Method and system for evaluating and repairing graph data quality based on multi-agent collaboration

Abstract

The application provides a kind of based on multi-agent cooperation's graph data quality evaluation and repair method and system, comprising: with the node attribute of the graph data set to be evaluated, edge relationship as input, extract attribute and topological dimension characteristics, construct multidimensional quality characteristic matrix and initial prior score vector;Rely on graph structure quality rule base to carry out automatic rule check, obtain layered quality problem list and score penalty increment;Combined with two-stage quality weighted purification view iteration and graph neural network, distinguish rule violation and neighborhood pollution defects, locate implicit topological defects;Based on graph database log and historical data, by whitening correction, mahalanobis distance calculation and double threshold mechanism to generate structured alarm event;Through four types of agent collaborative adaptive repair, determine convergence with Pareto vector, and construct autonomous quality assurance loop.The application solves the problems of topological perception loss, neighborhood pollution blindness, threshold reliability deficiency and poor adaptive repair strategy.

Description

Technical Field

[0001] This invention belongs to the field of graph data management and data quality engineering in computer science. Specifically, it relates to a graph data quality assessment and repair method and system based on multi-agent collaboration. It is applicable to general graph data scenarios such as social networks, financial transaction networks, Internet of Things, supply chain networks, and bioinformatics networks. It comprehensively utilizes graph neural networks, large-scale language model multi-agent frameworks, conformal risk control, and whitening normalization technologies to construct an intelligent quality control system covering the entire lifecycle of graph data. Background Technology

[0002] Graph data has been widely used in social networks, financial transactions, the Internet of Things, supply chains, and bioinformatics networks, becoming a core data form for connecting complex relationships between entities. Unlike relational data, the quality issues of graph data are not only reflected at the node attribute level (missing values, format errors, duplicate records), but also at the graph structure level—violations of graph-specific constraints such as edge direction validity, degree distribution rationality, and topological connectivity can have a profound impact on downstream graph analysis and graph neural network modeling.

[0003] Existing graph data management solutions have the following key shortcomings, which constitute the core technical problem to be solved by this invention:

[0004] First, the lack of graph topology quality awareness and the blind spots caused by neighborhood contamination. Existing data quality detection frameworks are designed for relational data and lack a dedicated constraint rule library for graph topology, making it impossible to systematically identify graph structure-level defects such as edge direction violations, abnormal degree distributions, and isolated nodes exceeding thresholds. A deeper problem is that the neighborhood aggregation mechanism of graph neural networks defaults to treating all neighboring nodes as equally weighted and trustworthy sources. When neighboring nodes have quality defects, their low-quality features will contaminate the representation vector of the target node during the aggregation operation, causing normal nodes to be misjudged as abnormal due to "neighborhood contamination," forming a hidden blind spot that traditional rule engines cannot detect.

[0005] Second, fixed threshold monitoring suffers from insufficient reliability, and remediation strategies lack adaptability. Existing graph data monitoring solutions trigger alarms with fixed statistical thresholds, which cannot adapt to the normal evolution and drift of the graph as business grows, leading to systemic false alarms. Alarm decisions lack quantifiable statistical reliability guarantees, and operations personnel cannot assess the upper bounds of the false positive and false negative rates for each alarm. On the remediation side, existing solutions rely on manually preset static remediation rules, failing to dynamically adjust remediation strategies based on defect type, severity, and graph business semantics. Furthermore, remediation results lack independent verification mechanisms, failing to form a self-closing quality assurance loop. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art by providing a graph data quality assessment and repair method and system based on multi-agent collaboration, in order to solve technical problems such as lack of graph topology quality perception, blinding due to neighborhood contamination, insufficient reliability of fixed threshold monitoring, and lack of adaptability of repair strategies.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a method for graph data quality assessment and repair based on multi-agent collaboration, comprising:

[0009] Using the set of node attributes and the set of edge relationships of the graph dataset to be evaluated as input, attribute dimension features and topology dimension features are extracted to construct a multidimensional quality feature matrix. Based on the attribute completeness rate and degree quantile reasonableness score of each node, an initial prior score vector is generated.

[0010] Based on the multidimensional quality feature matrix and the domain semantic rule definition file, an automated rule check is performed on the multidimensional quality feature matrix according to the graph structure-specific quality rule library to obtain a hierarchical quality problem list and rule-marked scoring penalty increments.

[0011] Based on the graph dataset to be evaluated, the initial prior score vector, and the rule-labeled score penalty increment, the implicit topological defect location report is obtained by iteratively constructing a quality-weighted clean view and conducting collaborative analysis with graph neural networks to distinguish between rule violation defects and neighborhood contamination blinding defects.

[0012] Using the graph database change log stream and historical normal distribution data of the graph dataset to be evaluated as input, and combined with the affected node range in the implicit topology defect location report, an alarm decision with adaptive benchmark correction and statistical reliability guarantee is performed on the multi-dimensional monitoring indicators to obtain structured alarm events.

[0013] Based on the hierarchical quality problem list, the implicit topological defect location report, and the structured alarm events, adaptive repair and iterative convergence verification are performed collaboratively by multiple agents to output the final repair report.

[0014] Optionally, the step of scoring the attribute completeness rate and degree quantile reasonableness of each node includes:

[0015] For the attribute dimension data in the multidimensional quality feature matrix, the proportion of non-empty attribute fields in each node to all fields that should be filled is calculated to obtain the attribute completeness rate of each node.

[0016] Based on the degree feature dimension data in the multidimensional quality feature matrix and the full-graph degree distribution histogram, using the full-graph degree distribution histogram as a benchmark, the cumulative distribution quantile position of each node's degree value in the full-graph degree distribution histogram is calculated to obtain the degree quantile position of each node; nodes whose degree quantile position exceeds the first quantile threshold are assigned a low reasonableness score, and nodes whose degree quantile position is lower than the second quantile threshold are assigned a low reasonableness score, wherein the first quantile threshold is greater than the second quantile threshold, to obtain the degree quantile reasonableness score of each node;

[0017] The attribute completeness rate and the degree quantile reasonableness score are weighted and linearly combined to obtain the unsupervised quality prior score of each node, which constitutes the initial prior score vector.

[0018] Optionally, the graph structure-specific quality rule base includes:

[0019] Attribute rules are used to perform non-null constraint checks, data type compliance checks, and value range compliance checks on the attribute fields of each node.

[0020] Direction rules are used to perform directed or undirected graph operations on the direction attributes of each edge. Figure 1 Adaptability testing;

[0021] Topology rules are used to construct the overall graph topology based on the set of node attributes and the set of edge relationships. The rules perform constraints on the overall graph topology, including checks on the proportion of isolated nodes, the interval of graph density, and the degree distribution power law exponent being within the interval formed by the lower and upper bounds of the power law exponent.

[0022] Optionally, the automated rule check is performed on the multidimensional quality feature matrix based on the graph structure-specific quality rule base to obtain a hierarchical quality problem list and rule-marked scoring penalty increments, including:

[0023] Using the attribute rules in the graph structure-specific quality rule library and the attribute dimension data in the multidimensional quality feature matrix as input, the attribute quality assertion execution engine is invoked to perform rule assertion checks on each node attribute field, and an attribute class rule execution report is obtained.

[0024] Based on the directional rules and topology rules in the graph structure-specific quality rule library, and combined with the topology dimension data in the multidimensional quality feature matrix, a graph query language is called to perform a topology statistical query to obtain a structure-class rule execution report.

[0025] Based on the attribute-type rule execution report and the structure-type rule execution report, a fixed attenuation amount is assigned to the node marked by any rule to obtain the rule marking score penalty increment. The problem records in the attribute-type rule execution report and the structure-type rule execution report are classified and summarized according to severity level to obtain the hierarchical quality problem list containing severity level, warning level and information level.

[0026] Optionally, the step of distinguishing between rule-violation defects and neighborhood contamination-induced blinding defects through quality-weighted cleanup view iterative construction and graph neural network collaborative analysis to obtain an implicit topological defect location report includes:

[0027] Based on the initial prior score vector and the rule-marked score penalty increment, the initial prior score of each node is added to the corresponding rule-marked score penalty increment and restricted to a closed numerical range of zero to one. When the rule-marked score penalty increment is zero, it degenerates into the initial prior score, thus obtaining the node quality score of each node.

[0028] Based on the node quality scores and the adjacency matrix of the graph dataset to be evaluated, the adjacency matrix is ​​quality-weighted by the node quality scores to obtain a quality-weighted adjacency matrix; a first-generation cleaned view is constructed based on the quality-weighted adjacency matrix and the set of node attributes, and a graph neural network model is trained on the first-generation cleaned view to obtain the embedding vector of each node;

[0029] Based on the embedding vectors of each node, the degree to which the embedding vectors of each node deviate from the normal distribution of the embedding space is evaluated. Nodes whose deviation exceeds the embedding deviation judgment threshold have their corresponding node quality scores reduced to obtain a refined quality score. The second-generation cleaned view is reconstructed based on the refined quality score and the adjacency matrix.

[0030] Based on the second-generation cleanup view and the hierarchical quality problem list, for each subgraph in the second-generation cleanup view, the subgraph performance deviation between the baseline accuracy and the accuracy on the cleanup view is calculated, and defective subgraphs with deviations exceeding the graph neural network performance deviation judgment threshold are filtered out. Combined with cross-validation of the hierarchical quality problem list, defective subgraphs that match the rule records in the hierarchical quality problem list are marked as rule violation defects, and defective subgraphs that do not match the rule records in the hierarchical quality problem list are marked as neighborhood contamination blinding defects, thus obtaining the implicit topological defect location report.

[0031] Optionally, the alarm decision-making process that performs adaptive benchmark correction and statistical reliability guarantees on multi-dimensional monitoring indicators includes the following steps:

[0032] Using the normal graph snapshot index sequence within the sliding window in the historical normal distribution data as input, the mean vector of the normal graph snapshot index sequence is calculated, and diagonal loading regularization is applied to the covariance matrix of the normal graph snapshot index sequence to obtain the mean vector and the covariance matrix with diagonal loading regularization.

[0033] Based on the comparison between the number of normal image snapshots within the sliding window and the minimum snapshot number threshold, when the number of normal image snapshots is less than the minimum snapshot number threshold, diagonal whitening is performed by replacing the covariance matrix with diagonal loading regularization with dimension-wise independent normalization to obtain a whitening transformation matrix; when the number of normal image snapshots is not less than the minimum snapshot number threshold, full covariance whitening is performed based on the covariance matrix with diagonal loading regularization to obtain a whitening transformation matrix.

[0034] Based on the multi-dimensional monitoring indicator vector, the mean vector, and the whitening transformation matrix corresponding to the current time window of the graph database change log stream, the whitened Mahalanobis distance anomaly score is calculated to obtain the whitened Mahalanobis distance anomaly score for the current time window.

[0035] Optionally, the alarm decision-making process for performing adaptive benchmark correction and statistical reliability guarantee on multi-dimensional monitoring indicators includes:

[0036] Using the historical normal albinism abnormality score sequence, the user-specified acceptable false negative rate, and the user-specified acceptable false positive rate as inputs, and based on the quantile distribution of the historical normal albinism abnormality score sequence, the false negative rate threshold satisfying the first confidence level and the false positive rate threshold satisfying the second confidence level are determined respectively.

[0037] Based on the aberration score of the whitened Mahalanobis distance in the current time window, the false negative rate threshold, and the false positive rate threshold, an alarm trigger state is output when the aberration score of the whitened Mahalanobis distance exceeds the false positive rate threshold; an alarm suppression state is output when the aberration score of the whitened Mahalanobis distance is lower than the false negative rate threshold; and an uncertain state is output when the aberration score of the whitened Mahalanobis distance is between the false negative rate threshold and the false positive rate threshold, thus obtaining an alarm status marker.

[0038] Based on the alarm status marker, the alarm timestamp, the whitened Mahalanobis distance anomaly score, the alarm type, and the range of affected nodes from the implicit topological defect location report are recorded to generate the structured alarm event.

[0039] Optionally, the step of performing adaptive repair and iterative convergence verification through multi-agent collaborative execution, and outputting a final repair report, includes:

[0040] Using the hierarchical quality problem list, the implicit topological defect location report, and the structured alarm events as input, the detection agent extracts the defect node set, defect edge set, affected subgraph range, and problem priority sorting to obtain a defect location summary;

[0041] Based on the defect location summary, the planning agent sorts the defect node set, defect edge set and problem priority in the defect location summary according to the repair strategy knowledge base, constrains the type field, target node field and repair parameter field of the repair operation in JSON format, generates a candidate repair operation set, and selects a repair operation sequence from the candidate repair operation set through a search strategy.

[0042] Based on the repair operation sequence, the action agent executes each repair operation in the repair operation sequence in order of the problem priority to obtain the repaired image;

[0043] Based on the repaired graph, the evaluation agent calculates the Pareto quality improvement vector, which includes three decreasing components: attribute integrity rate improvement, rule violation reduction, and defective subgraph elimination.

[0044] Based on the Pareto quality improvement vector, determine whether the convergence condition is met: if the L1 norm of the Pareto quality improvement vector is less than the convergence threshold for two consecutive rounds and each decreasing component does not exceed the tolerance threshold, then convergence is determined and the final repair report is output; otherwise, replace the dataset of the graph to be evaluated with the repaired graph and re-execute the collaborative steps of the detection agent, planning agent, action agent and evaluation agent.

[0045] Optionally, the step of having the action agent execute each repair operation in the repair operation sequence according to the problem priority to obtain the repaired graph includes:

[0046] Based on the repair operation sequence and the refinement quality score, for each repair operation in the repair operation sequence, according to the mapping relationship between node type and attribute field, and combined with the refinement quality score, the target fill value corresponding to each repair operation is determined, thus obtaining a repair operation sequence carrying the target fill value;

[0047] According to the repair operation sequence carrying the target fill value, the repair operations are performed in order according to the problem priority. According to the type field of each repair operation, the missing attribute field is filled, the erroneous attribute field is overwritten, the isolated node is reconnected, and the low-quality node is softly deleted, so as to obtain the repaired image.

[0048] Based on the repaired graph and the repair operation sequence carrying the target fill value, an operation log is generated for each executed repair operation in the repaired graph, recording the operation type, target node identifier, attribute value before repair, and attribute value after repair, thus obtaining the repair operation audit record.

[0049] This invention also provides a graph data quality assessment and repair system based on multi-agent collaboration, comprising:

[0050] The multidimensional feature construction module is used to extract attribute dimension features and topological dimension features from the set of node attributes and the set of edge relationships of the graph dataset to be evaluated, construct a multidimensional quality feature matrix, and generate an initial prior score vector based on the attribute completeness rate and degree quantile reasonableness score of each node.

[0051] The rule checking module is used to perform automated rule checking on the multidimensional quality feature matrix based on the multidimensional quality feature matrix and the domain semantic rule definition file, according to the graph structure-specific quality rule library, to obtain a hierarchical quality problem list and rule tag scoring penalty increment;

[0052] The implicit defect localization module is used to distinguish between rule violation defects and neighborhood contamination blinding defects based on the graph dataset to be evaluated, the initial prior score vector, and the rule-labeled score penalty increment, through quality-weighted cleanup view iterative construction and graph neural network collaborative analysis, and to obtain an implicit topological defect localization report.

[0053] The alarm decision module is used to take the graph database change log stream and historical normal distribution data of the graph dataset to be evaluated as input, and combine the affected node range in the implicit topology defect location report to perform adaptive benchmark correction and statistical reliability guarantee alarm decision on multi-dimensional monitoring indicators to obtain structured alarm events.

[0054] The multi-agent repair module is used to perform adaptive repair and iterative convergence verification through multi-agent collaboration based on the hierarchical quality problem list, the implicit topological defect location report, and the structured alarm events, and output the final repair report.

[0055] The technical effects of this invention are as follows:

[0056] 1. This invention proposes a graph structure-specific quality rule library and an automated detection engine, breaking through the traditional relational data quality detection framework. It is the first to realize the systematic, configurable, and automated detection of graph structure constraints. The rule execution results simultaneously generate node quality score penalty increments, providing structured priors for subsequent clean view mechanisms and eliminating the overhead of duplicate detection.

[0057] 2. This invention proposes a graph neural network implicit defect perception mechanism for iterative refinement of the clean view in two stages. In the first stage, an initial quality-weighted clean view is constructed by superimposing unsupervised prior scores with rule label penalties. In the cold start scenario, it degenerates into pure prior scores to ensure that the mechanism is always effective. In the second stage, the quality score is refined by embedding spatial bias in the graph neural network and the second-generation clean view is reconstructed. Through two iterations, the pollution effect of low-quality neighbors on node representation is completely eliminated, and two types of implicit structural defects are distinguished: "rule violation defects" and "neighborhood pollution blinding defects", filling the detection blind spots of traditional rule engines.

[0058] 3. This invention proposes a sliding window whitening correction and shape preservation dual-threshold statistical alarm mechanism. The whitening correction mechanism enables the alarm benchmark to be automatically adjusted as the graph evolves normally, fundamentally eliminating systemic false alarms caused by business growth. The shape preservation risk control dual-threshold mechanism attaches quantifiable upper bounds for false negative and false positive rates to each alarm decision, upgrading graph quality monitoring from experience threshold triggering to a reliable alarm system driven by statistical assurance.

[0059] 4. This invention proposes an adaptive repair planning-verification closed loop with JSON template constraints for a large language model multi-agent system. The planning agent replaces the hard-coded static rules with large language model inference, and the JSON structured operation template eliminates the semantic gap. The evaluation agent introduces Pareto quality improvement vectors to determine convergence and prevents multi-index oscillations caused by single-index optimization. When the repair acceptance is not met, the next iteration is automatically triggered, forming a self-closing full lifecycle graph data quality assurance loop. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating the graph data quality assessment and repair method based on multi-agent collaboration described in this invention.

[0062] Figure 2 This is a schematic diagram of the two-stage purification view iterative refinement and graph neural network collaborative analysis process described in this invention;

[0063] Figure 3 This is a schematic diagram of the graph data quality assessment and repair system based on multi-agent collaboration described in this invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0065] Example 1

[0066] like Figure 1 As shown, this invention provides a method for graph data quality assessment and repair based on multi-agent collaboration, including:

[0067] Step S1: Using the set of node attributes and the set of edge relationships of the graph dataset to be evaluated as input, extract attribute dimension features and topology dimension features, construct a multidimensional quality feature matrix, and generate an initial prior score vector based on the attribute completeness rate and degree quantile reasonableness score of each node.

[0068] Specifically, the graph dataset to be evaluated is stored in a graph database (such as Neo4j or JanusGraph), containing a set of node attributes and a set of edge relationships. The system extracts all node attribute data and edge relationship data from the graph database using a batch query method, and groups and organizes them according to node type and edge type to form a structured input. The extraction of attribute dimension features includes the attribute field values, missing data, and data type compliance of each node, forming an attribute matrix; the extraction of topological dimension features includes the in-degree, out-degree, and histogram of the degree distribution of the entire graph for each node, forming a degree feature matrix. The attribute matrix and the degree feature matrix are concatenated horizontally to obtain a multidimensional quality feature matrix. The number of rows in this matrix is ​​equal to the total number of nodes in the graph, and the number of columns covers the sum of the number of attribute dimension fields and the number of topological dimension fields. Based on this, according to the attribute completeness rate of each node (i.e., the proportion of non-empty attribute fields to all required fields) and the degree quantile reasonableness score (i.e., the degree to which the node degree value is within a reasonable quantile range in the degree distribution of the entire graph), an unsupervised quality prior score is generated for each node through a weighted linear combination. The prior scores of all nodes constitute the initial prior score vector. The initial prior score vector takes values ​​in a closed numerical range from zero to one, with higher scores indicating better node quality.

[0069] Step S2: Based on the multidimensional quality feature matrix and the domain semantic rule definition file, perform automated rule checks on the multidimensional quality feature matrix according to the graph structure-specific quality rule library to obtain a hierarchical quality problem list and rule-marked scoring penalty increments.

[0070] Specifically, the domain semantic rule definition file is pre-configured by domain experts, describing attribute constraints, edge direction constraints, and topological constraints in the graph business scenario using structured YAML format. After loading the domain semantic rule definition file, the graph structure-specific quality rule library contains three types of rules: attribute rules, direction rules, and topological rules. The rule engine takes the attribute dimension data from the multi-dimensional quality feature matrix as input, calls the attribute quality assertion execution engine to perform assertion checks on each node's attribute fields for non-empty constraints, data type compliance, and value range compliance, generating an attribute-type rule execution report. Simultaneously, it calls the graph query language to perform topological statistical queries on the overall graph topology, including isolated node proportion constraints, graph density interval constraints, and degree distribution power-law exponential interval constraints, generating a structure-type rule execution report. A fixed decay amount is assigned to nodes marked by any rule as the rule marking score penalty increment, and all problem records are categorized and summarized according to severity, warning, and information levels to obtain a hierarchical quality problem list.

[0071] Step S3: Based on the graph dataset to be evaluated, the initial prior score vector, and the rule-labeled score penalty increment, the implicit topological defect location report is obtained by iteratively constructing a quality-weighted clean view and conducting collaborative analysis with graph neural networks to distinguish between rule violation defects and neighborhood contamination blinding defects.

[0072] Specifically, the initial prior score vector is added node-by-node to the rule-labeled score penalty increment, and the sum is restricted to a closed numerical range between zero and one to obtain the node quality score for each node. When the rule-labeled score penalty increment for a node is zero, the node quality score degenerates to the initial prior score, ensuring that the mechanism remains effective in a cold start scenario where the rule execution report is empty. The adjacency matrix of the graph dataset to be evaluated is quality-weighted based on the node quality score to construct a quality-weighted adjacency matrix, and the first-generation cleaned view is generated accordingly. A graph neural network model is trained on the first-generation cleaned view to obtain the embedding vectors of each node. Then, the degree to which the embedding vectors of each node deviate from the normal distribution of the embedding space is evaluated. The node quality scores of nodes whose deviation exceeds the embedding deviation judgment threshold are further reduced to obtain refined quality scores, and the second-generation cleaned view is reconstructed accordingly. On the second-generation clean view, the subgraph performance deviation between the baseline accuracy and the accuracy on the clean view is calculated for each subgraph. Defective subgraphs with performance deviation exceeding the performance deviation judgment threshold of the graph neural network are screened. Combined with the hierarchical quality problem list for cross-validation, defective subgraphs that hit the rule record are marked as rule violation defects, and defective subgraphs that do not hit the rule record are marked as neighborhood contamination blinding defects. Finally, an implicit topological defect location report is output.

[0073] Step S4: Using the graph database change log stream and historical normal distribution data of the graph dataset to be evaluated as input, and combining the affected node range in the implicit topology defect location report, perform adaptive benchmark correction and statistical reliability guarantee alarm decisions on the multi-dimensional monitoring indicators to obtain structured alarm events.

[0074] Specifically, the graph database change log stream continuously collects graph structure changes within the current time window using a change data capture mechanism, extracting multi-dimensional monitoring indicator vectors such as degree distribution KL divergence, core node edge change rate, and graph density change. Historical normal state distribution data stores a multi-dimensional indicator sequence of normal graph snapshots within a sliding window, used to adaptively estimate the mean vector and covariance structure of the current normal state. A whitening transformation with diagonal loading regularization normalizes the current multi-dimensional monitoring indicators, calculating the whitened Mahalanobis distance anomaly score, allowing the alarm baseline to automatically adjust with the normal business evolution of the graph, eliminating systemic false alarms. Based on this, the historical normal state whitening anomaly score sequence serves as a calibration set, estimating false negative and false positive rate thresholds through a conformal risk control framework, attaching quantifiable statistical upper bounds to alarm decisions. Combined with the affected node range in the implicit topology defect location report, alarm events are precisely located, outputting structured alarm events.

[0075] The change data capture mechanism refers to capturing write operation events such as node creation, node attribute modification, edge addition, and edge deletion in the graph database in near real-time by subscribing to the graph database's transaction log (such as Neo4j's Bolt protocol transaction flow or JanusGraph's JanusGraph-ES change subscription interface). These change events are then aggregated into structured change batches according to time windows and output to the subsequent multi-dimensional monitoring metric extraction process. Specifically, the length of each time window is determined by the system configuration parameter T. window The default timeframe is configurable to five minutes. At the end of each time window, the change data capture mechanism performs batch aggregation statistics on all change events within the window, calculating multi-dimensional monitoring indicators such as degree distribution KL divergence, core node edge change rate, and graph density change, forming a multi-dimensional monitoring indicator vector m for the current time window. t This is processed by the whitening correction subunit.

[0076] Step S5: Based on the hierarchical quality problem list, the implicit topological defect location report, and the structured alarm events, adaptive repair and iterative convergence verification are performed collaboratively by multiple agents to output the final repair report.

[0077] Specifically, the detection agent compiles a hierarchical quality problem list, implicit topological defect location reports, and structured alarm events, extracting the defect node set, defect edge set, affected subgraph range, and problem priority ranking to generate a defect location summary. The planning agent, leveraging the reasoning capabilities of a large-scale language model, generates a candidate repair operation set based on the defect location summary and graph business semantics using a structured JSON operation template, and selects a repair operation sequence from it through a search strategy. The action agent executes each repair operation in the repair operation sequence sequentially according to problem priority: filling missing attributes, overwriting erroneous attributes, reconnecting isolated nodes, and soft-deleting low-quality nodes, resulting in a repaired graph. The evaluation agent calculates the Pareto quality improvement vector. If the L1 norm of the Pareto quality improvement vector in two consecutive rounds is less than the convergence threshold and each decreasing component does not exceed the tolerance threshold, convergence is determined, and the final repair report is output; otherwise, the repaired graph replaces the graph dataset to be evaluated, and the above collaborative steps are re-executed, achieving a self-closing full lifecycle graph data quality assurance loop.

[0078] The search strategy refers to the algorithmic process of selecting an optimal combination sequence of repair operations that balances operation priority, expected quality improvement, and conflict-free constraints after the planning agent generates a set of candidate repair operations. When the number of candidate repair operations is small (e.g., no more than twenty), a greedy search strategy is adopted, which sorts the operations in descending order of expected quality improvement and adds them to the repair sequence sequentially. Simultaneously, it checks whether there is a target node conflict between each newly added operation and existing operations (i.e., two operations assigning different target fill values ​​to the same node attribute field). In case of conflict, the operation with higher priority is retained. When the number of candidate repair operations is large, a bundle search strategy is adopted, maintaining a candidate sequence set of size B (default B=5). In each step, one operation is added to each sequence from the candidate sequence set, retaining the B sequences with the highest comprehensive scores, and finally outputting the complete repair operation sequence with the highest comprehensive score.

[0079] Example 2

[0080] In this embodiment, the step of scoring the attribute completeness rate and degree quantile reasonableness of each node includes:

[0081] For the attribute dimension data in the multidimensional quality feature matrix, the proportion of non-empty attribute fields in each node to all fields that should be filled is calculated to obtain the attribute completeness rate of each node.

[0082] Specifically, the attribute dimension data in the multidimensional quality feature matrix records the actual filling status of each node in each attribute field, including whether the field is empty, the field type, and the field value. For any node j in the graph, the number of fields in all required attribute fields that actually have non-empty values ​​filled is counted, and the ratio of this number to the total number of required fields is the attribute completeness rate of node j, denoted as r. attr The attribute completeness rate ranges from zero to one, a closed numerical interval. A value of one indicates that all fields for that node have been filled, while a value of zero indicates that all fields for that node are empty. The set of fields to be filled for different node types is determined according to the schema definition of the graph database to ensure a consistent calculation method for completeness rate across node types. After the attribute completeness rate is calculated, it is aggregated node by node to form an attribute completeness rate vector, which serves as the input for subsequent weighted linear combinations.

[0083] When determining the required field set for each node type, the schema metadata of the graph database is read (for example, using Neo4j, this can be obtained through the `CALL apoc.meta.schema()` interface or the metadata query interface of the database management system). All attribute fields defined for each node type are extracted, and these are combined with the mandatory field markers for each node type in the domain semantic rule definition file. Fields marked as mandatory are then included in the required field set. For fields not explicitly marked as mandatory in the schema, the default global policy in the domain semantic rule definition file is applied (e.g., the default policy is "business primary key field is mandatory, other fields are optional"). The required field sets for different node types are independent of each other. For example, in a social network graph, the required field set for a user node type might include user identifier, registration time, and account status, while the required field set for a content node type might include content identifier, publication time, and content type. The attribute completeness calculation methods for both types do not interfere with each other, ensuring that the cross-node type completeness comparison is business-interpretable.

[0084] Based on the degree feature dimension data in the multidimensional quality feature matrix and the full-graph degree distribution histogram, using the full-graph degree distribution histogram as a benchmark, the cumulative distribution quantile position of the degree value of each node in the full-graph degree distribution histogram is calculated to obtain the degree quantile position of each node; nodes whose degree quantile position exceeds the first quantile threshold are assigned a low reasonableness score, and nodes whose degree quantile position is lower than the second quantile threshold are assigned a low reasonableness score, wherein the first quantile threshold is greater than the second quantile threshold, to obtain the degree quantile reasonableness score of each node.

[0085] Specifically, the degree feature dimension data records the in-degree and out-degree of each node, with the sum of the in-degree and out-degree as the node's total degree value. The full-graph degree distribution histogram is constructed by statistically analyzing the frequency distribution of the total degree values ​​of all nodes in the graph. The horizontal axis represents the degree value range, and the vertical axis represents the normalized frequency, covering the entire range from the minimum to the maximum degree value. For any node j in the graph, the corresponding cumulative distribution quantile position is found in the full-graph degree distribution histogram based on its total degree value, that is, the cumulative frequency value accumulated from the minimum degree value to the degree value corresponding to that node, denoted as . The rules for assigning the degree quantile reasonableness score are as follows: when the cumulative distribution quantile position F of node j is... j When the degree value exceeds the first quantile threshold (e.g., 0.95, indicating an extremely high degree value, located at the right tail end of the overall graph degree distribution), a low reasonableness score is assigned, indicating that the node may be a super node or that the degree value is abnormally high due to data errors; when the cumulative distribution quantile position F of node j is... j When the degree value is below the second quantile threshold (e.g., 0.05, indicating an extremely low degree value and located at the left tail end of the overall graph degree distribution), a low reasonableness score is assigned, indicating that the node may be an isolated or nearly isolated node; when the cumulative distribution quantile position F of node j is... j When the value falls between the second quantile threshold and the first quantile threshold, a normal reasonableness score is assigned. The first quantile threshold is greater than the second quantile threshold; both thresholds jointly define the boundary of the reasonableness score range. The reasonableness score of all nodes in the quantile is denoted as r. deg ,j, constitute the degree quantile rationality score vector.

[0086] The attribute completeness rate and the degree quantile reasonableness score are weighted and linearly combined to obtain the unsupervised quality prior score of each node, which constitutes the initial prior score vector.

[0087] Specifically, for any node j in the graph, its attribute completeness r attr ,j and degree quantile reasonableness score r deg The unsupervised quality prior score of node j is obtained by weighting the nodes according to the preset weight hyperparameters and performing a linear combination. The calculation formula is as follows:

[0088]

[0089] in, For node j, assign an unsupervised quality prior score; r attr ,j is the attribute completeness rate of node j; r degLet j be the degree quantile reasonableness score of node j; λ1 be the weighted hyperparameter of attribute completeness; λ2 be the weighted hyperparameter of degree quantile reasonableness score; λ1 and λ2 are both positive real numbers and satisfy λ1 + λ2 = 1. Their specific values ​​are configured by domain experts based on the actual business characteristics of the graph dataset. For example, in scenarios with rich attribute fields, λ1 can be set to a larger value to emphasize attribute quality; in scenarios with strict topology structure, λ2 can be set to a larger value to emphasize topology quality. Unsupervised prior quality scores for all nodes. Arranged in order of node number, forming the initial prior score vector q (0) The vector dimension is equal to the total number of nodes in the graph, and the values ​​of each component fall within a closed numerical range from zero to one.

[0090] Example 3

[0091] In this embodiment, the graph-specific quality rule base refers to a set of constraint rules specifically designed for graph data quality assessment scenarios. Essentially, it systematically and configurably encodes business semantic constraints and graph topological constraints into a computer-executable assertion rule system. The fundamental difference from traditional relational data quality rule bases lies in the fact that traditional rule bases only focus on the compliance of attribute fields of individual records in a row-column structure, while the graph-specific quality rule base simultaneously covers three layers: node attribute layer, edge semantic layer, and full graph topology layer. The edge semantic layer and full graph topology layer rules are constraint dimensions unique to graph data and completely missing in traditional relational data quality frameworks. Due to the absence of these two layers, existing data quality tools often miss many detections when processing graph data. For example, an error in an e-commerce graph where the "buy" edge points from the product node to the user node (reversed direction) cannot be detected by any relational data quality rule. Similarly, in a financial transaction graph, a power-law exponent in the degree distribution exceeding the normal range often indicates the injection of a large number of fraudulent accounts. These global topological anomalies also exceed the perception range of traditional row-by-row inspection frameworks.

[0092] The design principle of the graph structure-specific quality rule base follows the idea of ​​"starting from the natural constraints of graph data": as a relational data structure, the root cause of quality defects in graph data lies in two major types of constraint violations—one is semantic constraint violation, that is, the edge direction does not conform to the semantics of entity relationships in the business domain; the other is statistical constraint violation, that is, the overall topological statistics of the graph (such as the proportion of isolated nodes, graph density, degree distribution) deviate from the statistical regularities that real graph data should possess. Based on this design concept, the rule base organically combines three types of rules: attribute rules, direction rules, and topological rules. Attribute rules ensure the integrity of node-level data; direction rules ensure the correctness of edge-level semantics through edge direction consistency checks driven by domain semantics; and topological rules ensure the statistical rationality of the entire graph through interval constraint checks on the proportion of isolated nodes, graph density, and degree distribution power law exponent. These three types of rules form a hierarchical quality perception system from micro to macro, covering all manifestations of graph data quality problems.

[0093] Regarding the configurability design of the rule base, the domain semantic rule definition file adopts a structured YAML format, allowing domain experts to flexibly adjust rule parameters without modifying the system code. Taking a financial transaction graph scenario as an example, domain experts can configure the following constraints in the domain semantic rule definition file: the account balance field of the account node is of floating-point type and its value is greater than or equal to zero (attribute rule); the source node of the transfer edge must be an account type node, the target node must be an account type node, and self-loops are not allowed (direction rule); the isolated node ratio threshold θ iso Set to 0.02 (meaning the proportion of account nodes with no transaction records in the graph should not exceed two percent); graph density range [ρ min , ρ max The density distribution is set to [0.0001, 0.01] based on the historical normal transaction graph; the degree distribution power law exponent is configured within a reasonable range of [2, 3] (consistent with the power law statistical law of real financial transaction networks). All the above rule parameters are dynamically loaded through the domain semantic rule definition file, which can respond to changes in business rules without redeploying the system, achieving full lifecycle maintainability of the rule base.

[0094] At the implementation level of the automated detection engine, the attribute quality assertion execution engine uses open-source data quality frameworks such as Apache Griffin, Great Expectations, or AWS Deequ as its underlying execution foundation. It automatically converts YAML-formatted attribute rules into assertion objects for the corresponding frameworks and performs batch assertions on the attribute dimension data in the multi-dimensional quality feature matrix in a parallel execution mode. After the assertion execution is complete, it outputs an attribute-class rule execution report containing the identifier of the violating node, the number of the violating rule, and details of the violating field. Topology rules and direction rules are executed using a graph query language (Cypher or Gremlin) in a declarative query manner. This leverages the graph database's built-in graph traversal engine to efficiently handle topology statistical queries on large-scale graphs, avoiding the data transfer overhead of exporting the entire graph data to an external storage system for statistical calculations. The execution results of both types of detection engines are represented by a rule-labeled scoring penalty increment vector Δq. rule It is tightly coupled with the subsequent clean view mechanism: all nodes marked by any rule receive a negative decay penalty. This structured prior information will serve as the quality input for the downstream two-stage clean view construction, achieving seamless integration of rule detection and GNN collaborative analysis, and eliminating the overhead of independent and repetitive processing of rule detection and model training in traditional schemes.

[0095] In this embodiment, the graph structure-specific quality rule base includes:

[0096] Attribute rules are used to perform non-null constraint checks, data type compliance checks, and value range compliance checks on the attribute fields of each node.

[0097] Specifically, attribute rules are a set of basic constraint rules at the node attribute field level in the graph structure-specific quality rule library. The non-null constraint check verifies whether the node ID field, business primary key field, and attribute fields marked as required by domain experts in the domain semantic rule definition file contain null values. Nodes corresponding to fields with null values ​​are recorded as attribute rule violation nodes and written into the attribute class rule execution report. The data type compliance check verifies whether the actual storage type of each attribute field is consistent with the expected type defined in the graph database schema. For example, whether the amount field is a numeric type and the timestamp field is a date / time type. Nodes corresponding to fields with mismatched types are recorded as type violation nodes. The value range compliance check verifies whether the actual value of each numeric or enumerated attribute field falls within the legal value range configured in the domain semantic rule definition file. For example, whether the age field is between zero and one hundred and fifty and the transaction amount field is greater than zero. Nodes corresponding to fields with out-of-bounds values ​​are recorded as range violation nodes. All three types of attribute rules are described in structured YAML configuration files, loaded by the attribute quality assertion execution engine, and automatically converted into executable assertions for execution one by one. The check results are summarized in the attribute class rule execution report.

[0098] Direction rules are used to perform directed or undirected graph operations on the direction attributes of each edge. Figure 1 Causality test.

[0099] Specifically, direction rules are a set of semantic constraint rules for edge direction attributes within a graph structure-specific quality rule library. For directed graph scenarios, direction rules, based on the edge type direction constraints configured in the domain semantic rule definition file, verify each edge's source node type and target node type to ensure they conform to the expected direction semantics. For example, in an e-commerce graph, the "buy" edge is only allowed to point from a user type node to a product type node; in a financial graph, the source node of the "transfer" edge must be an account type node. Edges violating direction rules are recorded with their edge identifier, actual direction, and expected direction, and written to the structure-class rule execution report. For undirected graph scenarios, direction rules verify whether any edges in the edge set have been manually labeled with direction attributes. Edges with such labels are recorded as having direction consistency violations and written to the structure-class rule execution report. Direction rules are executed using a graph query language, batch-checking the direction attributes in the edge set through graph structure traversal to ensure the graph as a whole conforms to the semantic constraints of the business domain at the edge direction level.

[0100] Topology rules are used to construct the overall graph topology based on the set of node attributes and the set of edge relationships. The rules perform constraints on the overall graph topology, including checks on the proportion of isolated nodes, the interval of graph density, and the degree distribution power law exponent being within the interval formed by the lower and upper bounds of the power law exponent.

[0101] Specifically, topology rules are a set of high-level constraint rules in a graph-specific quality rule library, addressing the overall structural rationality of the graph. First, the overall graph topology is constructed based on the set of node attributes and edge relationships, and the degree distribution of each node and the overall graph connectivity statistics are calculated. The isolated node proportion constraint checks the proportion of nodes with a degree of zero (i.e., no edge connections) to the total number of nodes in the graph, verifying whether this proportion does not exceed the isolated node proportion threshold θ. iso If the threshold is exceeded, a topology violation is recorded. The graph density interval constraint check calculates the actual edge density of the current graph (i.e., the ratio of the actual number of edges to the maximum possible number of edges) and verifies whether this density falls within the minimum graph density threshold ρ. min With the maximum graph density threshold ρ maxWithin the legally defined interval, any violation of the interval constraints is recorded as a topological violation. The degree distribution power-law exponent constraint check fits a power-law distribution to the entire graph's degree distribution data, estimates the power-law exponent α, and verifies whether α falls within the legal interval defined by the lower and upper bounds of the power-law exponent (based on the statistical regularity of real graph data, a reasonable interval is typically configured as [2, 3]). If α exceeds this interval, it indicates an abnormal deviation in the graph's degree distribution, recorded as a topological anomaly. All the above topological rules are executed using a graph query language to perform topological statistical queries, and the check results are summarized in the structure rule execution report.

[0102] Among them, the power-law distribution fitting adopts the maximum likelihood estimation method: first, the lower bound x of the power-law fitting is determined. min The x-values ​​are automatically estimated from the full map degree value sequence using the minimization of the Kolmogorov-Smirnov statistic as the criterion. min ; when the degree value is not less than x min On a subsequence of , estimate the power law exponent using the maximum likelihood formula for a continuous power law distribution. The formula is Where n is a degree value not less than x min The number of nodes, x i Let be the degree value of the i-th node that satisfies the condition. To address the computational overhead of full fitting when the graph data is large, all nodes in the graph can be sampled hierarchically according to their degree values. A degree value sample set can be constructed with a sampling ratio of no less than 30% of the total number of nodes in the graph. The maximum likelihood estimation described above is then performed on the sample set. The deviation between the estimated power law exponent α and the full fitting is usually no more than 5% in experimental verification, which meets the engineering accuracy requirements.

[0103] Example 4

[0104] In this embodiment, the automated rule check is performed on the multidimensional quality feature matrix based on the graph structure-specific quality rule base to obtain a hierarchical quality problem list and rule-marked scoring penalty increments, including:

[0105] Using the attribute rules in the graph structure-specific quality rule library and the attribute dimension data in the multidimensional quality feature matrix as input, the attribute quality assertion execution engine is invoked to perform rule assertion checks on each node attribute field, resulting in an attribute class rule execution report.

[0106] Specifically, the attribute quality assertion execution engine takes the attribute rule configuration set in the graph-structured dedicated quality rule library and the attribute dimension data in the multidimensional quality feature matrix as input, automatically converting each rule in the attribute rule configuration set into an executable assertion. For each node in the multidimensional quality feature matrix, it runs non-empty constraint assertions, data type compliance assertions, and value range compliance assertions one by one, recording the pass or fail result of the assertion on each rule for each node. After the assertion execution is completed, the assertion execution results of all nodes are summarized and organized to generate an attribute-type rule execution report. The report includes the identifier of the violating node, the number of the violating rule, the name of the violating field, the actual field value, and the description of the violation. The data structure of the attribute-type rule execution report supports fast indexing by node identifier or rule number, facilitating the subsequent generation of a hierarchical quality problem list and the calculation of rule marking scoring penalty increments.

[0107] Based on the directional rules and topology rules in the graph structure-specific quality rule library, and combined with the topology dimension data in the multidimensional quality feature matrix, a graph query language is invoked to perform a topology statistical query, resulting in a structure-class rule execution report.

[0108] Specifically, based on the directional and topological rule configuration sets in the graph structure-specific quality rule library, and combined with the topological dimension data (including the in-degree, out-degree, and degree distribution statistics of each node) in the multidimensional quality feature matrix, a graph query language (such as Cypher or Gremlin) is invoked to perform batch statistical queries on the overall graph topology. The directional rule query checks each edge's source node type against the target node type using graph traversal statements to ensure directional constraints are met, returning a list of violating edges. The topological rule query calculates the proportion of isolated nodes, graph density, and degree distribution power-law exponent using aggregated statistical statements, comparing them with corresponding thresholds to return topological constraint violation flags. The directional and topological rule query results are merged and compiled to generate a structural rule execution report. The report includes detailed records of violating edges or topological features, violation rule numbers, and violation descriptions.

[0109] Based on the attribute-type rule execution report and the structure-type rule execution report, a fixed attenuation amount is assigned to the node marked by any rule to obtain the rule marking score penalty increment. The problem records in the attribute-type rule execution report and the structure-type rule execution report are classified and summarized according to severity level to obtain the hierarchical quality problem list containing severity level, warning level and information level.

[0110] Specifically, taking attribute-based rule execution reports and structure-based rule execution reports as input, we iterate through all violation records in both reports and extract the set of node identifiers marked as violations by any rule. For each node j in the set, we assign its rule-marked score penalty increment. Set to a fixed negative attenuation (e.g.) = -0.3, indicating that the node's quality score decreased by 0.3 due to rule violation. The rule-marking score penalty increment for nodes not marked by any rules is set to zero, thus obtaining the rule-marking score penalty increment vector for all nodes in the graph. Simultaneously, based on the severity level label configured for each rule in the domain semantic rule definition file, all issue records in the attribute-type rule execution reports and structure-type rule execution reports are categorized and summarized into three levels: severity level, warning level, and information level, forming a severe issue sublist, a warning issue sublist, and an information issue sublist, which together constitute a hierarchical quality issue list. The hierarchical quality issue list is indexed by both node identifier and severity level, supporting cross-validation queries in the subsequent implicit topology defect location report generation step.

[0111] The selection of the fixed attenuation amount follows these principles: First, the absolute value of the attenuation amount should be large enough to effectively suppress the weight contribution of the rule-marked violation node to neighborhood aggregation in the subsequent cleanup view construction; typically, the absolute value should not be less than 0.2. Second, the absolute value of the attenuation amount should not be too large to avoid excessive penalties for quality scores in rule false positive scenarios (i.e., a node is mistakenly marked as a violation due to improper rule configuration); typically, the absolute value should not exceed 0.5. Third, for rule violation scenarios of different severity levels, domain experts can configure differentiated attenuation amounts for rules of different severity levels in the domain semantic rule definition file. For example, a severe rule violation corresponds to an attenuation amount of -0.4, a warning rule violation corresponds to an attenuation amount of -0.2, and an informational rule violation corresponds to an attenuation amount of -0.1, to achieve refined quality score penalty control. It should be noted that the specific value of the attenuation amount can be optimized through hyperparameter search on the historical labeled dataset (containing known violation node labels), with the node classification accuracy of the neural network in the final cleanup view as the optimization objective, selecting the attenuation amount configuration that optimizes the accuracy.

[0112] Example 5

[0113] In this embodiment, the neighborhood contamination blinding defect refers to the phenomenon where, during the neighborhood aggregation operation, the target node's embedding vector is contaminated by the features of its low-quality neighbors due to the presence of defective nodes (such as nodes with numerous missing attribute fields, degree values ​​severely deviating from the normal range, or phantom nodes inconsistent with business semantics) among the target node's neighboring nodes. This ultimately manifests in downstream tasks (node ​​classification, link prediction, anomaly detection, etc.) as normal nodes being misclassified as anomalies or anomalies being missed. The core difference between the neighborhood contamination blinding defect and the rule violation defect is that the rule violation defect can be directly detected by performing rule assertions on the attribute fields of a single node or the direction attribute of a single edge; while the neighborhood contamination blinding defect does not manifest on any single node or edge. Its essence is an indirect quality impairment caused by the propagation of graph structure relationships, which can only be discovered by analyzing the performance difference of the graph neural network before and after purification. This is precisely the detection blind spot that traditional rule engines cannot cover.

[0114] The design principle of the two-stage cleanup view iterative refinement mechanism is based on the following technical logic: In the neighborhood aggregation operation of a graph neural network, the embedding vector of node i is determined by the weighted sum of its own features and the features of its neighboring nodes. If the quality of neighboring node j has serious defects (i.e., q...), the embedding vector of node i is determined by the weighted sum of its own features and the features of its neighboring nodes. j If the adjacency matrix approaches zero, then a quality weighting is applied to the adjacency matrix. This approach compresses the contribution weight of node j in the aggregation operation to near zero, ensuring that the embedding vector of node i is primarily determined by the features of high-quality neighbors, thereby suppressing the contamination of node representations by low-quality neighbors. The first stage constructs an initial cleaned view based on a combination of unsupervised prior scoring and rule-based penalties. Its advantage lies in its ability to start without any historical labeled data. In system cold-start scenarios (empty rule execution reports and no historical quality labels), it degenerates into a purely prior-scored cleaned view, ensuring the mechanism remains effective. The second stage uses the embedding vectors obtained from training a graph neural network on the initial cleaned view. It then utilizes the Mahalanobis distance deviation score in the embedding space to further refine the hidden low-quality nodes that the first stage may not have fully identified, reconstructing the second-generation cleaned view. This two-step iteration completely eliminates residual contamination effects.

[0115] It should be noted that subgraph performance deviation analysis ( The criterion for detecting defective subgraphs is based on the following theoretical basis: if there are no quality defects in a subgraph that would affect the performance of the graph neural network, then the cleaned-up view (by quality-weighted compression of the aggregate contribution of low-quality nodes) should have an impact on the model performance of that subgraph that is close to zero. i≈0; If a subgraph contains a neighborhood pollution blinding defect (i.e., some nodes' embedding vectors deviate from the normal distribution due to aggregation pollution from low-quality neighbors), then after suppressing the contribution of low-quality neighbors by cleaning the view, the model's performance on that subgraph should be significantly improved, i.e., accuracy. >acc base,i ,thereby i <0 (meaning performance improved after purification, with a negative deviation). Therefore, screening... i The absolute value exceeds the threshold for judging the performance deviation of the graph neural network. By analyzing the subgraph, regions with neighborhood contamination blinding defects can be accurately located. Furthermore, through cross-validation with a hierarchical quality issue list, it is possible to distinguish whether performance differences stem from explicit rule violations (rule violation defects) or from implicit contamination that the rule engine cannot detect (neighborhood contamination blinding defects), thus achieving accurate classification and location of both types of defects.

[0116] Taking a social network graph data quality assessment scenario as an example, the graph contains five million user nodes and eighty million social relationship edges. Within a certain user group (approximately 20,000 nodes), there are batch-injected bot account nodes. These bot account nodes have extremely low attribute field fill rates (attribute completeness rate of only 0.15), causing their refined quality score to approach zero. Through a two-stage iterative refinement using a cleaned view, in the first-generation cleaned view, the neighborhood aggregation weights of normal user nodes directly connected to the bot account nodes have been initially compressed. In the second-generation cleaned view, the embedding deviation assessment further identifies some bot account nodes with slightly higher attribute fill rates but whose embedding vectors still significantly deviate from the normal distribution, and performs a second reduction on their refined quality score. Finally, in the subgraph performance deviation analysis on the second-generation cleaned view, the performance deviation of the subgraph corresponding to this user group is... i Exceed Furthermore, cross-validation revealed that most defective nodes in the subgraph did not match the rule records in the hierarchical quality problem list (because although these machine accounts had low fill rates, they did not trigger strict non-empty constraints on required fields), and were accurately marked as neighborhood contamination blinding defects, providing precise defect location input for the subsequent multi-agent repair module.

[0117] like Figure 2 As shown, in this embodiment, the step of distinguishing between rule-violation defects and neighborhood contamination-induced blinding defects through quality-weighted cleanup view iterative construction and graph neural network collaborative analysis to obtain an implicit topological defect location report includes:

[0118] Step S31: Based on the initial prior score vector and the rule-marked score penalty increment, add the initial prior score and the corresponding rule-marked score penalty increment for each node and limit it to a closed numerical range of zero to one. When the rule-marked score penalty increment is zero, it degenerates into the initial prior score, thus obtaining the node quality score of each node.

[0119] Specifically, for any node j in the graph, its initial prior score q Corresponding rule-based scoring penalty increment q The results are summed and then truncated to a closed numerical range between zero and one using a truncation function to obtain the node quality score for node j. The calculation formula is as follows:

[0120]

[0121] Where, q j Score the node quality of node j; q The initial prior score for node j; The rule-marked score penalty increment for node j is set, with a negative value for nodes that violate the rules and a zero value for nodes that do not violate the rules. This is a truncation function that restricts the input value to a closed numerical range between zero and one. When When =0, the formula degenerates into This means the node quality score equals the initial prior score, ensuring the mechanism remains effective even in cold start scenarios (where the rule execution report is empty). The node quality score q for all nodes... j The node quality score vector q is formed, and its dimension is equal to the total number of nodes in the graph.

[0122] Step S32: Based on the node quality score and the adjacency matrix of the graph dataset to be evaluated, the adjacency matrix is ​​quality-weighted by the node quality score to obtain a quality-weighted adjacency matrix; a first-generation cleaned view is constructed based on the quality-weighted adjacency matrix and the node attribute set; a graph neural network model is trained on the first-generation cleaned view to obtain the embedding vector of each node.

[0123] Specifically, let A be the adjacency matrix of the graph dataset to be evaluated, where A ij Let be the edge weight between node i and node j, zero when there is no edge, and either zero or one when there is an edge. The adjacency matrix A is then weighted by the node quality score vector q to obtain the quality-weighted adjacency matrix A. (1) The calculation formula is as follows:

[0124]

[0125] in, A represents the weighted edge value from node i to node j in the quality-weighted adjacency matrix. ij q represents the edge weights in the original adjacency matrix. j The quality score of target node j is assigned. This weighting method compresses the contributions of low-quality neighbors during neighborhood aggregation while preserving the contributions of high-quality neighbors, thereby suppressing the contamination effect of low-quality neighbors on the target node's representation vector. This is based on the quality-weighted adjacency matrix A. (1) The first-generation cleanup view G is constructed using the node attribute set. (1) In G (1) The above uses graph neural network models such as GraphSAGE to perform node representation learning and training, and outputs the embedding vector of each node, denoted as H. (1) The dimension is |V|×d h Where |V| is the total number of nodes, d h For the embedded dimension.

[0126] Step S33: Based on the embedding vectors of each node, evaluate the degree to which the embedding vectors of each node deviate from the normal distribution of the embedding space, and reduce the corresponding node quality score of nodes whose deviation exceeds the embedding deviation judgment threshold to obtain a refined quality score; reconstruct the second-generation clean view based on the refined quality score and the adjacency matrix.

[0127] Specifically, using the embedding vector H of each node (1) Using the input as input, the mean vector and covariance structure of the embedding space are calculated. For each node's embedding vector, the degree to which it deviates from the normal distribution of the embedding space is calculated (measured by Mahalanobis distance or reconstruction error, etc.), resulting in an embedding deviation score for each node. Nodes with embedding deviation scores exceeding an embedding deviation threshold are then considered. Nodes marked as anomalous nodes in the embedding space are further degraded in terms of node quality score, resulting in a refined quality score q. (1) The refinement quality score is also limited to a closed numerical range of zero to one. Based on the refinement quality score q... (1) The second-generation cleaned view G is reconstructed using a quality-weighted approach, based on the original adjacency matrix A. (2) This allows for a more accurate quality score to further suppress the aggregation effect of low-quality neighbors, completely eliminating the pollution effect that remains before the two-stage iteration.

[0128] In measuring the degree of deviation in the embedding space, the Mahalanobis distance method based on the statistical distribution of the embedding space is preferred: First, take the first-generation cleaned view G... (1) The embedding vectors H of all nodes obtained from the training above (1) Given a sample set, compute the sample mean vector in the embedding space. (Dimension d) h ) and sample covariance matrix (Dimension d) h ×d h ), and on Applying diagonal loading regularization yields To ensure invertibility; for each node j, the embedding vector h (Dimension d) h ), calculate its Mahalanobis distance to the mean of the embedding space:

[0129]

[0130] The obtained score This refers to the embedding deviation score of node j, which physically represents the standardized deviation of node j's embedding vector from the normal distribution in the embedding space. An embedding deviation score exceeding the embedding deviation judgment threshold is considered a threshold. Nodes marked as anomalous nodes in the embedding space are further reduced in terms of their node quality score. The reduction is proportional to the embedding deviation score (i.e., the greater the deviation, the greater the reduction), thus obtaining a refined quality score q. (1) Furthermore, the refinement quality score is also limited to a closed numerical range of zero to one.

[0131] Step S34: Based on the second-generation cleanup view and the hierarchical quality problem list, for each subgraph in the second-generation cleanup view, calculate the subgraph performance deviation between the baseline accuracy and the accuracy on the cleanup view, and filter out defective subgraphs whose deviation exceeds the graph neural network performance deviation judgment threshold; combine the hierarchical quality problem list for cross-validation, mark the defective subgraphs that hit the rule records in the hierarchical quality problem list as rule violation defects, and mark the defective subgraphs that do not hit the rule records in the hierarchical quality problem list as neighborhood contamination blinding defects, and obtain the implicit topological defect location report.

[0132] Specifically, regarding the second-generation purification view G (2) Divide the graph into several subgraphs based on connectivity or business partitions, and calculate the baseline accuracy for each subgraph i on the original graph. (i.e., the classification or link prediction accuracy of the unpurified graph neural network on this subgraph), and in the second-generation purified view G (2) The accuracy of the clean view calculation is as follows: The subgraph performance deviation is calculated as follows:

[0133]

[0134] in, The performance deviation of subgraph i; The baseline accuracy of subplot i on the original plot; acc The accuracy of subgraph i on the second-generation clean view. Filtering. Exceeding the performance deviation threshold of graph neural networks Subgraphs are identified and marked as defective subgraphs. For each defective subgraph, a cross-validation query is performed on the hierarchical quality issue list using its set of node identifiers: if a node in the defective subgraph matches a rule record in the hierarchical quality issue list, the subgraph is marked as a rule violation defect; if no node matches any rule record in the hierarchical quality issue list, the subgraph is marked as a neighborhood contamination blinding defect, indicating that the performance degradation of the subgraph is due to the aggregation of low-quality neighbors rather than explicit rule violations, belonging to a hidden blind spot that traditional rule engines cannot detect. Finally, the type markings, node ranges, performance deviations, and defect causes of all defective subgraphs are summarized, and an implicit topology defect location report D is output. impl .

[0135] Baseline accuracy acc base,i This refers to the accuracy metric obtained by performing node classification (for supervised scenarios with node labels) or link prediction (for unsupervised scenarios without node labels) on subgraph i using the same graph neural network model architecture on the original, unwashed graph. In the node classification scenario, the test set is composed of labeled nodes in subgraph i, and the accuracy is calculated based on the proportion of nodes whose predicted labels match the true labels. In the link prediction scenario, the test set is composed of randomly sampled positive edges and an equal number of negative edges (random node pairs) in subgraph i, and the accuracy is calculated based on the proportion of samples whose link prediction probability exceeds 0.5 and are correctly predicted. It should be noted that the baseline accuracy and the washed-out view accuracy use the same test set partitioning and the same model evaluation protocol in their calculations to ensure that subgraph performance deviations are not compromised. It only reflects the impact of clean view construction on model performance, not the random fluctuations caused by differences in the test set.

[0136] Example 6

[0137] In this embodiment, the sliding window whitening correction mechanism is technically defined as follows: Whitening correction refers to a monitoring benchmark correction method that performs a linear transformation on a multi-dimensional monitoring indicator vector to ensure that the transformed indicator vector satisfies the standardization condition that the variance of each dimension is one and the covariance between dimensions is zero under normal distribution. It also adaptively estimates the transformation parameters based on historical normal graph snapshots within the current sliding window, allowing the alarm criteria to dynamically adjust with the normal business evolution of the graph. The fundamental difference between whitening correction and traditional fixed threshold monitoring lies in the fact that traditional fixed threshold monitoring uses graph statistics at the deployment time as the benchmark. When the graph undergoes normal evolutionary drift due to business growth (such as increased graph size and naturally increased node average degree), the fixed threshold cannot distinguish between normal drift and true anomalies, leading to persistent system false alarms. In contrast, the sliding window whitening correction mechanism re-estimates the mean vector and covariance structure using the latest historical normal graph snapshot in each time window, dynamically absorbing the trend components of normal evolution from the monitoring benchmark. This ensures that the Mahalanobis distance anomaly score after whitening always uses the current normal state as the reference, fundamentally eliminating systemic false alarms caused by normal evolutionary drift.

[0138] The technical principle of the conformal risk control dual-threshold mechanism is as follows: Conformal risk control is a nonparametric statistical inference framework based on historical sample quantiles. Its core idea is to use historical normal-state observation samples as a calibration set, and estimate the tail probability of the current test statistic (i.e., the whitened Mahalanobis distance outlier score) under the normal distribution by calculating the quantile statistics of the calibration set samples. Then, using user-specified acceptable false positive and false negative rates as constraints, the corresponding alarm decision threshold is determined, and a statistical guarantee (i.e., finite sample coverage guarantee) is attached to the determined threshold. The dual-threshold mechanism (λ) FNR , λ FPR The statistical advantages of the dual-threshold alarm mechanism compared to the traditional single-threshold binary alarm mechanism are reflected in the following two aspects: First, the single-threshold mechanism requires a trade-off between the false positive rate and the false negative rate, while the dual-threshold mechanism introduces an "uncertain state" interval to explicitly mark the areas that cannot be clearly determined statistically, avoiding forced alarm or suppression decisions based on insufficient confidence levels; Second, the dual-threshold mechanism can independently control the upper bound of the false positive rate α2 and the upper bound of the false negative rate α1, enabling operation and maintenance personnel to asymmetrically configure alarm risk parameters according to the different tolerance levels of false alarms and missed alarms for the business. For example, for financial risk control scenarios, α1 can be set to be smaller (not allowing missed alarms) and α2 to be larger (tolerating a certain number of false alarms), while for general data quality monitoring scenarios, α1=α2 can be symmetrically configured.

[0139] It is worth noting that the introduction of the diagonal loading regularization coefficient ε is a key design element to ensure the invertibility of the covariance matrix. In practical graph monitoring scenarios, some monitoring metric dimensions may exhibit high linear correlation (for example, the change in graph density and the rate of change in the connection edges of core nodes have a strong positive correlation in certain business graph scenarios), causing the sample covariance matrix to approach singularity, making the inverse calculation of the matrix's square root unstable. By loading positive real numbers on the diagonal of the covariance matrix... (like =10 -3 ), can ensure t The whitening transformation matrix W is a positive definite matrix under any sample size and any correlation of monitoring indicators. t The calculated values ​​remain consistently stable. The value of can be determined through cross-validation on historical normal datasets, typically... The magnitude of the regularization bias is between one-tenth and one-hundredth of the smallest eigenvalue of the covariance matrix, in order to minimize the impact of the regularization bias on the whitening effect while ensuring reversibility.

[0140] The operation of the mechanism is illustrated using a quality monitoring scenario of a supply chain network graph. A supply chain management system maintains a supply chain graph containing approximately 200,000 nodes (suppliers, warehouses, stores) and 1.5 million edges (supply relationships, logistics relationships). During daily operations, the graph size naturally grows at a rate of approximately 3% per month. Initially, the fixed-threshold monitoring scheme set the alarm threshold for graph density changes to 0.001. As the graph size increased, the graph density naturally decreased. After six months of deployment, the fixed-threshold monitoring began generating numerous false alarms, forcing maintenance personnel to manually adjust the alarm threshold every few months. After replacing this with the sliding window whitening correction mechanism, the system adaptively estimates the current normal-state parameters using the most recent ninety normal graph snapshots (covering a sliding window of approximately three months). The whitened Mahalanobis distance anomaly score remains stably low during normal business evolution. However, when a batch of supplier node data is imported incorrectly, causing some logistics edges to be reversed, the whitened Mahalanobis distance anomaly score d... t It suddenly increases to λ within this time window FPR The system triggers a critical P1 alarm with a false positive rate upper bound of α2=0.05 (meaning the false alarm probability of this alarm does not exceed five percent), accurately locating the affected node range (involving approximately eight hundred warehouse nodes and related logistics edges), providing effective input for the subsequent multi-agent repair module.

[0141] In this embodiment, the alarm decision-making process that performs adaptive benchmark correction and statistical reliability guarantee on multi-dimensional monitoring indicators includes the following steps:

[0142] Using the normal graph snapshot index sequence within the sliding window in the historical normal distribution data as input, the mean vector of the normal graph snapshot index sequence is calculated, and diagonal loading regularization is applied to the covariance matrix of the normal graph snapshot index sequence to obtain the mean vector and the covariance matrix with diagonal loading regularization.

[0143] Specifically, historical normal distribution data is maintained using a sliding window approach, with each window storing a multi-dimensional monitoring indicator vector corresponding to the n most recent normal graph snapshots, denoted as {m}. , m , …, m } where each vector has dimension p, and multiple monitoring indicators such as coverage distribution KL divergence, core node connection edge change rate, and graph density change are included. The mean vector of the normal graph snapshot indicator sequence within the sliding window is calculated as follows:

[0144]

[0145] in, t The mean vector is the vector of the current time window t; n is the number of normal image snapshots within the sliding window; m Let be the multidimensional monitoring indicator vector for the k-th normal image snapshot. Calculate the sample covariance matrix and apply diagonal loading regularization to obtain the covariance matrix with diagonal loading regularization:

[0146] in, The covariance matrix is ​​diagonally loaded with regularization; The sample covariance matrix is ​​calculated from the normal image snapshot index sequence within the sliding window; The regularization coefficients are diagonal loading coefficients, taken as positive real numbers (e.g., ...). =10 -3 , which is a p×p dimensional identity matrix. The introduction of diagonal loading regularization ensures that the covariance matrix is ​​invertible with any sample size, avoiding numerical singularity problems in the subsequent whitening transformation.

[0147] Based on the comparison between the number of normal image snapshots within the sliding window and the minimum snapshot number threshold, when the number of normal image snapshots is less than the minimum snapshot number threshold, diagonal whitening is performed by replacing the covariance matrix with diagonal loading regularization with dimension-wise independent normalization to obtain a whitening transformation matrix; when the number of normal image snapshots is not less than the minimum snapshot number threshold, full covariance whitening is performed based on the covariance matrix with diagonal loading regularization to obtain a whitening transformation matrix.

[0148] Specifically, let the minimum number of snapshots threshold be n. min Compare the number of normal image snapshots n within the sliding window with n minThe size relationship is considered, and the whitening method is selected based on the comparison result. When n < n min When the sample size is insufficient to reliably estimate the total covariance structure, diagonal whitening is performed using independent standardization along each dimension, with the whitening transformation matrix W... This is a diagonal matrix, where the k-th diagonal element is the reciprocal of the sample standard deviation for the corresponding monitoring indicator dimension:

[0149]

[0150] Among them, W This represents the k-th diagonal element of the diagonal whitening transformation matrix; t,k Let be the sample standard deviation of the k-th monitoring indicator dimension in the normal graph snapshot indicator sequence within the sliding window; all off-diagonal elements are zero. At that time, with a sufficient sample size, the covariance matrix based on diagonal loading regularization was used. t Perform full covariance whitening; the whitening transformation matrix is:

[0151] Among them, W t This is the total covariance whitening transformation matrix; for The inverse of the square root of the matrix is ​​calculated through eigenvalue decomposition or Cholesky decomposition, ensuring that the linear correlation between the dimensions of each monitoring indicator is eliminated after whitening, thereby improving the ability to discriminate anomaly scores in subsequent Mahalanobis distance tests.

[0152] Based on the multi-dimensional monitoring indicator vector, the mean vector, and the whitening transformation matrix corresponding to the current time window of the graph database change log stream, the whitened Mahalanobis distance anomaly score is calculated to obtain the whitened Mahalanobis distance anomaly score for the current time window.

[0153] Specifically, using the graph database change log stream as the data source, a multi-dimensional monitoring indicator vector m is extracted from the change records within the current time window t. t Based on mean vector t and whitening transformation matrix W t (or W) (Depending on the number of snapshots compared), calculate the whitened Mahalanobis distance anomaly score for the current time window:

[0154]

[0155] Where, d t m is the whitened Mahalanobis distance anomaly score for the current time window t; t This is a vector of multi-dimensional monitoring indicators for the current time window; tW is the mean vector; t Here is the whitening transformation matrix; d is the Euclidean norm. t The larger the value, the greater the deviation of the multi-dimensional monitoring indicators of the current graph snapshot from the normal state, providing input for subsequent alarm decisions to ensure statistical reliability. Since the mean vector and whitening transformation matrix are adaptively re-estimated in each time window t as the normal graph snapshot is updated within the sliding window, the alarm benchmark can be automatically adjusted with the normal business evolution of the graph (such as the natural migration of the degree distribution caused by the growth of graph size), fundamentally eliminating systematic false alarms caused by the drift of fixed threshold monitoring due to normal evolution.

[0156] Example 7

[0157] In this embodiment, the alarm decision that performs adaptive benchmark correction and statistical reliability guarantee on multi-dimensional monitoring indicators includes:

[0158] Using the historical normal albinism abnormality score sequence, the user-specified acceptable false negative rate, and the user-specified acceptable false positive rate as inputs, and based on the quantile distribution of the historical normal albinism abnormality score sequence, the false negative rate threshold satisfying the first confidence level and the false positive rate threshold satisfying the second confidence level are determined respectively.

[0159] Specifically, the historical normal albinism abnormality score sequence This is a set of abnormal score records obtained by whitening correction of normal image snapshots within a historical time window, where N is the sequence length, and it is updated by sliding over time, every T. recalib Each step involves recalibrating once. The user-specified acceptable false negative rate α1 represents the maximum probability of missed detection that the user can tolerate, and the user-specified acceptable false positive rate α2 represents the maximum probability of false positive that the user can tolerate. Based on the quantile distribution of historical normal whitening abnormality score sequences, the first confidence level is determined using a conformity-preserving risk control framework. False negative rate threshold λ FNR And satisfying the second confidence level (1- 2 False positive rate threshold λ FPR This ensures that the following statistics hold true:

[0160] and

[0161] Where, λ FNR λ is the threshold for the false negative rate. FPR α1 represents the false positive rate threshold; α2 represents the acceptable false negative rate specified by the user. 1. Error parameters at the first confidence level; 2. Error parameters at the second confidence level; FNR is the actual false negative rate; FPR is the actual false positive rate. λ FNR λ is determined by the low quantiles of historical normal-state albinism aberration fraction sequences. FPR Determined by the high quantiles of historical normal-state albinism aberration fraction sequences, both satisfying λ. FNR ≤ λ FPR .

[0162] Based on the aberration score of the whitened Mahalanobis distance in the current time window, the false negative rate threshold, and the false positive rate threshold, an alarm trigger state is output when the aberration score of the whitened Mahalanobis distance exceeds the false positive rate threshold; an alarm suppression state is output when the aberration score of the whitened Mahalanobis distance is lower than the false negative rate threshold; and an uncertain state is output when the aberration score of the whitened Mahalanobis distance is between the false negative rate threshold and the false positive rate threshold, thus obtaining an alarm status marker.

[0163] Specifically, the whitened Mahalanobis distance anomaly score d within the current time window t. t False negative rate threshold λ FNR and false positive rate threshold λ FPR As input, perform a three-stage alarm status determination: when d t > λ FPR When the multi-dimensional monitoring indicators of the current snapshot are determined to deviate significantly from the normal state, and the abnormal score exceeds the upper bound of the false positive rate statistical control, an alarm trigger status is output; when d t < λ FNR When the multi-dimensional monitoring indicators of the current snapshot are determined to be within the normal range and the anomaly score is below the lower bound of the false negative rate statistical control, an alarm suppression status is output; when λ FNR ≤ d t ≤ λ FPR When the current snapshot's monitoring status is determined to be in a statistically uncertain range, making it impossible to make a clear alarm or suppression decision based on the user-specified confidence level, an uncertain state is output, which is then further confirmed by manual analysis or by waiting for subsequent time window data. This three-stage decision logic upgrades a single-threshold binary alarm to a dual-threshold tri-state alarm, ensuring that each alarm decision is accompanied by clear upper bounds for false positive and false negative rates, achieving statistically guaranteed reliable alarms.

[0164] Based on the alarm status marker, the alarm timestamp, the whitened Mahalanobis distance anomaly score, the alarm type, and the range of affected nodes from the implicit topological defect location report are recorded to generate the structured alarm event.

[0165] Specifically, for time windows marked as alarm triggered or indeterminate states, the system automatically generates structured alarm event records, including: alarm timestamp (start and end times of the current time window t), and whitened Mahalanobis distance anomaly score d. t (Quantitative indicators reflecting the degree of abnormality), alarm type (determined as P1 severe alarm or P2 warning level based on the interval of the abnormal score), statistical confidence level (1- The value 2 represents the confidence level corresponding to the upper bound of the false positive rate statistics for this alarm, as well as the range of affected nodes derived from the implicit topology defect location report (i.e., the set of defective nodes and defective subgraph identifiers associated with the current time window change). All fields are stored in a structured format, supporting direct parsing and consumption by downstream multi-agent repair modules without manual conversion. For time windows marked as alarm suppression states, no alarm event records are generated; only the anomaly score for that time window is retained in the system log for auditing and traceability.

[0166] Example 8

[0167] In this embodiment, the adaptive repair planning-verification closed loop with LLM multi-agent JSON template constraints refers to a graph data quality repair process executed sequentially and collaboratively by four dedicated functional agents: a detection agent, a planning agent, an action agent, and an evaluation agent. The planning agent replaces the hard-coded strategy of the traditional static repair rule base with the natural language reasoning capabilities of a large language model, generating an adaptive repair scheme that incorporates graph business semantics. The action agent uses the structured JSON operation instructions output by the planning agent as the execution basis, accurately translating the repair scheme into graph database write operations. The evaluation agent uses the Pareto quality improvement vector as a multi-objective convergence criterion, driving the repair iteration loop until the quality improvement stabilizes. The "autonomous closure" characteristic of the closed loop means that when the repair convergence condition is not met, the output of the evaluation agent directly triggers the automatic initiation of the next round of detection-planning-action-evaluation collaborative steps, completing multiple iterations without manual intervention until convergence or reaching the maximum iteration protection threshold.

[0168] The design motivation for the JSON structured operation template constraint mechanism is to solve the uncertainty problem of natural language output of large language models: When a planning agent generates a repair plan using reasoning from a large language model, if the output is in free-format natural language (such as "supplement the age field of node X with the average age of its neighboring nodes"), the action agent needs to perform natural language understanding and parameter parsing to convert it into an executable graph database write operation. This conversion process is ambiguous and prone to errors. However, by requiring the large language model to output each repair operation in a structured format with predefined JSON Schema constraints (including the op_type enumeration field, target_ids list field, model_params parameter field, priority integer field, and expected_delta_Q expected improvement vector field), the action agent can directly parse the JSON fields to execute the corresponding graph database API call, reducing the uncertainty of semantic parsing to zero and fundamentally eliminating the semantic gap between the natural language output of the large language model and the executable graph operation instructions. The JSON Schema uses enumeration constraints (the valid value set of the op_type field is {attribute filling, attribute overriding, edge completion, soft deletion}), non-empty constraints (the target_ids field cannot be an empty list), and type constraints (the edge completion probability threshold in the model_params field). link Three types of constraints are imposed: the data type must be floating-point and the value range must be [0,1], to enforce normalization of the output of large language models.

[0169] The theoretical basis for introducing the Pareto quality improvement vector into multi-objective judgment is as follows: In the process of graph data quality repair, multiple quality dimensions (attribute completeness rate, number of rule violations, and number of defective subgraphs) need to be improved simultaneously. If a single quality indicator (such as attribute completeness rate) is used as the convergence criterion, the repair optimization process may fall into multi-indicator oscillations. For example, in one iteration, the planning agent performs a large number of attribute filling operations to improve the attribute completeness rate, resulting in some filled values ​​not conforming to business constraints (causing new rule violations). In the next iteration, attribute overwriting operations are performed to eliminate rule violations, which may cause the attribute completeness rate of some nodes to decrease, thus generating continuous oscillations between the two indicators of attribute completeness rate and number of rule violations. By introducing a Pareto quality improvement vector containing three components... With "two consecutive rounds" And the absolute value of each decreasing component does not exceed tol "As a dual convergence condition, it can simultaneously ensure that all three quality dimensions have become stable (L1 norm constraint) and that no quality dimension has deteriorated significantly (decline component constraint), thereby preventing multi-index oscillations from the mechanism design level."

[0170] Regarding the selection of the convergence threshold, the convergence threshold... conv The value is determined based on the overall quality baseline of the image before repair: if the percentage of defective nodes in the overall image attribute integrity rate before repair is... (For example =0.05, meaning that 5% of the nodes have missing attributes), then conv Can be set to ×0.1 (meaning the expected quality improvement in each iteration is no less than one-tenth of the initial defect amount) ensures that convergence is triggered only after sufficient repair work is completed, rather than terminating prematurely when there is still significant room for improvement; tolerance threshold tol Sets the maximum degradation that a single repair operation may cause to a single quality metric. Typically, it can be set to 0.01 (i.e., the degradation of any quality metric in any single iteration does not exceed one percent) to allow for slight quality fluctuations caused by business constraints in the repair operation.

[0171] Taking a quality repair scenario of a financial knowledge graph as an example, this paper illustrates the complete iterative process of a four-agent closed-loop repair. A financial institution maintains a financial knowledge graph containing 120,000 entity nodes and 350,000 relation edges. It detects approximately 3,000 nodes with missing attributes (an average attribute completeness rate of 0.72), approximately 200 edges with directional violations (0.06% of the total edges), and approximately five subgraphs marked as having neighborhood contamination blinding defects. In the first round of repair iterations, the planning agent uses a large-scale language model to analyze defect location summaries and combines them with financial entity semantics (e.g., the establishment date field of a legal entity node can be filled using business registration data sources, and the shareholding ratio field can be filled using relevant regulatory disclosure documents) to generate 87 candidate repair operations; after the action agent executes these operations, the attribute completeness rate improves to 0.89 (…). Q1=0.17), the number of rule violations decreased to forty-one. Q2=159), the number of defect subgraphs eliminated is three ( Q3=3); the L1 norm of the Pareto quality improvement vector is 0.17+159+3=162.17, which far exceeds the convergence threshold, indicating non-convergence and triggering the second iteration. In the second iteration, the planning agent generates targeted repair operations for the remaining forty-one rule violations (edge ​​direction problem) and two remaining defective subgraphs; after the action agent executes these operations, all rule violations are eliminated. Q2=41), eliminate one more defective subgraph ( Q3=1), the attribute completeness rate slightly improved to 0.91 ( Q1=0.02); the L1 norm of the Pareto quality improvement vector in the second round is 41+1+0.02=42.02, which still exceeds the convergence threshold; after the third iteration, the improvement of all three indicators approaches zero, and the L1 norm is less than... conv The absolute value of each decreasing component does not exceed tol The system determines convergence, outputs the final repair report, and records the repair results of 131 repair operations performed in three iterations, increasing the full graph attribute integrity rate from 0.72 to 0.91, eliminating all rule violations and four neighborhood pollution-induced blinding defect subgraphs.

[0172] In this embodiment, the step of performing adaptive repair and iterative convergence verification through multi-agent collaborative execution, and outputting a final repair report, includes:

[0173] Using the hierarchical quality problem list, the implicit topological defect location report, and the structured alarm events as input, the detection agent extracts the defect node set, defect edge set, affected subgraph range, and problem priority sorting to obtain a defect location summary.

[0174] Specifically, the detection agent takes a hierarchical quality problem list, implicit topological defect location reports, and structured alarm events as input and performs multi-source defect information fusion. It extracts all problem records of severity and warning levels from the hierarchical quality problem list, obtaining the corresponding defect node identifier set and defect edge identifier set; it extracts the node ranges of rule-violation defect subgraphs and neighborhood contamination-induced blinding defect subgraphs from the implicit topological defect location reports, supplementing them to the defect node set; and it extracts the affected node ranges from the structured alarm events, further expanding the defect node set. It then performs deduplication and merging on the above defect node set, defect edge set, and affected subgraph ranges, and constructs a problem priority ranking based on defect severity level (severity level takes precedence over warning level), defect type (attribute level, structural level, and implicit defects are sorted by impact range), and alarm priority (P1 severe alarms take precedence over P2 warning levels), outputting a structured defect location summary for the planning agent to consume.

[0175] Based on the defect location summary, the planning agent sorts the defect node set, defect edge set, and problem priority in the defect location summary according to the repair strategy knowledge base, and constrains the type field, target node field, and repair parameter field of the repair operation in JSON format to generate a candidate repair operation set. Then, a repair operation sequence is selected from the candidate repair operation set through a search strategy.

[0176] Specifically, the intelligent agent loads a repair strategy knowledge base (containing a set of repair operation templates corresponding to various defect types). Using the reasoning capabilities of a large language model, it generates candidate repair operations for each defect based on the defect node set, defect edge set, problem priority ranking, and graph business semantics in the defect localization summary. Each candidate repair operation is output in JSON format, containing the following fields: type field (op_type, with values ​​including attribute correction, edge completion, node annotation, etc.), target node field (target_ids, a list of identifiers for target nodes or edges), and repair parameter field (model_params, containing specific repair parameters such as edge completion probability thresholds). The candidate repair operations are categorized into three fields: priority (aligned with problem priority) and expected quality improvement (expected_delta_Q, estimating the improvement of each quality metric). The JSON structure of all candidate repair operations is validated using a predefined JSON Schema constraint to ensure that the type field enumeration values ​​are valid, the target node field is not empty, and the repair parameter field is complete, thus eliminating the semantic gap between the natural language output of the large language model and the executable instructions of the action agent. After generating the candidate repair operation set, the planning agent selects a repair operation sequence O={o1, …, o} from the candidate repair operation set using a search strategy (such as greedy search or bundle search). k Prioritize operation combinations that are expected to improve quality significantly and have high priority, while avoiding conflicts between operations.

[0177] The access method for the large language model in the planning agent is as follows: Structured prompts are sent to the large language model via a standardized large language model call interface (such as an OpenAI compatible API or a local inference service interface of an open-source model), and the candidate repair operation description text returned by the large language model is received. The prompts are designed according to the following structure: the system prompts describe the current task background (graph data quality repair scenario), graph business semantics (node ​​and edge type descriptions automatically generated from graph database schema metadata), and output format constraints (requiring the large language model to strictly output candidate repair operations according to a predefined JSON schema); the user prompts input a defect location summary in structured text form, including defect node identifiers, defect types, problem priority ranking, and graph business semantic annotations. To prevent large language models from outputting text that does not conform to JSON Schema constraints, the planning agent adopts the following two-layer constraint mechanism: The first layer is the prompt word constraint, which explicitly requires the large language model to output only operation descriptions conforming to JSON format and not to output any interpretive natural language; the second layer is output validation, which performs JSON Schema compliance validation on the text returned by the large language model. If the validation fails, the prompt word is automatically resent with an additional explanation of the validation failure. The process is retried up to three times. If all three retries fail, the process is downgraded to directly generating candidate operations based on predefined rule templates in the repair strategy knowledge base, ensuring that the process is not interrupted due to abnormal output from the large language model.

[0178] Based on the repair operation sequence, the action agent executes each repair operation in the repair operation sequence in order of the problem priority to obtain the repaired image.

[0179] Specifically, the action agent receives the repair operation sequence O={o1, …, o} output by the planning agent. k The system executes repair operations sequentially according to problem priority. For attribute correction operations, it calls the graph database write interface to fill in missing attribute fields of the target node or overwrite erroneous attribute fields; for edge completion operations, it calls the link prediction model (such as the SEAL model) to perform edge completion on the target node pairs, based on the link prediction probability exceeding the edge completion probability threshold. Candidate edges are added to the criteria; for reconnecting isolated nodes, the optimal edge target is selected based on a weighted average of node importance scores (such as PageRank scores); for soft deletion marking operations on low-quality nodes, a soft deletion mark field is added to the node attributes instead of physical deletion to preserve audit traceability. After all repair operations are executed, the repaired graph G' is output.

[0180] Based on the repaired graph, the evaluation agent calculates the Pareto quality improvement vector, which includes three decreasing components: attribute integrity improvement, rule violation reduction, and defect subgraph elimination.

[0181] Specifically, the performance of the agent in re-triggering the rule engine check on the repaired graph G' (reusing the rule check logic of Example 4) and the cleaned view graph neural network performance test (reusing the cleaned view construction and subgraph performance deviation analysis logic of Example 5) is evaluated. The current values ​​of the repaired graph on the three quality dimensions of attribute completeness rate, number of rule violations, and number of defective subgraphs are calculated respectively, and the difference between these values ​​and the corresponding values ​​before repair is obtained to obtain the Pareto quality improvement vector. ,in Q1 represents the improvement in attribute completeness. Q2 represents the reduction in rule violations (a positive value indicates a reduction in the number of violations). Q3 represents the amount of defective subgraphs eliminated (a positive value indicates a reduction in the number of defective subgraphs). If any of the three components is negative (i.e., a certain quality indicator has decreased), the planning agent will perform corresponding repair operations on the decreased indicator in the next iteration to prevent multi-indicator oscillations caused by single-indicator optimization.

[0182] Based on the Pareto quality improvement vector, determine whether the convergence condition is met: if the L1 norm of the Pareto quality improvement vector is less than the convergence threshold for two consecutive rounds and each decreasing component does not exceed the tolerance threshold, then convergence is determined and the final repair report is output; otherwise, replace the dataset of the graph to be evaluated with the repaired graph and re-execute the collaborative steps of the detection agent, planning agent, action agent and evaluation agent.

[0183] Specifically, the convergence criterion is based on the dual conditions of the L1 norm and the descent component of the Pareto quality improvement vector: if the Pareto quality improvement vector in two consecutive iterations is... Q satisfies (The L1 norm being less than the convergence threshold indicates that the quality indicators have basically stabilized) and each decreasing component Q i The absolute values ​​of all values ​​less than 0 do not exceed the tolerance threshold. tol If no quality indicators show significant deterioration, then the current repair iteration is considered converged, the iteration loop is terminated, and the final repair report is output (including all repair operation records, statistics on the improvement of each quality indicator, and a summary of residual problems). If the convergence condition is not met, the repaired graph G' replaces the dataset to be evaluated, the four-agent collaborative execution step is restarted, and the next round of repair iteration begins. When the number of iterations exceeds the maximum number of iterations N... maxWhen the iteration is terminated, regardless of whether it has converged, any remaining unresolved quality issues will be pushed into the manual review queue, and issues requiring manual intervention will be marked in the final repair report to ensure a safety net.

[0184] Example 9

[0185] In this embodiment, the step of having the action agent execute each repair operation in the repair operation sequence according to the problem priority to obtain the repaired image includes:

[0186] Based on the repair operation sequence and the refinement quality score, for each repair operation in the repair operation sequence, the target fill value corresponding to each repair operation is determined according to the mapping relationship between node type and attribute field, combined with the refinement quality score, so as to obtain a repair operation sequence carrying the target fill value.

[0187] Specifically, the action agent uses the repair operation sequence output by the planning agent and the refinement quality score q obtained in Example 5. (1) As input, for each repair operation in the repair operation sequence, the calculation method for the target fill value is determined based on the mapping relationship between node type and attribute field (predefined in the graph database schema, specifying the set of attribute fields corresponding to each node type and the field value constraints). For fill operations with missing attribute fields, the target fill value is determined by comprehensively considering the following two types of information: First, the distribution of attribute values ​​of neighboring nodes of the same type (i.e., nodes directly connected to the target node and of the same node type) on that attribute field, in order to refine the quality score q. The attribute values ​​of each neighbor node k are weighted, with higher quality scores for neighbor nodes having greater weights on their attribute values. This makes the target fill value more likely to be a reference value from a high-quality neighbor. Secondly, business semantic constraints on attribute fields (such as the legal range of numeric fields and the legal value set of enumerated fields) are applied. The target fill value determined by combining the above two types of information is written into the target_value field of the corresponding repair operation, resulting in a repair operation sequence carrying the target fill value. This ensures that each repair operation carries complete and directly executable fill parameters.

[0188] The calculation method for weighted filling of neighbor attribute values ​​based on refined quality scores is as follows: Let the set of direct neighbor nodes of the same type for target node j be . (i.e., the set of nodes directly connected to node j and of the same type), for For each neighbor node k, take its refined quality score q. As a weight, the target fill value v for attribute field f j,f Calculate using the following formula: The summation range is limited to neighboring nodes whose attribute field f is not empty, ensuring that only valid attribute values ​​participate in the weighting. For empty set or When all neighboring nodes have null values ​​in attribute field f (i.e., no reference fill value can be obtained from neighboring nodes), the system degenerates into using the weighted mode (discrete field) or weighted mean (numerical field) of the non-null values ​​in the global attribute field f for that node type as the target fill value. The weights are also taken from the refined quality score of the corresponding node. The above weighted fill strategy ensures that the target fill value tends to a reasonable range under the guidance of high-quality neighboring node attribute values, while handling the edge case where all neighbors are null.

[0189] According to the repair operation sequence carrying the target fill value, the repair operations are executed in order of priority according to the problem. According to the type field of each repair operation, the missing attribute field is filled, the erroneous attribute field is overwritten, the isolated node is reconnected, and the low-quality node is softly deleted, so as to obtain the repaired image.

[0190] Specifically, the action agent sequentially traverses the sequence of repair operations carrying target fill values ​​according to problem priority, and executes the corresponding graph database write operation based on the type field of each repair operation. For missing attribute field fill operations (op_type is attribute fill), the missing attribute field value of the target node is filled with the target fill value through the graph database write interface using the target fill value in the target_value field; for erroneous attribute field overwrite operations (op_type is attribute overwrite), the erroneous attribute field value of the target node is overwritten with the correct attribute value in the target_value field; for isolated node re-edge connection operations (op_type is edge completion), the edge completion probability threshold configured in the model_params field is used. Based on the criterion, the link prediction model is invoked to predict the probability of connections between the target isolated node and its candidate neighbor nodes. For nodes with predicted probabilities exceeding [a certain threshold], [the model is then used to] predict the probability of connections between them. Candidate edges are added to the node pairs; for low-quality nodes, a soft deletion flag (op_type is soft deletion) is added to the target node's attributes (e.g., is_soft_deleted=true), so that the node is automatically filtered in subsequent graph analysis and graph neural network training, while retaining the node's original attribute data for auditing and traceability. After all the above repair operations are completed, the repaired graph G' is output.

[0191] Based on the repaired graph and the repair operation sequence carrying the target fill value, an operation log is generated for each executed repair operation in the repaired graph, recording the operation type, target node identifier, attribute value before repair, and attribute value after repair, thus obtaining the repair operation audit record.

[0192] Specifically, after the action agent completes the output of the repaired graph G', it generates an operation log entry for each successfully executed repair operation in the repair operation sequence carrying the target fill value. Each operation log entry contains the following fields: operation type (consistent with the op_type field of the repair operation), target node identifier (consistent with the target_ids field of the repair operation), attribute value before repair (the original field value read from the graph database before the repair operation is executed; for edge addition operations, it is recorded as empty), and attribute value after repair (the new field value written to the graph database after the repair operation is executed; for soft deletion mark operations, it records the key-value pair of the soft deletion mark field). All operation log entries are sorted in ascending order by execution timestamp and summarized into repair operation audit records. These records are persistently stored in an audit log library in a structured format, supporting fast retrieval by target node identifier, operation type, and time range, providing a reliable data foundation for subsequent quality backtracking, compliance auditing, and repair effect review.

[0193] Example 10

[0194] like Figure 3 As shown, the present invention also provides a graph data quality assessment and repair system based on multi-agent collaboration, comprising:

[0195] The multidimensional feature construction module 301 is used to extract attribute dimension features and topological dimension features from the set of node attributes and the set of edge relationships of the graph dataset to be evaluated, construct a multidimensional quality feature matrix, and generate an initial prior score vector based on the attribute completeness rate and degree quantile reasonableness score of each node.

[0196] The rule checking module 302 is used to perform automated rule checking on the multidimensional quality feature matrix based on the multidimensional quality feature matrix and the domain semantic rule definition file, according to the graph structure-specific quality rule library, to obtain a hierarchical quality problem list and rule tag scoring penalty increment.

[0197] The implicit defect localization module 303 is used to distinguish between rule violation defects and neighborhood contamination blinding defects based on the graph dataset to be evaluated, the initial prior score vector, and the rule-labeled score penalty increment, through quality-weighted clean view iterative construction and graph neural network collaborative analysis, and to obtain an implicit topological defect localization report.

[0198] The alarm decision module 304 is used to take the graph database change log stream and historical normal distribution data of the graph dataset to be evaluated as input, and combine the affected node range in the implicit topology defect location report to perform adaptive benchmark correction and statistical reliability guarantee alarm decision on multi-dimensional monitoring indicators to obtain structured alarm events.

[0199] The multi-agent repair module 305 is used to perform adaptive repair and iterative convergence verification through multi-agent collaboration based on the hierarchical quality problem list, the implicit topological defect location report and the structured alarm events, and output the final repair report.

[0200] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of this application are not directed to any particular programming language. It should be understood that the content of this application described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of this application.

[0201] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0202] Similarly, it should be understood that, in order to simplify this application and aid in understanding one or more of the various aspects of the invention, in the above description of exemplary embodiments of this application, various features of the embodiments of this application are sometimes grouped together into a single embodiment, figure, or description thereof.

[0203] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying abstract and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying abstract and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0204] It should be noted that the above embodiments are illustrative of this application and not restrictive, and those skilled in the art can design alternative embodiments without departing from the scope. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

[0205] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for graph data quality assessment and repair based on multi-agent collaboration, characterized in that, include: Using the set of node attributes and the set of edge relationships of the graph dataset to be evaluated as input, attribute dimension features and topology dimension features are extracted to construct a multidimensional quality feature matrix. Based on the attribute completeness rate and degree quantile reasonableness score of each node, an initial prior score vector is generated. Based on the multidimensional quality feature matrix and the domain semantic rule definition file, an automated rule check is performed on the multidimensional quality feature matrix according to the graph structure-specific quality rule library to obtain a hierarchical quality problem list and rule-marked scoring penalty increments. Based on the graph dataset to be evaluated, the initial prior score vector, and the rule-labeled score penalty increment, a quality-weighted cleanup view is iteratively constructed and a graph neural network is used for collaborative analysis to distinguish between rule violation defects and neighborhood contamination-induced blinding defects, resulting in an implicit topological defect location report. Specifically, this includes: for each node, based on the initial prior score vector and the rule-labeled score penalty increment, adding the initial prior score to the corresponding rule-labeled score penalty increment and limiting it to a closed numerical range of zero to one; when the rule-labeled score penalty increment is zero, it degenerates into the initial prior score, obtaining the node quality score for each node; based on the node quality score and the adjacency matrix of the graph dataset to be evaluated, the adjacency matrix is ​​quality-weighted using the node quality score to obtain a quality-weighted adjacency matrix; a first-generation cleanup view is constructed based on the quality-weighted adjacency matrix and the node attribute set, and a graph neural network is trained on the first-generation cleanup view. The network model is used to obtain the embedding vectors of each node. Based on the embedding vectors, the degree to which the embedding vectors deviate from the normal distribution of the embedding space is evaluated. Nodes whose deviation exceeds the embedding deviation judgment threshold have their corresponding node quality scores reduced to obtain a refined quality score. The second-generation cleaned view is reconstructed based on the refined quality score and the adjacency matrix. Based on the second-generation cleaned view and the hierarchical quality problem list, for each subgraph in the second-generation cleaned view, the subgraph performance deviation between the baseline accuracy and the accuracy on the cleaned view is calculated, and defective subgraphs with deviations exceeding the graph neural network performance deviation judgment threshold are selected. Combined with cross-validation of the hierarchical quality problem list, defective subgraphs that hit the rule records in the hierarchical quality problem list are marked as rule violation defects, and defective subgraphs that do not hit the rule records in the hierarchical quality problem list are marked as neighborhood contamination blinding defects, thus obtaining the implicit topological defect location report. Using the graph database change log stream and historical normal distribution data of the graph dataset to be evaluated as input, and combining the affected node range in the implicit topology defect location report, adaptive benchmark correction and statistical reliability guarantee alarm decisions are performed on multi-dimensional monitoring indicators to obtain structured alarm events. The adaptive benchmark correction includes: using the normal graph snapshot indicator sequence within the sliding window in the historical normal distribution data as input, calculating the mean vector of the normal graph snapshot indicator sequence, and applying diagonal loading regularization to the covariance matrix of the normal graph snapshot indicator sequence to obtain the mean vector and the covariance matrix with diagonal loading regularization; based on the normal graph snapshot within the sliding window... Based on the comparison between the number of normal graph snapshots and the minimum snapshot number threshold, when the number of normal graph snapshots is less than the minimum snapshot number threshold, diagonal whitening is performed by replacing the covariance matrix with diagonal loading regularization with dimension-wise independent standardization to obtain a whitening transformation matrix; when the number of normal graph snapshots is not less than the minimum snapshot number threshold, full covariance whitening is performed based on the covariance matrix with diagonal loading regularization to obtain a whitening transformation matrix; based on the multidimensional monitoring indicator vector of the current time window corresponding to the graph database change log stream, the mean vector, and the whitening transformation matrix, the whitened Mahalanobis distance anomaly score is calculated to obtain the whitened Mahalanobis distance anomaly score of the current time window; Based on the hierarchical quality problem list, the implicit topological defect location report, and the structured alarm events, adaptive repair and iterative convergence verification are performed collaboratively by multiple agents to output the final repair report.

2. The method according to claim 1, characterized in that, The reasonableness score based on the attribute completeness rate and degree quantile of each node includes: For the attribute dimension data in the multidimensional quality feature matrix, the proportion of non-empty attribute fields in each node to all fields that should be filled is calculated to obtain the attribute completeness rate of each node. Based on the degree feature dimension data in the multidimensional quality feature matrix and the full-graph degree distribution histogram, using the full-graph degree distribution histogram as a benchmark, the cumulative distribution quantile position of each node's degree value in the full-graph degree distribution histogram is calculated to obtain the degree quantile position of each node; nodes whose degree quantile position exceeds the first quantile threshold are assigned a low reasonableness score, and nodes whose degree quantile position is lower than the second quantile threshold are assigned a low reasonableness score, wherein the first quantile threshold is greater than the second quantile threshold, to obtain the degree quantile reasonableness score of each node; The attribute completeness rate and the degree quantile reasonableness score are weighted and linearly combined to obtain the unsupervised quality prior score of each node, which constitutes the initial prior score vector.

3. The method according to claim 1, characterized in that, The graph structure-specific quality rule base includes: Attribute rules are used to perform non-null constraint checks, data type compliance checks, and value range compliance checks on the attribute fields of each node. Direction rules are used to perform consistency checks on the direction attributes of each edge in directed or undirected graphs. Topology rules are used to construct the overall graph topology based on the set of node attributes and the set of edge relationships. The rules perform constraints on the overall graph topology, including checks on the proportion of isolated nodes, the interval of graph density, and the degree distribution power law exponent being within the interval formed by the lower and upper bounds of the power law exponent.

4. The method according to claim 3, characterized in that, The graph-structured, dedicated quality rule base performs automated rule checks on the multidimensional quality feature matrix to obtain a hierarchical quality problem list and rule-marked scoring penalty increments, including: Using the attribute rules in the graph structure-specific quality rule library and the attribute dimension data in the multidimensional quality feature matrix as input, the attribute quality assertion execution engine is invoked to perform rule assertion checks on each node attribute field, and an attribute class rule execution report is obtained. Based on the directional rules and topology rules in the graph structure-specific quality rule library, and combined with the topology dimension data in the multidimensional quality feature matrix, a graph query language is called to perform a topology statistical query to obtain a structure-class rule execution report. Based on the attribute-type rule execution report and the structure-type rule execution report, a fixed attenuation amount is assigned to the node marked by any rule to obtain the rule marking score penalty increment. The problem records in the attribute-type rule execution report and the structure-type rule execution report are classified and summarized according to severity level to obtain the hierarchical quality problem list containing severity level, warning level and information level.

5. The method according to claim 1, characterized in that, The alarm decision-making process for performing adaptive benchmark correction and statistical reliability guarantee on multi-dimensional monitoring indicators includes: Using the historical normal albinism abnormality score sequence, the user-specified acceptable false negative rate, and the user-specified acceptable false positive rate as inputs, and based on the quantile distribution of the historical normal albinism abnormality score sequence, the false negative rate threshold satisfying the first confidence level and the false positive rate threshold satisfying the second confidence level are determined respectively. Based on the aberration score of the whitened Mahalanobis distance in the current time window, the false negative rate threshold, and the false positive rate threshold, an alarm trigger state is output when the aberration score of the whitened Mahalanobis distance exceeds the false positive rate threshold; an alarm suppression state is output when the aberration score of the whitened Mahalanobis distance is lower than the false negative rate threshold; and an uncertain state is output when the aberration score of the whitened Mahalanobis distance is between the false negative rate threshold and the false positive rate threshold, thus obtaining an alarm status marker. Based on the alarm status marker, the alarm timestamp, the whitened Mahalanobis distance anomaly score, the alarm type, and the range of affected nodes from the implicit topological defect location report are recorded to generate the structured alarm event.

6. The method according to claim 1, characterized in that, The process involves multi-agent collaborative execution of adaptive repair and iterative convergence verification, outputting a final repair report, including: Using the hierarchical quality problem list, the implicit topological defect location report, and the structured alarm events as input, the detection agent extracts the defect node set, defect edge set, affected subgraph range, and problem priority sorting to obtain a defect location summary; Based on the defect location summary, the planning agent sorts the defect node set, defect edge set and problem priority in the defect location summary according to the repair strategy knowledge base, constrains the type field, target node field and repair parameter field of the repair operation in JSON format, generates a candidate repair operation set, and selects a repair operation sequence from the candidate repair operation set through a search strategy. Based on the repair operation sequence, the action agent executes each repair operation in the repair operation sequence in order of the problem priority to obtain the repaired image; Based on the repaired graph, the evaluation agent calculates the Pareto quality improvement vector, which includes three decreasing components: attribute integrity rate improvement, rule violation reduction, and defective subgraph elimination. Based on the Pareto quality improvement vector, determine whether the convergence condition is met: if the L1 norm of the Pareto quality improvement vector is less than the convergence threshold for two consecutive rounds and each decreasing component does not exceed the tolerance threshold, then convergence is determined and the final repair report is output; otherwise, replace the dataset of the graph to be evaluated with the repaired graph and re-execute the collaborative steps of the detection agent, planning agent, action agent and evaluation agent.

7. The method according to claim 6, characterized in that, The process involves the action agent sequentially executing each repair operation in the repair operation sequence according to the problem priority, resulting in a repaired image, including: Based on the repair operation sequence and the refinement quality score, for each repair operation in the repair operation sequence, according to the mapping relationship between node type and attribute field, and combined with the refinement quality score, the target fill value corresponding to each repair operation is determined, thus obtaining a repair operation sequence carrying the target fill value; According to the repair operation sequence carrying the target fill value, the repair operations are performed in order according to the problem priority. According to the type field of each repair operation, the missing attribute field is filled, the erroneous attribute field is overwritten, the isolated node is reconnected, and the low-quality node is softly deleted, so as to obtain the repaired image. Based on the repaired graph and the repair operation sequence carrying the target fill value, an operation log is generated for each executed repair operation in the repaired graph, recording the operation type, target node identifier, attribute value before repair, and attribute value after repair, thus obtaining the repair operation audit record.

8. A graph data quality assessment and repair system based on multi-agent collaboration, using the graph data quality assessment and repair method based on multi-agent collaboration as described in any one of claims 1-7, characterized in that, Includes the following modules: The multidimensional feature construction module is used to extract attribute dimension features and topological dimension features from the set of node attributes and the set of edge relationships of the graph dataset to be evaluated, construct a multidimensional quality feature matrix, and generate an initial prior score vector based on the attribute completeness rate and degree quantile reasonableness score of each node. The rule checking module is used to perform automated rule checking on the multidimensional quality feature matrix based on the multidimensional quality feature matrix and the domain semantic rule definition file, according to the graph structure-specific quality rule library, to obtain a hierarchical quality problem list and rule tag scoring penalty increment; An implicit defect localization module is used to distinguish between rule-violation defects and neighborhood contamination-induced blinding defects based on the graph dataset to be evaluated, the initial prior score vector, and the rule-labeled score penalty increment. This is achieved through iterative construction of a quality-weighted cleanup view and collaborative analysis with a graph neural network, resulting in an implicit topological defect localization report. Specifically, this includes: for each node, adding the initial prior score to the corresponding rule-labeled score penalty increment based on the initial prior score vector and the rule-labeled score penalty increment, limiting the sum to a closed numerical range of zero to one. When the rule-labeled score penalty increment is zero, it degenerates into the initial prior score, obtaining a node quality score for each node; based on the node quality scores and the adjacency matrix of the graph dataset to be evaluated, quality-weighting the adjacency matrix using the node quality scores to obtain a quality-weighted adjacency matrix; constructing a first-generation cleanup view based on the quality-weighted adjacency matrix and the node attribute set; and then performing a cleanup on the first-generation cleanup view. A graph neural network model is trained to obtain the embedding vectors of each node. Based on the embedding vectors, the degree to which the embedding vectors deviate from the normal distribution of the embedding space is evaluated. Nodes whose deviation exceeds the embedding deviation judgment threshold have their corresponding node quality scores reduced to obtain a refined quality score. A second-generation cleaned view is reconstructed based on the refined quality score and the adjacency matrix. Based on the second-generation cleaned view and the hierarchical quality problem list, for each subgraph in the second-generation cleaned view, the subgraph performance deviation between the baseline accuracy and the accuracy on the cleaned view is calculated, and defective subgraphs whose deviation exceeds the graph neural network performance deviation judgment threshold are selected. Combined with cross-validation using the hierarchical quality problem list, defective subgraphs that match the rule records in the hierarchical quality problem list are marked as rule violation defects, and defective subgraphs that do not match the rule records in the hierarchical quality problem list are marked as neighborhood contamination blinding defects, thus obtaining the implicit topological defect location report. The alarm decision module is used to take the graph database change log stream and historical normal distribution data of the graph dataset to be evaluated as input, and combine the affected node range in the implicit topology defect location report to perform adaptive benchmark correction and statistical reliability guarantee alarm decision on multi-dimensional monitoring indicators to obtain structured alarm events. The adaptive benchmark correction includes: taking the normal graph snapshot indicator sequence within the sliding window in the historical normal distribution data as input, calculating the mean vector of the normal graph snapshot indicator sequence, and applying diagonal loading regularization to the covariance matrix of the normal graph snapshot indicator sequence to obtain the mean vector and the covariance matrix with diagonal loading regularization; based on the sliding window... The comparison results between the number of normal graph snapshots and the minimum snapshot number threshold are as follows: when the number of normal graph snapshots is less than the minimum snapshot number threshold, diagonal whitening is performed by replacing the covariance matrix with diagonal loading regularization with dimension-wise independent standardization to obtain a whitening transformation matrix; when the number of normal graph snapshots is not less than the minimum snapshot number threshold, full covariance whitening is performed based on the covariance matrix with diagonal loading regularization to obtain a whitening transformation matrix; based on the multidimensional monitoring indicator vector of the current time window corresponding to the graph database change log stream, the mean vector, and the whitening transformation matrix, the whitened Mahalanobis distance anomaly score is calculated to obtain the whitened Mahalanobis distance anomaly score of the current time window; The multi-agent repair module is used to perform adaptive repair and iterative convergence verification through multi-agent collaboration based on the hierarchical quality problem list, the implicit topological defect location report, and the structured alarm events, and output the final repair report.

Images