A method and device for generating a holographic personnel profile relationship map

Abstract

The application relates to the field of artificial intelligence and big data processing technology, and discloses a personnel holographic archive relationship graph generation method and device, which comprises the following steps: obtaining target object archive record data, converting the target object archive record data into an evidence set composed of multi-dimensional time sequence evidence units containing time intervals and evidence states; constructing a scene constraint matrix group containing a role mutual exclusion and a time overlap tolerance matrix according to a business rule; extracting key attributes as blocking keys to perform clustering division, calculating attribute feature similarity of candidate node evidence units in a candidate object cluster, generating a candidate object subgraph containing identity similarity and source conflict edges, and mapping the candidate object subgraph into a multi-dimensional time sequence feature tensor; inputting the multi-dimensional time sequence feature tensor into a graph attention network, identifying conflict merging edges by using the matrix group, applying a loss penalty to calculate a fusion probability, and mapping relationship edges into different confidence levels to generate a multi-layer object relationship subgraph. The application realizes strict checking of fusion logic and hierarchical management and control of multi-dimensional evidence quality of a graph.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and big data processing technology, and in particular to a method and apparatus for generating a holographic profile of personnel relationships. Background Technology

[0002] With the deep application of big data and artificial intelligence technologies, constructing a holographic profile relationship graph centered on individuals has become a key infrastructure in social governance and business management scenarios such as financial anti-fraud and background checks of personnel in large enterprises. Holographic profiles aim to break down data silos between various business systems, effectively aggregating behavioral records, identity attributes, and relationships scattered across different time points, spatial regions, and dimensions to form a network topology that comprehensively depicts the target object. However, in real-world business environments, raw profile data often exhibits characteristics of multi-source heterogeneity, semantic ambiguity, and difficulty in distinguishing authenticity. Traditional graph generation methods mostly rely on string matching or shallow machine learning models, judging entity identity and performing graph fusion solely based on the mathematical distribution or literal similarity of node attributes. This purely data-driven computational paradigm severely deviates from the objective laws of the physical world and business logic. For example, a system might arbitrarily merge two activity trajectories occurring at the same time but thousands of kilometers apart into the same node simply because two people share the same name and place of origin. This erroneous fusion, lacking common-sense verification, not only pollutes the underlying graph topology but also seriously misleads subsequent clue analysis and business decisions.

[0003] Furthermore, to handle complex nonlinear graph topological relationships, existing graph construction techniques have begun to incorporate deep learning models such as graph neural networks for entity alignment and relationship prediction. However, a resulting technical bottleneck lies in the inherent black-box nature of deep learning models. Their internal high-dimensional tensor computations and attention weight allocation mechanisms are difficult to explain clearly to external users. When a graph network model outputs the probability of merging two entities, business personnel cannot know which underlying evidence chains this decision is based on. If an algorithmic misjudgment occurs or merging fails due to the inherent mutual exclusion of business roles, the system cannot provide a quantitative explanation down to the time slice and specific business rules. This makes it difficult to implement the generated holographic graph in practical scenarios where the certainty of evidence is extremely high. The resulting technical situation is that the system's computing power is consumed by a large number of invalid global similarity comparisons, while the generated graph network is a blind box lacking interpretability.

[0004] More importantly, after completing the initial node similarity calculation and probability inference, existing graph database underlying construction mechanisms generally lack dynamic decoupling capabilities based on evidence quality. Traditional knowledge graphs, upon receiving fusion instructions, often indiscriminately stack all in-degree and out-degree edges inherited by the newly merged node, along with the massive evidence chains attached to the underlying layer, on the same flat-dimensional topological plane. This haphazard storage mechanism mixes solid objective facts with questionable, contradictory, and even one-sided business statements, causing the graph itself to lose its ability to express data confidence. When faced with massive amounts of heterogeneous and contradictory archival information, the lack of effective isolation mechanisms and hierarchical routing strategies makes it difficult for business personnel to filter out invalid noise during retrieval, significantly weakening the graph database's flexible application capabilities and security compliance level in complex clue tracking and factual reasoning analysis. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and apparatus for generating a holographic knowledge graph of personnel records. This method introduces dynamic business rule constraints and a confidence level decoupling mechanism during the fusion of heterogeneous record data, thereby generating a high-quality holographic knowledge graph with high logical rigor, interpretability, and hierarchical relationship capabilities.

[0006] Firstly, this application provides a method for generating a holographic personnel file relationship map, the method comprising: Acquire the archival record data of the target object, and convert the archival record data into multi-dimensional temporal evidence units containing time intervals and evidence status to form the evidence set of the target object; A scenario constraint matrix group is constructed based on preset business rules. The scenario constraint matrix group includes a role mutual exclusion matrix and a time overlap tolerance matrix. Key attributes are extracted from the evidence set of the target object as blocking keys. Based on the blocking keys, the evidence set of the target object is clustered to obtain candidate object clusters. Multidimensional time-series evidence units within the candidate object clusters are used as candidate nodes. The attribute feature similarity of each candidate node is calculated. Based on the attribute feature similarity, a candidate object subgraph containing relation edges is generated. The relation edges include identity similarity edges and source conflict edges. Map candidate nodes and relation edges in the candidate object subgraph to a multidimensional temporal feature tensor; The multidimensional temporal feature tensor is input into the graph attention network. During the graph attention operation, the scene constraint matrix group is used to identify conflicting and merging edges that violate the scene constraint matrix group in the identity similarity edges of the relation edges, and a loss penalty is applied to the conflicting and merging edges to calculate the fusion probability of the target object. Based on the fusion probability of the target object, the relation edges in the candidate object subgraph are mapped to relation edges of different confidence levels, generating a multi-layer object relation subgraph as an archive relation graph.

[0007] Secondly, this application provides a device for generating a holographic personnel file relationship map, the device comprising: The evidence conversion module is used to acquire the archival record data of the target object and convert the archival record data into multi-dimensional temporal evidence units containing time intervals and evidence status, thus forming the evidence set of the target object. The constraint construction module is used to build a scenario constraint matrix group according to preset business rules. The scenario constraint matrix group includes a role mutual exclusion matrix and a time overlap tolerance matrix. The clustering analysis module is used to extract key attributes from the evidence set of the target object as blocking keys, and to cluster the evidence set of the target object based on the blocking keys to obtain candidate object clusters. The multidimensional time-series evidence units within the candidate object clusters are used as candidate nodes. The attribute feature similarity of each candidate node is calculated, and a candidate object subgraph containing relation edges is generated based on the attribute feature similarity. The relation edges include identity similarity edges and source conflict edges. The feature mapping module is used to map candidate nodes and relation edges in the candidate object subgraph into multidimensional temporal feature tensors; The fusion computing module is used to input multi-dimensional temporal feature tensors into the graph attention network. During the graph attention operation, the scene constraint matrix group is used to identify conflicting and merging edges that violate the scene constraint matrix group in the identity similarity edges of the relation edges, and a loss penalty is applied to the conflicting and merging edges to calculate the fusion probability of the target object. The hierarchical graph module is used to map the relational edges in the candidate object subgraph to relational edges at different confidence levels based on the fusion probability of the target object, thereby generating a multi-layer object relational subgraph as an archive relational graph.

[0008] Compared with the prior art, the beneficial effects of the present invention are at least as follows: 1. This application achieves a deep integration of data-driven and knowledge-driven approaches in the underlying architecture of graph neural networks, breaking through the technical limitations of relying solely on mathematical similarity in the traditional knowledge graph generation process. By pre-constructing a set of scene constraint matrices containing physical common sense such as role mutual exclusion and time overlap tolerance, and transforming them into dynamic loss penalty terms in the forward computation process of the graph attention network, the deep learning model can proactively perceive and adhere to the bottom line of real-world business logic when performing high-concurrency tensor operations. This mechanism of writing rules into the loss function blocks node erroneous merging caused by pure literal similarity from the source of the algorithm mechanism, significantly improving the absolute accuracy of graph fusion. At the same time, by calculating numerical contribution through reverse attribution and generating conflict explanation vectors, it endows the AI ​​model with extremely strong quantitative interpretability.

[0009] 2. This application optimizes the computational architecture and processing efficiency of the underlying data of the holographic map. Through pre-processing techniques such as blocking key extraction and clustering, the massive and chaotic global file comparison is reduced to a highly cohesive cluster of candidate objects for internal computation, effectively overcoming the computational bottleneck caused by massive entity parsing. Simultaneously, this application proposes a standardized dimensionality reduction mapping mechanism, which strictly aligns and encapsulates discrete multidimensional temporal evidence units with dual relational topologies containing identity similarity and source conflict into a multidimensional temporal feature tensor containing time series dimensions, node feature dimensions, and graph topology dimensions. This not only preserves the spatiotemporal evolution trajectory and contradictory clues of the personnel file status without loss, but also provides a regular and efficient computational foundation for subsequent graph attention networks, significantly enhancing the engineering feasibility for handling high-concurrency map generation tasks.

[0010] 3. This application constructs a novel multi-level object relationship subgraph storage and display paradigm, solving the technical challenge of hierarchical control over the reliability of relationships after data fusion in traditional graphs. By introducing a strict dual-threshold judgment mechanism based on the fusion probability of target objects, this application can not only automatically and securely perform physical topology reconstruction of high-confidence graph nodes, but also formulate exclusive priority routing rules by penetrating and analyzing the authority, continuity, cross-verification, and conflict status of the underlying evidence chain. These rules successfully deconstruct the originally flat and chaotic complex graph topology and map it into four mutually isolated view planes: the core layer, the auxiliary layer, the pending layer, and the abnormal layer. This multi-dimensional layered management design of evidence quality ensures the absolute purity of core facts while preserving exploration space for peripheral and questionable clues, enhancing the practical effectiveness of the holographic archive system in complex case reasoning and accurate judgment scenarios. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the steps of a method for generating a holographic personnel file relationship map in an embodiment of this application. Figure 2 This is a schematic diagram of the logical architecture for constructing the scenario constraint matrix group and dividing the blocking key clusters in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the process of fusing multidimensional temporal feature tensor mapping and graph attention network in the embodiments of this application; Figure 4 This is a structural diagram of a personnel holographic file relationship map generation device according to an embodiment of this application. Detailed Implementation

[0013] This application provides a method and apparatus for generating a holographic personnel file relationship map. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0014] Example 1: For ease of understanding, the specific process of the embodiments of this application is described below, such as... Figure 1 The method for generating a holographic personnel file relationship map in the embodiment of this application shown includes: S1. Obtain the archival record data of the target object, and convert the archival record data into multi-dimensional temporal evidence units containing time intervals and evidence status to form the evidence set of the target object.

[0015] The process involves converting archival record data into multi-dimensional temporal evidence units containing time intervals and evidence states to form a target object evidence set. This includes: extracting the start and end times of the archival record data to construct time intervals; extracting candidate object identifiers, event types, associated objects, and source numbers from the archival record data; acquiring external verification data from different business systems that contains similarity features to the candidate object identifiers greater than or equal to a preset identifier threshold; cross-validating the archival record data with the external verification data to determine the evidence state; and combining the time interval, evidence state, candidate object identifiers, event types, associated objects, and source numbers to generate a six-element data set as a multi-dimensional temporal evidence unit. Multiple multi-dimensional temporal evidence units are then aggregated to form the target object evidence set.

[0016] Specifically, in the process of constructing a holographic personnel relationship map, the data generated by the original business system often suffers from technical problems such as multi-source heterogeneity, ambiguous temporal semantics, and difficulty in distinguishing isolated evidence. This results in low accuracy and lack of factual basis for subsequent map generation. In order to solve the problems of unstructured original data and unknown credibility, this embodiment provides a specific implementation method for data preprocessing and multi-dimensional feature extraction to transform the underlying business data into a standard form that supports map calculation, as follows.

[0017] The process involves acquiring archival record data of the target object and converting it into multi-dimensional temporal evidence units containing time intervals and evidence states, thus forming the target object's evidence set. Each multi-dimensional temporal evidence unit is a structured data carrier with time span attributes and authenticity verification labels, providing the smallest granularity of credible factual underlying elements for subsequent graph construction. In the specific conversion process, the start and end times of the archival record data are extracted to construct time intervals, and object identifier candidates, event types, associated objects, and source numbers are extracted, thereby decomposing the unstructured archival text into standard attribute fields. Subsequently, external verification data from different business systems is acquired, containing similarity features to the object identifier candidates that are greater than or equal to a preset identifier threshold. When calculating similarity features, a character-level N-gram algorithm is used to convert the object identifier candidates in the archival record data and the identifier fields in the external system data into text numerical vectors, and the cosine score of the angle between the two text numerical vectors is calculated. When the cosine score is greater than or equal to the preset identifier threshold, it is determined that the two data points point to the same potential object, and the external system data is extracted as external verification data.

[0018] After acquiring external verification data, the archival record data is cross-validated with the external verification data to determine the evidence status. The evidence status characterizes the authenticity and level of factual confirmation of the current archival record data under multi-source cross-comparison. During the cross-validation calculation, corresponding fields of the archival record data and the external verification data are extracted respectively, and the time interval overlap ratio, event type consistency score, and cross-union ratio of related objects are calculated as basic comparison features. For example, the event type consistency score is obtained by comparing the event type hierarchical dictionary. Since the above basic comparison features have been mapped to the dimensionless standard numerical range of 0 to 1 through proportionalization, corresponding weights are directly assigned to the time interval overlap ratio, event type consistency score, and cross-union ratio of related objects. The coefficients are calculated, and the sum of the weight coefficients is 1. Then, each basic comparison feature is multiplied by its corresponding weight coefficient and summed to calculate the comprehensive cross-validation score. The system presets three continuous and non-overlapping judgment intervals, namely [0, first confidence threshold), [first confidence threshold, second confidence threshold) and [second confidence threshold, 1]. The archive record data whose comprehensive cross-validation score falls into the above three intervals are respectively assigned feature labels containing conflict anomalies, unilateral statements or multi-source corroboration, so as to determine the final evidence status. Finally, the time interval, evidence status, object identification candidate, event type, associated object and source number are combined to generate a six-element data group, which is used as a multi-dimensional temporal evidence unit. Multiple multi-dimensional temporal evidence units are aggregated to form the target object evidence set.

[0019] The above implementation transforms isolated and heterogeneous archival data into six-element evidence units with a unified time scale and multi-source verification status through text vectorization matching and dimensionless weighted cross-validation algorithms, without making invasive modifications to the original underlying system. This solution effectively filters out noise and false conflicts in business data, providing high-purity and structured underlying factual support for the subsequent construction of node topology and confidence inference of relation edges in the graph network.

[0020] S2. Construct a scenario constraint matrix group based on preset business rules. The scenario constraint matrix group includes a role mutual exclusion matrix and a time overlap tolerance matrix.

[0021] like Figure 2 The diagram illustrates the logical architecture for constructing the scenario constraint matrix group and the blocking key clustering. The construction of the scenario constraint matrix group based on preset business rules includes: parsing the preset business rules to extract the set of business roles and the set of event types; determining the logical exclusion relationship between any two business roles in the set of business roles, and binarizing the corresponding coordinate positions of the initial two-dimensional square matrix according to the logical exclusion relationship to generate a role mutual exclusion matrix; extracting the allowed overlap time threshold set for the event type set in the preset business rules, calculating the time tolerance coefficient between any two event types based on the allowed overlap time threshold, and using the time tolerance coefficient to fill matrix elements to construct a time overlap tolerance matrix; and combining the role mutual exclusion matrix and the time overlap tolerance matrix to form the scenario constraint matrix group.

[0022] Specifically, after converting the underlying archival records into multi-dimensional temporal evidence units through the above steps, the problem of how to identify conflicts between physical common sense and logical levels among massive evidence units is faced. In real social activities, different archival records of the same person are subject to objective social rules and spatiotemporal laws, and there are strict time exclusion periods between certain specific event types. This step aims to solve the technical problem that the graph generation algorithm lacks a machine-readable common sense verification mechanism when merging nodes. By converting implicit business common sense into an explicit mathematical matrix structure, it provides rigid constraint boundaries for the subsequent fusion operation of graph network nodes.

[0023] A scenario constraint matrix group is constructed based on preset business rules. The scenario constraint matrix group includes a role mutual exclusion matrix and a time overlap tolerance matrix. In specific implementation, the preset business rules are first parsed to extract the set of business roles and the set of event types. The preset business rules refer to the common sense verification knowledge base form pre-configured in the business system. By performing structured text parsing on the business rules, all potential identity tags of personnel in each business scenario are extracted to form the set of business roles. At the same time, various behavioral record tags are extracted to form the set of event types. Among them, the time overlap tolerance matrix is ​​used to determine whether the actual time intersection of two nodes exceeds the tolerance coefficient benchmark of their corresponding event type in the subsequent graph attention operation.

[0024] After obtaining the above set, the logical exclusion relationship between any two business roles in the set is determined. Based on the logical exclusion relationship, the corresponding coordinate positions of the initial two-dimensional matrix are binarized and assigned values ​​to generate a role mutual exclusion matrix. Here, the logical exclusion relationship refers to the conflicting attributes of business roles that are absolutely mutually exclusive under the framework of laws or business regulations. In the specific construction action, the number of elements contained in the set of business roles is used as the dimension, and the row and column indices of the business role elements are used as the row and column indices to build the initial two-dimensional matrix. Each row and column intersection coordinate in the matrix represents a combination pair of two business roles. All combination pairs in the matrix are traversed. When it is determined that there is a logical exclusion relationship between the two business roles corresponding to the coordinates, the value 1 is assigned to the coordinate position; when there is no logical exclusion relationship, the value 0 is assigned to the coordinate position. Through the above coordinate traversal and assignment operation of non-zero or one, the discrete text constraint rules are transformed into a sparse structure that can be computed by a computer, generating a role mutual exclusion matrix.

[0025] Simultaneously, the allowed overlap time thresholds set for event type sets in the preset business rules are extracted. Time tolerance coefficients between pairs of event types are calculated based on these thresholds, and matrix elements are filled using these coefficients to construct a time overlap tolerance matrix. Since the original units of the allowed overlap time thresholds for different event type combinations are absolute time units, to eliminate dimensional differences and adapt to subsequent network tensor operations, the maximum value among all allowed overlap time thresholds for all event type combinations is extracted as the global baseline threshold. Then, the allowed overlap time thresholds between pairs of event types are divided by this global baseline threshold for linear normalization, and dimensionless ratios between 0 and 1 are calculated and used as time tolerance coefficients. An event matrix is ​​constructed using the event type set as the dimension baseline, and the calculated time tolerance coefficients are filled into the corresponding row and column coordinates to construct the time overlap tolerance matrix. Finally, the completed role mutual exclusion matrix and time overlap tolerance matrix are combined logically and mathematically to form a scene constraint matrix group used to constrain feature node fusion.

[0026] The aforementioned implementation quantifies discrete humanities and physical common sense and business rules into dimensionless structured tensor matrices, solving the technical problem of lacking hard constraints of business rules when the algorithm model processes the merging of graph nodes. Through binary coordinate assignment and time dimension normalization calculation, a multi-dimensional scene constraint matrix group is constructed, effectively eliminating the semantic barriers of cross-system text rules. This common sense constraint mechanism with matrix as carrier can be seamlessly embedded into the subsequent deep learning network architecture as prior knowledge, avoiding the generation of absurd correlations and providing rigorous logical verification support for the accurate inference of personnel holographic graphs.

[0027] like Figure 2 The diagram shows the logical architecture of the scenario constraint matrix group construction and blocking key clustering. S3: Extract the key attributes from the target object evidence set as blocking keys, and cluster the target object evidence set based on the blocking keys to obtain candidate object clusters. Take the multi-dimensional temporal evidence units in the candidate object clusters as candidate nodes, calculate the attribute feature similarity of each candidate node, and generate a candidate object subgraph containing relation edges based on the attribute feature similarity. The relation edges include identity similarity edges and source conflict edges.

[0028] The process involves calculating the attribute feature similarity of each candidate node and generating a candidate object subgraph containing relational edges based on the attribute feature similarity. This includes: extracting at least one key attribute from each multidimensional time-series evidence unit in the target object evidence set, such as document hash value, pinyin encoding, contact information time period, organizational path, and address grid, as a blocking key; dividing and aggregating the target object evidence set into blocks according to the key-value matching results of the blocking key to obtain multiple independently divided candidate object clusters; mapping each multidimensional time-series evidence unit within the same candidate object cluster to candidate nodes, and comparing each candidate node pairwise to calculate attribute feature similarity, while simultaneously parsing the data source identifier and mutually exclusive attributes contained in the multidimensional time-series evidence unit corresponding to each candidate node; establishing an identity similarity edge between two candidate nodes with attribute feature similarity greater than or equal to a preset similarity threshold, and establishing a source conflict edge between two candidate nodes with attribute feature similarity greater than or equal to a preset similarity threshold and different data source identifiers or conflicting mutually exclusive attributes; and combining candidate nodes and relational edges to form a candidate object subgraph.

[0029] Specifically, after completing the temporal unit transformation of the underlying evidence and the construction of the scene constraint matrix, directly performing entity identity comparison on a massive set of target object evidence in the global scope would generate huge computational resource overhead. This step aims to solve the technical problems of excessive computational power consumption caused by global comparison during the graph generation process and the lack of underlying topological connectivity in the initial network. By introducing blocking block technology, computational dimensionality reduction is achieved, and a basic graph sub-topology structure representing data homology and contradiction is established.

[0030] Key attributes are extracted from the target object evidence set as blocking keys. Based on these blocking keys, the target object evidence set is clustered to obtain candidate object clusters. Here, the blocking key refers to a coarse-screening index field used in entity parsing techniques to divide a large-scale discrete dataset into multiple highly cohesive data blocks. In practice, the basic information carried by each multi-dimensional time-series evidence unit within the target object evidence set is parsed, and at least one feature (i.e., key attribute) from among document hash values, pinyin encoding, contact information time periods, organizational paths, and address grids is extracted and used as the blocking key. The target object evidence set is then segmented and aggregated according to the key-value matching results of the blocking keys to obtain multiple independently divided candidate object clusters. The document hash value is obtained by calling Locality-Sensitive Hash (LSH). The Locality Sensitive Hashing (LSH) algorithm generates a fixed-length string after feature mapping operations on the ID number. Unlike traditional cryptographic hash functions with avalanche effects, LSH ensures that similar input items in high-dimensional space still have a high probability of hash collisions after dimensionality reduction mapping. Therefore, even if the original ID number has individual character errors due to manual entry or device recognition deviations, the generated LSH value can still maintain a high degree of feature similarity. Pinyin encoding converts the person's name into Pinyin and then encodes it, which can handle homophones and variants of the name. Contact information time period refers to the effective time range of the contact information. By the overlap of time periods, it is possible to quickly determine whether two contact methods may belong to the same person. Organizational path refers to the location path of the person in the organizational hierarchy. For example, the department path of an employee is the group headquarters, R&D center, and software development department. Address grid is the result of converting geographical addresses into standardized grid encoding. Specifically, the hash bucketing logic is executed to classify multi-dimensional time-series evidence units with the same blocking key value into the same data slice, breaking the global comparison space and dividing the huge target object evidence set into multiple independent candidate object clusters.

[0031] Each multidimensional temporal evidence unit within the same candidate object cluster is mapped to a candidate node, and pairwise comparisons are made between candidate nodes to calculate attribute feature similarity. The Jaccard similarity algorithm is used to calculate the overlap ratio of the corresponding sets, i.e., by dividing the number of intersection elements of the feature sets corresponding to two nodes by the total number of union elements, the sub-feature similarity value in the range of 0 to 1 is obtained. Since the sub-feature similarities are naturally in the same dimensionless range, they are directly assigned corresponding constant weight coefficients and the sum of the weight coefficients is set to 1. The sub-feature similarity is multiplied by the corresponding weight coefficients in turn and summed to calculate the final attribute feature similarity. At the same time, the data source identifier and mutual exclusion attributes contained in the multidimensional temporal evidence units corresponding to each candidate node are analyzed. Specifically, the source system number of the underlying archive data of the candidate node is extracted as the data source identifier, and the business role label recorded in the candidate node that can trigger business exclusion logic is extracted as the mutual exclusion attribute.

[0032] After completing the similarity calculation and feature parsing, it is determined whether the conditions for establishing topological edges are met. An identity similarity edge is established between two candidate nodes whose attribute feature similarity is greater than or equal to a preset similarity threshold, to indicate that they are likely to point to the same real person. At the same time, under the premise that the attribute feature similarity is greater than or equal to the preset similarity threshold, it is determined whether there is a contradiction in the data source or logical attributes of the two nodes: if the comparison finds that the data source identifiers of the two candidate nodes are different, or if it is determined that their mutually exclusive attributes conflict, the mutually exclusive attributes of the two candidate nodes are used as row and column indices to query the role mutual exclusion matrix in the scene constraint matrix group. If the corresponding coordinate is assigned a value of 1, it is determined that there is a mutually exclusive attribute conflict, and a source conflict edge is established between the two candidate nodes. Through the above operations, there may be an identity similarity edge and a source conflict edge between two similar candidate nodes. Finally, all candidate nodes in the candidate object cluster are combined with the identity similarity edge and the source conflict edge established above to form a candidate object subgraph with a connected form.

[0033] The aforementioned implementation introduces blocking key clustering and multi-element graph edge construction mechanisms, solving the computational bottleneck caused by global computation of massive nodes and the technical challenge of lacking structured support in graph conflict tracing. By using blocking block technology to reduce global comparison to local computation within clusters, the time complexity of the algorithm is effectively reduced. On this basis, by using quantified attribute feature similarity calculation, not only are similar edges representing entity identity constructed, but conflict edges representing logical contradictions are also established through joint verification of source and business mutually exclusive attributes. This construction of a multi-edge topology structure preserves the clues of common origin and contradiction in the original data, providing a solid connectivity framework for subsequent graph network mining of hidden relationships and resolution of file conflicts, and significantly improving the completeness of graph inference.

[0034] S4. Map the candidate nodes and relation edges in the candidate object subgraph to a multidimensional temporal feature tensor.

[0035] The process of mapping candidate nodes and relation edges in the candidate object subgraph to a multidimensional temporal feature tensor includes: extracting attribute features and timestamp records contained in each candidate node; converting the attribute features into numerical vectors and arranging them according to the temporal order of the timestamp records to construct a temporal node feature matrix; extracting identity-similar edges and source-conflict edges in the candidate object subgraph; constructing a first adjacency matrix based on the topological associations formed by identity-similar edges between each candidate node; constructing a second adjacency matrix based on the topological associations formed by source-conflict edges between each candidate node; superimposing and concatenating the first and second adjacency matrices in the topological space dimension to obtain a multi-relationship graph structure matrix; and aligning and combining the temporal node feature matrix and the multi-relationship graph structure matrix in dimensions to generate a multidimensional temporal feature tensor containing time series dimension, node feature dimension, and graph topology dimension.

[0036] Specifically, after the preliminary construction of the candidate object subgraph is completed through the preceding steps, network topology data composed of nodes and dual relation edges is obtained. However, such unstructured graph data is difficult to be directly ingested by subsequent graph computation tasks or deep learning models. This step aims to solve the technical problem that complex unstructured graph network topology and node temporal attributes cannot be directly applied to high-concurrency matrix operations by the model. Through feature dimensionality reduction and tensor mapping techniques, the graph elements are transformed into a standard high-dimensional tensor data base that conforms to the computation specifications.

[0037] Candidate nodes and relation edges in the candidate object subgraph are mapped to multi-dimensional temporal feature tensors. In practice, the attribute features and timestamp records contained in each candidate node are extracted, the attribute features are converted into numerical vectors and arranged in chronological order according to the timestamp records to construct a temporal node feature matrix. The multi-dimensional temporal evidence units associated with the candidate nodes are analyzed, and character attribute features such as candidate object identifiers are extracted. Through numerical encoding algorithms such as feature hashing, discrete text attributes are mapped into fixed-length, computer-readable dense numerical vectors. Then, using the extracted timestamp records as a time sorting scale, the numerical vectors belonging to the same candidate node are stacked in chronological order of events. Thus, a three-dimensional temporal node feature matrix is ​​constructed with the total number of candidate nodes as the first dimension, the length of the temporal record as the second dimension, and the length of the attribute feature vector as the third dimension. This matrix fully depicts the dynamic sequence of the evolution of the attribute features of each candidate node over time.

[0038] Simultaneously, extract each identity-similar edge and each source conflict edge from the candidate object subgraph. Construct a first adjacency matrix based on the topological associations formed by identity-similar edges between candidate nodes, and construct a second adjacency matrix based on the topological associations formed by source conflict edges between candidate nodes. Here, the adjacency matrix is ​​a mathematical tool in graph theory algorithms specifically used to represent the connectivity state between network nodes in the form of a square matrix. Initialize two two-dimensional square matrices with all zeros, using the total number of candidate nodes in the candidate object subgraph as the row and column dimensions. For the first adjacency matrix, traverse all candidate node pairs in the subgraph. If it is determined that there is an identity-similar edge between two candidate nodes, assign a value of 1 to the row and column intersection coordinate position of the first adjacency matrix corresponding to the two node indices; otherwise, keep the value 0. Similarly, for the second adjacency matrix, check whether there is a source conflict edge between each node pair. If there is, assign a value of 1 to the row and column intersection coordinate position of the second adjacency matrix; otherwise, keep it as 0.

[0039] Subsequently, the first and second adjacency matrices are superimposed and concatenated along the topological space dimension to obtain a multi-relationship graph structure matrix. The two two-dimensional square matrices containing connected state values ​​are then superimposed in parallel along the channel dimension representing the relationship type. This stacking and concatenation operation upgrades the originally independent two-dimensional matrices into a three-dimensional multi-relationship graph structure matrix containing two relationship channels, ensuring that the two basic skeleton structures of similar topologies and logically conflicting topologies can be carried without interference within the same data structure. Finally, the temporal node feature matrix and the multi-relationship graph structure matrix are dimensionally aligned and combined to generate a multi-dimensional temporal feature tensor containing time series dimension, node feature dimension, and graph topology dimension. Here, the multi-dimensional temporal feature tensor does not refer to a single dense matrix in the traditional sense, but rather a composite graph tensor object used in the graph computing framework to represent heterogeneous graph structured data.

[0040] When performing specific mapping and combination, the first step is to perform a dimension alignment operation, which involves extracting the temporal node feature matrix. The dimension is typically the node index arrangement order in [number of nodes, time series length, feature vector dimension]. This order is then used to strictly recalibrate the multi-relationship graph structure matrix, whose dimension is typically the row and column index in [number of relationship channels, number of nodes, number of nodes]. This ensures that in tensor slicing operations, the same index coordinates absolutely point to the same entity candidate node in the real world. After completing the index space alignment, a combination and encapsulation operation is performed: instead of a simple arithmetic concatenation, the data construction mechanism of the graph deep learning framework is invoked. The temporal node feature matrix is ​​used as the node feature load input of the tensor, and the multi-relationship graph structure matrix is ​​used as the edge topology load input of the tensor. The two are encapsulated and constructed into a unified high-order composite data structure, i.e., generating a multi-dimensional temporal feature tensor. After mapping, this multi-dimensional temporal feature tensor not only preserves the spatiotemporal evolution trajectory of the personnel file status without loss, but also clearly depicts the network topology from the dual perspectives of identity similarity and source tracing conflict, providing a standard and regular computational foundation for subsequent graph attention networks.

[0041] After obtaining the multidimensional temporal feature tensor, the graph attention network needs to perform in-depth feature interaction calculations on the edge relationships between nodes. Specifically, based on the preset feature decoding logic, it extracts the identity similarity features, time overlap features, role mutual exclusion features, and source credibility features between candidate nodes from the multidimensional temporal feature tensor. Since the multidimensional temporal feature tensor is encapsulated from underlying multidimensional data, the above extraction process is actually a slice reading and cross-calculation of different dimensional subspaces within the tensor: First, for the identity similarity features and role mutual exclusion features, the "graph topology dimension" (i.e., the slice of the multi-relationship graph structure matrix) in the multidimensional temporal feature tensor is parsed; the edge weight values ​​corresponding to the "identity similarity edge" channel in the graph topology dimension are read and directly used as the identity similarity features. This weight value carries the Jaccard similarity calculated in the previous step; at the same time, the edge weight values ​​corresponding to the "identity similarity edge" channel in the graph topology dimension are read. The topological connectivity of the "source conflict edge" channel is analyzed, and combined with the mutual exclusion attribute field in the node feature dimension, the role mutual exclusion feature is mapped to represent whether there is an exclusion relationship between the two in terms of business roles. Secondly, for the time overlap feature, the "time series dimension" in the multi-dimensional time series feature tensor is analyzed. Based on the edge index provided by the graph topology dimension, the effective time window segments of the two candidate nodes that generate the connectivity relationship are extracted in the time series dimension. By calculating the intersection-union ratio or overlap duration of these two time window segments, the time overlap feature of the two is derived and extracted. Finally, for the source credibility feature, the "node feature dimension" in the multi-dimensional time series feature tensor is analyzed, the data source identifier embedded in the connected candidate nodes, such as the source system number, is extracted, and the source right reset information table preset by the system is queried to map the identifier to the corresponding source credibility feature to evaluate the authority of the underlying archive data.

[0042] S5. Input the multi-dimensional temporal feature tensor into the graph attention network. During the graph attention operation, use the scene constraint matrix set to identify conflicting and merging edges that violate the scene constraint matrix set within the identity similarity edges of the relationship edges, and apply a loss penalty to the conflicting and merging edges to calculate the target object fusion probability, such as... Figure 3 The diagram shows the process of fusing multidimensional temporal feature tensor mapping with graph attention network.

[0043] The process involves inputting a multidimensional temporal feature tensor into a graph attention network. During graph attention operations, a scene constraint matrix is ​​used to identify conflicting and merging edges that violate the scene constraint matrix within identity similarity edges in relational edges. A loss penalty is applied to these conflicting and merging edges to calculate the fusion probability of the target object. This includes extracting identity similarity features, temporal overlap features, role mutual exclusion features, and source credibility features from the multidimensional temporal feature tensor, and parsing the time slice sequence and role attribute data corresponding to the candidate nodes. Positive weight coefficients are assigned to identity similarity features and source credibility features, while negative weight coefficients are assigned to role mutual exclusion features. These are then combined with the temporal overlap feature for weighted grouping. The comprehensive attention weight values ​​between candidate nodes are calculated. Using the role mutual exclusion matrix, the role attribute data corresponding to the candidate nodes in overlapping time slice sequences are logically compared, and identity similarity edges that violate the mutual exclusion rules in the role mutual exclusion matrix are judged as conflict merging edges. A penalty term that is directly proportional to the magnitude of the role mutual exclusion feature value is constructed and added to the loss function of the graph attention network as a loss penalty. The multi-dimensional temporal feature tensor is aggregated through multiple rounds of network iteration based on the comprehensive attention weight values ​​to generate fusion representation vectors for candidate nodes. The vector similarity between the fusion representation vectors corresponding to different candidate nodes is calculated, and the vector similarity is normalized and used as the fusion probability of the target object.

[0044] The process includes, after calculating the target object fusion probability, extracting the role mutual exclusion feature with negative weight coefficients and the penalty term as negative association data, and calculating the numerical contribution of the negative association data to the decrease in the target object fusion probability; generating a conflict explanation vector based on the numerical contribution, and outputting the business conflict reasons that prevent different candidate nodes from merging through the conflict explanation vector.

[0045] Specifically, in the preliminary steps, graph elements containing multiple relationships and dynamic attributes have been transformed into standard high-dimensional feature tensors. However, in actual scenarios of personnel holographic file fusion, if graph network aggregation is based solely on the mathematical similarity of node attributes, it is easy to overlook the implicit mutual exclusion of business logic within a specific time slice. For example, the same individual cannot appear at two locations that are physically far apart at the same time, or simultaneously possess specific business identities that cannot coexist. This step aims to solve the technical problem that existing graph network fusion algorithms lack the ability to perceive implicit business conflict constraints, which easily leads to abnormal erroneous fusion of nodes and makes it difficult to provide a quantitative explanation for the reasons for fusion failure.

[0046] A multidimensional temporal feature tensor is input into a graph attention network (Graph Attention Network), which performs graph attention operations. During these operations, a scene constraint matrix is ​​used to identify conflicting and merging edges that violate the scene constraint matrix within identity-similar edges. A loss penalty is applied to these conflicting and merging edges to calculate the target object fusion probability. A Graph Attention Network is a deep learning model specifically designed for processing graph-structured data. Its model structure includes multiple layers of self-attention mechanisms and feedforward neural networks. The input data is a multidimensional temporal feature tensor, and the output data consists of the target object fusion probability and conflict explanation vector between nodes. The training process is based on a graph dataset containing historical known fusion archives and known conflict archives. The basic loss function is the cross-entropy that minimizes the error between the predicted probability and the true label. The backpropagation algorithm iteratively updates the network's internal parameters. In the specific implementation of graph attention operations, the input multidimensional temporal feature tensor is first sliced ​​and parsed according to the channel dimension. Feature channel data representing node attributes is read from the tensor, and the feature vector distance between node pairs is calculated to extract identity similarity features. Channel data representing timestamp sequences is read from the tensor, and the intersection ratio of node pair lifecycles is calculated to extract temporal overlap features. Role encoding channel data is read from the tensor, and role mutual exclusion features are extracted based on a pre-defined dictionary mapping. Source credibility features are extracted based on the historical accuracy of the source system traced by the data source identifier carried in the tensor. Simultaneously, the temporal dimension of the tensor is decoded in reverse, parsing out the time slice sequence and role attribute data corresponding to each candidate node.

[0047] Subsequently, a weighted combination calculation of multi-dimensional features is performed. Specifically, this calculation process is not a simple linear accumulation, but rather employs a non-linear aggregation strategy based on a time-aware gating mechanism. The time overlap feature serves as a dynamic adjustment factor in the generation of the final edge weights. First, the basic interaction score between candidate nodes is calculated: the identity similarity feature and the source credibility feature are multiplied by a preset positive weight coefficient, representing positive gain, and the role mutual exclusion feature is multiplied by a preset negative weight coefficient, representing conflict penalty. These three results are then linearly summed to obtain the basic interaction score. Second... A dynamic adjustment calculation is performed by introducing temporal overlap features: The extracted temporal overlap features are input into a preset normalization mapping function, such as the Sigmoid function or max-min normalization, and mapped to a time-gated scalar with values ​​in the interval (0, 1). Finally, the basic interaction score is multiplied by this time-gated scalar, i.e., a dot product operation. Under this mechanism, the specific logic of the temporal overlap features in the calculation is as follows: when the temporal overlap between two candidate nodes is high, i.e., the scalar approaches 1, the basic interaction score is fully preserved or even amplified; when the temporal overlap is extremely low or nonexistent, the basic interaction score is reduced. When the scalar approaches zero, a time decay penalty is applied to the basic interaction score, significantly weakening the attention weights of the two nodes in the message passing process of the graph neural network. This weighted combination calculation ultimately outputs a time-sensitive integrated initial edge weight between nodes. To eliminate differences in physical dimensions and numerical scales between different features, a maximum-minimum normalization algorithm is used to uniformly map the extracted feature values ​​to a dimensionless range of 0 to 1. After unifying the dimensions, positive constants with values ​​greater than 0 are assigned to the identity similarity feature and source credibility feature that facilitate node merging. The weights are negative constants with values ​​less than 0 assigned to the role mutual exclusion feature that hinders node merging. During calculation, the normalized identity similarity, source credibility, and role mutual exclusion are multiplied by their corresponding weight coefficients and summed to obtain the basic attention score. Then, the basic attention score is multiplied by the normalized time overlap feature value. The principle of this operation is that only when two nodes have an intersection on the time axis do the positive and negative effects of their attributes have real business constraint significance. The final output product is the comprehensive attention weight value between candidate nodes.

[0048] In the forward computation and parameter optimization of the graph attention network, a role mutual exclusion matrix is ​​used to logically compare the role attribute data of candidate nodes in overlapping time slice sequences. Edges with similar identities that violate the mutual exclusion rules in the role mutual exclusion matrix are identified as conflicting and merged edges. The role mutual exclusion matrix is ​​the core two-dimensional lookup table in the scene constraint matrix group, used to define the legality of coexistence of different business roles in the same time dimension. By comparing the time slice sequences of two connected candidate nodes, the overlapping time period is extracted, and the role attribute data of both parties within that time period is obtained. The role mutual exclusion matrix is ​​then queried using these two role attribute codes as row and column indices. If the corresponding coordinate value in the matrix is ​​1, it indicates the existence of... Business mutual exclusion; at this point, even if two nodes have similar identities due to similar attributes, they are still forcibly marked as conflict merging edges; based on this, a penalty term that is directly proportional to the magnitude of the role mutual exclusion feature value is constructed and added to the loss function of the graph attention network as a loss penalty; a fixed penalty multiplier coefficient is set and multiplied by the aforementioned normalized role mutual exclusion feature value to form a penalty term. When calculating the total network loss, this penalty term is added to the basic cross-entropy loss function in an additive form; by applying the loss penalty, the network will actively and greatly reduce the aggregation attention weights of such logically conflicting nodes when optimizing parameters, thereby suppressing erroneous fusion from the algorithm mechanism.

[0049] Based on the comprehensive attention weight value, the multidimensional temporal feature tensor is iteratively aggregated through multiple rounds of networks to generate fusion representation vectors for candidate nodes. Under the topological connectivity constraints of the multi-relationship graph structure matrix, a self-attention layer is used to weight and sum the features of neighboring nodes according to the proportion of the comprehensive attention weight value, and then concatenate them with the features of the central node. After mapping through multiple layers of nonlinear activation functions, a high-dimensional fusion representation vector that integrates local and global contextual information is output. Next, the vector similarity between the fusion representation vectors corresponding to different candidate nodes is calculated, and the vector similarity is normalized and used as the fusion probability of the target object. The cosine similarity formula is used to calculate the cosine value of the angle between the two fusion representation vectors in space as the initial similarity. Then, the normalized exponential function is applied to map the initial similarity to a probability distribution value between 0 and 1 to obtain the final fusion probability of the target object.

[0050] After calculating the fusion probability of the target object, the mutual exclusion feature of the role with negative weight coefficient and the penalty term are extracted as negative association data, and the numerical contribution of the negative association data to the decrease in the fusion probability of the target object is calculated. The absolute value of the sum of the negative feature weight values ​​and the loss penalty term values ​​generated in the operation is retrieved and divided by the sum of the absolute values ​​of all positive feature weight values, negative feature weight values ​​and penalty terms. The resulting ratio is the numerical contribution. A conflict explanation vector is generated by concatenating the numerical contribution and outputting the business conflict reasons that prevent different candidate nodes from merging through the conflict explanation vector. The role code, overlapping time period and numerical contribution of the conflict are concatenated into strings according to the preset bit width to generate a structured conflict explanation vector. The internal mapping dictionary is called to parse and translate it into natural language text, outputting the underlying reasons such as the existence of specific business mutual exclusion during the time overlap period.

[0051] The aforementioned implementation constructs a dynamic loss penalty mechanism based on temporal overlap and business scenario matrices within the graph attention model. Combined with reverse attribution calculation, this addresses the technical challenges of existing graph fusion technologies, such as frequent business logic violations and a lack of interpretability. By extracting multi-dimensional temporal and topological features and performing positive and negative weighted calculations under a unified scale, the graph attention network acquires the ability to accurately quantify factors that promote and hinder fusion. The introduction of a role mutual exclusion matrix for logical comparison and triggering dynamic penalties in the loss function fundamentally prevents erroneous node merging caused by relying solely on mathematical similarity, ensuring the rigor and business compliance of the holographic archival graph generation results. Simultaneously, by calculating the numerical contribution of negatively correlated data and generating conflict explanation vectors, the black-box graph neural network is endowed with interpretable output capabilities, providing quantitative evidence support down to the time slice and specific rules for subsequent manual review and archival data governance.

[0052] S6. Based on the fusion probability of the target object, map the relation edges in the candidate object subgraph to relation edges at different confidence levels, and generate a multi-layer object relation subgraph as an archive relation graph.

[0053] Specifically, based on the target object fusion probability, the relational edges in the candidate object subgraph are mapped to relational edges of different confidence levels to generate a multi-layer object relational subgraph. This includes: comparing the target object fusion probability with preset high confidence thresholds and low confidence thresholds; if the target object fusion probability is greater than the high confidence threshold, then the candidate node pairs corresponding to the target object fusion probability in the candidate object subgraph are merged into a unified object node, and the unified object node inherits the multi-dimensional temporal evidence unit and relational edges of the candidate node pairs before merging; if the target object fusion probability is between the low confidence threshold and the high confidence threshold, then the candidate node pairs corresponding to the target object fusion probability in the candidate object subgraph are retained, and the relational edges between the candidate node pairs are marked as edges to be verified.

[0054] The process of generating a multi-layered object relationship subgraph includes: parsing the multi-dimensional temporal evidence units inherited by unified object nodes to extract the data source authority, temporal continuity features, cross-verification status, and evidence conflict status corresponding to the relationship edges; based on the data source authority, temporal continuity features, cross-verification status, and evidence conflict status, mapping the relationship edges inherited by unified object nodes to strong fact edges in the core layer, stage fact edges in the auxiliary layer, pending verification edges in the pending review layer, and conflict retention edges in the abnormal layer, to jointly construct a multi-layered object relationship subgraph; wherein, relationship edges with multi-source cross-verification status are mapped to strong fact edges, relationship edges with a single data source or time discontinuity status are mapped to stage fact edges, relationship edges lacking independent verification basis are mapped to pending verification edges, and relationship edges containing evidence conflict statuses that cannot be determined as right or wrong are mapped to conflict retention edges.

[0055] Specifically, in the preliminary step, the graph attention network operation has already output the quantitative value of the fusion probability of target objects between candidate nodes; however, obtaining only the probability value cannot directly complete the physical merging of the underlying topology of the graph network; at the same time, if the massive amount of related data inherited by the nodes after fusion is not filtered and classified according to quality level, it will lead to the mixing of solid and objective archival facts with doubtful and contradictory data in the same dimension of the flat network, making it difficult to support subsequent accurate background checks and clue analysis; this step aims to solve the technical problems of existing graph generation methods lacking a secure topology reconstruction mechanism based on quantitative confidence and lacking multi-dimensional evidence quality hierarchical management of graph relationships.

[0056] Based on the target object fusion probability, the relationship edges in the candidate object subgraph are mapped to relationship edges of different confidence levels, generating a multi-layer object relationship subgraph. In practice, the target object fusion probability is compared numerically with preset high-confidence and low-confidence thresholds. To ensure the mutual exclusion and completeness of the judgment logic, the high-confidence threshold is set to a constant value that is strictly greater than the low-confidence threshold. Since the target object fusion probability is a dimensionless value between 0 and 1 after normalization, it is directly compared with the threshold in the same range. If the target object fusion probability is strictly greater than the high-confidence threshold, the corresponding candidate node pair is determined to point to the same real-world physical entity with a very high probability. Then, a graph topology reconstruction operation is performed, merging the candidate node pairs with the corresponding target object fusion probabilities in the candidate object subgraph into a unified pair. Image nodes; In the graph computing engine, merging is completed by rewriting the globally unique identifier of the nodes, and the unified object node fully inherits the multi-dimensional temporal evidence units and relational edges of the candidate node pairs before merging; All in-degree and out-degree topological edges originally connected to the two independent nodes are re-bridged to the unified object node, and the underlying supporting files originally belonging to the two nodes are integrated into the evidence set of the unified object node; If the target object fusion probability is greater than or equal to the low confidence threshold and less than or equal to the high confidence threshold, that is, between the low confidence threshold and the high confidence threshold, a conservative non-merging strategy is adopted, the original appearance of the candidate node pairs corresponding to the target object fusion probability in the candidate object subgraph is preserved, and the correlation edges between the candidate node pairs are marked as edges to be verified in the graph structure attributes to freeze their state and hand them over to subsequent manual review.

[0057] To generate a multi-layered object relationship subgraph, data parsing is performed on the multi-dimensional temporal evidence units inherited by unified object nodes to extract the data source authority, temporal continuity features, cross-verification status, and evidence conflict status corresponding to the relationship edges. Since each relationship edge in the graph carries a temporal evidence chain proving its existence at its underlying level, the evidence records attached to a specific relationship edge are traversed and read. For the extraction process, the data source identifier field in the evidence record is read, and the data source authority value corresponding to that record is obtained and extracted by querying a preset data quality scoring dictionary. Multiple pieces of evidence under the same relationship edge are arranged in chronological order of their timestamps. The system records and calculates the time difference between adjacent records. If the maximum time difference is lower than a preset continuity interval, it is extracted as a continuous temporal feature with continuity; otherwise, it is extracted as a discontinuous temporal feature. It then uses a natural language semantic encoding algorithm to calculate the vector distance between the descriptive texts of evidence from different sources. If the vector distance is lower than a preset semantic similarity threshold (i.e., the content is consistent), it extracts a state with multi-source cross-verification. If there are records with mutually exclusive semantic contradictions within the same time slice, and the authority values ​​of the corresponding data sources are equal and cannot be automatically arbitrated by the system, it extracts a state of conflicting evidence containing contradictions.

[0058] After extracting the above features, based on data source authority, temporal continuity, cross-verification status, and evidence conflict status, the relationship edges inherited by unified object nodes are mapped to strong fact edges in the core layer, stage fact edges in the auxiliary layer, unverified edges in the pending layer, and conflict-preserving edges in the anomaly layer, respectively, to jointly construct a multi-layered object relationship subgraph. When performing mapping and hierarchical routing, priority judgment logic is followed to prevent state overlap; conflict states are prioritized for investigation, and relationship edges containing evidence conflicts that cannot be determined are directly mapped to conflict-preserving edges and isolated and distributed in the anomaly layer; after excluding conflict states, relationship edges with multi-source cross-verification status are mapped to strong fact edges, falling into the core layer of the graph for representation. The system identifies key relationships in personnel files with irrefutable certainty. For relationship edges that are conflict-free and lack multi-source verification, their continuity is examined. Relationship edges that are provided by a single data source or have temporal discontinuities are mapped to stage fact edges and placed in the auxiliary layer, indicating that the relationship is partially supported by facts but lacks multi-dimensional completeness. Finally, relationship edges with data source authority values ​​below the passing threshold and lacking independent verification are mapped to edges to be verified and placed in the review layer. Through the above feature evaluation and routing mapping mechanism, the complex network relationships that were originally on a single plane are deconstructed, and a multi-layer object relationship subgraph composed of four independent and related view planes—core, auxiliary, review, and anomaly—is successfully constructed.

[0059] The aforementioned implementation introduces a graph node secure merging mechanism based on probabilistic dual thresholds. Combined with priority judgment and analysis of the underlying evidence chain state, it achieves automatic decoupling mapping of confidence levels in complex graph relationships, solving the technical challenge of tiered control over data certainty after holographic graph fusion. By setting strict dual threshold intervals for high and low confidence levels, it effectively isolates questionable nodes while ensuring automated and accurate merging of high-probability nodes in the topology, curbing data corruption caused by forced and crude fusion from the underlying computation stage of graph generation. Furthermore, by penetrating and analyzing the multi-dimensional temporal evidence units attached to the underlying relation edges, it extracts and calculates four key aspects: authority, continuity, corroboration, and conflict. The core metrics break through the limitations of homogeneous and flattened storage of relation edges in traditional knowledge graphs. They scientifically, accurately, and mutually exclusively distribute complex holographic archive relationships to the core layer, auxiliary layer, pending review layer, and anomaly layer. This multi-level decoupled graph structure design enhances the graph database's compatibility with massive amounts of heterogeneous and contradictory archive information. In actual business queries and case reasoning applications, the caller can, based on the rigor requirements of specific scenarios, penetrate and call data at different confidence levels as needed. This ensures the absolute accuracy of core business decisions while preserving ample auxiliary space for expanding clue mining, significantly improving the practical management efficiency and flexible application capabilities of the holographic archive system.

[0060] Through the coordination of steps S1 to S6 above, this application improves the practical effectiveness and data control security of personnel holograms in complex business analysis and evidence tracing scenarios.

[0061] Example 2: The above describes a method for generating a holographic personnel file relationship map according to an embodiment of this application. The following describes a device for generating a holographic personnel file relationship map according to an embodiment of this application. Figure 4 As shown, a personnel holographic file relationship map generation device in this application embodiment includes: The evidence conversion module is used to acquire the archival record data of the target object and convert the archival record data into multi-dimensional time-series evidence units containing time intervals and evidence status, thus forming the evidence set of the target object.

[0062] The constraint construction module is used to build a scenario constraint matrix group according to preset business rules. The scenario constraint matrix group includes a role mutual exclusion matrix and a time overlap tolerance matrix.

[0063] The clustering analysis module is used to extract key attributes from the evidence set of the target object as blocking keys, and to cluster the evidence set of the target object to obtain candidate object clusters based on the blocking keys. The multidimensional time-series evidence units within the candidate object clusters are used as candidate nodes. The attribute feature similarity of each candidate node is calculated, and a candidate object subgraph containing relation edges is generated based on the attribute feature similarity. The relation edges include identity similarity edges and source conflict edges.

[0064] The feature mapping module is used to map candidate nodes and relation edges in the candidate object subgraph into multidimensional temporal feature tensors.

[0065] The fusion computing module is used to input multi-dimensional temporal feature tensors into the graph attention network. During the graph attention operation, the scene constraint matrix group is used to identify conflicting and merging edges that violate the scene constraint matrix group in the identity similarity edges of the relation edges, and a loss penalty is applied to the conflicting and merging edges to calculate the fusion probability of the target object.

[0066] The hierarchical graph module is used to map the relational edges in the candidate object subgraph to relational edges at different confidence levels based on the fusion probability of the target object, thereby generating a multi-layer object relational subgraph as an archive relational graph.

[0067] Through the collaborative efforts of the aforementioned components, this application further enhances the practical effectiveness and data control security of personnel holographic mapping in complex business analysis and evidence tracing scenarios.

[0068] In summary, this application provides a method and apparatus for generating a holographic personnel file relationship graph, fundamentally reconstructing the graph fusion paradigm of multi-source heterogeneous file data. This scheme quantifies discrete humanistic and physical common sense and business rules into a set of scenario constraint matrices and deeply embeds them as dynamic penalty terms into the loss function of the graph attention network. This enables the deep learning model to not only have a powerful ability to capture topological features but also to adhere to the common sense bottom line of real business logic. At the same time, with the combination of the pre-blocking block dimensionality reduction technology and the post-confidence dual threshold decoupling mechanism, this scheme successfully routes and reconstructs massive amounts of file association data with difficulty in distinguishing between true and false data into a multi-layer relationship topology with core layer, auxiliary layer, pending review layer and abnormal layer stacked. This application improves the practical effectiveness and data control security of holographic personnel graphs in complex business judgment and evidence tracing scenarios.

[0069] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0070] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0071] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for generating a holographic personnel file relationship map, characterized in that, The method includes: Acquire the archival record data of the target object, and convert the archival record data into multi-dimensional temporal evidence units containing time intervals and evidence status to form the evidence set of the target object; A scenario constraint matrix group is constructed based on preset business rules. The scenario constraint matrix group includes a role mutual exclusion matrix and a time overlap tolerance matrix. Key attributes are extracted from the evidence set of the target object as blocking keys. Based on the blocking keys, the evidence set of the target object is clustered to obtain candidate object clusters. Multidimensional time-series evidence units within the candidate object clusters are used as candidate nodes. The attribute feature similarity of each candidate node is calculated. Based on the attribute feature similarity, a candidate object subgraph containing relation edges is generated. The relation edges include identity similarity edges and source conflict edges. Map candidate nodes and relation edges in the candidate object subgraph to a multidimensional temporal feature tensor; The multidimensional temporal feature tensor is input into the graph attention network. During the graph attention operation, the scene constraint matrix group is used to identify conflicting and merging edges that violate the scene constraint matrix group in the identity similarity edges of the relation edges, and a loss penalty is applied to the conflicting and merging edges to calculate the fusion probability of the target object. Based on the fusion probability of the target object, the relation edges in the candidate object subgraph are mapped to relation edges of different confidence levels, generating a multi-layer object relation subgraph as an archive relation graph.

2. The method for generating a holographic personnel file relationship map according to claim 1, characterized in that, The archival record data is converted into multi-dimensional temporal evidence units containing time intervals and evidentiary status, forming the evidence set of the target object, including: Extract the start and end times of the archival record data to construct a time interval, and extract the candidate object identifiers, event types, associated objects, and source numbers from the archival record data; Obtain external verification data from different business systems that contains similarity features to the candidate object identifiers with a similarity greater than or equal to a preset identifier threshold; Cross-validate archival records with external verification data to determine the status of the evidence; The six-element data set is generated by combining the time interval, evidence status, candidate object identifier, event type, associated object, and source number. This data set serves as a multi-dimensional temporal evidence unit, and multiple multi-dimensional temporal evidence units are aggregated to form the target object evidence set.

3. The method for generating a holographic personnel file relationship map according to claim 1, characterized in that, Construct a scenario constraint matrix group based on preset business rules, including: Parse the preset business rules and extract the set of business roles and event types; Determine the logical exclusion relationship between any two business roles in the set of business roles, and assign binary values ​​to the corresponding coordinate positions of the initial two-dimensional square matrix according to the logical exclusion relationship to generate a role mutual exclusion matrix; Extract the allowed overlap time thresholds set for the event type set in the preset business rules, calculate the time tolerance coefficient between each pair of event types based on the allowed overlap time thresholds, and use the time tolerance coefficients to fill matrix elements to construct a time overlap tolerance matrix; The joint role mutual exclusion matrix and the time overlap tolerance matrix constitute the scene constraint matrix group.

4. The method for generating a holographic personnel file relationship map according to claim 1, characterized in that, Calculate the attribute feature similarity of each candidate node, and generate a candidate object subgraph containing relation edges based on the attribute feature similarity, including: Extract at least one feature from the document hash value, pinyin code, contact information time period, organizational path, and address grid of each multi-dimensional temporal evidence unit in the target object evidence set, and use it as a blocking key; The evidence set of the target object is divided and aggregated according to the key value matching results of the blocking key to obtain multiple independently divided candidate object clusters; Each multidimensional temporal evidence unit within the same candidate object cluster is mapped to a candidate node, and the candidate nodes are compared pairwise to calculate the similarity of attribute features. At the same time, the data source identifier and mutually exclusive attributes contained in the multidimensional temporal evidence unit corresponding to each candidate node are analyzed. Establish identity similarity edges between two candidate nodes whose attribute feature similarity is greater than or equal to a preset similarity threshold, and establish source conflict edges between two candidate nodes whose attribute feature similarity is greater than or equal to a preset similarity threshold and whose data source identifiers are different or whose mutually exclusive attributes conflict. The candidate object subgraph is formed by combining candidate nodes and relation edges.

5. The method for generating a holographic personnel file relationship map according to claim 1, characterized in that, The candidate nodes and relation edges in the candidate object subgraph are mapped to a multidimensional temporal feature tensor, including: Extract the attribute features and timestamp records contained in each candidate node, convert the attribute features into numerical vectors and arrange them in chronological order according to the timestamp records to construct a time-series node feature matrix; Extract each identity-similar edge and each source conflict edge from the candidate object subgraph. Construct a first adjacency matrix based on the topological associations formed by identity-similar edges between each candidate node, and construct a second adjacency matrix based on the topological associations formed by source conflict edges between each candidate node. By superimposing and concatenating the first adjacency matrix and the second adjacency matrix in the topological space dimension, a multi-relationship graph structure matrix is ​​obtained; The temporal node feature matrix and the multi-relationship graph structure matrix are dimensionally aligned and combined to generate a multi-dimensional temporal feature tensor containing time series dimension, node feature dimension and graph topology dimension.

6. The method for generating a holographic personnel file relationship map according to claim 1, characterized in that, A multidimensional temporal feature tensor is input into a graph attention network. During the graph attention operation, a scene constraint matrix is ​​used to identify conflicting and merging edges that violate the scene constraint matrix within identity similarity edges in relational edges. A loss penalty is applied to these conflicting and merging edges, and the fusion probability of the target object is calculated, including: The identity similarity feature, time overlap feature, role mutual exclusion feature and source credibility feature between candidate nodes are extracted from the multidimensional temporal feature tensor, and the time slice sequence and role attribute data corresponding to the candidate nodes are parsed out. The identity similarity feature and source credibility feature are assigned positive weight coefficients, the role mutual exclusion feature is assigned negative weight coefficients, and the time overlap feature is combined for weighted combination calculation to obtain the comprehensive attention weight value between candidate nodes; using the role mutual exclusion matrix, the role attribute data corresponding to the candidate nodes in the overlapping time slice sequence are logically compared, and identity similarity edges that violate the mutual exclusion rules in the role mutual exclusion matrix are judged as conflict merging edges. A penalty term that is directly proportional to the magnitude of the mutual exclusion feature is constructed and added to the loss function of the graph attention network as a loss penalty. The multidimensional temporal feature tensor is aggregated through multiple rounds of network iteration based on the comprehensive attention weight value to generate a fusion representation vector for the candidate nodes. The vector similarity between the fusion representation vectors corresponding to different candidate nodes is calculated and normalized to serve as the fusion probability of the target object.

7. The method for generating a holographic personnel file relationship map according to claim 6, characterized in that, After calculating the fusion probability of the target object, the following is also included: Extract the role mutual exclusion feature with negative weight coefficient and the penalty term as negative association data, and calculate the numerical contribution of the negative association data to the decrease in the fusion probability of the target object; generate a conflict explanation vector by concatenating the numerical contribution, and output the business conflict reasons that cause different candidate nodes to fail to merge through the conflict explanation vector.

8. The method for generating a holographic personnel file relationship map according to claim 1, characterized in that, Based on the fusion probability of the target object, the relationship edges in the candidate object subgraph are mapped to relationship edges of different confidence levels, generating a multi-layer object relationship subgraph, including: The fusion probability of the target object is compared with the preset high confidence threshold and low confidence threshold respectively. If the fusion probability of the target object is greater than the high confidence threshold, the candidate node pairs with the corresponding fusion probability of the target object in the candidate object subgraph are merged into a unified object node, and the unified object node inherits the multi-dimensional temporal evidence unit and relation edge of the candidate node pair before merging. If the fusion probability of the target object is between the low confidence threshold and the high confidence threshold, retain the candidate node pairs in the candidate object subgraph that correspond to the fusion probability of the target object, and mark the relationship edges between the candidate node pairs as edges to be verified.

9. The method for generating a holographic personnel file relationship map according to claim 8, characterized in that, Generate a multi-level object relationship subgraph, including: Data parsing is performed on the multi-dimensional temporal evidence units inherited by the unified object node to extract the data source authority, temporal continuity features, cross-verification status, and evidence conflict status corresponding to the relation edges; Based on the authority of the data source, the continuity of time, the cross-verification status, and the evidence conflict status, the relationship edges inherited by the unified object node are mapped to strong fact edges in the core layer, stage fact edges in the auxiliary layer, unverified edges in the pending layer, and conflict-preserving edges in the abnormal layer, so as to jointly construct a multi-layer object relationship subgraph. Among them, relationship edges with multi-source cross-verification status are mapped to strong fact edges, relationship edges with single data source or time discontinuity status are mapped to stage fact edges, relationship edges lacking independent verification basis are mapped to unverified edges, and relationship edges containing evidence conflict status that cannot be determined as conflict-preserving edges are mapped to conflict-preserving edges.

10. A personnel holographic file relationship map generation device, used to implement the personnel holographic file relationship map generation method as described in any one of claims 1-9, characterized in that, The device includes: The evidence conversion module is used to acquire the archival record data of the target object and convert the archival record data into multi-dimensional temporal evidence units containing time intervals and evidence status, thus forming the evidence set of the target object. The constraint construction module is used to build a scenario constraint matrix group according to preset business rules. The scenario constraint matrix group includes a role mutual exclusion matrix and a time overlap tolerance matrix. The clustering analysis module is used to extract key attributes from the evidence set of the target object as blocking keys, and to cluster the evidence set of the target object based on the blocking keys to obtain candidate object clusters. The multidimensional time-series evidence units within the candidate object clusters are used as candidate nodes. The attribute feature similarity of each candidate node is calculated, and a candidate object subgraph containing relation edges is generated based on the attribute feature similarity. The relation edges include identity similarity edges and source conflict edges. The feature mapping module is used to map candidate nodes and relation edges in the candidate object subgraph into multidimensional temporal feature tensors; The fusion computing module is used to input multi-dimensional temporal feature tensors into the graph attention network. During the graph attention operation, the scene constraint matrix group is used to identify conflicting and merging edges that violate the scene constraint matrix group in the identity similarity edges of the relation edges, and a loss penalty is applied to the conflicting and merging edges to calculate the fusion probability of the target object. The hierarchical graph module is used to map the relational edges in the candidate object subgraph to relational edges at different confidence levels based on the fusion probability of the target object, thereby generating a multi-layer object relational subgraph as an archive relational graph.

Images