysis;

[0014] Step S4: Continuously monitor the dynamic risk index, and generate and push early warning information when the dynamic risk index exceeds the preset risk threshold.

[0015] Preferably, step S1 includes:

[0016] Step S11: In response to the analysis request, collect the data to be analyzed from a preset heterogeneous data source;

[0017] Step S12: Based on the intelligence analysis model, extract information from the unstructured text in the data to be analyzed to obtain multiple entities and the key feature vectors of the corresponding entities.

[0018] Step S13: Calculate the cosine similarity of the key feature vectors of the entity to obtain the first similarity metric.

[0019] Step S14: Calculate the edit distance of the entity's string to obtain a second similarity metric.

[0020] Step S15: The first similarity metric and the second similarity metric are weighted and fused to obtain a comprehensive fused similarity metric.

[0021] Step S16: When the comprehensive fusion similarity metric exceeds the preset similarity threshold, it is determined to be the same entity and fused, and the fused entity is stored in the graph database to construct the knowledge graph.

[0022] Preferably, step S14 includes:

[0023] Step S141: Preprocess the entity, the preprocessing including at least data cleaning, to obtain a preprocessed entity;

[0024] Step S142: Calculate the edit distance between the strings of the two preprocessed entities to be compared to obtain the first edit distance;

[0025] Step S143: Normalize the first edit distance to obtain the second edit distance;

[0026] Step S144: The second edit distance of the entity and the matching weight of at least one auxiliary attribute entity corresponding to the entity are weighted to obtain the second similarity metric.

[0027] Preferably, step S2 includes:

[0028] Step S21: Construct the first preset prompt word, wherein the first preset prompt word includes at least one of role definition, fact constraint and output format specification;

[0029] Step S22: The intelligence analysis model generates a standard holographic data packet in a predetermined format based on the first preset prompt word and the knowledge graph;

[0030] Step S23: The front-end portrait rendering engine parses the standard holographic data package and generates a holographic portrait of the person of interest. The holographic portrait includes at least one of the following: a biographical sketch, a visual relationship graph, and a dynamic activity timeline.

[0031] Preferably, step S3 includes:

[0032] Step S31: Perform multi-factor analysis on the knowledge graph to obtain multi-factor analysis results. The multi-factors include at least one of static resume factors, behavioral dynamic factors, social network factors, and speech characteristic factors.

[0033] Step S32: Perform a weighted summation based on the results of the multi-factor analysis to determine the dynamic risk index.

[0034] Preferably, in step S31, the static resume factor is calculated as follows:

[0035]

[0036] Where W1 represents the weight corresponding to the previous level of involvement; S level W1 represents the normalized level of prior criminal record; W2 represents the weight corresponding to the number of criminal records; S count The normalized number of criminal records is represented by S; the static resume factor is represented by S.

[0037] The calculation method for the behavioral dynamic factor is as follows:

[0038]

[0039] Where k1 represents the weight corresponding to the recent abnormal travel frequency; F freq Indicates recent abnormal travel frequency; k2 represents the weight corresponding to the proportion of nighttime activities; R night B represents the percentage of nighttime activities; B represents the behavioral dynamic factor.

[0040] The calculation method for the social network factor is as follows:

[0041]

[0042] Where, N PR Indicates the centrality of individuals within a criminal network; N Path N represents the shortest path distance between individuals on the watchlist and those on the restricted list; N represents the social network factor.

[0043] The method for calculating the speech characteristic factors is as follows:

[0044]

[0045] Among them, Score senti Score represents the sentiment analysis index of the intelligence analysis model for the statements of the individuals of interest; α represents the weight of the sentiment analysis index; keywordC represents the total score obtained by matching all sensitive words in the statements of the person under observation from the graded sensitive word database; C represents the characteristic factor of the statement.

[0046] Preferably, the method for calculating the normalized prior case involvement level is as follows:

[0047]

[0048] Here, L indicates the level of prior criminal record of the person being monitored; L max Indicates the highest level of involvement in the case;

[0049] The method for calculating the normalized number of criminal records is as follows:

[0050]

[0051] Where n represents the total number of historical criminal records;

[0052] The calculation method for the recent abnormal travel frequency is as follows:

[0053]

[0054] Among them, C sens This indicates the number of times the personnel in question visited the target area within the first preset time period; T threshold Indicates the preset number of times threshold;

[0055] The method for calculating the proportion of nighttime activities is as follows:

[0056]

[0057] Where, N night N represents the amount of data collected by the monitored individuals within the second preset time period. total This indicates the total number of activities of the monitored individuals within the first preset time period;

[0058] The method for calculating the centrality of the individuals under surveillance within the criminal network is as follows:

[0059] A local social relationship subgraph is constructed based on the people being followed and their corresponding second-degree connections. The local social relationship subgraph includes nodes and edges. The edges represent the association between two nodes, and the weight of the edges is determined based on the interaction frequency between the nodes.

[0060] Calculate the page ranking value for each node;

[0061] The centrality of the person being monitored in the criminal network is obtained by comparing the webpage ranking value of the monitored person with the maximum webpage ranking value among all nodes in the local social relationship subgraph. The corresponding calculation formula is as follows:

[0062]

[0063] Among them, PR target PageRank (PR) represents the ranking value of a webpage based on the number of people following it. max This represents the maximum value among all the webpage ranking values ​​of all nodes in the local social relationship subgraph;

[0064] The method for calculating the shortest path distance between the individuals on the watch list and those on the restricted list is as follows:

[0065] Calculate the set of shortest hop counts between the individuals of interest and the nodes marked as restricted in the knowledge graph;

[0066] Based on the minimum value in the set of shortest hop counts, the shortest path distance between the monitored individuals and those on the restricted list is calculated using the distance decay formula. The corresponding calculation formula is as follows:

[0067]

[0068] Where, d min This represents the minimum value in the set of shortest hop counts.

[0069] Preferably, the early warning information includes an explanation of the causes of the risk, and the method for obtaining the explanation of the causes of the risk includes:

[0070] Step S41: When the dynamic risk index exceeds the preset risk threshold, extract the corresponding original fact fragments from the data to be analyzed based on the top k factors that contribute the most to the dynamic risk index, where k is a positive integer greater than 1.

[0071] Step S42: Input the top k factors with the highest contribution and their corresponding original fact fragments into the intelligence analysis model. The intelligence analysis model generates risk cause explanations based on the second preset prompt words, so as to generate early warning information containing the risk cause explanations.

[0072] Preferably, after step S4, the method further includes:

[0073] Receive user questions and transform them into question vectors;

[0074] Based on the question vector, a knowledge graph slice corresponding to the question vector is retrieved.

[0075] The question vector and the corresponding map slice are input into the intelligence analysis model to generate fact-based question answers.

[0076] On the other hand, a system for holographic profiling and dynamic risk assessment of individuals under surveillance is provided for implementing the aforementioned method for holographic profiling and dynamic risk assessment of individuals under surveillance, including:

[0077] The heterogeneous data fusion module is used to respond to analysis requests and use an intelligence analysis model based on a large language model to perform data fusion and entity alignment on the data to be analyzed, and to build a knowledge graph.

[0078] The graph storage engine, connected to the heterogeneous data fusion module, is used to store the knowledge graph based on a graph database;

[0079] A risk calculation engine, connected to the knowledge graph storage engine, is used to perform multi-factor analysis on the knowledge graph and determine a dynamic risk index based on the results of the multi-factor analysis.

[0080] The intelligence analysis model is also used to generate standard holographic data packets based on the first preset prompt words and the knowledge graph;

[0081] A front-end portrait rendering engine, connected to the graph storage engine, is used to generate holographic portraits of the individuals of interest based on the standard holographic data package.

[0082] The interactive terminal is connected to the risk calculation engine and the front-end portrait rendering engine respectively, and is used to continuously monitor the dynamic risk index. When the dynamic risk index exceeds a preset risk threshold, it generates and pushes early warning information.

[0083] The advantages or beneficial effects of the technical solution of this invention are as follows:

[0084] This invention leverages the semantic understanding capabilities of a large language model to achieve deep fusion and entity alignment of multi-source heterogeneous data, enabling the automated generation of holographic profiles of individuals of interest. Simultaneously, through multi-factor analysis of knowledge graphs, it achieves quantitative calculation of dynamic risk indices and provides dynamic, objective, and interpretable risk assessments, thereby enhancing the intelligence level of public safety intelligence work. Attached Figure Description

[0085] Figure 1 This is a flowchart illustrating the method for creating a holographic profile of a person and conducting a dynamic risk assessment, which is a preferred embodiment of the present invention.

[0086] Figure 2 This is a flowchart illustrating step S1 in a preferred embodiment of the present invention.

[0087] Figure 3 This is a flowchart illustrating step S14 in a preferred embodiment of the present invention.

[0088] Figure 4 This is a flowchart illustrating step S2 in a preferred embodiment of the present invention.

[0089] Figure 5 This is a flowchart illustrating step S3 in a preferred embodiment of the present invention.

[0090] Figure 6 This is a flowchart illustrating step S4 in a preferred embodiment of the present invention.

[0091] Figure 7 This is a schematic diagram of the process of enhancing the generation of question answers in a preferred embodiment of the present invention;

[0092] Figure 8 This is a schematic diagram of the overall process of the holographic portrait and dynamic risk assessment method for monitoring personnel, as shown in a preferred embodiment of the present invention.

[0093] Figure 9 This is a schematic diagram of the architecture of the holographic profile and dynamic risk assessment system for monitored personnel, which is a preferred embodiment of the present invention. Detailed Implementation

[0094] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0095] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0096] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0097] The fundamental purpose of this invention is to reshape the workflow of intelligence analysis by introducing large language model technology, transforming the traditional intelligence processing mode, which is mainly manual and passive, into a new intelligent paradigm of human-machine collaboration and proactive early warning. This greatly improves the efficiency, depth, and foresight of intelligence work, and provides strong technical support for the integrated intelligence, command, and action in the field of public security.

[0098] In a preferred embodiment of the present invention, based on the problems of low data fusion efficiency, insufficient relationship mining depth, and subjective and static risk assessment in the field of intelligence analysis of persons of concern, a method and system for holographic profiling and dynamic risk assessment of persons of concern is provided. This method and system are applied to the field of public security intelligence analysis. It utilizes a domain-adjusted large language model (LLM) to perform deep fusion and intelligent analysis of multi-source heterogeneous data, thereby realizing the automated generation of holographic profiles of persons of concern and providing dynamic, objective, and interpretable risk assessments to improve the intelligence level of public security intelligence work.

[0099] like Figure 1As shown, the method includes the following steps:

[0100] Step S1: In response to the analysis request, use the intelligence analysis model based on the large language model to perform data fusion and entity alignment on the data to be analyzed, and construct a knowledge graph;

[0101] Specifically, in response to the problems of data silos and the inadequacy of conventional fusion methods, this embodiment adopts an entity alignment method based on a mixture of semantic vectors and rules, and combines the semantic understanding capabilities of a large language model to achieve deep cleaning and fusion of multi-source heterogeneous data.

[0102] Furthermore, such as Figure 2 As shown, step S1 includes:

[0103] Step S11: In response to the analysis request, collect the data to be analyzed from a preset heterogeneous data source;

[0104] Specifically, such as Figure 8 As shown, in the multi-source data fusion and intelligent entity alignment stage 100, in response to the analysis request, the data to be analyzed is collected from the heterogeneous data sources preset in the relevant department's intranet, such as the previous case database, accommodation database, travel database, communication database, and online opinion database.

[0105] Step S12: Based on the intelligence analysis model, information is extracted from the unstructured text in the data to be analyzed to obtain multiple entities and their corresponding key feature vectors.

[0106] Specifically, an intelligence analysis model based on a large language model is used to extract information from unstructured text in the data to be analyzed.

[0107] Existing generalized large language models (LLMs) lack domain logic. If directly applied to specialized analysis, these general models struggle to understand specific business logic such as accomplices and past case connections, potentially leading to misinterpretations. Furthermore, the construction steps for an intelligence analysis model include: fine-tuning the large language model based on police case records and intelligence reports to obtain the intelligence analysis model.

[0108] This invention uses case records and intelligence reports from the police field to fine-tune a large language model, enabling the intelligence analysis model to extract specific elements related to the case.

[0109] After extracting the relevant entities such as personnel, relationships, and events, entity alignment and fusion are performed. In this embodiment, a hybrid alignment strategy is adopted. Specifically, a hybrid alignment strategy based on cosine similarity and edit distance is used. Cosine similarity is used to capture semantic-level associations; edit distance is used to capture the closeness of literal spelling.

[0110] Step S13: Calculate the cosine similarity of the key feature vectors of the entities to obtain the first similarity metric.

[0111] Specifically, firstly, cosine similarity is calculated using the key feature vectors of entities extracted by the intelligence analysis model to obtain the first similarity metric. The first similarity metric is the cosine of the angle between the key feature vectors of the two entities.

[0112] Step S14: Calculate the edit distance of the entity's string to obtain the second similarity metric.

[0113] Furthermore, such as Figure 3 As shown, step S14 includes:

[0114] Step S141: Preprocess the entity, including at least data cleaning, to obtain a preprocessed entity;

[0115] During the text preprocessing stage, the collected text fields such as personnel names, former names, and nicknames are cleaned to remove special symbols and spaces.

[0116] Step S142: Calculate the edit distance between the strings of the two preprocessed entities to be compared to obtain the first edit distance;

[0117] Specifically, for the first string S of the two strings to be compared... a Second string S b First, the standard edit distance (Levenshtein) algorithm is applied to calculate the first edit distance between the two strings. The first edit distance is the distance between the first string S and the second string S. a Convert to the second string S b The minimum number of single-character editing operations required.

[0118] Step S143: Normalize the first edit distance to obtain the second edit distance;

[0119] Subsequently, the first edit distance is normalized and transformed into a second edit distance between 0 and 1. Specifically, the second edit distance is calculated as follows: First, determine the first string S. a The length of the first string S b The maximum value in the length; then, calculate the first edit distance LD(S). a ,S b The second edit distance is obtained by dividing the length of the first string Sa by the ratio of the maximum length of the first string Sb to the length of the second string Sb. Subtracting this ratio from 1 gives the second edit distance, calculated as follows:

[0120]

[0121] Among them, LD(S) a ,S b ) represents the first string S to be compared. a Second string S b The first edit distance; len(S a ) represents the first string S a Length; len(S) b ) represents the first string S b Length; Sim edit Indicates the second edit distance.

[0122] Step S144: The second edit distance of the entity and the entity matching weight of at least one auxiliary attribute corresponding to the entity are weighted to obtain the second similarity metric.

[0123] Specifically, this embodiment does not simply use the standard edit distance algorithm, but rather employs an improved edit distance algorithm. Specifically, for entities such as names, in addition to calculating their normalized edit distance, auxiliary attributes such as the last four digits of the ID number and the first three digits of the mobile phone number are introduced. Through the precise matching weights of these auxiliary attributes, a second similarity metric is obtained, forming a comprehensive edit similarity. By organically combining multiple complementary dimensions, a multi-dimensional joint similarity metric is achieved, avoiding the one-sidedness and limitations of a single-dimensional metric.

[0124] Step S15: Weighted fusion of the first similarity measure and the second similarity measure to obtain a comprehensive fused similarity measure;

[0125] Specifically, in this embodiment, a strategy of fusing vector semantics and literal features is adopted, and the formula for calculating the comprehensive similarity metric is as follows:

[0126]

[0127] Where ω1 represents the weight corresponding to the first similarity metric; Sim vec The first similarity metric is the cosine similarity calculated based on feature vectors extracted from a large model. It is used to capture semantic-level associations, such as the consistency of reference between "Zhang San" and "Zhang Moumou" in the description of a specific case.

[0128] ω2 represents the weight corresponding to the second similarity metric; Sim edit Indicates the second similarity measure; Sim final This represents the overall similarity metric, ω1+ω2=1.

[0129] For example, ω1=0.6, ω2=0.4. Of course, this is just an example, and the present invention does not limit it.

[0130] Step S16: When the comprehensive fusion similarity metric exceeds the preset similarity threshold, it is determined to be the same entity and fused, and the fused entity is stored in the graph database to construct a knowledge graph.

[0131] In this embodiment, the preset similarity threshold can be set to 0.85. Of course, in other embodiments, a specific threshold value can be set as needed.

[0132] Two entities are identified as the same person only when the overall similarity metric exceeds a preset similarity threshold, and data fusion is performed to avoid misjudgment caused by relying on a single indicator.

[0133] Step S2: The intelligence analysis model generates a standard holographic data package based on the first preset prompt words and knowledge graph, and generates a holographic profile of the person of interest based on the standard holographic data package.

[0134] Furthermore, such as Figure 4 As shown, step S2 includes:

[0135] Step S21: Construct a first preset prompt word, which includes at least one of role definition, factual constraints, and output format specifications;

[0136] Step S22: The intelligence analysis model generates a standard holographic data packet in a predetermined format based on the first preset prompt and knowledge graph;

[0137] Step S23: The front-end portrait rendering engine parses the standard holographic data package and generates a holographic portrait of the person of interest. The holographic portrait includes at least one of the following: a biographical sketch, a visual relationship graph, and a dynamic activity timeline.

[0138] In the holographic portrait structuring generation stage 200, based on data from the graph database, structured prompts containing character definitions, factual constraints, and output format specifications are constructed. An intelligence analysis model is invoked to generate a standard holographic data package in JSON format based on these prompts. The front-end portrait rendering engine parses this JSON standard holographic data package and generates:

[0139] (a) Biographical sketches: Narrative texts generated by an intelligence analysis model based on a large language model, following chronological logic;

[0140] (b) Visualizing the Relationship Graph: Node and Edge data in the JSON standard holographic data package are parsed and drawn by the graph rendering engine; during this process, the intelligence analysis model is also used to identify core relationships and filter out irrelevant nodes;

[0141] (c) Dynamic Activity Timeline: Parses timestamp events in the JSON standard holographic data package and renders them according to the timeline component.

[0142] Step S3: Perform multi-factor analysis on the knowledge graph and determine the dynamic risk index based on the results of the multi-factor analysis;

[0143] Specifically, while general-purpose LLM models excel at generating text, they lack quantitative assessment standards and cannot directly output precise, quantifiable risk values, thus failing to trigger automated warnings. In this embodiment, the system incorporates a risk calculation engine that calculates a dynamic risk index R based on data from a graph database, enabling quantitative risk calculation.

[0144] Furthermore, such as Figure 5 As shown, step S3 includes: step S31, performing multi-factor analysis on the knowledge graph to obtain multi-factor analysis results, wherein the multi-factors include at least one of static resume factors, behavioral dynamic factors, social network factors, and speech characteristic factors; step S32, performing weighted summation based on the multi-factor analysis results to determine the dynamic risk index.

[0145] In the multi-factor quantitative risk assessment phase 300, the dynamic risk index R is calculated using the following formula:

[0146] R = λ*S + β*B + γ*N + δ*C

[0147] Where S represents the static resume factor; B represents the behavioral dynamic factor; N represents the social network factor; C represents the speech characteristic factor; λ, β, γ, and δ represent the weight coefficients corresponding to each factor, and λ+β+γ+δ=1.

[0148] The static resume factor S is derived by normalizing the level of past offenses and the number of criminal records. Furthermore, the calculation method for the static resume factor is as follows:

[0149]

[0150] Where W1 represents the weight corresponding to the previous level of involvement; S level W1 represents the normalized level of prior criminal record; W2 represents the weight corresponding to the number of criminal records, W1+W2=1, for example, W1=0.7, W2=0.3; S count This represents the number of criminal records after normalization; S represents the static resume factor.

[0151] The specific normalization calculation process is as follows:

[0152] The system has a pre-set crime severity level mapping table, which contains the severity levels corresponding to past crimes, for example: administrative penalty = 1, minor criminal offense = 3, serious criminal offense = 5. First, based on the past crimes of the target individuals in the knowledge graph, the past crimes are mapped to numerical levels, resulting in the past crime severity level L for the individuals. Then, the past crime severity level is normalized using the following formula:

[0153]

[0154] Among them, L max L indicates the highest level of involvement in the case. max =5.

[0155] To avoid numerical overflow due to an excessive number of criminal records, a variant of the Sigmoid function is used for mapping. Let n be the total number of historical criminal records, and the calculation formula is:

[0156]

[0157] Ensure that when n=0, the normalized number of criminal records S count The normalized number of criminal records S is 0; as the total number of historical criminal records n increases, the normalized number of criminal records S also increases. count The numerical smoothness approaches 1.

[0158] Furthermore, the behavioral dynamic factor B is calculated based on the recent frequency of abnormal travel and the proportion of nighttime activity. The calculation method for the behavioral dynamic factor B is as follows:

[0159]

[0160] Where k1 represents the weight corresponding to the recent abnormal travel frequency; F freq Indicates recent abnormal travel frequency; k2 represents the weight corresponding to the proportion of nighttime activities; R night B represents the percentage of nighttime activities; B represents the behavioral dynamic factor.

[0161] The specific data acquisition and calculation steps are as follows:

[0162] Retrieve from the travel database the number of times the target population traveled to the target area within a first preset time period (C). sens The calculation method for recent abnormal travel frequency is as follows:

[0163]

[0164] Among them, C sens This indicates the number of times the personnel in question visited the target area within the first preset time period; T threshold Indicates the preset number of times threshold;

[0165] In this embodiment, the first preset time period can be set as needed, such as the past 7 days, 30 days, 60 days, 90 days, or more. The target area can be key sensitive areas such as border areas or areas with high incidence of drug-related crimes, defined by a built-in list in the system. The preset threshold for the number of occurrences can be set to 3, meaning that exceeding 3 occurrences is considered a maximum risk score.

[0166] Simultaneously, all perceived data from the monitored individuals within the first preset time period will be collected, including but not limited to data on accommodation registration, internet access at internet cafes, and data captured by checkpoints, totaling N. total The occurrence time is within a second preset time period, such as from 22:00 to 05:00 the next day, and the amount of data between these periods is N. night The calculation method for the proportion of nighttime activities is as follows:

[0167]

[0168] Where, N night N represents the amount of data collected by the monitored individuals within the second preset time period. total This indicates the total number of activities of the monitored individuals within the first preset time period.

[0169] Furthermore, the social network factor N (Network) is the maximum of the centrality of the person under surveillance within the criminal network and the shortest path distance between the person under surveillance and those on the restricted list. The calculation method for the social network factor N is as follows:

[0170]

[0171] Where, N PR Indicates the centrality of individuals within a criminal network; N Path This represents the shortest path distance between individuals on the watchlist and those on the restricted list; N represents the social network factor.

[0172] The centrality N of individuals of interest in a criminal network is calculated using the PageRank graph algorithm. PR Specifically, a local social relationship subgraph is constructed based on the target audience and their corresponding second-degree connections. This subgraph includes nodes and edges; nodes are considered individuals, and edges represent the relationships between two nodes. The weight of each edge is determined based on the frequency of interaction between nodes. A weighted PageRank algorithm is then run to iteratively calculate the PageRank (PR) value for each node. Finally, the PageRank values ​​of the target audience are used to determine the final ranking. target The maximum PageRank (PR) among all nodes in the local social relationship subgraph max The ratio of these ratios yields the centrality of the individuals of interest within the criminal network; the corresponding calculation formula is:

[0173]

[0174] Among them, PR target PageRank (PR) represents the ranking value of a webpage based on the number of people following it. max This represents the maximum webpage ranking value among all nodes in the local social relationship subgraph.

[0175] Using the shortest path algorithm for graph databases, calculate the shortest hop count set {d1, d2, ...} between the target user node and nodes marked as restricted (fugitive / priority) in the knowledge graph; then take the minimum value from the shortest hop count set {d1, d2, ...}.

[0176]

[0177] Based on the minimum value in the set of shortest hop counts, the shortest path distance between the monitored individuals and those on the restricted list is calculated using the distance decay formula. The corresponding formula is as follows:

[0178]

[0179] Where, d min This represents the minimum value in the set of shortest hop counts.

[0180] If there is no connecting path, then the shortest path distance N between the monitored personnel and the restricted list personnel is... Path =0. The closer the distance, the smaller the minimum value d in the set of shortest hops. min The smaller the value, the higher the risk.

[0181] Furthermore, the content feature factor C is obtained by performing sentiment analysis and weighting of sensitive words on online speech. The calculation method for the content feature factor C is as follows:

[0182]

[0183] Among them, Score senti Score represents the sentiment analysis index of the intelligence analysis model for the statements of the individuals of interest; α represents the weight of the sentiment analysis index; keyword C represents the total score of all sensitive words in the statements of the person being monitored, obtained by matching them from the graded sensitive word database; C represents the statement feature factor.

[0184] Using a large language model, we infer the recent public online statements of the target audience and output a sentiment analysis index ranging from 0 to 1, where 0 represents positive / neutral and 1 represents extremely negative / violent.

[0185] The system has a pre-built tiered sensitive word library, which includes multiple tiers of sensitive words and the weight of each tier. For example, a first-level word, such as "explosion," has a weight of 50; a second-level word, such as "retaliation" or "purchase," has a weight of 10. In this embodiment, the system performs word segmentation matching on the recent public online statements of the target user, and accumulates the weights of all matched words to obtain the total score. keyword .

[0186] Step S4: Continuously monitor the dynamic risk index, and generate and push early warning information when the dynamic risk index exceeds the preset risk threshold.

[0187] Furthermore, the early warning information includes explanations of the causes of the risk, such as... Figure 6 As shown, the methods for obtaining explanations of risk causes include:

[0188] Step S41: When the dynamic risk index exceeds the preset risk threshold, extract the corresponding original fact fragments from the data to be analyzed based on the top k factors that contribute the most to the dynamic risk index, where k is a positive integer greater than 1.

[0189] Step S42: Input the top k factors with the highest contribution and their corresponding original fact fragments into the intelligence analysis model. The intelligence analysis model generates risk cause explanations based on the second preset prompt words, so as to generate early warning information containing risk cause explanations.

[0190] Specifically, in this embodiment, during the interactive early warning and retrieval enhancement question-and-answer generation stage 400, after calculating the dynamic risk index R, the system inputs the top k risk factors with the highest contribution and their corresponding original fact fragments into a dedicated large model. Through a chain of thought prompting strategy, it generates a natural language explanation of the risk's causes. For example, the risk cause explanation could be "The surge in the risk index is mainly caused by factor B, specifically manifested as...", thus achieving a qualitative explanation of the risk.

[0191] The system backend continuously monitors the dynamic risk index R. Once the dynamic risk index R exceeds the preset risk threshold R... TH1 Or the change in the dynamic risk index R, ∆R, exceeds the preset risk change threshold R. TH2 Immediately push out a warning message containing the model's explanation.

[0192] To address the issue of subjectivity in risk assessment, this invention constructs a dual-modal assessment system that combines quantitative calculation and qualitative interpretation, clarifies the risk index calculation formula, and achieves a unity between objective quantification and subjective interpretation.

[0193] To address the issues of complex models and insufficient versatility, this invention employs a single, finely tuned intelligence analysis model based on a large language model. Through different prompting strategies, it uniformly executes information extraction, profile generation, and risk interpretation tasks, thereby reducing system complexity and improving the accuracy of business logic.

[0194] Furthermore, such as Figure 7 As shown, after step S4, the following steps are also included:

[0195] S5 receives user questions and transforms them into question vectors.

[0196] S6. Based on the question vector, retrieve the knowledge graph slice corresponding to the question vector.

[0197] S7 inputs the question vector and the corresponding map slice into the intelligence analysis model to generate fact-based question answers.

[0198] Specifically, in this embodiment, a natural language question-and-answer interface can also be provided. Using retrieval augmented generation (RAG) technology, user questions are converted into vectors, relevant graph slices are retrieved from the vector database, and the questions are input into a dedicated large model, which generates fact-based answers.

[0199] This invention provides a system for creating holographic profiles and dynamic risk assessments of individuals of interest, used to implement the aforementioned method for creating holographic profiles and dynamic risk assessments of individuals of interest. Figure 9 As shown, it includes:

[0200] The heterogeneous data fusion module 10 is used to respond to analysis requests and use an intelligence analysis model based on a large language model to perform data fusion and entity alignment on the data to be analyzed, and to construct a knowledge graph.

[0201] The graph storage engine 20 is connected to the heterogeneous data fusion module 10 and is used to store knowledge graphs based on graph databases.

[0202] The risk calculation engine 40 is connected to the knowledge graph storage engine 20, which is used to perform multi-factor analysis on the knowledge graph and determine the dynamic risk index based on the results of the multi-factor analysis.

[0203] The intelligence analysis model is also used to generate standard holographic data packets based on the first preset prompt words and knowledge graph;

[0204] The front-end portrait rendering engine 30 and the connected graph storage engine 20 are used to generate holographic portraits of the people of interest based on standard holographic data packages.

[0205] The interactive terminal 50 is connected to the risk calculation engine 40 and the front-end portrait rendering engine 30 respectively. It is used to continuously monitor the dynamic risk index and generate and push early warning information when the dynamic risk index exceeds the preset risk threshold.

[0206] Specifically, the heterogeneous data fusion module 10 connects to an external data source layer. External data sources include, but are not limited to, existing case databases, accommodation databases, travel databases, communication databases, and online opinion databases. The heterogeneous data fusion module 10 incorporates model interfaces, database connection interfaces, and entity alignment algorithms. It employs a Large Language Model (LLM) to drive the processing pipeline, based on Named Entity Recognition (NER), Relation Extraction (RE), and Event Extraction (Event). The system integrates Extraction (EE) and entity alignment to achieve cleaning, extraction, and graph construction of multi-source data; the graph storage engine 20 uses a graph database to store knowledge graphs, storing entities such as people, relationships, and events; the front-end profiling and rendering engine 30 includes a JSON parser and front-end visualization plugins, such as ECharts or D3.js, to convert the structured data output by the model into charts, rather than relying on the model to directly draw graphs; the risk calculation engine 40 has a built-in graph algorithm library (for calculating social network factor N) and a risk formula calculator; the interactive terminal 50 provides a natural language processing interface, using Retrieval Augmentation (RAG) technology to convert user questions into vectors, retrieve relevant graph slices from the vector database, input the questions along with a dedicated large model, and the model generates fact-based answers to respond to users in real time.

[0207] Furthermore, the system also includes a user interaction layer that connects with the responsible personnel or case handlers to receive early warning information containing explanations of the causes of risks, query holographic profile reports or ask questions in natural language, and provide data support for real-time decision-making.

[0208] Furthermore, the system also includes a unified model service gateway for managing the invocation of intelligence analysis models, and a built-in prompt library that can dynamically load corresponding system prompts based on different upstream requests, such as extraction, profiling, and interpretation.

[0209] The following example demonstrates how to use an intelligence analysis model to assess a target person of interest, "Zhang San".

[0210] Data Access and Cleaning: The system retrieves a record from the accommodation database: "Zhang San checked into a certain hotel on the 10th." Simultaneously, it reads unstructured text from the interrogation record database: "Suspect 'Lao San' (ID number ending in 201X) mentioned a transaction at a certain hotel." In the entity alignment stage, the system first clarifies the relationship through ID number matching, then uses an edit distance algorithm to identify that "Zhang San" and "Lao San" have a similarity of 0.92, indicating identity in a specific context, thus deciding to fuse the data. In the information extraction stage, the system calls the intelligence analysis model and inputs a prompt: "You are a criminal investigation expert, please extract entities and events from the following text..." The intelligence analysis model outputs structured event data.

[0211] Profile Generation: The system aggregates the merged data, constructs prompt words, and then calls the intelligence analysis model and inputs the constructed prompt words: "Based on the above facts, please generate Zhang San's profile data in JSON format"; the intelligence analysis model will output JSON data containing {"timeline":[...],"relations":[...]}; the front-end profile rendering engine 30 reads the JSON data and draws Zhang San's timeline and the relationship between Zhang San and Li Si.

[0212] Risk Calculation and Interpretation: In the quantitative calculation stage, the system extracts Zhang San's data and substitutes it into the dynamic risk index calculation formula. Among these factors, the static resume factor S: due to Zhang San's past involvement in theft, the normalized score is 0.6; the behavioral dynamic factor B: due to Zhang San's recent nocturnal lifestyle, the score is 0.8; the social network factor N: Zhang San had two phone calls with fugitive Li Si, the score is 0.7; and the speech characteristic factor C: no abnormalities were found, the score is 0.1. Assuming each factor has a weight of 0.25, then R = 0.25 × (0.6 + 0.8 + 0.7 + 0.1) = 0.55. In the qualitative interpretation stage, the system detects that the behavioral dynamic factor B and the social network factor N contribute the most, constructing the prompt: "The dynamic risk index of target person Zhang San is 0.55, mainly due to abnormal work and rest schedules and contact with fugitives. Please generate a briefing." The intelligence analysis model generates the following content: "Zhang San's risk has increased recently, mainly due to frequent nighttime activities and resuming contact with fugitive Li Si. Close monitoring is recommended."

[0213] Warning push: The system will push the above briefing to the terminal.

[0214] The methods and systems of this invention achieve a leapfrog improvement in intelligence processing and analysis efficiency, significantly saving labor costs. Addressing the time-consuming and labor-intensive problem of data integration in existing technologies, this invention, through automated data fusion and information extraction, compresses manual data processing work that previously took hours or even days to minutes. Empirical studies show that using a similar AI natural language processing system for emergency call analysis can reduce the extraction time of key information from an average of 2.7 minutes to 38 seconds, improving efficiency by 86%. In daily work, this invention can automatically integrate voice, video, and dispatch data into case reports, saving frontline personnel several hours of paperwork time daily. Moreover, AI-enhanced operations can reduce the manpower required for routine monitoring by 35%, allowing organizations to reallocate up to 420 work hours for higher-value tasks within a month.

[0215] The human-computer interaction experience and early warning response speed have been revolutionaryly optimized, improving the efficiency and effectiveness of practical decision-making. Addressing the issues of complex operation and delayed early warning in traditional systems, the natural language interface provided by this invention significantly lowers the barrier to entry. Comparative studies show that compared to traditional multi-field search interfaces, the natural language search interface similar to this invention not only completes tasks significantly faster but also has a higher accuracy rate, and users rate its ease of use far more highly. Simultaneously, the automated real-time early warning mechanism of this invention transforms intelligence work from passive querying to proactive pushing, ensuring that risks are addressed immediately upon discovery. Application data in related fields show that AI-assisted dispatch systems can reduce the average arrival time of emergency response units by 4.3 minutes. This time advantage is crucial in rapidly changing public safety incidents, effectively saving lives and preventing property loss.

[0216] Reduced resource consumption and improved maintainability. Compared to using multiple independent models, this invention significantly reduces memory usage and deployment costs through multi-tasking technology on a single model, with an estimated reduction in hardware costs of 60%, while avoiding information loss during context transfer between multiple models.

[0217] The scientific rigor and interpretability of risk assessment. Specific mathematical formulas quantify risk, resolving the issue of inconsistent standards in traditional manual assessments; combined with the semantic interpretation of a large-scale model, cold, hard numbers acquire business meaning, increasing the decision adoption rate to over 90%.

[0218] The detection rate of deep relationships has been improved. Leveraging the reasoning capabilities of the fine-tuned large model, it is able to identify implicit, non-directly connected group relationships. Experimental data shows that, on the same dataset, this system uncovers 45% more valid association clues than traditional rule-based systems.

[0219] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. A method for holographic profiling and dynamic risk assessment of personnel, characterized in that, include: Step S1: In response to the analysis request, use the intelligence analysis model based on the large language model to perform data fusion and entity alignment on the data to be analyzed, and construct a knowledge graph; Step S2: The intelligence analysis model generates a standard holographic data package based on the first preset prompt word and the knowledge graph, so as to generate a holographic portrait of the person of interest based on the standard holographic data package; Step S3: Perform multi-factor analysis on the knowledge graph and determine the dynamic risk index based on the results of the multi-factor analysis; Step S4: Continuously monitor the dynamic risk index, and generate and push early warning information when the dynamic risk index exceeds the preset risk threshold.

2. The method for holographic profiling and dynamic risk assessment of individuals of interest according to claim 1, characterized in that, Step S1 includes: Step S11: In response to the analysis request, collect the data to be analyzed from a preset heterogeneous data source; Step S12: Based on the intelligence analysis model, extract information from the unstructured text in the data to be analyzed to obtain multiple entities and the key feature vectors of the corresponding entities. Step S13: Calculate the cosine similarity of the key feature vectors of the entity to obtain the first similarity metric. Step S14: Calculate the edit distance of the entity's string to obtain a second similarity metric. Step S15: The first similarity metric and the second similarity metric are weighted and fused to obtain a comprehensive fused similarity metric. Step S16: When the comprehensive fusion similarity metric exceeds the preset similarity threshold, it is determined to be the same entity and fused, and the fused entity is stored in the graph database to construct the knowledge graph.

3. The method for holographic profiling and dynamic risk assessment of individuals of interest according to claim 2, characterized in that, Step S14 includes: Step S141: Preprocess the entity, the preprocessing including at least data cleaning, to obtain a preprocessed entity; Step S142: Calculate the edit distance between the strings of the two preprocessed entities to be compared to obtain the first edit distance; Step S143: Normalize the first edit distance to obtain the second edit distance; Step S144: The second edit distance of the entity and the matching weight of at least one auxiliary attribute entity corresponding to the entity are weighted to obtain the second similarity metric.

4. The method for holographic profiling and dynamic risk assessment of individuals of interest according to claim 1, characterized in that, Step S2 includes: Step S21: Construct the first preset prompt word, wherein the first preset prompt word includes at least one of role definition, fact constraint and output format specification; Step S22: The intelligence analysis model generates a standard holographic data packet in a predetermined format based on the first preset prompt word and the knowledge graph; Step S23: The front-end portrait rendering engine parses the standard holographic data package and generates a holographic portrait of the person of interest. The holographic portrait includes at least one of the following: a biographical sketch, a visual relationship graph, and a dynamic activity timeline.

5. The method for holographic profiling and dynamic risk assessment of individuals of interest according to claim 1, characterized in that, Step S3 includes: Step S31: Perform multi-factor analysis on the knowledge graph to obtain multi-factor analysis results. The multi-factors include at least one of static resume factors, behavioral dynamic factors, social network factors, and speech characteristic factors. Step S32: Perform a weighted summation based on the results of the multi-factor analysis to determine the dynamic risk index.

6. The method for holographic profiling and dynamic risk assessment of individuals of interest according to claim 1, characterized in that, In step S31, the static resume factor is calculated as follows: ； Where W1 represents the weight corresponding to the previous level of involvement; S level W1 represents the normalized level of prior criminal record; W2 represents the weight corresponding to the number of criminal records; S count The normalized number of criminal records is represented by S; the static resume factor is represented by S. The calculation method for the behavioral dynamic factor is as follows: ； Where k1 represents the weight corresponding to the recent abnormal travel frequency; F freq Indicates recent abnormal travel frequency; k2 represents the weight corresponding to the proportion of nighttime activities; R night B represents the percentage of nighttime activities; B represents the behavioral dynamic factor. The calculation method for the social network factor is as follows: ； Where, N PR Indicates the centrality of individuals within a criminal network; N Path N represents the shortest path distance between individuals on the watchlist and those on the restricted list; N represents the social network factor. The method for calculating the speech characteristic factors is as follows: ； Among them, Score senti Score represents the sentiment analysis index of the intelligence analysis model for the statements of the individuals of interest; α represents the weight of the sentiment analysis index; keyword C represents the total score obtained by matching all sensitive words in the statements of the person under observation from the graded sensitive word database; C represents the characteristic factor of the statement.

7. The method for holographic profiling and dynamic risk assessment of individuals of interest according to claim 6, characterized in that, The method for calculating the normalized prior offense level is as follows: ； Here, L indicates the level of prior criminal record of the person being monitored; L max Indicates the highest level of involvement in the case; The method for calculating the normalized number of criminal records is as follows: ； Where n represents the total number of historical criminal records; The calculation method for the recent abnormal travel frequency is as follows: ； Among them, C sens This indicates the number of times the personnel in question visited the target area within the first preset time period; T threshold Indicates the preset number of times threshold; The method for calculating the proportion of nighttime activities is as follows: ； Where, N night N represents the amount of data collected by the monitored individuals within the second preset time period. total This indicates the total number of activities of the monitored individuals within the first preset time period; The method for calculating the centrality of the individuals under surveillance within the criminal network is as follows: A local social relationship subgraph is constructed based on the people being followed and their corresponding second-degree connections. The local social relationship subgraph includes nodes and edges. The edges represent the association between two nodes, and the weight of the edges is determined based on the interaction frequency between the nodes. Calculate the page ranking value for each node; The centrality of the person being monitored in the criminal network is obtained by comparing the webpage ranking value of the monitored person with the maximum webpage ranking value among all nodes in the local social relationship subgraph. The corresponding calculation formula is as follows: ； Among them, PR target PageRank (PR) represents the ranking value of a webpage based on the number of people following it. max This represents the maximum value among all the webpage ranking values ​​of all nodes in the local social relationship subgraph; The method for calculating the shortest path distance between the individuals on the watch list and those on the restricted list is as follows: Calculate the set of shortest hop counts between the individuals of interest and the nodes marked as restricted in the knowledge graph; Based on the minimum value in the set of shortest hop counts, the shortest path distance between the monitored individuals and those on the restricted list is calculated using the distance decay formula. The corresponding calculation formula is as follows: ； Where, d min This represents the minimum value in the set of shortest hop counts.

8. The method for holographic profiling and dynamic risk assessment of individuals of interest according to claim 1, characterized in that, The warning information includes an explanation of the causes of the risk, and the method for obtaining the explanation of the causes of the risk includes: Step S41: When the dynamic risk index exceeds the preset risk threshold, extract the corresponding original fact fragments from the data to be analyzed based on the top k factors that contribute the most to the dynamic risk index, where k is a positive integer greater than 1. Step S42: Input the top k factors with the highest contribution and their corresponding original fact fragments into the intelligence analysis model. The intelligence analysis model generates risk cause explanations based on the second preset prompt words, so as to generate early warning information containing the risk cause explanations.

9. The method for holographic profiling and dynamic risk assessment of individuals of interest according to claim 1, characterized in that, Following step S4, the following is also included: Receive user questions and transform them into question vectors; Based on the question vector, a knowledge graph slice corresponding to the question vector is retrieved. The question vector and the corresponding map slice are input into the intelligence analysis model to generate fact-based question answers.

10. A system for holographic profiling and dynamic risk assessment of individuals, characterized in that, For implementing the method for holographic profiling and dynamic risk assessment of individuals of concern as described in any one of claims 1-9, including: The heterogeneous data fusion module is used to respond to analysis requests and use an intelligence analysis model based on a large language model to perform data fusion and entity alignment on the data to be analyzed, and to build a knowledge graph. The graph storage engine, connected to the heterogeneous data fusion module, is used to store the knowledge graph based on a graph database; A risk calculation engine, connected to the knowledge graph storage engine, is used to perform multi-factor analysis on the knowledge graph and determine a dynamic risk index based on the results of the multi-factor analysis. The intelligence analysis model is also used to generate standard holographic data packets based on the first preset prompt words and the knowledge graph; A front-end portrait rendering engine, connected to the graph storage engine, is used to generate holographic portraits of the individuals of interest based on the standard holographic data package. The interactive terminal is connected to the risk calculation engine and the front-end portrait rendering engine respectively, and is used to continuously monitor the dynamic risk index. When the dynamic risk index exceeds a preset risk threshold, it generates and pushes early warning information.

Images