A method for time-series evidence archiving of real estate registration based on dynamic heterogeneous graphs
By constructing dynamic heterogeneous graphs and using multi-path coding technology, the problem of identifying pseudo-continuous links and abnormal paths in real estate registration archiving has been solved, achieving efficient case file-level archiving judgment and reducing the cost of manual review and the consumption of computing resources.
Patent Information
- Application Number
- CN202611546008.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-08
- Publication Date
- 2026-11-17
AI Technical Summary
The existing real estate registration and archiving technology lacks a unified propagation mechanism for cross-type relationships, making it difficult to identify pseudo-continuous archiving and abnormal paths, resulting in a large workload for manual review and high costs for anomaly tracing.
A dynamic heterogeneous graph-based approach is adopted to construct an evidence storage chain subgraph, a version evolution subgraph, and a cross-dosage reference subgraph. By combining node multi-way coding and hyperedge high-order propagation, dosage-level archiving decision representations are generated through temporal edge sorting learning and graph-comparative chain consistency learning.
It improves the identification and recall rate of pseudo-continuous links and time-reversed edges, reduces the workload of manual review and the cost of anomaly tracing, and reduces the consumption of computing resources and the latency of archiving processing.
Smart Images

Figure CN123118035A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of electronic archives and intelligent computing technology, and in particular to a method for time-series archiving of real estate registration based on dynamic heterogeneous graphs. Background Technology
[0002] Real estate registration archiving involves multiple types of objects and their temporal relationships, including application forms, case files, business entities, real estate units, certificates, and approval processes. Existing archiving detection methods mostly use single documents or single states as the judgment object, lacking a unified propagation mechanism for cross-type relationships, making it difficult to identify pseudo-continuous archives that appear continuous but have broken links. Although graph model matching schemes introduce node relationship representation, they do not take learning the sequence of events as a core task, making it difficult to distinguish between normal version continuation and rollback forged paths. Ordinary binary edges cannot bear the high-order business facts of a single registration action simultaneously involving multiple entities, multiple materials, and multiple approval processes, making it difficult to accurately locate cross-case erroneous references, material replacements, and broken inheritance chains. Manual reviewers need to reverse-engineer from massive logs and case file materials, resulting in a large workload for archiving verification and high costs for anomaly tracing. Summary of the Invention
[0003] The purpose of this invention is to provide a method for time-series evidence archiving of real estate registration based on dynamic heterogeneous graphs, to solve the problems of insufficient link-level judgment capability, difficulty in locating abnormal paths, and high manual review costs caused by simplifying real estate registration archiving to file-level detection, static graph matching, or rule verification in existing technologies. To achieve the above objective, this invention provides a method for time-series evidence archiving of real estate registration based on dynamic heterogeneous graphs, comprising: Step S1: Obtain the registration business objects and their temporal relationships, map the application form, file, business entity, real estate unit, certificate, approval process and time snapshot into a dynamic heterogeneous graph, and generate the evidence storage chain subgraph, version evolution subgraph and cross-file reference subgraph; Step S2: Perform structured field encoding, text digest encoding, and log time location encoding on various types of nodes in the dynamic heterogeneous graph, inject time encoding into various types of edges, and expand the relationship between a single business association of multiple entities and multiple materials into a hyperedge structure. Step S3: Generate a dossier-level archiving decision representation based on heterogeneous relation propagation, temporal edge sorting learning, graph comparison chain consistency learning, and key edge reconstruction learning; Step S4: Perform local subgraph incremental update on newly entered registration events, and output the archiving probability, chain integrity score, and abnormal edge location set based on the dossier-level archiving determination representation; Step S5: Generate archiving processing instructions based on the archiving probability, chain integrity score, and abnormal edge location set, and trigger the corresponding archiving action to generate structured archiving results.
[0004] Preferably, constructing a dynamic heterogeneous graph and generating a storage chain subgraph, a version evolution subgraph, and a cross-volume reference subgraph includes: reading the application form, file, right holder, obligor, business personnel, real estate unit, certificate, approval process, and time snapshot within the time window to be determined, to obtain a set of registration business objects; constructing a dynamic heterogeneous graph based on the set of registration business objects, and extracting relationship views corresponding to different business constraints from the dynamic heterogeneous graph using the file number as the anchor point; obtaining the storage chain graph, version evolution subgraph, and cross-volume reference subgraph according to the relationship views, and establishing a time-series index for the file, subject, and real estate unit.
[0005] Preferably, constructing the dynamic heterogeneous graph further includes: mapping application forms, file files, rights holders, obligors, business personnel, real estate units, certificates, approval processes, and time snapshots in the registration business object set to different node types to obtain a node set; mapping submission, review, reference, change, inheritance, cancellation, supplementation, version continuation, and hash anchoring actions to different edge types and recording the generation time of each edge to obtain an edge set; and establishing node type mapping, edge type mapping, and timestamp functions based on the node set, the edge set, and the generation time of each edge to obtain the dynamic heterogeneous graph.
[0006] Preferably, generating the evidence storage chain subgraph, version evolution subgraph, and cross-volume reference subgraph includes: using the volume number as an anchor point, extracting the closed-loop trace relationship between the application form, certificate, approval process, hash anchor, and volume from the dynamic heterogeneous graph to obtain the evidence storage chain subgraph; extracting the relationships related to supplementation, version continuation, change, and cancellation from the dynamic heterogeneous graph to obtain the version evolution subgraph; extracting the relationships related to inheritance, reference, and real estate unit sharing from the dynamic heterogeneous graph to obtain the cross-volume reference subgraph; when the current volume has multiple time snapshots, generating links between snapshots in the version evolution subgraph; when the current volume has only a single time snapshot, retaining the single snapshot view and skipping the construction of positive samples between snapshots to obtain the evidence storage chain subgraph, version evolution subgraph, and cross-volume reference subgraph.
[0007] Preferably, structured field encoding, text digest encoding, and log time location encoding are performed on various types of nodes in the dynamic heterogeneous graph, including: extracting the dossier number, business type, certificate type, status code, approval conclusion, version number, and field hash of each node, performing structured field encoding to obtain a structured field representation; extracting the application summary, certificate summary, and approval opinion, performing text digest encoding to obtain a text digest representation, wherein when the original material only contains full text, the title, summary paragraph, and approval opinion paragraph are trimmed from the original text according to a preset trimming rule; extracting the event occurrence sequence, the interval between adjacent business actions, and the time snapshot position, performing log time location encoding to obtain a log time location representation; and concatenating the structured field representation, text digest representation, and log time location representation to obtain a unified node representation.
[0008] Preferably, time encoding is injected into various types of edges, and the relationship between multiple entities and multiple materials associated with a single business transaction is expanded into a hyperedge structure. This includes: performing time encoding on various types of edges based on the time difference between the generation time of each edge and the current snapshot reference time to obtain an edge time representation. When log times are missing or conflicting, the time difference or additional missing time flag is generated according to adjacent known time interpolation, interval median replacement, standard time sequence template replacement, and master log priority rules; identifying multiple rights holders, multiple obligors, multiple supporting materials, and approval stages jointly involved in the same business action, establishing a higher-order association carrier, and obtaining a hyperedge structure; and performing higher-order propagation based on the unified node representation, edge time representation, and hyperedge structure to obtain a node representation with time relationships and higher-order business relationships.
[0009] Preferably, a dossier-level archiving decision representation is generated based on heterogeneous relationship propagation, temporal edge ranking learning, graph comparative chain consistency learning, and key edge reconstruction learning. This includes: based on the node representation with temporal and higher-order business relationships, heterogeneous relationship propagation is performed according to joint constraints of node type, edge type, and time encoding; read aggregation is performed on the dossier-related node set to obtain a dossier-level intermediate representation; legitimate sequential relationships are extracted from the real event chain of the same dossier as positive sample edges, and difficult negative sample edges are extracted from candidate relationships that are temporally close but have inconsistent business links or similar fields but conflicting entity mappings; temporal edge ranking learning is performed; different time snapshots of the same dossier during its legitimate evolution process are constructed as positive sample pairs, and snapshots with tampering, page missing, rollback operations, and erroneous cross-dosier references are constructed as difficult negative sample pairs; graph comparative chain consistency learning is performed, and masking reconstruction is performed on reference relationships, review relationships, and change relationships; joint learning is performed based on the archiving stage classification objective, ranking objective, comparison objective, and reconstruction objective to obtain the dossier-level archiving decision representation.
[0010] Preferably, the output of the archiving probability, chain integrity score, and abnormal edge location set includes: outputting the archiving probability based on the dossier-level archiving determination representation; calculating the key edge reconstruction consistency score, temporal sequence consistency score, cross-snapshot consistency score, and anchor closure score, and fusing them according to preset sub-item weights to obtain the chain integrity score; calculating the edge reconstruction deviation, edge sequence scoring deviation, and cross-snapshot consistency deviation for each candidate edge, and fusing them according to preset deviation fusion weights to obtain the abnormal edge comprehensive score; obtaining the abnormal edge location set through any of the following methods: filtering abnormal edges according to the abnormal edge comprehensive score and a preset abnormal threshold; and sorting and outputting the abnormal edges according to the abnormal edge comprehensive score when the number of edges that meet the threshold exceeds the preset upper limit of the number of abnormal edges.
[0011] Preferably, an archiving processing instruction is generated based on the archiving probability, chain integrity score, and abnormal edge location set, and a corresponding archiving action is triggered to generate a structured archiving result. This includes: when the archiving probability is higher than a preset archiving threshold, the chain integrity score is higher than a preset trusted chain integrity threshold, and the abnormal edge location set is empty or contains only edges with an abnormal edge comprehensive score lower than the preset abnormal threshold, an automatic archiving instruction is generated, and electronic case file encapsulation, archiving index generation, and evidence storage result writing are performed; when any of the following conditions are met, a manual review instruction is generated, and an abnormal path description and review queue are written: the archiving probability is lower than a preset low-trust threshold; the chain integrity score is lower than a preset low-trust chain integrity threshold; there are edges in the abnormal edge location set with order deviation exceeding a threshold and inconsistent cross-case source; in other cases, a delayed archiving instruction is generated, and a pending completion status is written; the structured archiving result containing case file number, archiving probability, chain integrity score, abnormal edge location set, archiving action type, and abnormal path summary is output.
[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above method.
[0013] The beneficial effects of this invention are as follows: By uniformly mapping the various types of objects and temporal relationships in registration business into a dynamic heterogeneous graph and extracting three types of subgraphs—evidence storage chains, version evolution, and cross-volume references—and combining node multi-path coding and high-order propagation of superedges, and through joint training via temporal edge sorting learning, graph-based chain consistency learning, and key edge reconstruction learning, incremental updates of the dossier-level archiving judgment representation are achieved. The above technical solution improves the identification recall rate of pseudo-continuous links and temporally reversed edges, reduces the workload of manual review and the cost of anomaly tracing, and reduces the computational resource consumption and archiving processing latency caused by full graph recalculation. Attached Figure Description
[0014] Figure 1 The present invention provides an overall flowchart of a time-series evidence archiving method for real estate registration based on dynamic heterogeneous graphs.
[0015] Figure 2 This is a schematic diagram of a time-series evidence archiving method for real estate registration based on dynamic heterogeneous graphs, provided by the present invention.
[0016] Figure 3 This is a flowchart of the dynamic graph construction and multi-view subgraph generation steps in an embodiment of the present invention.
[0017] Figure 4 This is a flowchart of the node / edge encoding and hyperedge construction steps in an embodiment of the present invention.
[0018] Figure 5 This is a flowchart of the multi-target learning and decision representation steps in an embodiment of the present invention.
[0019] Figure 6 This is a flowchart of the incremental reasoning and archiving action mapping steps in an embodiment of the present invention.
[0020] Figure 7 This is a structural diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings.
[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0023] Example 1: Overall Flow of a Time-Sequence Evidence Preservation and Archiving Method for Real Estate Registration Based on Dynamic Heterogeneous Graphs
[0024] like Figure 1 As shown in the figure, the real estate registration time-series evidence archiving method based on dynamic heterogeneous graphs provided in this embodiment is executed by a real estate registration archiving server device. The archiving server device includes a processor, a memory, a network interface, and an archive database interface. The processor reads registration business events, archive metadata, log records, and version snapshot data from the memory, and receives registration acceptance, review, supplementary entry, change, cancellation, and archiving feedback events through the business system interface. It also performs electronic case file packaging, archive index writing, pending completion status writing, and manual review queue writing through the archive system interface.
[0025] In this embodiment, the structured field encoding dimension In engineering terms, this refers to a unified embedding dimension used to carry discrete or semi-structured fields such as dossier number, business type, certificate type, status code, approval conclusion, version number, and field hash. This parameter can be obtained through offline calibration using historical dossier field type statistics and model capacity balancing. In practical engineering applications, its typical value is 64. (Text summary encoding dimension) This metric measures the semantic capacity of text such as application summaries, certificate summaries, and approval opinions in node representations. It can be determined jointly by text classification stability on the validation set and memory overhead, with a typical value of 128. (Log time location encoding dimension) The engineering meaning of "dimensionality" refers to the temporal expression width that describes the sequence of events, the interval between adjacent actions, and the snapshot position. It is typically calibrated based on the historical business time granularity and edge temporal complexity, with a typical value of 32. Total feature dimension This is formed by splicing together the aforementioned three types of features. The heterogeneous graph transformation network adopts a 3-layer structure, with each layer having a hidden dimension of 192 and 4 attention heads; these parameters are jointly determined by the average node size of the historical dossier, the complexity of relation types, and the training stability of a single 24GB graphics processing unit. The hypergraph propagation layer is 1, and the set-to-set readout output dimension is 256. The learning rate is... The training rounds are 60, and the training batch size is [not specified]. The above parameters were determined by grouping and searching a verification set consisting of historically archived files and manually reviewed files.
[0026] like Figure 2 The diagram illustrates the principle of a time-series evidence archiving method for real estate registration based on dynamic heterogeneous graphs, as provided in an embodiment of the present invention. It showcases the entire process from inputting original registration business data to outputting the final structured archiving result. The overall processing logic can be divided into five core steps:
[0027] First, in step S1 (constructing a dynamic heterogeneous graph and multi-view subgraphs), this invention receives registration business objects such as application forms, file files, and business entities. Through S101, these heterogeneous entities and their interactions are mapped as nodes and edges, constructing a basic dynamic heterogeneous graph. Subsequently, in S102, using file files as anchor points, a storage chain subgraph, a version evolution subgraph, and a cross-file reference subgraph are extracted and generated. Figure 3 Multi-view subgraphs in multiple dimensions; and time-series indexes for dossiers, entities, and real estate units are established in S103, thus laying a solid graph data foundation for subsequent link-level analysis.
[0028] Next, proceed to step S2 (node / edge multi-path encoding and hyperedge construction). To fully extract features, S201 performs structured field encoding (dimension) on each node. Text digest encoding () ) and log time location encoding ( S202 generates a unified node representation with dimension 224 by splicing the nodes together; S202 calculates the time difference. Injecting time coding into edges It also has a mechanism for handling time-outs or conflicts; S203 innovatively unfolds the complex relationship between multiple subjects and materials associated with a single business into a hyper-edge structure, performs high-order propagation, and enables the node representation to fully integrate time relationships and high-order business collaboration features.
[0029] Subsequently, step S3 (multi-objective learning to generate dossier-level decision representation) is the core representation stage of the model. S301 generates a 256-dimensional dossier-level representation through heterogeneous relation propagation and readout aggregation. S302 introduces temporal edge ranking learning, which captures causal order by distinguishing between legitimate sequential edges (positive samples) and difficult negative edges; S303 combines positive and negative samples with execution graph contrast learning and key edge occlusion reconstruction; finally, in S304, classification loss, ranking loss, contrast loss and reconstruction loss are jointly learned to output a highly robust dossier-level archiving decision representation.
[0030] Then, in step S4 (incremental update and result output), for the continuous registration event stream, S401 adopts a local subgraph incremental update mechanism based on 2-hop neighbor sampling, which greatly reduces the computational overhead of full graph reconstruction; S402 integrates reconstruction consistency, sequential consistency, cross-snapshot consistency and anchor closure to calculate the high-confidence archiving probability and chain integrity score; S403 calculates and sorts the anomaly scores of candidate edges, accurately locates and outputs the set of anomaly edges.
[0031] Finally, step S5 (archiving instructions and structured result output) intelligently triggers the corresponding archiving action in S501 based on the above judgment results (automatic archiving if high confidence is met, and manual review or delayed archiving if serious anomalies exist); and generates a structured result in S502 containing information such as dossier number, archiving probability, chain integrity score, anomaly edge location set, and anomaly path summary, and finally writes it to the archive library or monitoring interface. The entire schematic diagram intuitively shows how this invention leaps from traditional static file-level verification to dynamic, time-series, chain-level intelligent archiving judgment.
[0032] Step S1: Obtain the registration business objects and their temporal relationships, map the application form, file, business entity, real estate unit, certificate, approval process and time snapshot into a dynamic heterogeneous graph, and generate the evidence storage chain subgraph, version evolution subgraph and cross-file reference subgraph.
[0033] In this embodiment, this step is performed by the processor of the archiving server device. Figure 3The process shown is used to organize the objects and their relationships that were originally scattered in the business database, archive database, and log database into a graph structure that can be learned later.
[0034] Step S101: Construction of dynamic heterogeneous graph
[0035] The system reads the application forms, file files, rights holders, obligors, business personnel, real estate units, certificates, approval processes, and time snapshots within the time window to be determined, thus obtaining a set of registration business objects. Based on the set of registration business objects, a dynamic heterogeneous graph is constructed, and relationship views corresponding to different business constraints are extracted from the dynamic heterogeneous graph using the file number as the anchor point. According to the relationship views, the evidence storage chain subgraph, version evolution subgraph, and cross-file reference subgraph are obtained respectively, and a time-series index of files, subjects, and real estate units is established.
[0036] The application forms, file files, rights holders, obligors, business personnel, real estate units, certificates, approval processes, and time snapshots in the registration business object set are mapped to different node types to obtain a node set; the actions of submission, review, reference, change, inheritance, cancellation, supplementation, version continuation, and hash anchoring are mapped to different edge types, and the generation time of each edge is recorded to obtain an edge set; based on the node set, the edge set, and the generation time of each edge, a node type mapping, an edge type mapping, and a timestamp function are established to obtain the dynamic heterogeneous graph.
[0037] In this embodiment, the processor first reads application forms, file files, rights holders, obligors, business personnel, real estate units, certificates, approval processes, and time snapshot data within the time window to be determined from the business database, and maps them as nodes. For business entities, these include three subordinate objects: rights holders, obligors, and business personnel. For certificates, these may include real estate ownership certificates, identity documents, inheritance certificates, marriage certificates, or tax payment certificates, etc. For approval processes, these may include acceptance, preliminary review, secondary review, final review, or archiving confirmation, etc. This supports the concrete implementation of higher-level concepts such as "business entities," "certificates," and "approval processes."
[0038] To clarify the entity, relation, and time ternary constraints of the dynamic graph, the processor establishes the following dynamic heterogeneous graph representation: ,in, Indicates time The set of nodes; Indicates time The set of edges; This is a node type mapping function used to map each node to one of the following: application node, case file node, rights holder node, obligor node, business personnel node, real estate unit node, certificate node, approval process node, or time snapshot node. This is an edge type mapping function used to map each edge to one of the following: submission edge, review edge, reference edge, change edge, inheritance edge, cancellation edge, supplementary entry edge, version continuation edge, or hash anchored edge. This is a timestamp function used to record the generation time of edges. This formula corresponds to the processing step of "obtaining registered business objects and their temporal relationships and mapping them to a dynamic heterogeneous graph" in this invention. The input data comes from the business system event stream, archived metadata, and log records, and the output result is the basic graph structure for subsequent graph propagation and subgraph extraction.
[0039] In this embodiment, the time window to be determined is used to limit the range of data currently participating in the archiving determination. In order to avoid introducing too much historical noise unrelated to the current file during the inference stage, this step introduces a preset time window, which represents the data extraction span participating in one archiving determination. It is usually determined based on offline statistics of the average processing cycle of historical business at the local registration center. In actual engineering applications, the preset time window is typically 30 days before and after, and can be reasonably configured between 15 and 60 days in a high-concurrency business environment.
[0040] Step S102: Generation of multi-view subgraphs
[0041] Using the case file number as the anchor point, the closed-loop traceability relationship between application forms, certificates, approval processes, hash anchors, and case files is extracted from the dynamic heterogeneous graph to obtain the evidence storage chain subgraph; the relationships related to supplementary entries, version continuation, changes, and cancellation are extracted from the dynamic heterogeneous graph to obtain the version evolution subgraph; the relationships related to inheritance, reference, and real estate unit sharing are extracted from the dynamic heterogeneous graph to obtain the cross-case reference subgraph; when the current case file has multiple time snapshots, a snapshot link is generated in the version evolution subgraph; when the current case file has only a single time snapshot, the single snapshot view is retained and the construction of positive samples between snapshots is skipped, thus obtaining the evidence storage chain subgraph, the version evolution subgraph, and the cross-case reference subgraph.
[0042] Specifically, to achieve sub-view learning under different business constraints, the processor uses the dossier number as an anchor point to perform relationship extraction on the full dynamic graph. The evidence storage chain subgraph mainly retains the closed-loop trace relationship between application forms, certificates, approval processes, hash anchors, and dossiers, used for subsequent calculation of anchor closure and key edge reconstruction consistency; the version evolution subgraph mainly retains the relationships of supplementary entries, version continuation, changes, and cancellations, used for subsequent processing of legal snapshot continuation and abnormal snapshot separation; the cross-dossier reference subgraph mainly retains the relationships of inheritance, reference, and real estate unit sharing, used to detect source consistency and dangling reference issues between different dossiers. After this processing, different anomaly patterns are mapped to more targeted structural views, which can reduce the probability of rare anomaly edges being overwhelmed by high-frequency regular edges.
[0043] In one alternative implementation, when multiple time snapshots exist in the current dossier, the processor generates inter-snapshot links for adjacent valid snapshots in the version evolution subgraph for constructing positive sample pairs for subsequent graph comparison learning; when only a single time snapshot exists, the processor retains the single snapshot view and skips the construction of inter-snapshot positive sample pairs, and uses the consistency results of historical template snapshots as a reference when calculating the chain integrity score in the subsequent process.
[0044] Step S103: Time Series Index Establishment
[0045] Specifically, it involves establishing time-series indexes for case files, main bodies, and real estate units, including: establishing time-series indexes for case files, main bodies, and real estate units.
[0046] In this embodiment, the processor further establishes three types of time-series indexes. For dossiers, an index queue of "dossier number - version number - time snapshot number" is established; for real estate units, a reference index of "unit number - associated dossier set" is established; and for entities, a participation index of "identity identifier - business participation record" is established. This index is not only used for database retrieval, but also for generating difficult negative sample edges during the training phase and for quickly searching for cross-dossier candidate relationships during the inference phase.
[0047] When a dossier has only a single snapshot, the version evolution subgraph degenerates into a single snapshot view. The processor skips the construction of positive samples between snapshots, retaining only subsequent classification targets and key edge reconstruction targets. When a dossier lacks hash anchor edges, it can still enter the training and inference process, but its subsequent anchor closure score will be constrained, thus reducing the upper limit of the chain integrity score. The above indexing and boundary processing constitute the temporal indexes of dossiers, subjects, and real estate units mentioned above.
[0048] Step S2 involves performing structured field encoding, text digest encoding, and log time location encoding on various types of nodes in the dynamic heterogeneous graph, injecting time encoding into various types of edges, and expanding the relationship between a single business transaction and multiple entities and materials into a hyperedge structure.
[0049] like Figure 4 As shown, the goal of this step is to convert the original fields, text content, and time-series logs in the heterogeneous graph into a unified representation that can participate in graph propagation, and to supplement the higher-order co-occurrence constraints in a business action.
[0050] Step S201: Node multicoding
[0051] Extract the dossier number, business type, certificate type, status code, approval conclusion, version number, and field hash of each node, and perform structured field encoding to obtain a structured field representation; extract the application summary, certificate summary, and approval opinion, and perform text summary encoding to obtain a text summary representation. When the original material only contains full text, the title, summary paragraph, and approval opinion paragraph are trimmed from the original text according to preset trimming rules; extract the event occurrence sequence, the interval between adjacent business actions, and the time snapshot position, and perform log time position encoding to obtain a log time position representation; concatenate the structured field representation, text summary representation, and log time position representation to obtain a unified node representation.
[0052] In this embodiment, the processor extracts three types of features from each node and organizes them into a batch input tensor. The structured field encoded input is denoted as... The text summary encoding input is denoted as The log time location encoding input is denoted as ,in To train the number of dossier subgraphs within a batch, This represents the upper limit of the number of nodes sampled for each subgraph. The three features are concatenated by channel to form the total input features. This method uses concatenation based on the last dimension, rather than adding element by element, to avoid features from different semantic sources overlapping each other in the early stages of encoding.
[0053] In this embodiment, structured field encoding can be implemented using a Multi-Layer Perceptron (MLP) combined with a discrete field embedding table. Specifically, for discrete fields such as case file number, business type, certificate type, status code, approval conclusion, and version number, a fixed-length vector is first obtained by looking up the embedding table; for field hashes, the hash value can be mapped to a fixed-dimensional sparse index before embedding; then, each discrete field vector is concatenated with the necessary numerical fields, and then input into two layers of the Multi-Layer Perceptron to obtain a dimension of... The structured field representation is used. The text summary encoding preferably employs a lightweight Chinese Transformer encoder, which segments, truncates, and encodes the application summary, certificate summary, and approval comments separately. Then, the classification tag bits or mean pooling results are used to obtain the dimension. The text digest representation. Log time location encoding is used to carry event sequence, adjacent action interval, and time snapshot location. It can be obtained by concatenating time slice embedding, sequence index embedding, and action interval bucket embedding, and then performing a linear mapping to obtain a dimension of... The expression .
[0054] To ensure feasibility in full-text scenarios, this step introduces a preset pruning rule. In engineering terms, this means extracting stable local text fragments relevant to archiving decisions from the full text when the original material lacks a structured summary field. This rule can be determined offline through correlation analysis of text fields in historical files and manual review conclusions. In practical engineering applications, the typical length for retaining a single field in the title, summary, and approval opinion sections is between 64 and 256 Chinese characters. When the upper limit is exceeded, pruning is performed according to the priority of "title—conclusion—key explanations." After this processing, text inputs of different node types are mapped to a uniform length scale, facilitating subsequent graph calculations.
[0055] Step S202: Side-time encoding injection
[0056] Based on the time difference between the generation time of each edge and the current snapshot reference time, time encoding is performed on various types of edges to obtain edge time representations. When log times are missing or conflicting, the time difference or additional missing time flags are generated according to adjacent known time interpolation, interval median replacement, standard time sequence template replacement, and master log priority rules.
[0057] In practical engineering applications, this step specifically includes: the processor, using the current snapshot reference time as a reference, calculates the time difference between the generation time of each edge and the reference time, and maps it to a periodic time representation that can participate in relation propagation. Corresponding to the "injecting time encoding into various types of edges" processing step in this invention, the input data are the edge generation time, the current snapshot reference time, and a time encoding parameter set. The technical sub-problem it solves is distinguishing between relation edges that are close in time but have an illegal order and those that are legally sequential. The output result is the time representation of each edge. This time encoding can be expressed as: ,in, For the edge The time-encoded vector has the same dimensions as the subsequent time-mapping channels; For the edge The time difference between the generation time and the current snapshot reference time can be in hours or days, depending on the granularity of the system log. This is a set of frequency parameters used to map different time scales onto multiple periodic bases; This is the phase offset parameter set; This is an element-wise cosine mapping. and This can be learned through joint learning during the training phase, or it can be initialized as a geometric frequency group first and then fine-tuned. At that time, the time coding degenerates into It can still participate in propagation normally. If log time is missing, the processor prioritizes interpolating adjacent known times within the same approval process; if the missing interval spans approval nodes, it prioritizes using the completion time of the upstream node and the start time of the downstream node to form a constrained interval, and takes the midpoint of the interval as the replacement; when there are many consecutive missing entries, the processor marks the segment as a missing time fragment and replaces it with a standard time sequence template time, while adding a missing time flag. The above time encoding realizes the aforementioned edge time representation.
[0058] To determine whether there are non-negligible conflicts in the timestamps of multiple log sources, this step introduces a preset conflict threshold, which represents the boundary for determining whether there are non-negligible conflicts in the timestamps of multiple log sources. It is usually determined based on the historical deviation distribution between the main log of the handling system, the time of file storage, and the message bus timestamp, with a typical value of 30 minutes. When the time difference of multiple log sources exceeds the preset conflict threshold, the time of the main log of the handling system is used first, followed by the time of file storage, and a conflict flag is added to the structured field.
[0059] Step S203: Hyperedge Unfolding and Higher-Order Propagation
[0060] Identify multiple rights holders, multiple obligors, multiple supporting documents and approval processes involved in the same business action, establish a high-order association carrier, and obtain a hyperedge structure; perform high-order propagation based on the unified node representation, the edge time representation and the hyperedge structure to obtain a node representation with time relationship and high-order business relationship.
[0061] In practical engineering applications, this step involves the following processes: the processor identifies the nodes that participate in a single business action, constructs a high-order association vector, and then uses a Hypergraph Neural Network (HGNN) for propagation. The hyperedge association matrix is denoted as... ,in For the number of nodes, Let be the number of hyperedges, if the th The node belongs to the node. If there is a super edge, then Subsequently, the processor performs a higher-order propagation on the node representation based on the hyperedge relationship, obtaining a node representation with higher-order business relationships, the computational relationship of which is as follows: ,in, This represents the node representation after hypergraph propagation. For the output dimension of dissemination; The activation function can be either ReLU or GELU; It is a diagonal matrix of node degree, whose diagonal elements are obtained by counting the number of hyperedges that each node participates in; It is a hyperedge degree diagonal matrix, whose diagonal elements are obtained by counting the number of nodes contained in each hyperedge; The weight matrix is a hyperedge weight matrix, which can be learned through training or assigned different initial values according to the type of approval action during the initialization phase; The input mapping matrix; This is the unified node representation obtained in step S201. This formula corresponds to the processing step of "expanding the relationship between multiple entities and multiple materials associated with a single business into a hyperedge structure and performing higher-order propagation". The inputs are the unified node representation, the hyperedge association matrix and weight parameters, and the output is the node representation that retains the co-occurrence of multiple entities and the co-attraction constraints of multiple materials.
[0062] In this embodiment, if the same business action is associated with only a single node, the hyperedge degenerates into a self-associated hyperedge, retaining only its self-updating function; if If zero elements exist, a smoothing term is added during the inversion. To avoid numerical instability, a preset minimum positive constant is used in this embodiment to smooth the case where the hyperedge degree is zero. This value can be obtained through numerical stability testing, and its typical value is [value to be filled in]. to When the number of nodes associated with the same business action exceeds the super-edge limit, the core node is retained according to the priority of "file node - approval node - real estate unit node - main node - material node", and the remaining nodes are recorded as subordinate ordinary edges. Here, the preset super-edge limit is used to control the computational scale of a single high-order propagation. It can be calibrated offline based on the average number of participating objects per file and the upper limit of video memory. Its typical value is 8 to 16 nodes.
[0063] The construction and propagation process of the aforementioned unified node representation, edge temporal representation, and hyperedge structure constitutes the aforementioned node representation with temporal and higher-order business relationships.
[0064] Step S3: Generate a dossier-level archiving decision representation based on heterogeneous relation propagation, temporal edge sorting learning, graph-comparative chain consistency learning, and key edge reconstruction learning.
[0065] like Figure 5 As shown, this step is performed by the processor on the graph representation backbone shared by the training and inference phases, with the goal of forming a unified decision representation at the dossier level that can be used for both classification and anomaly edge localization.
[0066] Step S301: Heterogeneous Relationship Propagation and Dossier-Level Readout
[0067] Based on the node representation with time and higher-order business relationships, heterogeneous relationship propagation is performed according to the joint constraints of node type, edge type and time encoding, and read aggregation is performed on the dossier-related node set to obtain the dossier-level intermediate representation.
[0068] In this embodiment, the processor employs a Heterogeneous Graph Transformer (HGT) network as the backbone propagation model. For different node and edge types, type-related message transformation matrices are set, and edge temporal encoding is injected into the propagation channel as part of the relation message. The node at the... The update relationship of a layer can be represented as:
[0069] ,
[0070] in, For nodes In the Layer representation; and The target node and the source node are respectively in the th... Layer representation; Pointing to a node The set of adjacent triples; The attention weight is determined by the source node, edge type, and target node. This is the message transformation matrix corresponding to the source node type, edge type, and target node type; This is a time mapping matrix related to edge type; The edge time encoding obtained in step S202; This is a layer normalization operation. The technical sub-problem this formula solves is how to simultaneously propagate heterogeneous object relationships, edge type relationships, and temporal constraints within the same network, and the output is a node representation that balances the validity of local relationships and the global link structure.
[0071] After completing multi-level propagation, the processor performs a two-level read of the set of related nodes in the dossier. First, it performs a set-to-set read to obtain an intermediate representation. Subsequently, using the dossier node representation as the query vector, attention pooling is performed on the set of dossier-related nodes to obtain the final dossier-level archiving determination representation. The dossier-related nodes here include dossier nodes, approval process nodes, time snapshot nodes, hash anchor nodes, directly associated real estate unit nodes, and key certificate nodes. After this readout, local edge legality, version continuation trajectory, and cross-dossier dependencies are aggregated into a unified dossier-level representation, which is then shared by subsequent classification, sorting, comparison, and reconstruction targets.
[0072] Step S302: Learning Temporal Edge Sorting
[0073] Legitimate sequential relationships are extracted from the real event chain of the same dossier as positive sample edges, and difficult negative sample edges are extracted from candidate relationships that are close in time but inconsistent in business links and similar in fields but conflicting in entity mapping. Temporal edge ranking learning is then performed.
[0074] In this embodiment, the processor samples legitimate sequential edges as positive samples from the real event chain of the same dossier, such as "submit application form → preliminary review", "preliminary review → final review", "version continuation → hash anchoring", etc.; at the same time, it samples difficult negative sample edges from the candidate relationships. "Time proximity" means that the time difference between the candidate edge and the corresponding positive sample edge is close; "business link inconsistency" means that the source node and target node of the candidate edge belong to different dossiers or different real estate units, and there is no legitimate inheritance or reference relationship; "field similarity but entity mapping conflict" means that the digest hash is similar, but at least one of the following is inconsistent: subject identity identifier, dossier primary key, or real estate unit primary key. These positive and negative edges constitute the aforementioned legitimate sequential relationships and difficult negative sample edges.
[0075] Corresponding to the core constraint of this step, the processor performs a sorting loss calculation on the scores of positive and negative edges: ,in, The loss is the time-series edge sorting loss; The preset sorting interval is used to require that the score of a legal edge be at least a certain margin higher than that of a difficult negative edge; Scoring is given for positive sample edges; The edge score is calculated for difficult negative samples. The edge score is obtained by inputting the node representations at both ends of the edge, the edge type embedding, and the temporal encoding into an edge scoring network. This formula addresses the difficulty in separating legitimate sequential edges from out-of-order forged edges, and the output is an explicit penalty for temporal validity.
[0076] In this embodiment, the preset sorting interval is the minimum safe interval that real sequential edges and difficult negative edges need to maintain in the scoring space. It is typically obtained offline based on the median interval of the historical score distributions of legal edges and difficult negative edges, with a typical value of 0.25 to 0.45. In this embodiment, the preset sorting interval is specifically set to 0.35. The time proximity window is used to limit the candidate range of difficult negative edges. It can be calibrated by the 90th quantile of the time difference between adjacent events of similar historical business transactions, with a typical value of 4 to 24 hours, adjusted according to the business type. The field similarity threshold is used to identify candidate relationships where "text is similar but entities are inconsistent." It can be determined by manually reviewing the intersection of the summary similarity between normal and incorrect citations in the sample, with a typical value of 0.80.
[0077] When a dossier cannot construct a difficult negative edge within the current batch, the processor supplements the difficult negative edge from other dossiers of the same business type, similar time window, and different real estate units. When the input required for edge score calculation is missing, the current sample skips the ranking items and only retains the remaining loss items. The above ranking learning process realizes the aforementioned temporal edge ranking learning.
[0078] Step S303: Graph Comparative Chain Consistency Learning and Key Edge Reconstruction
[0079] Snapshots of the same dossier at different times during its legitimate evolution are constructed as positive sample pairs, while snapshots with tampering, missing pages, rollback operations, and erroneous cross-dossier references are constructed as difficult negative sample pairs. Graph-comparative chain consistency learning is performed, and masking reconstruction is performed on reference relationships, audit relationships, and change relationships.
[0080] In this embodiment, the processor uses snapshots from different times during the legitimate evolution of the same dossier as positive sample pairs, and snapshots exhibiting tampering, missing pages, rollback operations, and erroneous cross-dossier references as difficult negative sample pairs. Specifically, a tampered snapshot refers to a snapshot where field values have changed but lack legitimate change edges or hash anchor edges; a missing page snapshot refers to a snapshot where key review materials, cited materials, or approval conclusion pages that should be present in the dossier directory are missing; a rollback snapshot refers to a snapshot where a later state appears prematurely in an earlier snapshot, or a later approval result covers an earlier approval state and lacks legitimate rollback edges; and an erroneous cross-dossier reference snapshot refers to a snapshot where the cited certificate, approval conclusion, or dossier fragment originates from another dossier, but there is no legitimate inheritance, shared unit, or formal citation basis between the source dossier and the current dossier. All of the above state quantities belong to the risk or estimation category, and their reasonableness can be verified through historical manually reviewed and labeled samples or rule-injected samples.
[0081] Graph-based contrastive chain consensus learning uses information-noise contrastive estimation of loss, expressed as: ,in, To compare the learning loss; A graph-level representation of the current volume snapshot; This represents a positive sample within the same legitimate evidence storage chain. To compare with other sample representations in the set; For the similarity function, cosine similarity can be used; A preset contrast temperature coefficient is used to adjust the smoothness of the similarity distribution. This formula is used to solve the problem of difficulty in separating snapshots when "fields are similar but links are inconsistent," and the output is a representation space constraint for the aggregation of valid chain snapshots and the separation of abnormal chain snapshots.
[0082] To improve the ability to determine the closure of key relationships, the processor also randomly occludes referencing edges, auditing edges, and modified edges. The occlusion ratio controls the number of hidden key edges in each reconstruction task and can be determined jointly by reconstruction accuracy and training stability, typically ranging from 10% to 30%. Key edge reconstruction is then performed. The reconstruction loss is denoted as: ,in, For reconstruction losses; The set of key edges that are hidden; This is the model's prediction matrix for the probability of the existence of the occluded key edge, with values ranging from 0 to 1; This is the true adjacency matrix of the occluded key edges, with elements taking values of 0 or 1; This is the binary cross-entropy. The negative sample set of the occluded key edge consists of candidate edges of the same type as the edge and the same type of target node, but from different dossiers or with invalid temporal order. This reconstruction process forces the model to understand the structural dependencies between key edges, rather than just a few local fields.
[0083] In this embodiment, a preset contrast temperature coefficient is used to balance the aggregation degree of the same file across snapshots and the separation degree of different files, which can be obtained by validation set search; its typical value is 0.15 to 0.30, and in this embodiment it is 0.20.
[0084] When the comparison set contains only one positive sample and no other negative samples, the processor loads historical representations of the same business type from the cross-batch cache queue as supplementary negative samples; when a subgraph has no reconstructable key edges after occlusion, the reconstruction loss of the current sample is recorded as zero. The above process constitutes the aforementioned graph comparison-based chain consistency learning and key edge reconstruction learning.
[0085] Step S304: Jointly learn to generate dossier-level archiving decision representation
[0086] Joint learning is performed based on the classification, sorting, comparison, and reconstruction objectives of the archiving stage to obtain the dossier-level archiving determination representation.
[0087] In this embodiment, the processor constructs a classification loss based on the archive category output. Dossier-level archive decision representation. The archived category distribution is output after a fully connected layer and Softmax, and the categories include at least automatic archiving, delayed archiving, and manual review. The classification loss is expressed as cross-entropy, and its calculation relationship is as follows: ,in, Classify losses during the archiving phase; The distribution of archive categories predicted by the model; These are the actual archived labels. This loss corresponds to the availability constraint of the results, ensuring that the model output is consistent with the historical manually verified archive labels.
[0088] The processor then weights and fuses the aforementioned classification loss, ranking loss, contrastive loss, and reconstruction loss to obtain the overall training objective: ,in, Total training loss; , , and These are preset loss weight coefficients for the classification objective, ranking objective, comparison objective, and reconstruction objective, respectively. The four weights are jointly determined through validation set grid search and human review consistency, and are used to balance the final archive classification stability, anomaly edge localization accuracy, and human review hit rate.
[0089] In this embodiment, the preset loss weight coefficient represents the proportion of contribution of different training objectives to the final model behavior. It can be jointly calibrated using three metrics: automatic archiving accuracy, anomaly edge localization accuracy, and manual review hit rate on the validation set. Its typical actual value is: , , , When a loss term in a batch becomes zero due to insufficient sample construction, the total loss degenerates into a weighted sum of the remaining non-zero terms. When any loss term has an abnormally large value, the processor performs gradient pruning on that term to avoid training instability. The joint optimization of the above four loss terms generates the aforementioned dossier-level archiving decision representation.
[0090] Step S4: Perform local subgraph incremental update on newly entered registration events, and output the archiving probability, chain integrity score, and abnormal edge location set based on the dossier-level archiving determination representation.
[0091] like Figure 6 As shown, this step corresponds to the main execution flow of the online inference stage, and is used to apply the model obtained from offline training to newly entered registration events.
[0092] Step S401: Local subgraph incremental update
[0093] Specifically, for newly entered registration events, incremental updates to the local subgraph are performed. In this embodiment, after receiving a newly entered registration event, the processor does not recalculate the entire dynamic graph. Instead, it extracts a local subgraph centered on the current file, real estate unit, and directly related entities, according to a preset neighbor sampling depth, and performs incremental inference on this local subgraph. Specifically, the processor uses a graph sampling aggregation method to complete the local update inference, thus avoiding triggering full graph recoding each time a new edge is added or reviewed.
[0094] To cover direct business dependencies and suppress irrelevant distant neighbor noise, this step introduces a preset neighbor sampling depth. This parameter can be obtained offline based on the average shortest path length of key anomaly edges in historical dossiers; in practical engineering applications, its typical value is 2 hops, and in this embodiment, 2 hops are used. The upper limit of a single dossier node is used to control the maximum number of nodes participating in propagation during incremental inference. It can be determined based on the average dossier node size, memory constraints, and single dossier latency requirements, and its typical value is 80 to 150, and in this embodiment, 120 are used. In this embodiment, the maximum number of neighbors per hop is preferably 20.
[0095] In practical engineering applications, this step involves the following specific processing steps: The first hop prioritizes retaining approval nodes, certificate nodes, time snapshot nodes, and hash anchor nodes directly connected to the current dossier; the second hop prioritizes retaining real estate unit nodes, main nodes, and cross-dossier reference nodes related to the aforementioned nodes; when the number of nodes at the same distance exceeds the upper limit, they are retained according to the priority of "time closer to the current snapshot, edge type belonging to approval / reference / version continuation, higher node degree, and more nodes sharing primary keys with the current dossier." If time window pruning is effective, nodes within a preset number of days before and after the current snapshot are prioritized. Nodes exceeding the window but belonging to legitimate version inheritance ancestor nodes can retain one level of ancestor nodes. As a specific engineering example, the preset number of days can range from 15 to 45 days, and in this embodiment, 30 days is preferred. The above sampling process enables incremental updates of the most relevant substructures of the current dossier without reconstructing the entire graph.
[0096] To control the propagation range of local incremental inference, this step introduces a preset neighbor sampling depth, which represents the furthest business dependency level that needs to be retained during local inference. It is usually calibrated based on the average propagation radius of abnormal edges in historical volumes, and its typical value is 2 hops. The aforementioned upper limit of nodes per volume is specifically set to 120, which is determined based on the peak memory constraint of a single 24GB graphics processor and the average node size per volume.
[0097] Step S402: Output the archiving probability, chain integrity score, and anomaly edge location set.
[0098] Based on the dossier-level archiving determination, the archiving probability is output; the critical edge reconstruction consistency score, temporal sequence consistency score, cross-snapshot consistency score, and anchor closure score are calculated and fused according to preset sub-item weights to obtain the chain integrity score; for each candidate edge, the edge reconstruction deviation, edge sequence scoring deviation, and cross-snapshot consistency deviation are calculated and fused according to preset deviation fusion weights to obtain the abnormal edge comprehensive score; the abnormal edge location set is obtained through any of the following methods: abnormal edges are filtered according to the abnormal edge comprehensive score and a preset abnormal threshold; when the number of edges that meet the threshold exceeds the preset abnormal edge number limit, they are sorted and output according to the abnormal edge comprehensive score.
[0099] In this embodiment, the processor first inputs the file-level archiving determination representation into the archiving probability output header to obtain the probability that the current file meets the archiving conditions after time-series evidence storage. Subsequently, the processor calculates the consensus score for each key edge reconstruction. Temporal sequence consistency score Cross-snapshot consistency score And anchored closed fractions .in, It is obtained by the degree of consistency between the predicted probability of the occluded key edge and the existence of the actual edge; It is obtained from the proportion of the true first-to-last edge score being higher than the difficult negative edge score; It is determined by the similarity of the representation of the current snapshot with the legitimate snapshots on the same chain and the degree of separation from the abnormal snapshots; The reliability of the link is determined by whether the hash anchor edge, version continuation edge, and approval closure edge form a closed loop. These four scores each characterize different aspects of the link's reliability.
[0100] To obtain a unified chain integrity score, the processor performs fusion according to preset component weights: ,in, The chain integrity score; , , , The weights are assigned to individual components, with a total weight of 1. Each weight is jointly determined by the validation set and human review consistency to balance the contributions of reconstruction consistency, sequential validity, snapshot continuity consistency, and anchoring closure to the final credibility. In this embodiment, the engineering meaning of the aforementioned preset weights is the importance ratio of each component of chain integrity, which can be obtained by manually reviewing the files through the score distribution of the dossier; its typical value is... , , , When the dossier is missing a hash anchor edge, It will be limited to a lower range; when the file has only a single snapshot, The calculation is replaced by the consistency between the current snapshot and the historical template snapshot.
[0101] To locate abnormal edges, the processor calculates the edge reconstruction deviation for each candidate edge. Side order scoring deviation and cross-snapshot consistency deviation Then, the edges are fused according to the preset deviation fusion weights to obtain the comprehensive score of the abnormal edges: ,in, For the edge The overall score for abnormal edges; , , The bias fusion weights are set such that their sum is 1. The preset bias fusion weights are preferably calibrated based on the contribution of reconstruction bias, sequence bias, and cross-snapshot consistency bias to anomaly identification in manually reviewed samples. Typical values may be: , , . This reflects the degree of deviation of the edge during the reconstruction task; This reflects the degree of anomaly in the sequential scoring of the edge; This reflects the consistency deviation between the snapshot containing the edge and the legitimate snapshot. This formula is used to map edge anomaly evidence from different sources to the same scoring scale, outputting anomaly edge scores that can be used for filtering and ranking.
[0102] The preset anomaly threshold is used to distinguish between normal and abnormal edges. It can be obtained through offline testing of the intersection of the comprehensive scores of normal and abnormal edges in the manually reviewed samples. Its typical value is 0.55 to 0.70; in practical engineering applications, 0.60 can be used. The preset upper limit for the number of abnormal edges is used to control the number of abnormal edges that need to be prioritized by humans in a single output. It can be determined based on the workload of manual review and the average complexity of the dossier, with a typical value of 3 to 10 edges. If the number of edges that meet the threshold exceeds the upper limit, the first few edges are output according to the comprehensive score of abnormal edges from high to low. The above calculation process of chain integrity score and abnormal edge comprehensive score constitutes the aforementioned chain integrity score and abnormal edge location set, respectively.
[0103] like Figure 6 As shown, in step S5, an archiving processing instruction is generated based on the archiving probability, chain integrity score, and abnormal edge location set, and the corresponding archiving action is triggered to generate a structured archiving result.
[0104] Step S501: Generate archiving processing instructions
[0105] When the archiving probability is higher than a preset archiving threshold, the chain integrity score is higher than a preset trusted chain integrity threshold, and the abnormal edge location set is empty or contains only edges with an abnormal edge comprehensive score lower than the preset abnormal threshold, an automatic archiving instruction is generated, and electronic case file encapsulation, archiving index generation, and evidence storage result writing are executed. When any of the following conditions are met, a manual review instruction is generated, and an abnormal path description and review queue are written: the archiving probability is lower than a preset low trusted threshold; the chain integrity score is lower than a preset low trusted chain integrity threshold; there are edges in the abnormal edge location set with order deviation exceeding the threshold and inconsistent cross-case source. In other cases, a delayed archiving instruction is generated, and a pending completion status is written. The structured archiving result containing case file number, archiving probability, chain integrity score, abnormal edge location set, archiving action type, and abnormal path summary is output.
[0106] In this embodiment, the preset archiving threshold is defined as the minimum archiving probability requirement that allows the system to automatically initiate the archiving action. It is typically determined offline based on the balance between the cost of misjudgment and the cost of missed judgment in historical automatic archiving samples, with a typical value of 0.80 to 0.90; in this embodiment, it is 0.85. The preset trust chain integrity threshold is used to measure whether the overall trustworthiness of the current case file's evidence chain meets the requirements for automatic archiving. It can be determined manually through manual review using the lower quartiles of the case file's chain integrity distribution, with a typical value of 0.75 to 0.85; in this embodiment, it is 0.80. To distinguish low-trust cases, this step also introduces a preset low-trust threshold and a preset low-trust chain integrity threshold. These represent the lower bound of the archiving probability and the lower bound of chain integrity, respectively, and are typically calibrated offline based on the tolerance for erroneous archiving risk. The preset low-trust threshold limits the lower bound of the manual review trigger for the archiving probability, and the preset low-trust chain integrity threshold limits the lower bound of the manual review trigger for the chain integrity score; both typically have a value of 0.60.
[0107] In terms of judgment logic, if the abnormal edge location set is empty or contains only low-risk edges, the processor generates an automatic archiving instruction; if the abnormal edge location set contains high-risk edges with order deviations exceeding the threshold and inconsistent cross-volume sources, the processor generates a manual review instruction; otherwise, a delayed archiving instruction is generated and written to the pending completion status. If the archiving probability conflicts with the judgment given by the chain integrity score, the lower confidence level takes precedence, i.e., manual review takes precedence over delayed archiving, and delayed archiving takes precedence over automatic archiving. This avoids erroneously triggering automatic archiving based solely on a high local classification probability when the chain evidence is insufficient.
[0108] When a local subgraph is empty, the processor directly outputs a manual review instruction to avoid making automatic archiving decisions without link-based evidence. When the number of nodes in a local subgraph exceeds the node limit, only the nearest neighbor node to the current volume and the most relevant node within the time window is retained. The above logic implements the aforementioned archiving instruction generation.
[0109] Step S502: Archive action triggered
[0110] Specifically, triggering the corresponding archiving actions includes: in this embodiment, the processor performs electronic case file encapsulation, archive index generation, and evidence storage result writing for automatic archiving instructions; writes a pending completion status for delayed archiving instructions; and writes an exception path description and review queue for manual review instructions. During automatic archiving, the archive system encapsulation interface is called to write case file metadata, summary index, and evidence storage summary; during delayed archiving, the pending completion status interface is called to write back the case file gap type; during manual review, the review queue interface is called to write the case file number, exception edge location set, exception path summary, and suggested review order.
[0111] In one optional implementation, the archiving instruction frame includes an instruction number, a dossier number, a generation timestamp, an archiving action type, an archiving probability, a chain integrity score, a number of abnormal edges, an abnormal path summary, and an integrity checksum. The archival system interface determines whether the archiving instruction is within the allowed write window based on the generation timestamp and whether the archiving instruction has been truncated, tampered with, or out of order during transmission based on the integrity checksum. This ensures that the model output is stably mapped to a data frame format executable by the archival system.
[0112] Step S503: Output of Structured Archiving Results
[0113] The output is a structured archiving result containing the dossier number, archiving probability, chain integrity score, anomaly edge location set, archiving action type, and anomaly path summary.
[0114] In this embodiment, the structured archiving results are output in the form of structured records or message bodies, and can be written to an archiving database, a manual review queue, or an external regulatory interface. Archiving action types include at least three categories: automatic archiving, delayed archiving, and manual review. Anomaly path summaries are used to further transform the anomaly path location set into natural language descriptions for manual review, such as "supplementary materials were inserted after the final review" or "there is no legal inheritance relationship between the cited source file and the current file." These fields collectively constitute the aforementioned structured archiving results.
[0115] Test and Effect Description of Example 1
[0116] The training, validation, and testing data in this embodiment are constructed using the historical business structure of a municipal real estate registration center as a template, including the actual business field structure, actual approval process configuration, actual case file directory rules, and actual log field format. Without exposing original personal information and the full text of the original case files, the actual business samples are anonymized and structurally extracted. These are then combined with manual review records to label normal case files, supplementary contaminated case files, case files in reverse chronological order, case files with erroneous cross-case references, and case files with missing critical edges. The testing environment is an archive server environment consisting of a single 24GB graphics processing unit, 128GB of RAM, and a multi-core server processor. Inference time statistics use a single-case serial inference caliber, graphics memory usage statistics use peak graphics memory caliber, and the proportion of manual review workload is calculated as "number of case files entering the manual review queue / total number of test case files".
[0117] The test was conducted on a digital sample set containing 120,000 historical case files, including 96,000 training files, 12,000 validation files, and 12,000 test files, with a test period of 90 consecutive days of event stream registration. Comparison scheme one uses a database serial number plus rule verification scheme, and comparison scheme two uses a single document classification scheme. Test results show that under the test conditions, the automatic archiving accuracy of this embodiment is 96.8%, and the archiving reliability misjudgment rate is 3.2%; compared to 91.3% and 8.7% for comparison scheme one, respectively. This demonstrates that the coupled design of dynamic heterogeneous graph modeling, temporal edge sorting learning, and graph-comparative chain consistency learning can improve the accuracy of archiving reliability judgment and reduce the probability of misjudgment. Furthermore, under the local subgraph incremental update mechanism, compared to the full graph recalculation scheme, the peak memory usage for a single case file inference decreases from 8.9GB to 5.1GB, and the average inference time decreases from 214 milliseconds to 96 milliseconds. After this processing, newly entering events do not need to trigger full graph reconstruction, and the subsequent system can continuously complete archiving judgment with lower overhead.
[0118] Example 2: Anomaly Identification in Scenarios with Data Supplementation Contamination and Time Reversal Edges
[0119] This embodiment illustrates how the present invention outputs manual review instructions and abnormal path descriptions when both data corruption and time reversal edges occur simultaneously.
[0120] In this embodiment, file B017 contains 18 nodes and 27 edges, including one supplementary recording edge and one reverse-order review edge. The preset archiving threshold is 0.85, the preset trusted chain integrity threshold is 0.80, the preset sorting interval is 0.35, and the preset neighbor sampling depth is 2 hops. After the processor constructs the dynamic heterogeneous graph and three types of subgraphs corresponding to the file in step S1, it obtains the shape in step S2. The unified node input features are used, and time codes are injected into the reverse review edge and the supplementary recording edge respectively.
[0121] Subsequently, in step S3, the processor constructs a sorted sample from the real edge "preliminary review → final review" and the negative edge "final review → preliminary review", and constructs another sorted sample from the "legitimate version continuation edge" and the "post-final review supplementary insertion edge". Since the source approval node time of the reverse review edge is later than the target approval node time, and the supplementary insertion edge appears after the final review is completed, the model incurs significant penalties in both sorting loss and graph comparison loss. The processor also constructs a positive sample pair from the current snapshot and the previous legitimate snapshot, and constructs a difficult negative sample pair from a supplementary insertion pollution rule sample snapshot, thereby amplifying the abnormal pattern of "similar fields but inconsistent links".
[0122] In step S4, the model outputs the archiving probability. Chain integrity score The abnormal edge location set contains two edges: one supplementary entry edge and one time-reverse review edge. Since the archiving probability is lower than the preset low-confidence threshold and a high-risk abnormal edge exists, the processor generates a manual review instruction and outputs an abnormal path description: "Supplementary material insertion occurred after final review; a time-reverse review edge exists in the approval chain." Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Example 1, completing operations such as writing the manual review queue and outputting structured archiving results. The specific execution logic is consistent with Example 1 and will not be repeated here.
[0123] An additional 15,000 abnormal case files were injected into the test set. Of these, 40% were due to data corruption, 35% were due to temporal reversal, and 25% were a combination of both. The rationality of the injection was verified by manual review. The control scheme was a static graph matching classification scheme. Test results showed that the present invention achieved a recall rate of 93.6% for identifying abnormal temporal chains, compared to 82.1% for the control scheme; the accuracy rate for locating temporal reversal edges was 92.5%, compared to 78.4% for the control scheme. This demonstrates that through the aforementioned temporal edge sorting learning and chain consistency learning, it can effectively identify case files with continuous serial numbers in traditional databases but abnormal business causal chains, and pinpoint the anomalies to the specific edge level.
[0124] Example 3: Delayed archiving in scenarios involving cross-volume erroneous references and multi-subject higher-order associations
[0125] This embodiment illustrates how the present invention utilizes hyperedge propagation and cross-case reference subgraphs to identify anomalies when multiple entities, multiple documents, and multiple case files are cross-referenced in inheritance registration, property division registration, or historical legacy certificate supplementation business.
[0126] Case file C233 is an inheritance registration, involving 3 rights holders, 2 obligors, 7 supporting documents, 2 related case files, and 1 real estate unit, forming a total of 29 nodes and 46 edges, including 5 cross-case reference edges. The maximum number of nodes per case file is set to 120, the preset trusted chain integrity threshold is set to 0.80, and the preset comparison temperature coefficient is set to 0.20. In step S2, the processor identifies that an inheritance relationship confirmation action simultaneously associates 3 rights holders, 2 historical supporting documents, and 1 approval node. Therefore, it constructs a superedge covering 8 nodes and preserves the overall relationship of "jointly participating in the same inheritance confirmation action" through high-order propagation.
[0127] In step S3, the processor constructs a positive sample pair between dossier C233 and its previous valid snapshot, and constructs a difficult negative sample for a forged snapshot that "references other dossier documents but has highly similar fields". Since the forged snapshot is close to the valid snapshot in terms of field content, single-document classification schemes usually have difficulty distinguishing it; however, this embodiment enables the model to identify structural features that are inconsistent in source dossier and lack valid version continuation anchoring by using cross-dossier reference subgraphs and hyperedge co-occurrence relationships.
[0128] In step S4, the model outputs the archiving probability. Chain integrity score The abnormal edge location set contains three cross-volume reference edges, two of which are erroneous document references, and one is a dangling reference caused by a missing version inheritance. Since the archiving probability and chain integrity score do not meet the automatic archiving requirements but are not below the low-confidence threshold, the processor generates a delayed archiving instruction and provides an abnormal path summary: "Volume C233 references the document materials of Volume C198 during the inheritance confirmation process, failing to form a legitimate version continuation anchor; there is a dangling reference edge between Volume C233 and Volume C205." Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Example 1, completing operations such as writing the pending completion state and outputting the structured archiving results. The specific execution logic is consistent with Example 1 and will not be repeated here.
[0129] The test was conducted on 20,000 multi-subject, multi-volume cases, including 8,000 cases containing cross-volume erroneous references, 4,000 cases containing missing key review edges, and the remainder being normal samples. A single-document classification scheme was used as the control. The test results show that the manual review workload of this invention is 11.6%, compared to 27.8% for the control scheme; the cross-volume erroneous reference detection rate is 90.1%, compared to 74.2% for the control scheme; and the missing key review edge detection rate is 91.4%, compared to 79.5% for the control scheme. These results indicate that expressing the interaction relationships among multiple subjects and materials through a high-order association carrier and outputting the abnormal edge location set as a verifiable path result can reduce the workload of manual document review and improve the ability to detect erroneous references.
[0130] Example 4: An electronic device
[0131] like Figure 7 As shown, this embodiment provides an electronic device, including a memory, a processor, a communication interface, and a computer program stored in the memory and running on the processor.
[0132] The processor can be a general-purpose central processing unit, a graphics processing unit, a tensor processor, or a combination thereof; the memory can be random access memory, read-only memory, flash memory, or solid-state memory; the communication interface is used for data interaction with the business system, the archive system, the log system, and the database interface. When the computer program is executed by the processor, it implements the steps of the real estate registration time-series evidence archiving method based on dynamic heterogeneous graphs as described in any one of Embodiments 1 to 3, including obtaining the registration business objects and their time-series relationships, constructing a dynamic heterogeneous graph and multi-view subgraphs, performing node encoding and time encoding, performing heterogeneous relationship propagation and joint learning, performing incremental updates of local subgraphs, and outputting archiving processing instructions and structured archiving results.
[0133] In one specific implementation, the electronic device is deployed in the real estate registration center's archiving server environment, equipped with a single 24GB graphics processing unit, 128GB of RAM, and a multi-core server processor to support the aforementioned model training and online inference. After this hardware and software co-configuration, the electronic device can complete dossier-level archiving determination and abnormal path location in a continuous registration event stream.
[0134] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.
Claims
1. A dynamic heterogeneous graph-based real estate registration time sequence evidence archiving method, characterized in that, include: Step S1: Obtain the registration business objects and their temporal relationships, map the application form, file, business entity, real estate unit, certificate, approval process and time snapshot into a dynamic heterogeneous graph, and generate the evidence storage chain subgraph, version evolution subgraph and cross-file reference subgraph; Step S2: Perform structured field encoding, text digest encoding, and log time location encoding on various types of nodes in the dynamic heterogeneous graph, inject time encoding into various types of edges, and expand the relationship between a single business association of multiple entities and multiple materials into a hyperedge structure. Step S3: Generate a dossier-level archiving decision representation based on heterogeneous relation propagation, temporal edge sorting learning, graph comparison chain consistency learning, and key edge reconstruction learning; Step S4: Perform local subgraph incremental update on newly entered registration events, and output the archiving probability, chain integrity score, and abnormal edge location set based on the dossier-level archiving determination representation; Step S5: Generate archiving processing instructions based on the archiving probability, chain integrity score, and abnormal edge location set, and trigger the corresponding archiving action to generate structured archiving results.
2. The method of claim 1, wherein, Construct a dynamic heterogeneous graph and generate a subgraph of evidence storage chains, a subgraph of version evolution, and a subgraph of cross-volume references, including: Read the application forms, file files, rights holders, obligors, business personnel, real estate units, certificates, approval processes, and time snapshots within the pending judgment time window to obtain the set of registration business objects; A dynamic heterogeneous graph is constructed based on the set of registered business objects, and a relational view corresponding to different business constraints is extracted from the dynamic heterogeneous graph using the case file number as the anchor point. Based on the relationship view, the evidence storage chain subgraph, version evolution subgraph, and cross-volume reference subgraph are obtained respectively, and a time-series index of the volume, subject, and real estate unit is established.
3. The method of claim 2, wherein the method further comprises: Constructing dynamic heterogeneous graphs also includes: The application forms, file files, rights holders, obligors, business personnel, real estate units, certificates, approval processes, and time snapshots in the set of registration business objects are mapped to different node types to obtain a node set; The actions of submission, review, reference, change, inheritance, cancellation, supplementation, version continuation, and hash anchoring are mapped to different edge types, and the generation time of each edge is recorded to obtain an edge set; Based on the set of nodes, the set of edges, and the generation time of each edge, a node type mapping, an edge type mapping, and a timestamp function are established to obtain the dynamic heterogeneous graph.
4. The method of claim 3, wherein, Generate evidence storage chain subgraphs, version evolution subgraphs, and cross-volume reference subgraphs, including: Using the case file number as the anchor point, the closed-loop trace relationship between the application form, certificate, approval process, hash anchor and case file is extracted from the dynamic heterogeneous graph to obtain the evidence storage chain graph. The relationship between supplementary entries, version continuation, changes, and cancellation is extracted from the dynamic heterogeneous graph to obtain the version evolution subgraph; The cross-volume reference subgraph is obtained by extracting inheritance, reference, and real estate unit sharing relationships from the dynamic heterogeneous graph. When the current dossier has multiple time snapshots, a link between snapshots is generated in the version evolution subgraph. When the current dossier has only a single time snapshot, the single snapshot view is retained and the construction of positive samples between snapshots is skipped, thus obtaining the evidence link subgraph, the version evolution subgraph, and the cross-dossier reference subgraph.
5. The method of claim 4, wherein, Structured field encoding, text digest encoding, and log time location encoding are performed on various types of nodes in the dynamic heterogeneous graph, including: Extract the dossier number, business type, certificate type, status code, approval conclusion, version number, and field hash of each node, and perform structured field encoding to obtain the structured field representation; Extract the application summary, certificate summary, and approval opinion, perform text summary coding, and obtain the text summary representation. When the original material only contains the full text, the title, summary paragraph, and approval opinion paragraph are trimmed from the original text according to the preset trimming rules. Extract the sequence of events, the interval between adjacent business actions, and the time snapshot position, perform log time position encoding to obtain the log time position representation; concatenate the structured field representation, text digest representation, and log time position representation to obtain a unified node representation.
6. The method of claim 5, wherein the method further comprises: Time-encoded for various edges, and the relationship between a single business transaction and multiple entities and materials is expanded into a hyper-edge structure, including: Based on the time difference between the generation time of each edge and the current snapshot reference time, time encoding is performed on various types of edges to obtain edge time representations. When log times are missing or conflicting, the time difference or additional missing time flags are generated according to adjacent known time interpolation, interval median replacement, standard time sequence template replacement, and the main log priority rule. Identify multiple rights holders, multiple obligors, multiple supporting documents and approval processes involved in the same business operation, establish a high-order association carrier, and obtain a hyper-edge structure; Based on the unified node representation, edge temporal representation, and hyperedge structure, higher-order propagation is performed to obtain node representations with temporal relationships and higher-order business relationships.
7. The method of claim 6, wherein the method further comprises: Based on heterogeneous relation propagation, temporal edge ranking learning, graph-based comparative chain consistency learning, and key edge reconstruction learning, a dossier-level archiving decision representation is generated, including: Based on the node representation with time relationship and higher-order business relationship, heterogeneous relationship propagation is performed according to the joint constraints of node type, edge type and time encoding, and read aggregation is performed on the dossier-related node set to obtain the dossier-level intermediate representation; Legitimate sequential relationships are extracted from the real event chain of the same file as positive sample edges, and difficult negative sample edges are extracted from candidate relationships that are close in time but inconsistent in business links and similar in fields but conflicting in entity mapping. Temporal edge ranking learning is then performed. Snapshots of the same dossier at different times during its legal evolution are constructed as positive sample pairs, and snapshots with tampering, missing pages, rollback operations, and erroneous cross-dossier references are constructed as difficult negative sample pairs. Graph comparison chain consistency learning is performed, and masking reconstruction is performed on reference relationships, audit relationships, and change relationships. Joint learning is performed based on the classification, sorting, comparison, and reconstruction objectives of the archiving stage to obtain the dossier-level archiving determination representation.
8. The method for time-series evidence archiving of real estate registration based on dynamic heterogeneous graphs according to claim 7, characterized in that, Output the archiving probability, chain integrity score, and anomaly edge location set, including: Based on the dossier-level archiving determination, the archiving probability is output; the critical edge reconstruction consistency score, temporal sequence consistency score, cross-snapshot consistency score, and anchor closure score are calculated and fused according to preset item weights to obtain the chain integrity score; For each candidate edge, the edge reconstruction bias, edge order scoring bias, and cross-snapshot consistency bias are calculated and fused according to the preset bias fusion weight to obtain the comprehensive score of the abnormal edge. The abnormal edge location set can be obtained by any of the following methods: filtering abnormal edges based on the comprehensive score of the abnormal edges and a preset abnormal threshold; and sorting and outputting the abnormal edges according to their comprehensive scores when the number of edges that meet the threshold exceeds the preset upper limit of the number of abnormal edges.
9. A method for time-series evidence archiving of real estate registration based on dynamic heterogeneous graphs according to claim 8, characterized in that, Based on the archiving probability, chain integrity score, and abnormal edge location set, an archiving processing instruction is generated, and the corresponding archiving action is triggered to generate a structured archiving result, including: When the archiving probability is higher than the preset archiving threshold, the chain integrity score is higher than the preset trusted chain integrity threshold, and the abnormal edge location set is empty or only contains edges with an abnormal edge comprehensive score lower than the preset abnormal threshold, an automatic archiving instruction is generated, and electronic case file encapsulation, archiving index generation, and evidence storage result writing are performed. A manual review instruction is generated and written to the abnormal path description and review queue when any of the following conditions are met: the archiving probability is lower than a preset low confidence threshold; the chain integrity score is lower than a preset low confidence chain integrity threshold; or there are edges in the abnormal edge location set that have an order deviation exceeding the threshold and inconsistent cross-volume source. In other cases, a delayed archiving instruction is generated and written to the pending completion status; The output is a structured archiving result containing the dossier number, archiving probability, chain integrity score, anomaly edge location set, archiving action type, and anomaly path summary.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the real estate registration time-series evidence archiving method based on dynamic heterogeneous graphs as described in any one of claims 1 to 9.
