Urban agent decision-making auditing method and system based on data consanguinity tracking

Abstract

The invention relates to the technical field of urban agent decision-making auditing, and discloses an urban agent decision-making auditing method and system based on data consanguinity tracking, and the method comprises the steps: collecting metadata of a whole urban agent decision-making process in a layered manner, integrating multiple layers of metadata, and carrying out the multi-dimensional marking and dependence recognition; constructing a time sequence blood relationship map, and performing storage optimization; performing decision contribution degree evaluation on the target decision node; restoring a key path of the target decision node, and analyzing a causal association relationship; performing multi-level auditing analysis on the decision-making process, and identifying a root cause node; generating a decision audit report; analyzing historical influence and future risk execution of the root cause problem; auditing information is extracted and integrated; matching a treatment scheme and pushing and executing the treatment scheme; tracking the treatment effect and continuously optimizing; according to the method, the blood relationship tracking system covering the whole process from data acquisition to decision output is constructed, so that the transparency and auditing of the urban agent decision process are realized.

Description

Technical Field

[0001] This invention relates to the field of urban intelligent agent decision auditing technology, and more specifically, to an urban intelligent agent decision auditing method and system based on data lineage tracing. Background Technology

[0002] With the deepening of smart city construction, urban emergency management systems are increasingly relying on intelligent agent technology for automated decision-making. Especially in emergency scenarios such as public health emergencies, natural disasters, and traffic accidents, urban intelligent agents need to comprehensively analyze multi-source heterogeneous data, including population flow data, medical resource distribution, epidemic monitoring information, traffic operation status, and meteorological environment, to automatically generate key decisions such as resource allocation plans, evacuation routes, and the delineation of isolation areas. However, existing urban intelligent agent decision-making systems generally suffer from a "black box" problem in the decision-making process, with complex decision-making chains and a lack of effective auditing mechanisms, leading to serious challenges to the credibility and traceability of decision results.

[0003] In current technologies, the decision-making process of intelligent agents involves multiple stages, including data acquisition, data processing, feature engineering, model inference, and decision execution. Each stage may introduce risk factors such as data quality issues, model bias, and rule conflicts. Due to the lack of effective monitoring and auditing capabilities for the entire decision-making process, it is difficult to quickly locate the root cause of any decision deviations or anomalies, let alone assess the scope of the problem's impact on historical decisions and its potential risks to future decisions. Existing auditing methods mainly rely on manual inspection and simple log analysis, which cannot meet the needs of large-scale, high-frequency intelligent agent decision auditing, nor can they handle complex data dependencies and causal propagation chains.

[0004] Furthermore, the diverse data sources, lengthy processing chains, and complex model combinations within urban intelligent systems make it difficult for traditional auditing methods to construct complete data lineages and accurately assess the contribution and impact of each element on the final decision. The lack of a systematic decision-making quality assurance mechanism not only affects the efficiency and accuracy of emergency response but may also lead to serious consequences such as improper resource allocation and ineffective emergency measures, threatening urban safety and public interests. Summary of the Invention

[0005] This invention provides a method and system for auditing urban intelligent agent decisions based on data lineage tracing, which solves the technical problems of lack of transparency, poor traceability, and difficulty in assessing decision quality in the decision-making process of related intelligent agents.

[0006] This invention provides a method for auditing urban intelligent agent decisions based on data lineage tracing, comprising the following steps: Metadata from the entire decision-making process of urban intelligent agents is collected in layers, and multi-layered metadata is integrated to perform multi-dimensional annotation and dependency identification, resulting in a multi-layered lineage graph structure. Based on a multi-level kinship graph structure, a temporal kinship graph is constructed and its storage is optimized to obtain an optimized kinship graph and target decision nodes; Based on the optimized kinship map, the decision contribution of the target decision nodes is evaluated to obtain the dynamic decision contribution vector of the elements. Based on the optimized lineage graph and the dynamic decision contribution vector of elements, the critical path of the target decision node is reconstructed, the causal relationship is analyzed, and the decision path structure with enhanced causality is obtained. Based on the decision path structure of causal reinforcement, a multi-level audit analysis is performed on the decision process to identify root cause nodes and obtain a root cause node set; a decision audit report is generated by combining the audit analysis results and the root cause node set. Based on the root cause node set, the historical impact and future risk of root cause problems are analyzed in the optimized lineage graph, resulting in a cross-time dimension impact analysis report; Extract and integrate audit information from decision audit reports and cross-time dimension impact analysis reports; match disposal plans and push them for execution; track disposal effects and continuously optimize them to obtain decision quality assurance plans.

[0007] In a preferred embodiment, the hierarchical collection of metadata for the entire decision-making process of the urban intelligent agent includes: Metadata acquisition agent technology is used to capture information across the entire decision-making process. The acquisition agent is deployed in the data access layer, data processing layer, feature calculation layer, model inference layer, and decision execution layer; and captures metadata of the data access layer, data processing layer, feature calculation layer, model inference layer, and decision execution layer, respectively.

[0008] In a preferred embodiment, the construction of the temporal kinship map and the optimization of its storage include: The element nodes in the multi-level kinship graph structure are mapped to vertices in the graph database, and attributes are set for each vertex to obtain the kinship graph vertex set; Map the dependent edges to the edges of the graph database, and set the dependency type, data transfer amount, and transfer delay attributes for each edge to obtain the bloodline graph edge set; An enumeration type index is created for the lineage hierarchy attribute of the vertex, a time range index is created for the timestamp attribute, and a hash index is created for the element identifier attribute, resulting in a multi-dimensional composite index structure; For high-frequency query scenarios, the subgraph structure involved in the path is pre-calculated and stored in the key-value caching system to obtain the path query cache.

[0009] In a preferred embodiment, the decision contribution assessment includes: Extract decision node identifiers from the optimized kinship graph, perform reverse graph traversal query, and construct a decision kinship subgraph; Multidimensional feature extraction is performed on each node in the decision lineage subgraph to obtain the initial feature vector of the node; A graph neural network model is constructed to encode and learn the decision lineage subgraph. The node representation is updated by aggregating the neighborhood features of the nodes through multi-layer graph convolution operations, resulting in a set of node embedding vectors. An attention mechanism is used to calculate the attention weight of the element node to the decision node, and the dot product similarity between the embedding vectors of the decision node and the element node is calculated. After softmax normalization, the dynamic decision contribution vector of the element is obtained.

[0010] In a preferred embodiment, the process of reconstructing the critical path of the target decision node and analyzing causal relationships includes: Extract the target decision node identifier from the optimized kinship graph, and execute the breadth-first reverse traversal algorithm to obtain the set of decision-dependent nodes and the set of dependencies. Based on the dynamic decision contribution vector of elements, the decision-dependent nodes are sorted in descending order of contribution value, and the nodes with contribution values ​​exceeding a preset threshold are retained as key nodes, thus obtaining the set of key decision path nodes. For adjacent nodes in the critical decision path, the conditional mutual information between nodes is calculated. Node pairs with conditional mutual information greater than a significance threshold are identified as having a causal relationship. Counterfactual reasoning is performed on key element nodes to construct counterfactual scenarios, simulate the propagation process of data changes through kinship paths, assess the impact of data changes on decision outcomes, and obtain a causal-enhanced decision path structure.

[0011] In a preferred embodiment, the multi-level audit analysis of the decision-making process to identify root cause nodes includes: For data source nodes in the causal reinforcement decision path structure, perform data quality audit checks, calculate data timeliness indicators, integrity indicators, and consistency indicators, compare the timeliness indicators, integrity indicators, and consistency indicators with preset thresholds, identify data source nodes with quality problems, and calculate risk scores. For model nodes, perform model running status audit checks, calculate the divergence value between the input feature distribution and the benchmark distribution to represent the distribution drift, extract inference confidence, verify version compliance, and calculate model risk score; For decision-making rule nodes, conduct decision compliance audits, check triggering conditions based on the business rule rationality knowledge base, check execution actions based on the business specification knowledge base, and calculate compliance risk scores; A multi-source evidence fusion algorithm is used to identify root cause nodes. The number of audit issues, risk scores, decision contribution, and causal effect assessment values ​​of the nodes are comprehensively considered. Nodes with confidence scores exceeding a preset threshold are selected to obtain the root cause node set.

[0012] In a preferred embodiment, the analysis of the historical impact and future risks of root cause problems includes: For root cause nodes, a forward graph traversal query is performed based on the optimized lineage graph to filter historical decision nodes and obtain the set of historical decision nodes affected by the root cause. Recalculate the contribution weight of the root cause node, combine the contribution weight with the problem impact coefficient, and calculate the score of the degree to which historical decisions are affected. Query the current usage status of the components involved in the root cause node, identify active decision-making links, and assess the potential impact on future decisions; A time-series graph neural network model is constructed to predict the future impact range and propagation path of root cause nodes, resulting in a cross-time dimension impact analysis report.

[0013] In a preferred embodiment, the extraction and integration of audit information from the decision audit report and the cross-time dimension impact analysis report includes: Based on the decision audit report, extract the root cause node's identifier, type, problem category, risk score, and confidence level assessment; Based on the cross-time dimension impact analysis report, extract the number of historical impact decisions, risk level distribution, high-risk decision list, and future risk scenario list to construct a set of key audit elements.

[0014] In a preferred embodiment, the tracking and continuous optimization of the treatment effect includes: Establish an audit effectiveness evaluation mechanism to track the implementation of root cause problem handling, compare changes in decision quality indicators before and after handling, and evaluate the effectiveness of auditing and handling. Based on the audit performance feedback data, the accuracy, false positive rate, and false negative rate of root cause identification were statistically analyzed, and the model parameters and threshold settings were adjusted, with false negative cases used as new training samples. Online learning algorithms are used to continuously optimize the audit model, and the graph neural network model is incrementally trained regularly to dynamically optimize parameter settings, thereby obtaining a decision quality assurance solution.

[0015] In a preferred embodiment, a city intelligent agent decision auditing system based on data lineage tracing is used to execute the above-described city intelligent agent decision auditing method based on data lineage tracing, including: The metadata collection module is used to collect metadata from the entire decision-making process of the urban intelligent agent in a hierarchical manner, integrate multi-level metadata, perform multi-dimensional annotation and dependency identification, and obtain a multi-level lineage graph structure. The kinship graph construction module constructs a temporal kinship graph based on a multi-level kinship graph structure and optimizes its storage to obtain an optimized kinship graph and target decision nodes. The contribution assessment module, based on the optimized kinship map, assesses the decision contribution of the target decision nodes and obtains the dynamic decision contribution vector of the elements. The causal analysis module, based on the optimized lineage graph and the dynamic decision contribution vector of elements, reconstructs the critical path of the target decision node, analyzes the causal relationship, and obtains the causal-enhanced decision path structure. The audit analysis module, based on the causal reinforcement decision path structure, performs multi-level audit analysis on the decision-making process, identifies root cause nodes, and obtains a root cause node set; it then combines the audit analysis results and the root cause node set to generate a decision audit report. The impact analysis module, based on the root cause node set, analyzes the historical impact and future risk of root cause issues in the optimized lineage graph, and obtains a cross-time dimension impact analysis report. The quality assurance module is used to extract and integrate audit information from decision audit reports and cross-time dimension impact analysis reports; match disposal plans and push them for execution; track disposal effects and continuously optimize them to obtain decision quality assurance plans.

[0016] The beneficial effects of this invention are as follows: This invention achieves transparency and auditability in the decision-making process of urban intelligent agents by constructing a lineage tracing system covering the entire process from data acquisition to decision output. Employing technologies such as hierarchical metadata collection, graph neural network modeling, and causal relationship analysis, it can accurately identify key elements and risk nodes in the decision-making chain, automatically locate the root causes of decision deviations, and quantitatively assess the true contribution of each element to the decision outcome. Compared to traditional manual auditing methods, this invention improves auditing efficiency and accuracy, can handle large-scale, high-frequency decision auditing tasks, and provides a reliable decision quality assurance mechanism for urban emergency management.

[0017] The cross-time dimension impact analysis capability established by this invention can not only trace the scope of the impact of root causes on historical decisions, but also predict potential risks in future decisions, providing decision-makers with comprehensive risk assessment and early warning services. Through an optimized lineage graph storage architecture and an intelligent audit engine, the system can automatically generate structured audit reports, match targeted response plans, and support continuous optimization and knowledge accumulation of the audit model; effectively solving the problem of lack of supervision and quality control in the decision-making process of urban intelligent agents, and improving the credibility of emergency decisions and the level of intelligence in urban governance. Attached Figure Description

[0018] Figure 1This is a flowchart of the main process of a city intelligent agent decision auditing method based on data lineage tracing in this invention; Figure 2 This is a detailed flowchart of a city intelligent agent decision auditing method based on data lineage tracing in this invention; Figure 3 This is a module diagram of an urban intelligent agent decision auditing system based on data lineage tracing in this invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a method for auditing urban intelligent agent decisions based on data lineage tracing, such as... Figures 1 to 2 As shown, it includes the following steps: Step 1: Collect metadata of the entire decision-making process of the urban intelligent agent in layers, integrate multi-layer metadata, perform multi-dimensional annotation and dependency identification, and obtain a multi-layer lineage graph structure; Step 1.1: Capture full-process metadata; Based on the decision requests received by the city intelligent agent, metadata collection agent technology is used to capture metadata of the entire decision-making process of the city intelligent agent. When the decision request is triggered, the distributed metadata collection agent is started. The collection agent is deployed in the data access layer, data processing layer, feature calculation layer, model inference layer and decision execution layer. The collection agent captures the processing events of each layer in real time through hook functions and log parsing technology.

[0021] Step 1.2: Collect metadata for each layer in a hierarchical manner; Based on the metadata collection agent deployed in step 1.1, metadata at each layer is collected in a hierarchical manner.

[0022] Specifically, for the data access layer, the system identifier, data table name, data field name, data collection timestamp, number of data records, and data freshness index of the original data source are captured to obtain the data source layer metadata. This data source layer metadata is used to identify the original data source and timeliness upon which decision-making depends. For example, in emergency decision-making scenarios, the data source layer metadata includes population flow data from the signaling data system of a mobile communication operator, a collection time of 30 minutes before the decision is triggered, and data freshness as real-time updates.

[0023] For the data processing layer, the metadata is obtained by capturing the types of data cleaning operators, the types of data transformation operators, the parameter configurations of the operators, the input data identifiers of the operators, the output data identifiers of the operators, and data quality change indicators. This metadata is used to track the evolution path and quality changes of data during the cleaning and transformation process. Data quality change indicators are obtained by comparing the completeness, consistency, and accuracy of the input and output data of the operators.

[0024] For the feature computation layer, the transformation logic description of the feature engineering operators, the set of input raw fields, the set of output feature fields, the textual description of the feature calculation formulas, the statistical distribution of feature values, and the feature importance score are captured to obtain feature layer metadata. Feature layer metadata is used to track the transformation process from raw data fields to model input features. For example, regional population density features are generated from raw population flow data using spatiotemporal aggregation operators, and the importance score of this feature is obtained through correlation analysis between the feature and historical decision results.

[0025] For the model inference layer, the following metadata is captured: model type identifier, model version number, list of model input features, model output prediction results, model inference confidence, distribution statistics of model input features, and baseline values ​​of feature distribution during model training. This metadata is used to track the model's inference process and assess the risk of distribution drift in the model's inputs. The model's inference confidence is obtained through the entropy value of the probability distribution of the model's output or the maximum predicted probability, while the distribution drift risk is calculated by measuring the divergence between the current input feature distribution and the baseline distribution during training.

[0026] For the decision execution layer, the metadata of the decision layer is obtained by capturing the identifier of the decision rule, the trigger condition expression of the decision rule, the description of the execution action of the decision rule, the fusion strategy of multiple model outputs, the result content of the decision output, and the confidence score of the decision. The decision layer metadata is used to track the rule reasoning process from model prediction to the final decision. For example, the trigger condition for the decision rule of delineating isolation areas is that the epidemic risk prediction model outputs a high-risk area and the population density of the area exceeds a threshold.

[0027] Step 1.3: Integrate metadata from each layer; Based on metadata from the data source layer, data processing layer, feature layer, model layer, and decision layer, a unified metadata record structure is constructed. Each metadata record contains an element identifier, element type, lineage level identifier, timestamp, version number, input dependency list, output flow list, quality indicator dictionary, and configuration parameter dictionary, resulting in a decision process metadata set with time-series annotations.

[0028] Step 1.4: Construct multidimensional labeled vectors and extract the corresponding quality indicators and feature attributes; Based on the decision process metadata set with time-series annotations obtained in step 1.3, for each metadata record, corresponding quality indicators and feature attributes are extracted according to its lineage hierarchy identifier and element type to construct a multi-dimensional annotation vector. For the metadata at the data source layer, attributes such as data freshness, completeness, and consistency are annotated; for the metadata at the data processing layer, attributes such as processing time and data quality improvement are annotated; for the metadata at the feature layer, attributes such as feature importance and feature distribution statistics are annotated; for the metadata at the model layer, attributes such as confidence and distribution drift are annotated; and for the metadata at the decision layer, attributes such as decision confidence and rule coverage are annotated, resulting in a multi-dimensional annotated metadata set.

[0029] Step 1.5: Construct directed dependency edges and generate a multi-level lineage graph structure; Based on the multidimensionally annotated metadata set obtained in step 1.4, each metadata record is traversed, its input dependency list is read, all predecessor elements of the current element are identified, and directed dependency edges from predecessor elements to the current element are established. The type, data transfer volume, and transfer latency of the dependency edges are recorded. The types of dependency edges include data flow dependency, control flow dependency, and configuration dependency. The data transfer volume represents the number of data records or data fields transferred through the dependency edge, and the transfer latency represents the time interval between the output of the predecessor element and the input of the current element. A directed graph structure containing all element nodes and dependency edges is constructed to obtain a multi-level lineage graph structure.

[0030] In some embodiments, since the decision-making process of urban intelligent agents may involve distributed data processing frameworks and microservice architectures, metadata collection faces challenges across systems and networks. An asynchronous metadata collection mechanism based on message queues can be adopted to achieve low-latency and highly reliable metadata collection. Specifically, metadata collection agents are deployed at each processing layer. These agents serialize the captured metadata records and send them to a central message queue. A dedicated metadata processing service consumes the metadata records from the message queue and performs standardized processing and storage. The message queue adopts a publish-subscribe model to support parallel processing by multiple consumers and employs a message persistence mechanism to ensure that metadata is not lost. A session identifier field is added to the metadata records to associate all metadata records for the same decision request. The session identifier aggregates the distributed metadata into a complete set of decision-making process metadata.

[0031] In some embodiments, due to differences in metadata formats and semantics between different data sources and processing systems, directly collected metadata may suffer from semantic inconsistencies. An ontology-based metadata semantic standardization method can be employed to achieve semantic alignment and unified representation of heterogeneous metadata. Specifically, a metadata ontology model for the urban intelligent agent domain is pre-constructed. This ontology model defines the semantic descriptions and relational constraints of concepts such as data sources, data fields, processing operators, features, models, and decisions. For the collected raw metadata records, ontology mapping rules are used to map the fields of the raw metadata to ontology concepts, identifying the semantic type of the metadata records. For data field-type metadata, through field name matching and data content analysis, the fields are mapped to standard data entities defined in the ontology. For example, different expressions such as "population count," "total population," and "population" are uniformly mapped to the ontology concept of "regional population." Through semantic standardization, a semantically consistent metadata set is obtained, supporting the accurate construction of subsequent lineage relationships.

[0032] Step 2: Based on the multi-level kinship graph structure, construct a temporal kinship graph and optimize its storage to obtain an optimized kinship graph and target decision nodes; Step 2.1, construct a temporal kinship map; Based on the multi-level kinship graph structure obtained in step 1, a native graph database system supporting temporal attributes and attribute indexes is selected. The element nodes in the multi-level kinship graph structure are mapped to vertices in the graph database. Attributes are set for each vertex, including element identifier, element type, kinship level, timestamp, version number, quality metric dictionary, and configuration parameter dictionary, resulting in a vertex set for the kinship graph. Dependency edges in the multi-level kinship graph structure are mapped to edges in the graph database. Attributes are set for each edge, including dependency type, data transfer amount, transfer delay, and start timestamp, resulting in an edge set for the kinship graph. Vertices and edges together constitute the temporal kinship graph stored in the graph database.

[0033] Step 2.2, create a multidimensional index; Based on the time-series kinship graph obtained in step 2.1, typical kinship tracing query patterns are analyzed. These patterns include positive impact analysis query, reverse root cause tracing query, time range filtering query, kinship hierarchy filtering query, and quality indicator filtering query. In the urban intelligent agent decision-making system, the nodes of the kinship graph are divided into five categories according to their functional characteristics: data source elements refer to the source systems and data tables of the original data, such as population flow data and medical resource data; processing operator elements refer to processing components that clean, transform, and aggregate the data; feature elements refer to the feature variables extracted or calculated from the original data for model input; model elements refer to machine learning models or rule engines used for prediction, classification, and reasoning; and decision elements refer to the final decision results and execution plans generated based on the model output. Specifically, the positive impact analysis query starts from a specified data source element or processing operator element and traverses forward along the lineage relationship to find all downstream nodes and final decision elements affected by that node, used to assess the potential impact range of data source changes or model updates on decisions; the reverse root cause tracing query starts from a decision element and traverses backward along the lineage relationship to trace all upstream data source elements, processing operator elements, feature elements, and model elements that affect the decision, used to analyze the data source and processing chain of the decision result; the time range filtering query filters the lineage relationship and decision nodes within a specific time window based on the timestamp attribute of the node; the lineage hierarchy filtering query filters nodes at specific levels according to the lineage hierarchy such as data access layer, data processing layer, feature calculation layer, model inference layer, and decision execution layer; the quality index filtering query filters nodes that meet the quality conditions based on the values ​​in the node's quality index dictionary, such as data freshness, completeness, and model confidence. Based on the query pattern analysis results, an enumeration type index is established for the lineage hierarchy attribute of vertices in the time-series lineage relationship graph, a time range index is established for the timestamp attribute of vertices, a unique hash index is established for the element identifier attribute of vertices, and a numerical range index is established for the key indicators in the quality indicator dictionary of vertices, resulting in a multi-dimensional composite index structure.

[0034] Step 2.3, build the query cache; Based on the temporal lineage graph obtained in step 2.1, common lineage tracing path patterns are identified by analyzing historical query logs for high-frequency query scenarios. Path patterns include complete paths tracing backward from decision nodes to data source nodes, paths tracing forward from specific data source nodes to all affected decision nodes, and upstream and downstream dependency paths of specific model types. For the identified high-frequency path patterns, a graph traversal algorithm is used to pre-calculate the subgraph structures involved in the path. The subgraph structures are serialized and stored in a key-value caching system, using the hash value of the path query conditions as the cache key to obtain the path query cache. A hierarchical caching strategy is designed to improve query performance. The hierarchical caching strategy consists of three layers: the first layer is a memory cache, storing hot query results within a recent time window, using an LRU eviction policy; the second layer is a distributed cache cluster, storing moderately popular query results and path query caches, using a consistent hashing algorithm to achieve distributed storage of cached data; the third layer is the query result cache of the graph database itself. Query requests search the cache layer by layer, and only when all cache layers are not found is a complete graph traversal query executed, resulting in an accelerated lineage query structure.

[0035] Step 2.4, implement tiered storage; Based on the temporal lineage graph obtained in step 2.1, a time-window-based partitioned storage strategy is adopted to address the issue of its continuously increasing data volume over time. According to the timestamp attributes of vertices and edges, the lineage graph is divided into multiple time partitions by day or week, with each time partition corresponding to an independent graph partition in the graph database. Recent time partitions are stored on high-performance SSD storage media, while historical time partitions are migrated to high-capacity HDD storage media or object storage systems, resulting in a hierarchical lineage graph architecture. Query requests are automatically routed to the corresponding time partition based on time range conditions. For cross-partition queries, a distributed query coordination mechanism is used to aggregate query results from each partition.

[0036] Step 2.5, establish version management; Based on the temporal lineage graph obtained in step 2.1, a version management mechanism for the lineage graph is established. When the data source or model is updated, the vertices and edges of the historical version are retained in the graph database, while the vertices and edges of the new version are added. Different versions are distinguished by the version number attribute. The lineage relationships of historical versions support auditing historical decisions, while the lineage relationships of new versions are used for current decisions. Versions are connected by special version evolution edges to identify the replacement relationship between old and new versions, resulting in the lineage graph storage architecture.

[0037] Step 2.6, integrate and optimize measures; Based on the time-series lineage graph obtained in step 2.1, the multi-dimensional composite index structure obtained in step 2.2, the lineage query acceleration structure obtained in step 2.3, the hierarchical storage lineage graph architecture obtained in step 2.4, and the lineage graph storage architecture obtained in step 2.5, an optimized lineage graph is integrated and constructed. It provides services externally through a unified query interface. Query requests are retrieved through a cache layer; if a match is not found, the data is quickly located through the index structure and automatically routed to the corresponding storage partition based on the time range. It supports cross-version lineage tracing and analysis. The optimized lineage graph possesses efficient graph data storage capabilities, multi-dimensional index query capabilities, multi-level caching acceleration capabilities, hierarchical storage management capabilities, and version evolution tracking capabilities.

[0038] Step 2.7: Identify and extract target decision nodes; Based on the optimized lineage map obtained in step 2.6, the target decision nodes that need to be audited are identified. The identification methods for target decision nodes include the following scenarios: Based on anomaly detection, automatic identification is performed on all decision-type element nodes in the optimized lineage graph. Attributes such as decision timestamp, decision type, and decision importance are extracted. Anomaly detection algorithms are used to identify decision nodes with abnormal decision quality indicators, abnormal decision times, or decision results that deviate from historical patterns. The detected abnormal decision nodes are marked as candidate target decision nodes.

[0039] Based on proactive identification of business rules, and according to preset audit rules and business importance assessment, decision nodes that meet the audit conditions are screened. The audit conditions include the impact of the decision exceeding a threshold, the decision involving the allocation of key resources, and the decision being in a high-risk time window. Decision nodes that meet the conditions are added to the target decision node candidate set.

[0040] Based on manual passive identification, it receives audit requests from decision management personnel or auditors. The requests include information such as the specified decision identifier, decision time range, and decision scenario type. According to the request conditions, it queries the optimized lineage graph for matching decision nodes and uses the query results as a candidate set of target decision nodes.

[0041] The decision nodes in the candidate set are sorted according to their audit priority. The priority is determined by comprehensively considering factors such as the business importance of the decision, the degree of abnormality, and the timeliness requirements. The decision node with the highest priority is selected as the target decision node for the current audit and serves as the clear audit objective.

[0042] In some embodiments, due to the high decision-making frequency of the urban intelligent agent system, the write throughput requirement for the kinship graph is very high; a batch asynchronous write mechanism can be adopted to improve the performance of kinship data ingestion. Specifically, a write buffer layer is added before the graph database. Metadata records are written to a memory buffer. When the buffer accumulates a certain number of records or after a certain time interval, they are batch submitted to the graph database for batch insertion operations. The batch insertion operation uses the transaction batch processing interface provided by the graph database, inserting multiple vertices and edges at once in a single transaction, reducing transaction overhead. The write buffer layer adopts a double buffering mechanism: one buffer receives new data, and the other buffer performs batch submissions. The two buffers work alternately to ensure the continuity of writing. For scenarios that require immediate querying of newly written kinship data, the query service checks both the write buffer and the graph database simultaneously and returns the results by merging them.

[0043] In some embodiments, since the pedigree graph may reach hundreds of millions of vertices and edges, the storage and computing power of a single-machine graph database may be insufficient; a distributed graph database cluster architecture can be adopted to achieve horizontal scaling and high availability of the pedigree graph. Specifically, a distributed graph database system is used, where the pedigree graph is hash-sharded according to the element identifier of the vertex and distributed across multiple graph database nodes. Each node is responsible for managing a portion of the vertices and their associated edges, and cross-node edges between vertices are processed through a distributed query protocol. For graph traversal queries, a distributed graph computing framework is used to decompose the traversal task into multiple subtasks that are executed in parallel on different nodes, and the execution of subtasks and result aggregation are coordinated through a message passing mechanism. A replica mechanism for the graph database is configured, with each data shard maintained on multiple nodes, providing fault tolerance and load balancing capabilities for read queries.

[0044] Step 3: Based on the optimized kinship map, evaluate the decision contribution of the target decision nodes to obtain the dynamic decision contribution vector of the elements; Step 3.1, extract the decision lineage subgraph; Based on the optimized lineage graph obtained in step 2, for the decision node representing the final decision output in the optimized lineage graph, the identifier of the decision node is extracted from the optimized lineage graph. A reverse graph traversal query is performed in the graph database, starting from the decision node and traversing in the reverse direction of the edges to visit all reachable predecessor nodes, thus obtaining the set of predecessor nodes for the decision. The set of predecessor nodes includes data source nodes, data processing nodes, feature nodes, model nodes, and the dependency edges between these nodes. A decision lineage subgraph from the data source to the decision result is constructed, with all nodes in the predecessor node set as vertices of the subgraph. At the same time, all dependency edges between these nodes in the original optimized lineage graph are extracted as edges of the subgraph. Step 3.2: Extract multidimensional features of nodes in the decision lineage subgraph; Based on the decision lineage subgraph obtained in step 3.1, multi-dimensional feature extraction is performed on each node in the decision lineage subgraph. Specifically, element type feature vectors are extracted according to the element type of the node. The element type feature vectors use one-hot encoding to represent the category of the node, such as data source, processing operator, feature, model, and decision. Numerical indicators from the node's quality indicator dictionary are extracted to construct a quality feature vector, which includes normalized values ​​such as data freshness, completeness, consistency, and model confidence. The timestamp attribute of the node is extracted to calculate temporal features, which include the node's absolute timestamp and the time difference relative to the decision trigger time. The structural features of the node in the lineage subgraph are extracted, which include the node's in-degree, out-degree, and the shortest path length to the decision node.

[0045] The initial feature vector of a node is obtained by concatenating the element type feature vector, quality feature vector, temporal feature, and structural feature.

[0046] Step 3.3: Extract the feature vectors of the edges in the decision lineage subgraph; Based on the decision lineage subgraph obtained in step 3.1, for each edge in the decision lineage subgraph, the dependency type attribute of the edge is extracted to construct an edge type feature. The dependency type includes categories such as data flow dependency, control flow dependency, and configuration dependency, and is represented using one-hot encoding. The data transmission quantity attribute and transmission delay attribute of the edge are extracted, normalized, and used as the numerical feature of the edge. The edge type feature and the numerical feature are concatenated to obtain the edge feature vector. Based on the initial node feature vector and the edge feature vector obtained in step 3.2, the feature-annotated decision lineage subgraph is constructed together.

[0047] Step 3.4: Construct the graph neural network model; Based on the decision lineage subgraph with feature annotations obtained in step 3.3, a graph neural network model is constructed to encode and learn it. The graph neural network model contains multiple graph convolutional layers. Each graph convolutional operation updates the node representation by aggregating the features of the neighboring nodes. Specifically, in the first graph convolutional layer, for each node in the decision lineage subgraph, the initial feature vectors of all its predecessor nodes and the feature vectors of the connecting edges are collected. The influence weight of each predecessor node on the current node is calculated through an attention mechanism. The influence weight is obtained by inputting the concatenated vector of predecessor node features, edge features, and current node features into a multilayer perceptron and then normalizing it through softmax. The predecessor node features are weighted and aggregated according to the influence weight to obtain aggregated features from the neighborhood. The aggregated features are concatenated with the initial features of the current node and then subjected to a nonlinear transformation to obtain the node features after the first graph convolution.

[0048] Step 3.5: Perform multi-layer graph convolution operation; The graph neural network model constructed in step 3.4 undergoes a complete multi-layer graph convolution computation process: Based on the node features after the first layer of graph convolution, neighborhood aggregation and feature update operations are repeatedly performed in subsequent graph convolution layers. Each layer of graph convolution allows the node's feature representation to fuse information from a wider range of neighborhoods. After multiple layers of graph convolution, the feature representation of the decision node fuses the structural information of the entire lineage subgraph and the feature information of all predecessor nodes. The feature vectors of all nodes in the decision lineage subgraph after the last layer of graph convolution are extracted to obtain a set of node embedding vectors. The node embedding vectors are high-dimensional representations learned in the lineage network structure, implying the node's position and role in the decision chain.

[0049] Step 3.6: Generate a dynamic decision contribution vector; Using the graph neural network model constructed in step 3.4 and the set of node embedding vectors obtained in step 3.5, an attention mechanism is employed to calculate the attention weights of element nodes on decision nodes, specifically for decision nodes and each element node. Element nodes refer to data source elements, processing operator elements, feature elements, and model elements in the decision lineage subgraph. Specifically, the embedding vectors of decision nodes and element nodes are extracted, and the dot product similarity between the two embedding vectors is calculated. The dot product similarity represents the association strength between the element node and the decision node in the embedding space. Softmax normalization is applied to the dot product similarity between all element nodes and the decision node in the decision lineage subgraph to obtain the attention weight for each element node. The attention weight reflects the comprehensive influence of an element on the decision through the lineage network, automatically considering multiple propagation paths, interactions with other elements, and nonlinear relationships, resulting in a dynamic decision contribution vector. The dimension of the dynamic decision contribution vector is equal to the number of element nodes in the decision lineage subgraph. Each element in the vector corresponds to a contribution weight value of an element node, with weight values ​​ranging from 0 to 1, and the sum of all weight values ​​is 1.

[0050] A node contribution index table is constructed by associating the contribution weight values ​​in the dynamic decision contribution vector of each element with the corresponding element node identifier. The node contribution index table uses the element node identifier as the key and the node's contribution weight value as the value, and is stored using a hash table or key-value data structure. The node contribution index table is stored in a memory cache system, allowing subsequent steps to quickly retrieve the corresponding contribution value using the node identifier. Simultaneously, the node contribution index table is persistently stored in the node attributes of the graph database, adding a current decision contribution attribute field to each element node. The attribute value is the node's contribution weight value in the current target decision, ensuring the traceability and consistency of contribution information.

[0051] Step 3.7, Model Training; Based on the element-based dynamic decision contribution vector, the graph neural network model parameters are trained using supervised learning. A lineage subgraph of historical decisions and decision quality labels are collected. These quality labels are obtained through post-audit or expert evaluation, identifying whether a decision is normal or abnormal. The historical decision lineage subgraph is input into the graph neural network model. Based on the node embedding vectors obtained in step 3.5, the quality category of the decision is predicted. The cross-entropy loss function is used to measure the difference between the predicted result and the true label. The parameters of the graph neural network are optimized using the backpropagation algorithm, enabling the model to learn effective node representations and contribution calculation methods, resulting in the trained graph neural network model.

[0052] In some embodiments, the computational complexity of graph neural networks is high because the decision lineage subgraph can be very large, containing hundreds of nodes and thousands of edges. A subgraph sampling and hierarchical computation strategy can be employed to reduce computational complexity and improve inference efficiency. Specifically, for large-scale decision lineage subgraphs, subgraph sampling is performed before graph convolution to retain key nodes and paths that contribute significantly to decision-making, while filtering out redundant nodes with low contribution. Subgraph sampling is achieved by ranking nodes based on importance, with node importance calculated comprehensively based on the number of paths from the node to the decision node, the node's in-degree and out-degree, and the node's quality metrics. The size of the sampled subgraph is controlled within an acceptable range of computational complexity. For multi-layer graph convolution operations, a hierarchical computation strategy is adopted. The first layer of graph convolution is performed on the complete subgraph, and subsequent layers are performed on progressively smaller subgraphs. Each layer retains only nodes with high attention weights for the next layer's computation, reducing computational load through hierarchical filtering.

[0053] In some embodiments, since the contribution distribution patterns of elements may differ across different decision-making scenarios, a single graph neural network model may lack sufficient generalization ability. Therefore, a multi-task learning graph neural network architecture can be adopted to improve the model's adaptability to different decision-making scenarios. Specifically, a multi-task graph neural network model is constructed, comprising shared graph convolutional layers and multiple task-specific output layers. The shared graph convolutional layers learn general node embedding representations, while the task-specific output layers perform specialized contribution calculations and decision quality predictions for different decision-making scenario types. During training, labeled data from multiple decision-making scenarios are used, with each scenario corresponding to a task. The weighted sum of the loss functions of multiple tasks serves as the total loss function. Through multi-task joint training, the shared graph convolutional layers learn general feature representations across scenarios, and the task-specific output layers learn scenario-specific patterns. During inference, the corresponding task-specific output layer is selected based on the current decision scenario type to calculate the contribution.

[0054] Step 4: Based on the optimized lineage graph and the dynamic decision contribution vector of elements, the critical path of the target decision node is reconstructed, the causal relationship is analyzed, and the decision path structure with enhanced causality is obtained. Step 4.1: Obtain the set of decision-dependent nodes and their dependencies; Based on the optimized lineage graph obtained in step 2, for the target decision node that needs to be audited, the node identifier is extracted from the optimized lineage graph. A breadth-first backward traversal algorithm is then executed in the graph database, visiting predecessor nodes hierarchically starting from the decision node. Specifically, the direct predecessor nodes of the decision node include model nodes and decision rule input nodes. The predecessor nodes of these nodes are then visited, including feature nodes and rule condition nodes. This process continues backward until the data source node is reached, resulting in a set of decision-dependent nodes containing all levels and a set of dependencies between nodes.

[0055] Step 4.2: Construct a set of key decision-making path nodes; Based on the dynamic decision contribution vector obtained in step 3, for each node in the decision dependency node set obtained in step 4.1, the corresponding contribution value of that node is queried. The nodes in the decision dependency node set are sorted in descending order of contribution value. A contribution threshold is set as the set of nodes whose contribution or cumulative contribution reaches a preset proportion. The specific value is determined based on the size of the decision lineage subgraph and the audit accuracy requirements. Nodes with contribution values ​​exceeding the threshold are retained as key nodes, while secondary nodes with lower contribution values ​​are filtered out, resulting in the key decision path node set. The key decision path node set focuses on the elements and processing steps that have the greatest impact on the decision, simplifying the complexity of subsequent causal analysis.

[0056] Step 4.3: Construct a decision path structure that labels the strength of causal relationships; Based on the set of key decision path nodes and their dependencies obtained in step 4.2, a key decision path graph structure is constructed. For each pair of adjacent nodes in the key decision path graph structure, it is analyzed whether a true causal relationship exists between the two nodes or merely a correlation exists. A causal discovery method based on conditional independence testing is used. For nodes A and B, if a statistical dependency still exists given other nodes, then a potential causal relationship exists between nodes A and B. By analyzing the conditional probability distribution of node values ​​in historical decision data, the conditional mutual information between nodes is calculated. Node pairs with conditional mutual information greater than a significance threshold are identified as having a causal relationship, thus obtaining a decision path structure labeled with the strength of causal relationships.

[0057] Step 4.4: Perform counterfactual reasoning to evaluate the causal effects of the elements on the decision outcome; Based on the set of key decision path nodes obtained in step 4.2, counterfactual reasoning is performed on important element nodes to evaluate their causal effect on the decision outcome. Specifically, an element node is selected, and a counterfactual scenario is constructed, assuming that the value of the element changes. The process of simulating data change propagating downstream through the kinship path is based on the optimized kinship graph obtained in step 2. According to the transformation logic of each processing node and model node in the kinship path, the impact of data change on downstream nodes is calculated step by step. Finally, the probability and degree of change in the decision outcome caused by data change are evaluated, and the causal effect evaluation value of the element is obtained. Counterfactual reasoning is performed on multiple important elements in the key decision path to obtain a set of causal effect evaluation values ​​for each element. The causal effect evaluation value quantifies the element's real influence on the decision, which is different from the contribution measure of correlation.

[0058] Step 4.5: Rank based on overall score; Based on the dynamic decision contribution vector of elements obtained in step 3, the decision path structure labeled with causal association strength obtained in step 4.3, and the set of causal effect evaluation values ​​for each element obtained in step 4.4, the nodes in the key decision path node set obtained in step 4.2 are comprehensively scored and ranked. The comprehensive score is calculated using a weighted summation method. The contribution reflects the statistical influence of the node, the causal association strength reflects the credibility of the causal connection between nodes, and the causal effect evaluation value reflects the counterfactual influence capability of the node. The combination of these three factors provides a multi-dimensional evaluation of node importance, resulting in the key decision path nodes after comprehensive scoring. Based on the comprehensive score, the nodes are prioritized, and the nodes with the highest priority and the paths connecting these nodes are retained, resulting in a causal-enhanced decision path structure.

[0059] Step 4.6, Visualization of the decision path; Based on the causal enhancement decision path structure obtained in step 4.5, a hierarchical graph layout algorithm is used to generate a visual representation of the decision path. The graph layout algorithm organizes the vertical positions of nodes according to their lineage hierarchy, with data source nodes at the top level, decision nodes at the bottom level, and data processing nodes, feature nodes, and model nodes in the middle layers. Nodes at the same lineage hierarchy are arranged horizontally, and dependencies are represented by directed edges connecting them. The visual attributes of a node map to its importance and status information; the size of a node maps to its overall score (higher scores result in larger nodes); the color of a node maps to its quality status (green for nodes of normal quality, red or yellow for nodes of abnormal quality); the visual attributes of an edge map to its causal relationship strength and data transfer volume; the thickness of an edge maps to the causal relationship strength or data transfer volume (stronger relationships or larger data volumes result in thicker edges); and the color of an edge maps to its dependency type (blue for data flow dependencies and gray for control flow dependencies). The generated visualization is presented using an interactive graphical interface, allowing users to click on nodes to view detailed attributes and historical data, and to adjust layout parameters and filtering conditions to obtain a visual representation of the decision path.

[0060] In some embodiments, since the identification of causal relationships relies on statistical analysis of historical data, the reliability of causal discovery decreases when the number of historical samples for certain nodes is insufficient. A causal relationship enhancement method incorporating domain knowledge can be employed to improve the accuracy and interpretability of causal identification. Specifically, a causal knowledge graph for the urban emergency decision-making domain is pre-constructed, defining known causal relationships between common elements, such as common-sense causal relationships like increased population density leading to a higher risk of epidemic transmission, and traffic congestion leading to prolonged emergency response time. During causal discovery, the statistical causal discovery results based on conditional independence tests are compared and fused with the domain causal knowledge graph. If a causal relationship discovered by the statistical method is supported by the knowledge graph, the confidence level of that causal relationship is increased; if a causal relationship not discovered by the statistical method but present in the knowledge graph is added as a priori causal relationship to the decision path. By fusing statistical causal discovery with domain knowledge, a more reliable and interpretable causal-enhanced decision path is obtained.

[0061] In some embodiments, since counterfactual reasoning requires simulating the propagation of data changes along a lineage path, and the reasoning logic of certain complex models is difficult to simulate precisely, an approximate counterfactual reasoning method based on a model surrogate can be adopted. The aim is to achieve feasible counterfactual analysis in complex model scenarios. Specifically, for black-box model nodes in a lineage path, an interpretable surrogate model is trained to approximate the input-output relationship of the black-box model. The surrogate model uses an interpretable model such as a decision tree or a linear model, trained on the training or inference data of the black-box model, so that the prediction results of the surrogate model are as close as possible to the black-box model. When performing counterfactual reasoning, the surrogate model is used instead of the black-box model to infer the impact of data changes. The interpretability of the surrogate model makes the impact inference process traceable and understandable. Although the surrogate model has approximation errors, in the relative comparison scenario of counterfactual reasoning, the surrogate model can provide sufficient reference for causal effect assessment. The approximation error of the surrogate model is evaluated and labeled, and error range information is included in the causal effect assessment results.

[0062] Step 5: Based on the causal reinforcement decision path structure, conduct multi-level audit analysis on the decision-making process, identify root cause nodes, and obtain a root cause node set; combine the audit analysis results and the root cause node set to generate a decision audit report; Step 5.1: Perform a data quality audit check on the data source node and calculate the comprehensive risk score; Based on the causal reinforcement decision path structure obtained in step 4, data quality audit checks are performed on the data source nodes. The timestamp attribute and current time of the data source nodes are extracted to calculate the data timeliness, i.e., the time interval between data collection and decision triggering. This timeliness is compared to a preset timeliness threshold for this type of data source; if the timeliness exceeds the threshold, the data source is identified as having a timeliness problem. The integrity index of the data source nodes is extracted. This index represents the proportion of missing data records or null values ​​in key fields. The integrity index is compared to a preset integrity threshold; if the integrity index is below the threshold, the data source is identified as having an integrity problem. The consistency index of the data source nodes is extracted. This index is calculated by comparing the differences in values ​​of the same data entity in different data sources or logical inconsistencies within the data. The consistency index is compared to a preset consistency threshold; if the consistency index is below the threshold, the data source is identified as having a consistency problem. For all data source nodes identified as having quality problems, a risk score is calculated based on the problem type and the node's comprehensive score in the decision path. The risk score comprehensively considers the degree of data quality deviation and the dynamic decision contribution vector of elements, resulting in the data quality audit results and risk scores.

[0063] Step 5.2: Perform an audit of the operational status of the model nodes and assess the risk of distribution drift and confidence level. Based on the causal reinforcement decision path structure obtained in step 4, a model runtime status audit is performed on the model nodes. The input feature distribution statistics and the baseline values ​​of the feature distribution during model training are extracted from the model nodes. The KL divergence or JS divergence between the current input feature distribution and the baseline distribution is calculated. The divergence value represents the degree of deviation of the feature distribution, i.e., the distribution drift. The distribution drift is compared with a preset drift threshold. If the drift exceeds the threshold, it indicates that the model has a risk of input distribution drift. The inference confidence of the model nodes is extracted. The inference confidence represents the model's confidence in the current prediction result. The inference confidence is compared with a preset confidence threshold. If the confidence is lower than the threshold, it indicates that the model's prediction result is not reliable enough. The version number of the model nodes is extracted, and the model version management system is queried to verify whether the version is a stable version that has undergone sufficient testing and approval. If the version number corresponds to an unapproved version or an experimental version, it indicates that the model has a version compliance issue. For all model nodes identified as having risks, a model risk score is calculated by comprehensively considering distribution drift, confidence bias, and version compliance issues, thus obtaining the model audit results.

[0064] Step 5.3: Perform a compliance audit of the decision rule nodes and evaluate the rationality and consistency of the rules; Based on the causal enhancement decision path structure obtained in step 4, a decision compliance audit is performed on the decision rule nodes. The trigger condition expressions of the decision rule nodes are extracted, and the condition variables and threshold settings involved in the trigger conditions are parsed. The trigger conditions are checked against a pre-set business rule rationality knowledge base to ensure they conform to business logic. For example, the threshold settings are checked to ensure they are within a reasonable range, and there are no logical contradictions in the combination of conditions. If the trigger conditions are unreasonable, the rule is identified as having a rationality issue. The execution action descriptions of the decision rule nodes are extracted, and the execution actions are checked against a pre-set business specification and regulatory requirement knowledge base to ensure they conform to business process specifications and legal requirements. For example, resource allocation decisions are checked to ensure they comply with emergency plan provisions, and isolation measures decisions are checked to ensure they comply with public health regulations. If the execution actions do not comply with the specifications, the rule is identified as having a compliance issue. For situations where multiple decision rules exist in the decision path, conflicts in trigger conditions or execution actions between rules are checked. If conflicts exist, the rule set is identified as having a consistency issue. For all decision rule nodes identified as having issues, a compliance risk score is calculated based on the severity and scope of the problem to obtain the decision compliance audit results.

[0065] Step 5.4: Construct a historical decision-making lineage dataset and train a baseline model for normal decision-making lineage patterns; Based on the optimized lineage graph obtained in step 2, all decisions and their lineage subgraphs within a historical time window are extracted to construct a historical decision lineage dataset. For each decision lineage subgraph in the historical decision lineage dataset, statistical features are extracted, including node type distribution, node contribution distribution, causal path pattern, data quality index distribution, and model confidence distribution, resulting in a set of lineage feature vectors for historical decisions. Historical decisions are categorized into normal decision sets and abnormal decision sets according to decision quality labels. The lineage feature vectors of the normal decision sets are used to train a normal pattern benchmark model. Unsupervised anomaly detection algorithms, such as Isolation Forest or Autoencoder, are used to train the model on the lineage feature vectors of the normal decision sets. The model learns the distribution pattern of lineage features for normal decisions, thus obtaining a normal pattern benchmark for decision lineage.

[0066] Step 5.5: Calculate the bloodline feature vector of the current decision and perform anomaly detection to assess the deviation. For the target decision in the current audit, its decision lineage subgraph is extracted and a lineage feature vector is calculated. The extraction method for the lineage feature vector is consistent with the feature extraction method for historical decisions in step 5.4. The lineage feature vector of the current decision is input into the normal pattern baseline model of decision lineage trained in step 5.4. The model calculates the deviation of the lineage feature vector from the normal pattern, i.e., the anomaly score. The anomaly score is obtained by calculating the distance or reconstruction error from the feature vector to the normal pattern decision cluster. The higher the anomaly score, the more the lineage pattern of the current decision deviates from the normal pattern, and the greater the possibility of anomaly risk. The anomaly score is compared with a preset anomaly threshold. If the anomaly score exceeds the threshold, it indicates that the current decision has an abnormal lineage pattern, and the anomaly detection result is obtained.

[0067] Step 5.6: Use a multi-source evidence fusion algorithm to identify root cause nodes and calculate root cause confidence. Based on the data quality audit results and risk scores obtained in step 5.1, the model audit results obtained in step 5.2, the decision compliance audit results obtained in step 5.3, and the anomaly detection results obtained in step 5.5, a multi-source evidence fusion algorithm is used to identify the root cause nodes leading to decision bias. Specifically, for each node in the causal enhancement decision path structure obtained in step 4, the problem identification and risk score of that node in each audit inspection are collected. If a node is identified as having problems in multiple audits, that node is more likely to be the root cause. The root cause confidence of the node is calculated. The root cause confidence comprehensively considers multiple dimensions of indicators such as the number of audit problems, risk score, decision contribution, and causal effect assessment value of the node, and a weighted fusion method is used to calculate the comprehensive confidence. All nodes are sorted in descending order of root cause confidence, and nodes with confidence exceeding a preset threshold are selected as the root cause node candidate set. For multiple nodes in the root cause node candidate set, the causal relationship between the nodes is analyzed. If there is a causal chain between multiple nodes, the upstream node of the causal chain is selected as the root cause, and downstream derivative problem nodes are filtered out to obtain the root cause node set.

[0068] Step 5.7: Generate a root cause analysis report and integrate audit evidence details to form a decision audit report; Based on the root cause node set obtained in step 5.6, a root cause analysis report is generated for each root cause node. The root cause analysis report includes the root cause node's identification information, node type, problem category, risk score, decision contribution, causal effect assessment value, and audit evidence details. The audit evidence details record the specific problems found in various audit checks for this node, such as the specific time difference for data timeliness delays, the specific divergence value for model distribution drift, and the specific logical description of unreasonable rule triggering conditions. By combining the analysis reports of all root cause nodes, a decision audit report containing root cause location and confidence level is generated.

[0069] In some embodiments, since preset audit rules and thresholds may not cover all anomalies, especially newly emerging anomaly patterns, an adaptive update mechanism for audit rules based on active learning can be adopted to improve the audit engine's ability to identify new anomalies. Specifically, the abnormal decisions and root cause analysis results identified by the audit engine are pushed to audit experts for manual review. Experts evaluate and provide feedback on the accuracy of root cause identification and the rationality of audit rules. False positives and false negatives reported by experts are collected. False positives indicate that the audit engine incorrectly classifies normal decisions as anomalies, while false negatives indicate that the audit engine fails to identify truly abnormal decisions. The characteristics of false positives and false negatives are analyzed to identify audit rule defects and inappropriate threshold settings that lead to false positives and false negatives. An active learning algorithm is then used to automatically adjust the audit rules and threshold parameters. For new anomaly patterns, a supplementary module of the anomaly detection model is trained using expert-annotated cases to incorporate the new patterns into the detection capability. Through continuous human-machine collaboration and model updates, the audit engine achieves adaptive evolution.

[0070] In some embodiments, since there may be multiple problem nodes in the decision-making path and causal propagation relationships between problems, simple independent audits may miss the problem propagation chain. A problem analysis method based on causal propagation networks can be adopted to identify the cascading propagation and amplification effects of problems. Specifically, based on a causal-enhanced decision-making path structure, a problem propagation network is constructed. Nodes in the network are audit problem nodes that identify existing problems, and edges between nodes represent causal relationships. Path analysis is performed on the problem propagation network to identify the complete propagation chain from upstream data source problems to downstream decision biases. The cumulative and amplification effects of problems in the propagation chain are analyzed; for example, a small decrease in the quality of upstream data can be gradually amplified through multi-layer model inference, leading to a serious bias in the final decision. The problem amplification coefficient of each link in the propagation chain is quantified, calculated by comparing the severity of the input problem and the severity of the output problem at that link. The key link with the highest amplification coefficient in the propagation chain is identified; this link is the main node for problem amplification, and optimization of this node can effectively reduce the impact of problem propagation. Based on the comprehensive results of the problem propagation network analysis, information on the problem propagation chain and amplification links is supplemented in the root cause analysis report.

[0071] Step 6: Based on the root cause node set, analyze the historical impact and future risk of root cause problems in the optimized lineage graph to obtain a cross-time dimension impact analysis report; Step 6.1: Perform a forward graph traversal query and obtain the set of historical decision nodes affected by the root cause; Based on the root cause node set obtained in step 5, for each root cause node, extract its identifier and type information, and perform a forward graph traversal query based on the optimized lineage graph obtained in step 2. Starting from the root cause node, traverse along the forward direction of the edges, visiting all successor nodes. Successor nodes include all downstream processing nodes, feature nodes, model nodes, and decision nodes that use or depend on the root cause node's data. For each traversed decision node, extract its timestamp attribute, and filter decision nodes whose timestamps are earlier than the current audit decision timestamp to obtain a set of historical decision nodes affected by the root cause. Sort the historical decision node set by timestamp, and record the decision type, decision scenario, decision result, and other attribute information for each historical decision to obtain a historical decision set with time-series information.

[0072] Step 6.2: Recalculate the root cause node contribution weights and assess the extent to which historical decisions were influenced by root causes; Based on the historical decision set with time-series information obtained in step 6.1, for each historical decision, the complete lineage subgraph of that historical decision is queried based on the optimized lineage graph obtained in step 2, locating the root cause node's position and propagation path within the lineage subgraph. The graph neural network method from step 3 is used to recalculate the contribution weight of the root cause node in the lineage path of that historical decision; the contribution weight reflects the degree of influence of the root cause node on that historical decision. Based on the problem category and risk score of the root cause node in step 5, a problem influence coefficient is defined according to the severity of the problem; the more severe the problem, the larger the influence coefficient. Combining the contribution weight of the root cause node and the problem influence coefficient, the degree of influence score of the historical decision on the root cause is calculated; the influence score represents the risk level of potential bias in the historical decision. The influence score is compared with a preset risk level threshold to classify the historical decision into high-risk, medium-risk, and low-risk levels, obtaining the historical decision impact assessment result.

[0073] Step 6.3: Compile the historical decision impact assessment results and generate a structured historical decision impact assessment report; Based on the historical decision impact assessment results obtained in step 6.2, the total number of historical decisions affected by the root cause, the distribution of the number of decisions at each risk level, the time span of the impact, and the distribution of the types of decision scenarios involved are statistically analyzed. A historical impact statistical report is generated, which displays the distribution of historical decisions affected by the root cause in the form of a time axis, the distribution of risk levels in the form of a pie chart, and a detailed list of high-risk historical decisions in the form of a table, resulting in a structured historical decision impact assessment report.

[0074] Step 6.4: Query the current usage status of the root cause node and identify future risks in active decision-making scenarios; Based on the root cause node set obtained in step 5, the current usage status of the data source or model component involved in the root cause node within the urban intelligent agent system is queried. For data source root cause nodes, based on the optimized lineage graph obtained in step 2, newly generated lineage relationships within the current time window are queried to identify active decision-making links still using the data source, thus obtaining a set of active decision-making scenarios dependent on that data source. For model root cause nodes, the deployment status and call records of the model are queried to identify decision-making processes still using that model version, thus obtaining a set of active decision-making scenarios dependent on that model. For each scenario in the active decision-making scenario set, the decision frequency, decision importance, and scope of impact of that scenario are analyzed, and the potential impact on future decisions of that scenario if the root cause problem persists is assessed, resulting in future decision risk identification results.

[0075] Step 6.5: Construct a temporal graph neural network model and learn the spatiotemporal evolution of blood relations; Based on the historical evolution data of the optimized kinship graph obtained in step 2, multiple time slice snapshots of the kinship graph within a past time window are extracted to analyze the evolutionary patterns of kinship relationships, including the type distribution of newly added nodes, the dependency patterns of newly added edges, and the changing trends of node attributes. A temporal graph neural network model is constructed to model the evolution of the kinship graph. The temporal graph neural network uses graph convolution to capture the neighborhood structure features of nodes in the spatial dimension, and uses recurrent neural networks or temporal convolution to capture the evolutionary patterns of the graph structure in the temporal dimension. The kinship graph of historical time slices is input into the temporal graph neural network for training. The model learns the spatiotemporal evolutionary patterns of kinship relationships, resulting in a trained kinship evolution prediction model.

[0076] Step 6.6: Use a bloodline evolution prediction model to predict the future propagation path and scope of impact of the root cause problem; Based on the kinship evolution prediction model trained in step 6.5, the propagation of root cause problems within future time windows is predicted. The current optimized kinship map obtained in step 2 and the root cause node set information obtained in step 5 are input into the prediction model. The model infers new decision nodes and propagation paths that root cause nodes may influence within future time windows based on learned evolutionary patterns. The prediction model outputs the set of decision nodes that may be affected by the root cause within the future time window and their probability of influence; the probability of influence represents the likelihood of the decision node being affected by the root cause. For the predicted affected decision nodes, the risk level of future decisions is assessed by combining their decision importance and the severity of the root cause problem, resulting in a future decision risk prediction result.

[0077] Step 6.7: Design multiple scenario assumptions to predict risks and provide risk evolution trends under different response measures; Based on the root cause node set obtained in step 5, multiple scenario assumptions are designed for risk prediction, considering the persistence and potential for worsening of the root cause problem. These scenario assumptions include a baseline scenario where the root cause problem maintains its current severity, a pessimistic scenario where the root cause problem gradually worsens, and an optimistic scenario where the root cause problem is partially alleviated. For each scenario, the severity parameter of the root cause node is adjusted, and the lineage evolution prediction model obtained in step 6.5 and the propagation prediction method in step 6.6 are used to predict the scope and risk level of future decisions under that scenario, respectively, yielding multi-scenario future risk prediction results. This multi-scenario analysis provides decision-makers with insights into the risk evolution trends under different response measures.

[0078] Step 6.8: Generate a comprehensive cross-time dimension impact analysis report and assess the severity of the impact; Based on the historical decision impact assessment results obtained in step 6.2, the structured historical decision impact assessment report obtained in step 6.3, the future decision risk identification results obtained in step 6.4, the future decision risk prediction results obtained in step 6.6, and the multi-scenario future risk prediction results obtained in step 6.7, a complete impact analysis report is generated. The impact analysis report includes a description of the impact timeframe, statistics on the number of historical impact decisions, the distribution of historical decision risk levels, a list of high-risk historical decisions, a list of future risk decision scenarios, a risk propagation path prediction diagram, multi-scenario risk evolution curves, and a comprehensive assessment of the severity of the impact, resulting in a structured, cross-time-dimensional impact analysis report.

[0079] In some embodiments, since training a bloodline evolution prediction model requires sufficient historical evolution data, and historical data is insufficient for newly deployed urban intelligent agent systems, transfer learning methods can be employed to utilize evolution data from other similar systems. The aim is to build effective evolution prediction capabilities even with limited data. Specifically, bloodline relationship graph evolution data from other urban or other domain intelligent agent systems are collected as source domain data. A time-series graph neural network model is pre-trained on the source domain data, and the model learns general bloodline evolution patterns. The pre-trained model is then transferred to the target urban intelligent agent system, and the model is fine-tuned using a small amount of historical data from the target system, adjusting the model parameters to adapt to the specific evolutionary patterns of the target system. By reducing the dependence on historical data of the target system through transfer learning, relatively accurate evolution predictions can still be achieved even with insufficient data.

[0080] In some embodiments, since the impact assessment of root cause issues needs to consider the business and social impacts of decisions, a purely technical analysis based on kinship may not be comprehensive enough. A comprehensive impact assessment method incorporating a business impact model can be adopted to provide impact analysis that better meets actual business needs. Specifically, a business impact assessment model for decisions is constructed. This model assesses the business importance and social impact of decisions based on business factors such as the type of decision, the scale of resources involved, the number of people affected, and the urgency of the decision. For historical decisions and future risk decisions influenced by root causes, in addition to calculating the technical impact score, the business impact assessment results of the decisions are queried, and the comprehensive risk priority of the decisions is calculated by combining the technical and business impacts. When prioritizing, both the technical probability of decision deviations caused by root causes and the severity of the business consequences of decision errors are considered, identifying key decisions with high technical risks and significant business impacts, which are highlighted in the impact analysis report.

[0081] Step 7: Extract and integrate audit information from the decision audit report and the cross-time dimension impact analysis report; match the disposal plan and push it for execution; track the disposal effect and continuously optimize it to obtain the decision quality assurance plan; Step 7.1: Extract key audit information; Based on the decision audit report obtained in step 5, key information is extracted, including the identification of root cause nodes, the type of root cause nodes, the category of root cause issues, the risk score of the root cause, the confidence level assessment of the root cause, and details of audit evidence. Based on the structured cross-time dimension impact analysis report obtained in step 6, key information is extracted, including the number of historical impact decisions, the distribution of historical decision risk levels, a list of high-risk historical decisions, a list of future risk decision scenarios, risk propagation path prediction, and multi-scenario risk evolution trends. The above information is integrated to construct a set of key audit elements, which includes three main categories of information: root cause identification, scope of impact, and risk assessment.

[0082] Step 7.2, Matching and handling solutions; Based on the root cause node types and problem categories in the audit key element set obtained in step 7.1, a pre-defined problem handling knowledge base is queried. This knowledge base stores common problem types and their corresponding solution templates. For data quality issues, the knowledge base provides suggestions for data source repair, data collection process optimization, and enhanced data quality monitoring. For model risk issues, the knowledge base provides suggestions for model retraining, enhanced model input validation, and model version rollback. For decision compliance issues, the knowledge base provides suggestions for rule revision, rule review processes, and manual review of decisions. Based on the specific problem characteristics of the root cause node, the most relevant solution template in the knowledge base is matched, and the improvement measures descriptions and implementation steps are extracted to obtain targeted improvement suggestions.

[0083] Based on the problem-solving knowledge base, best practice cases are retrieved. These cases record successful experiences and effectiveness evaluations of similar problems in the past. The best practice cases most similar to the current root cause problem are selected, and information such as the handling methods, implementation cycle, resource input, and effectiveness improvement are extracted to provide reference for improvement suggestions. Key information from the best practice cases is then added to the improvement suggestions to enhance their operability and credibility.

[0084] Step 7.3: Develop an action plan for handling the situation; Based on the severity and scope of impact of root causes in the audit key element set obtained in step 7.1, a priority ranking for issue handling is established. Root causes with high risk scores and a large number of historical decisions affected are assigned the highest priority and marked as urgent handling. Root causes with medium risk scores but high future propagation risk are assigned a high priority and marked as priority handling. Root causes with low risk scores and limited scope of impact are assigned a medium-low priority and marked as planned handling. Based on the priority ranking, a recommended handling time window is established for each root cause: urgent handling issues are recommended to be responded to within 24 hours, priority handling issues within 72 hours, and planned handling issues within one week, resulting in an action plan.

[0085] Step 7.4: Generate a comprehensive audit report; A template-based report generation framework is adopted, integrating the decision path visualization map obtained in step 4, the causal relationship analysis results, the root cause localization results and risk scores obtained in step 5, the structured cross-time dimension impact analysis report obtained in step 6, the improvement suggestions obtained in step 7.2, and the action plan obtained in step 7.3 to generate a structured audit report document. The audit report document includes an execution summary section, which concisely summarizes the basic information of the audit objective decision, the main problems identified, the scope of impact assessment, and the core recommendations, facilitating decision-makers' quick understanding of the audit conclusions. The audit report document includes a detailed analysis section, which elaborates on the audit process and analysis results in a detailed manner, following the logical order of problem discovery, root cause localization, impact assessment, and improvement suggestions, providing supporting evidence and data for each conclusion. The audit report document includes a visualization section, embedding the decision path visualization map, scope of impact statistical charts, risk evolution trend curves, and other visualization content generated in step 4, enhancing the report's intuitiveness and readability. The audit report document includes an appendix section, providing detailed audit data tables, historical decision lists, reference cases, and other supplementary information, supporting in-depth analysis and retrospective queries to obtain a complete decision audit report.

[0086] Step 7.5: Distribute reports to track effectiveness; Based on the complete decision audit report obtained in step 7.4, it is pushed to relevant roles such as decision management personnel, system operation and maintenance personnel, and data governance personnel through the message push service of the city intelligent body system. Different versions of the report are customized according to the recipient's role permissions and focus: decision management personnel receive the management version report focusing on business impact and improvement suggestions; system operation and maintenance personnel receive the technical version report focusing on technical issues and handling measures; and data governance personnel receive the data version report focusing on data quality and data source issues. The report push includes a tracking link for issue handling, allowing recipients to submit handling feedback and execution progress, supporting closed-loop management of audit issues.

[0087] Establish an audit effectiveness evaluation and feedback collection mechanism to track the usage of the audit report obtained in step 7.4 and the effectiveness of problem handling. For root cause problems identified in the audit report, track the implementation of problem handling, and record the implementation time, method, and result of the handling measures. After the problem handling is completed, compare the changes in decision quality indicators before and after the handling to evaluate the effectiveness of the audit and handling. Send feedback questionnaires to audit report users to collect user evaluations on the accuracy of root cause identification, the reasonableness of impact scope assessment, and the feasibility of improvement suggestions, thereby obtaining audit effectiveness feedback data.

[0088] Step 7.6, optimize the performance of the audit model; Based on the audit performance feedback data obtained in step 7.5, the performance and shortcomings of the audit model are analyzed. Performance indicators such as the accuracy, false positive rate, and false negative rate of root cause localization are statistically analyzed to identify weaknesses in the audit model. For false positive cases, the causes are analyzed, which may include overly strict audit rules, inaccurate decision boundaries of the anomaly detection model, or biases in the contribution calculation method. Corresponding model parameters and threshold settings are adjusted based on the causes of false positives. For false negative cases, the characteristics of missed anomaly patterns are analyzed. These false negative cases are used as new training samples to incrementally train the graph neural network model from step 3 and the anomaly detection model from step 5, expanding the model's anomaly recognition capabilities.

[0089] Based on the analysis results, an online learning algorithm is used to continuously optimize the audit model. This algorithm allows the model to learn new samples and patterns during operation. Newly collected decision lineage data, audit feedback data, and problem handling effect data are used as incremental training data to periodically train the graph neural network model from step 3, updating the model parameters to adapt to the evolution of lineage relationships and decision patterns. For the anomaly detection model from step 5, a sliding time window is used to update the normal pattern benchmark, keeping it synchronized with the current system state and avoiding false alarms caused by outdated benchmarks. An adaptive adjustment mechanism is adopted for audit rules and threshold parameters, dynamically optimizing parameter settings based on audit feedback to achieve self-evolution of the audit engine and obtain a decision quality assurance solution.

[0090] Step 7.7: Establish an audit knowledge accumulation mechanism and improve the intelligence level of the audit system; Based on the problem handling knowledge base queried in step 7.2, an audit knowledge accumulation mechanism is established. New problem patterns, effective handling solutions, and typical case experiences identified during each audit are structured and stored in the problem handling knowledge base. Audit experts and system maintenance personnel are regularly organized to review and summarize the accumulated audit cases, extract common patterns and best practices, and update the knowledge base content. Through knowledge accumulation and sharing, the intelligence level of the audit system and the decision-making quality management capabilities of the entire city's intelligent system are improved.

[0091] In some embodiments, due to differences in the focus and reading habits of different decision-makers regarding audit reports, fixed-format reports may not meet personalized needs. Natural language generation technology can be employed to provide customizable report generation services, aiming to improve the relevance and user experience of audit reports. Specifically, a template- and rule-based natural language generation system is constructed. The system automatically generates fluent audit report text based on the content of a set of key audit elements. The report generation system supports various report templates, including concise, detailed, technical, and management versions. Users can configure personalized parameters such as the level of detail, the type of issues to focus on, and the presentation of charts. The report generation system dynamically assembles report chapters based on user configuration, selects appropriate language and visualizations, and generates customized audit reports that meet user needs. For repetitive audit tasks, the system remembers user preferences and automatically applies them to subsequent report generation.

[0092] In some embodiments, since the optimization of the audit model relies on high-quality feedback data, but the quantity and quality of user feedback may be insufficient in actual use, an active learning strategy can be adopted to improve the efficiency of feedback data acquisition. The aim is to maximize the model optimization effect with limited human feedback resources. Specifically, the audit system calculates the confidence level of the root cause localization results for each audit. For uncertain cases with low confidence levels, the system actively requests feedback verification from audit experts. The system prioritizes manually annotating uncertain cases that have the greatest value for model improvement, such as cases near the decision boundary, cases representing novel anomaly patterns, and problem types with high historical false positive rates. Through the active learning strategy, limited human feedback resources are concentrated on the most valuable cases, improving the sample efficiency of model optimization. High-quality feedback cases annotated by experts are assigned higher training weights, accelerating model convergence and performance improvement.

[0093] A decision-making auditing system for urban intelligent agents based on data lineage tracing, such as Figure 3As shown, a method for auditing urban intelligent agent decisions based on data lineage tracing, as described above, includes: The metadata collection module is used to collect metadata from the entire decision-making process of the urban intelligent agent in a hierarchical manner, integrate multi-level metadata, perform multi-dimensional annotation and dependency identification, and obtain a multi-level lineage graph structure. The kinship graph construction module constructs a temporal kinship graph based on a multi-level kinship graph structure and optimizes its storage to obtain an optimized kinship graph and target decision nodes. The contribution assessment module, based on the optimized kinship map, assesses the decision contribution of the target decision nodes and obtains the dynamic decision contribution vector of the elements. The causal analysis module, based on the optimized lineage graph and the dynamic decision contribution vector of elements, reconstructs the critical path of the target decision node, analyzes the causal relationship, and obtains the causal-enhanced decision path structure. The audit analysis module, based on the causal reinforcement decision path structure, performs multi-level audit analysis on the decision-making process, identifies root cause nodes, and obtains a root cause node set; it then combines the audit analysis results and the root cause node set to generate a decision audit report. The impact analysis module, based on the root cause node set, analyzes the historical impact and future risk of root cause issues in the optimized lineage graph, and obtains a cross-time dimension impact analysis report. The quality assurance module is used to extract and integrate audit information from decision audit reports and cross-time dimension impact analysis reports; match disposal plans and push them for execution; track disposal effects and continuously optimize them to obtain decision quality assurance plans.

[0094] In one embodiment of the present invention, a specific example is provided: This study describes the application of intelligent decision-making and auditing in the emergency command center of a megacity during a public health emergency. The city has deployed an urban emergency intelligent system to comprehensively analyze multi-source data and generate decisions on resource allocation and prevention and control measures.

[0095] The city's emergency command center detected an outbreak risk signal in a certain area, and the emergency intelligent agent system automatically triggered a risk assessment and decision generation process. The system collected multi-source information, including population flow data, medical resource distribution data, historical epidemic data, transportation network data, and meteorological data from the area and surrounding regions. After data cleaning, feature engineering, risk prediction model inference, and decision rule engine processing, a comprehensive decision plan was generated, including the delineation of isolation areas, medical resource allocation plans, and traffic control measures. The decision plan recommended designating a certain commercial area and three surrounding communities as key prevention and control areas.

[0096] The key data acquired by the system is shown in Table 1: Table 1: Examples of key data acquired by the system;

[0097] The method of this invention was applied to audit and analyze this decision. The system constructed a decision lineage graph containing 237 nodes and 418 dependency edges, tracing the complete link from 5 data source systems to the final decision output. Graph neural network calculations revealed that population flow data had the greatest impact on the decision to delineate the quarantine area, with a weight of 0.34.

[0098] During the data quality audit, the system discovered that the timestamps of the population flow data showed the data actually came from 18 hours before the decision was triggered, rather than from the 30 minutes recorded by the system, indicating a serious timeliness issue. This timeliness issue caused the population flow characteristics to deviate from the actual situation, resulting in an input distribution drift of 0.67 for the risk prediction model, exceeding the preset threshold of 0.5. The system identified this timeliness issue as the root cause, with a root cause confidence level of 0.89.

[0099] Through cross-time dimension impact analysis, the system found that this data source was used in 312 historical decisions over the past 7 days, of which an estimated 48 high-risk historical decisions may be affected by the timeliness of this data. The system also identified 5 active decision-making scenarios that are still using this data source, and it is estimated that approximately 150 decisions will be at risk in the next 3 days.

[0100] The system-generated audit report clearly identified the root cause of the data timeliness issue, quantified the scope of impact, and provided suggestions for data source restoration and historical decision review. Based on the audit report, the emergency command center immediately corrected the data collection timestamp error, reassessed the affected historical decisions, and prevented the expansion of control measures based on erroneous data, effectively improving the accuracy and credibility of emergency decision-making.

[0101] By applying the method of this invention, the urban emergency intelligent agent system achieves full transparency and traceability in the decision-making process, establishes a sound decision quality assurance mechanism, and provides solid technical support for the intelligentization and credibility of urban emergency management.

[0102] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for auditing urban intelligent agent decisions based on data lineage tracing, characterized in that, Includes the following steps: Metadata from the entire decision-making process of urban intelligent agents is collected in layers, and multi-layered metadata is integrated to perform multi-dimensional annotation and dependency identification, resulting in a multi-layered lineage graph structure. Based on a multi-level kinship graph structure, a temporal kinship graph is constructed and its storage is optimized to obtain an optimized kinship graph and target decision nodes; Based on the optimized kinship map, the decision contribution of the target decision nodes is evaluated to obtain the dynamic decision contribution vector of the elements. Based on the optimized lineage graph and the dynamic decision contribution vector of elements, the critical path of the target decision node is reconstructed, the causal relationship is analyzed, and the decision path structure with enhanced causality is obtained. Based on the decision path structure of causal reinforcement, a multi-level audit analysis is performed on the decision process to identify root cause nodes and obtain a root cause node set; a decision audit report is generated by combining the audit analysis results and the root cause node set. Based on the root cause node set, the historical impact and future risk of root cause problems are analyzed in the optimized lineage graph, resulting in a cross-time dimension impact analysis report; Extract and integrate audit information from decision audit reports and cross-time dimension impact analysis reports; Match the response plan and push it for execution; Track the effectiveness of the measures and continuously optimize them to obtain a decision-making quality assurance plan.

2. The urban intelligent agent decision auditing method based on data lineage tracing according to claim 1, characterized in that, The hierarchical collection of metadata for the entire decision-making process of urban intelligent agents includes: Metadata acquisition agent technology is used to capture information across the entire decision-making process. The acquisition agent is deployed in the data access layer, data processing layer, feature calculation layer, model inference layer, and decision execution layer; and captures metadata of the data access layer, data processing layer, feature calculation layer, model inference layer, and decision execution layer, respectively.

3. The urban intelligent agent decision auditing method based on data lineage tracing according to claim 1, characterized in that, The construction of the temporal kinship map and its storage optimization include: The element nodes in the multi-level kinship graph structure are mapped to vertices in the graph database, and attributes are set for each vertex to obtain the kinship graph vertex set; Map the dependent edges to the edges of the graph database, and set the dependency type, data transfer amount, and transfer delay attributes for each edge to obtain the bloodline graph edge set; An enumeration type index is created for the lineage hierarchy attribute of the vertex, a time range index is created for the timestamp attribute, and a hash index is created for the element identifier attribute, resulting in a multi-dimensional composite index structure; For high-frequency query scenarios, the subgraph structure involved in the path is pre-calculated and stored in the key-value caching system to obtain the path query cache.

4. The urban intelligent agent decision auditing method based on data lineage tracing according to claim 1, characterized in that, The assessment of the contribution to decision-making includes: Extract decision node identifiers from the optimized kinship graph, perform reverse graph traversal query, and construct a decision kinship subgraph; Multidimensional feature extraction is performed on each node in the decision lineage subgraph to obtain the initial feature vector of the node; A graph neural network model is constructed to encode and learn the decision lineage subgraph. The node representation is updated by aggregating the neighborhood features of the nodes through multi-layer graph convolution operations, resulting in a set of node embedding vectors. An attention mechanism is used to calculate the attention weight of the element node to the decision node, and the dot product similarity between the embedding vectors of the decision node and the element node is calculated. After softmax normalization, the dynamic decision contribution vector of the element is obtained.

5. The urban intelligent agent decision auditing method based on data lineage tracing according to claim 1, characterized in that, The critical path of the target decision node is restored, and the causal relationship is analyzed, including: Extract the target decision node identifier from the optimized kinship graph, and execute the breadth-first reverse traversal algorithm to obtain the set of decision-dependent nodes and the set of dependencies. Based on the dynamic decision contribution vector of elements, the decision-dependent nodes are sorted in descending order of contribution value, and the nodes with contribution values ​​exceeding a preset threshold are retained as key nodes, thus obtaining the set of key decision path nodes. For adjacent nodes in the critical decision path, the conditional mutual information between nodes is calculated. Node pairs with conditional mutual information greater than a significance threshold are identified as having a causal relationship. Counterfactual reasoning is performed on key element nodes to construct counterfactual scenarios, simulate the propagation process of data changes through kinship paths, assess the impact of data changes on decision outcomes, and obtain a causal-enhanced decision path structure.

6. The urban intelligent agent decision auditing method based on data lineage tracing according to claim 1, characterized in that, The multi-level audit analysis of the decision-making process, including the identification of root cause nodes, includes: For data source nodes in the causal reinforcement decision path structure, perform data quality audit checks, calculate data timeliness indicators, integrity indicators, and consistency indicators, compare the timeliness indicators, integrity indicators, and consistency indicators with preset thresholds, identify data source nodes with quality problems, and calculate risk scores. For model nodes, perform model running status audit checks, calculate the divergence value between the input feature distribution and the benchmark distribution to represent the distribution drift, extract inference confidence, verify version compliance, and calculate model risk score; For decision-making rule nodes, conduct decision compliance audits, check triggering conditions based on the business rule rationality knowledge base, check execution actions based on the business specification knowledge base, and calculate compliance risk scores; A multi-source evidence fusion algorithm is used to identify root cause nodes. The number of audit issues, risk scores, decision contribution, and causal effect assessment values ​​of the nodes are comprehensively considered. Nodes with confidence scores exceeding a preset threshold are selected to obtain the root cause node set.

7. The urban intelligent agent decision auditing method based on data lineage tracing according to claim 1, characterized in that, The analysis of the historical impact and future risks of root cause problems includes: For root cause nodes, a forward graph traversal query is performed based on the optimized lineage graph to filter historical decision nodes and obtain the set of historical decision nodes affected by the root cause. Recalculate the contribution weight of the root cause node, combine the contribution weight with the problem impact coefficient, and calculate the score of the degree to which historical decisions are affected. Query the current usage status of the components involved in the root cause node, identify active decision-making links, and assess the potential impact on future decisions; A time-series graph neural network model is constructed to predict the future impact range and propagation path of root cause nodes, resulting in a cross-time dimension impact analysis report.

8. A method for auditing urban intelligent agent decisions based on data lineage tracing as described in claim 1, characterized in that, The extraction and integration of audit information from decision audit reports and cross-time dimension impact analysis reports includes: Based on the decision audit report, extract the root cause node's identifier, type, problem category, risk score, and confidence level assessment; Based on the cross-time dimension impact analysis report, extract the number of historical impact decisions, risk level distribution, high-risk decision list, and future risk scenario list to construct a set of key audit elements.

9. A method for auditing urban intelligent agent decisions based on data lineage tracing as described in claim 1, characterized in that, The tracking and continuous optimization of the treatment effect includes: Establish an audit effectiveness evaluation mechanism to track the implementation of root cause problem handling, compare changes in decision quality indicators before and after handling, and evaluate the effectiveness of auditing and handling. Based on the audit performance feedback data, the accuracy, false positive rate, and false negative rate of root cause identification were statistically analyzed, and the model parameters and threshold settings were adjusted, with false negative cases used as new training samples. Online learning algorithms are used to continuously optimize the audit model, and the graph neural network model is incrementally trained regularly to dynamically optimize parameter settings, thereby obtaining a decision quality assurance solution.

10. A city intelligent agent decision-making auditing system based on data lineage tracing, characterized in that, A method for implementing a city intelligent agent decision auditing method based on data lineage tracing as described in any one of claims 1-9, comprising: The metadata collection module is used to collect metadata from the entire decision-making process of the urban intelligent agent in a hierarchical manner, integrate multi-level metadata, perform multi-dimensional annotation and dependency identification, and obtain a multi-level lineage graph structure. The kinship graph construction module constructs a temporal kinship graph based on a multi-level kinship graph structure and optimizes its storage to obtain an optimized kinship graph and target decision nodes. The contribution assessment module, based on the optimized kinship map, assesses the decision contribution of the target decision nodes and obtains the dynamic decision contribution vector of the elements. The causal analysis module, based on the optimized pedigree graph and the dynamic decision contribution vector of elements, reconstructs the critical path of the target decision node, analyzes the causal relationship, and obtains the causal-enhanced decision path structure. The audit analysis module, based on the causal reinforcement decision path structure, performs multi-level audit analysis on the decision-making process, identifies root cause nodes, and obtains a root cause node set; it then combines the audit analysis results and the root cause node set to generate a decision audit report. The impact analysis module, based on the root cause node set, analyzes the historical impact and future risk of root cause issues in the optimized lineage graph, and obtains a cross-time dimension impact analysis report. The quality assurance module is used to extract and integrate audit information from decision audit reports and cross-time dimension impact analysis reports; match disposal plans and push them for execution; track disposal effects and continuously optimize them to obtain decision quality assurance plans.

Images