Big data operation supervision platform based on smart government affairs

Abstract

According to the big data operation supervision platform based on intelligent government affairs, a semantic vector set capable of expressing a policy goal is established, and a unified government affair behavior life cycle diagram is constructed, so that structured mapping between the policy goal and an operation behavior is realized; through fusion of semantic similarity and flow structure consistency calculation and in combination with a deviation punishment mechanism, a deviation path which does not conform to policy intention, flow specifications or compliance requirements is accurately identified, and structured labeling is carried out; for the identified deviation path, introducing a causal analysis method of laws and regulations priority, cross-department interaction strength and policy association constraint, constructing a causal chain of deviation propagation and positioning an initial responsibility node, and ensuring that the responsibility chain has interpretability and reproducibility; and finally, integrating the causal chain and responsibility node information into a unified abnormal event object, calculating a risk priority, and generating a supervision action which can be directly executed in combination with a compliance feasibility condition, thereby realizing closed-loop management from problem discovery and cause analysis to execution and disposal.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of big data supervision, and particularly relates to a big data operation supervision platform based on smart government affairs. BACKGROUND

[0002] With the continuous improvement of the informatization level of government affairs, government departments at all levels have generally established information platforms covering administrative approval, public services, social governance, financial management and other business lines, and have built a relatively complete e-government platform on this basis.

[0003] However, the rapid growth of the number of platforms, the continuous expansion of the data size, and the complication of the business processes make the data flow, behavior compliance and cross-departmental collaborative supervision in the operation process of the government affairs platform more prominent. Most of the current government affairs platforms have data collection and transaction processing capabilities, but the operation process supervision still mainly relies on fixed rules or static process models, which are difficult to adapt to frequent policy adjustments and diversified operation modes, and cannot effectively identify the behavior paths deviating from the policy goals. At the same time, government data types are highly heterogeneous, including structured tables, unstructured texts and multi-source logs, and existing technologies mostly use simple correlation analysis or log backtracking, lacking the ability to model and semantically align the whole life cycle of cross-platform and cross-departmental data. In addition, existing anomaly detection mostly stays in single-index threshold alarm, which is difficult to identify implicit problems caused by multi-platform and multi-process interaction, and lacks a mechanism to associate deviated results with responsibility chains, so as to provide clear basis for accountability and rectification.

[0004] Although some platforms have introduced intelligent monitoring components, they often only do statistical analysis on existing data, lacking structured mapping and closed-loop management from policy intent to actual operation behavior, resulting in obvious short boards in supervision capability in the dynamic and complex government operation environment. SUMMARY

[0005] The application proposes a big data operation supervision platform for smart government affairs, which can realize whole-link modeling and closed-loop supervision from policy goals to operation behavior in a dynamic environment.

[0006] In order to achieve the above purpose, the application provides a big data operation supervision platform based on smart government affairs, which comprises:

[0007] A data collection and processing module is used to collect policy text data, government operation logs and process definition files, construct a policy goal semantic vector set and a government behavior life cycle graph after processing, and send them to a semantic deviation path identification and labeling module and a deviated behavior cause-effect chain identification module.

[0008] The semantic deviation path identification and labeling module is configured to identify and label deviation paths existing in platform operation according to a policy target semantic vector set and a government affair behavior life cycle graph, introduce a compliance correction item, comprehensively determine the compliance of a path to a policy applicable condition of a legal process and a target population, obtain a deviation path set, and send the deviation path set to the deviation behavior causal chain identification module.

[0009] The deviation behavior causal chain identification module is configured to construct a causal chain reflecting a deviation propagation process according to the deviation path set and global structure information of the government affair behavior life cycle graph, extract an initial responsibility node, obtain a deviation causal chain set by summarizing an event sequence from the initial responsibility node to a deviation result node, record operation details, cross-department participation information and matching conditions with a policy target of each node on the path, form a complete responsibility path modeling result, and send the deviation causal chain set and the initial responsibility node information to the executable supervision action output module.

[0010] The executable supervision action output module is configured to convert an initial responsibility node set obtained by summarizing the deviation causal chain set and the initial responsibility node of each deviation path into a structured abnormal event object that can be directly executed, introduce a compliance feasibility mask, and generate a supervision action according to the compliance feasibility mask.

[0011] Preferably, in the construction process of the policy target semantic vector, the policy text data is obtained by sentence segmentation, word segmentation, word table mapping and context encoding, and the context vector of each policy clause is mapped to a fixed dimension.

[0012] Preferably, in the construction process of the government affair behavior life cycle graph, the government operation log and the process definition file are aligned by event and process node and dependency relationship to generate an aligned event node set and an edge set determined by time sequence and legal dependency.

[0013] Preferably, in the semantic deviation path identification and labeling module, all candidate paths are enumerated according to start and end nodes in the government affair behavior life cycle graph, the enumeration of the paths follows topological sorting, which is configured to ensure that the node order is consistent with time and the legal process, the operation type and object field of the node sequence of each path are spliced into a path description text according to time, and the path description text is input into the same government field Transformer encoder as the policy text data to obtain a path semantic vector.

[0014] As preferred, in the semantic deviation path identification and labeling module, the deviation degree calculation comprises a semantic similarity based on cosine similarity and a structure consistency weight, the structure consistency weight comprises a node coverage rate of the path, a key node integrity, and whether an applicable condition of a specific policy target is met, and a compliance correction term is introduced to amplify the influence of skipping mandatory nodes and key violation behaviors of unauthorized access.

[0015] As preferred, in the construction process of the causal chain, a causal analysis subgraph is established based on each deviation path and a corresponding government affair behavior life cycle graph, the causal analysis subgraph comprises nodes and edges e ij , the nodes comprise specific government operation events, and the government operation events comprise an operator ID, a timestamp, an operation type, and an object; the edges e ij are used to represent a direct dependency relationship between events v i and v j .

[0016] More preferably, in the construction process of the causal chain, a node with the maximum penalty term and the earliest time is selected from each deviation path as a candidate starting point, and a maximum comprehensive score path search is performed in the government affair behavior life cycle graph in combination with a comprehensive causal score value, so as to ensure that the path covers main trigger events and meets regulatory and policy constraints; a path robustness coefficient is further introduced, the path robustness coefficient is calculated based on structural similarity among multiple possible causal chains of the same deviation result, and the path robustness coefficient is also used as a regularization term of a final score of the path.

[0017] As preferred, in the executable regulatory action output module, a risk priority of each causal chain is calculated, the risk priority comprises a mean value of causal strength in the chain, a policy severity score, and a cross-department propagation range score, and a structured abnormal event object is generated based on the risk priority.

[0018] More preferably, the structured abnormal event object adopts a unified Schema configuration, and comprises an event identifier, a policy association, a behavior chain segment, responsibility information, a risk assessment, and a recommended action.

[0019] As preferred, the compliance feasibility mask is used to make a constraint adjustment on the risk to avoid producing an unexecutable or unauthorized action, and specifically, when a node where an initial responsibility node is located has a process suspension permission, the deviation causal chain does not touch an interruptible link, and a platform capability required by the action exists, the compliance feasibility mask takes a higher value; if any condition is not met, the compliance feasibility mask is reduced according to a rule; and the compliance feasibility mask takes a value in an interval of [0, 1].

[0020] The present application has at least the following beneficial technical effects:

[0021] The platform of the present application firstly establishes a semantic vector set that can express policy objectives, and constructs a unified government affair behavior life cycle graph, standardizes the abstraction of operation records, process calls and data interactions across platforms and departments, and realizes the structured mapping of policy objectives and running behaviors. On this basis, through the fusion of semantic similarity and process structure consistency calculation, combined with the deviation punishment mechanism, the deviated paths that do not conform to the policy intention, process specification or compliance requirements are accurately identified and structuredly labeled. For the identified deviated paths, the platform uses the causal analysis method of introducing the priority of regulations, the interaction intensity across departments and the policy correlation constraints to construct the causal chain of deviation propagation and locate the initial responsibility node, so as to ensure that the responsibility chain has explainability and reproducibility. Finally, the platform integrates the causal chain and the responsibility node information into a unified abnormal event object, calculates the risk priority, and generates executable supervision actions combined with the compliance feasibility conditions, so as to realize the closed-loop management from problem discovery, cause analysis to execution disposal. The scheme not only solves the problems that the traditional platform cannot adapt to policy changes and cannot accurately detect and trace the deviated behaviors, but also implements the analysis results to the actual supervision through the output of executable actions, which significantly improves the real-time performance, accuracy and operability of the operation supervision of smart government affairs. BRIEF DESCRIPTION OF DRAWINGS

[0022] The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled in the art, other drawings can be obtained without creative labor on the basis of the following drawings.

[0023] Figure 1 The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled in the art, other drawings can be obtained without creative labor on the basis of the following drawings. DETAILED DESCRIPTION

[0024] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation on the present application.

[0025] In one or more embodiments, as shown in Figure 1 a smart government affair-based big data operation supervision platform is disclosed, which comprises the following:

[0026] A data acquisition and processing module is used to acquire policy text data, government affair operation logs and process definition files, construct a policy objective semantic vector set and a government affair behavior life cycle graph after processing, and send them to a semantic deviation path identification and labeling module and a deviated behavior causal chain identification module;

[0027] In the big data operation supervision platform of smart government affairs, the policy target semantic vector set P and the government affair behavior life cycle graph G are established. These two are the core inputs of the whole platform, which bear the functions of "target constraint" and "behavior structure" respectively. The policy target vector set is used to express the operation requirements, compliance conditions and target orientation in the policy document, while the life cycle graph is used to describe the sequence and dependency relationship of the actual government operation events. In order to ensure that the output of this step can directly enter the subsequent analysis link, it is necessary to design special data collection, analysis and modeling process according to the actual data characteristics of government affairs scene.

[0028] Data collection is divided into three types of data sources:

[0029] First, the policy text data T p , from the official release channel of the government, such as provincial government service network, department regulations database or internal document platform. These data may exist in the form of PDF, Word, HTML or database records. In actual operation, it can be directly batch pulled through the policy document interface of government informationization platform, or the text analysis service can be used to extract the text from the existing files.

[0030] Second, the government operation log L r , usually stored in the database or log server of each business platform, records the events of approval, handling, data access, etc. For example, the approval platform may store table records in MySQL, each record contains applicant ID, approver ID, operation timestamp, operation type and operation object, etc. The log file of data sharing platform may record the time, request source and data category of each cross-department data exchange in JSON format. The log can be pulled through the secure API, or it can be exported regularly through the internal network batch processing task.

[0031] Finally, the process definition file F d , maintained by the government business process management (BPM) platform, describes the legal process steps and sequence.

[0032] For example, approving a construction project needs to go through four legal nodes of "acceptance → review → approval → archiving". These nodes and their dependency relationships are stored in the BPM platform in XML or BPMN format, which can be exported through the platform interface.

[0033] First, the policy text T pPreprocessing is performed, including removing format tags, sentence segmentation, and word segmentation. Proper nouns (such as department names and business terms) are mapped to a unified vocabulary to ensure lexical consistency in subsequent encoding stages. The processed text is then fed into a multi-layered Transformer encoding structure fine-tuned for the government domain. This structure is pre-trained on a general corpus and retrained using a large corpus of government documents to better capture the constraints and object descriptions in policies. During the encoding stage, each policy clause is independently fed into the encoder, and the output is a sequence of context vectors. Average pooling is used to obtain a fixed-dimensional representation of the clause. This results in P = {p1, p2, ..., p...} m In}, each p i This represents a policy objective, such as "approval must be completed within 5 working days" or "priority services for low-income groups." For example, in actual testing, a policy document regarding public housing approval, after going through this process, will generate several vectors, each carrying information such as time limits, target groups, and approval stages.

[0034] Secondly, regarding the runtime log L r The process involves parsing each record and mapping it to an event quadruple (u, t, a, o), where u is the user or platform account ID, t is the operation timestamp, a is the operation type (e.g., submit, review, modify, query), and o is the operation object (e.g., a case number or table name). These log events need to be correlated with the process definition file F. d Alignment, the alignment process includes two parts: first, matching events by node name and node type to assign them to the corresponding process nodes; second, using timestamps and F... d Directed edges are generated between events in the order of nodes in the lifecycle graph. For example, if the process definition file specifies that the "Acceptance" node must precede the "Review" node, and the corresponding "Acceptance" event time in the log is earlier than the "Review" event time, then an edge from the "Acceptance" node to the "Review" node is created in the lifecycle graph. In the resulting G=(V,E), V is the set of all actually occurring event nodes, and E is the set of sequence and dependency relationships between events. To avoid cross-platform time synchronization issues, all timestamps are uniformly timezoned and offset-corrected before construction.

[0035] The formula section is controlled into two core formulas:

[0036] P = Pool(Encoder(T) p )) (1) The Encoder uses a multi-layer Transformer encoding structure, with the processed policy text T as input. p Pool is an average pooling operation that maps the context vector of each policy clause to a fixed dimension; p iThe single policy clause is derived from the sentence segmentation, word segmentation, word table mapping and context encoding. The policy target semantic vector set P includes the semantic representation of multiple policy targets, which is directly used for subsequent alignment of behaviors.

[0037] G = (V, E) = Align (L r , F d ) (2) Align represents the alignment and dependency generation of events and process nodes, and the input is the running log L r and the process definition file F d ; V is derived from the aligned event node set; E is derived from the edge set jointly determined by the time sequence and the legal dependency. The government behavior life cycle graph G includes the actual running event nodes and their order relationship, which is used as the running track structure for subsequent analysis and tracing.

[0038] The data acquisition and processing module obtains the policy target semantic vector set P and the government behavior life cycle graph G, and sends them to the semantic deviation path identification and labeling module and the deviation behavior causal chain identification module.

[0039] The semantic deviation path identification and labeling module is used to identify and label the deviation paths existing in the platform running according to the policy target semantic vector set and the government behavior life cycle graph, and introduce compliance correction items, and comprehensively analyze the compliance of the path to the applicable conditions of the legal process and target population, to obtain the deviation path set, and send the deviation path set to the deviation behavior causal chain identification module.

[0040] In the smart government running supervision platform, the semantic deviation path identification and labeling module uses the policy target semantic vector set P and the government behavior life cycle graph G sent by the data acquisition and processing module to identify and label the deviation paths existing in the running. Different from conventional text matching, the semantic deviation path identification and labeling module integrates semantic alignment and process structure consistency analysis, and introduces specific compliance correction items for the government scene, so that the deviation detection not only considers the language similarity, but also comprehensively considers the compliance of the path to the applicable conditions of the legal process, target population and other policies, thereby reducing false positives and enhancing explainability.

[0041] In the policy target semantic vector set P, p i represents a fixed dimension vector of a single policy point, which is derived from the output of the government field Transformer encoder.

[0042] In the government behavior life cycle graph G = (V, E) constructed in the data acquisition and processing module, V is the event node set, and each node v jRecord the operator ID, timestamp, operation type, and operation object; E is the event dependency set, derived from the alignment of the process definition file and the running log.

[0043] First, enumerate all candidate paths r from the government affair behavior life cycle graph G by start and end nodes k . The enumeration of the path follows the topological sorting, which ensures that the node order is consistent with the time and legal process. The operation type and object fields of the node sequence of each path are spliced into a path description text by time, and input into the same government domain Transformer encoder in the data collection and processing module to obtain the path semantic vector B k . This ensures that B k and p i are in the same semantic space, which facilitates subsequent calculations.

[0044] In order to improve the detection of pertinence, the deviation calculation not only depends on Sim(p i , B k ) (semantic similarity based on cosine similarity), but also introduces a structural consistency weight ω k . ω k combines the node coverage rate of the path, the integrity of the key node, and whether it meets the applicable conditions of a specific policy goal (such as target population label, business category matching, etc.), and adds a compliance correction term λ·Penalty(r k ) in the calculation, which is used to amplify the impact of skipping mandatory nodes, unauthorized access, and other key violations.

[0045] δ ik = 1 - [Sim(p i , B k ) × (ω k - λ·Penalty(r k ))] (3)

[0046] Sim(p i , B k ) is a dimensionless ratio, if the semantic similarity and the dimension of the structure weight are different, then it has been normalized; ω k is derived from the structural matching calculation of the life cycle graph; Penalty(r k ) outputs a value between 0 and 1 based on the detection rules of key node missing and permission violation in the path; λ is the adjustment coefficient, which is obtained by historical case tuning; δ ik represents the deviation of the path r k from the policy p i , the larger the value, the more obvious the deviation.

[0047] Finally, compare δ ik with the threshold θ, if δ ikIf the value exceeds θ, it is determined to be a deviation from the path, and a deviation label is attached to the path object, including the deviation policy number, node sequence, and deviation type (such as timeout, unauthorized access, or missing approval step). This labeled data structure will be directly input into the deviation behavior causal chain identification module without repeated parsing.

[0048] The deviation behavior causal chain identification module is used to construct a causal chain reflecting the deviation propagation process based on the global structure information of the deviation path set and the government behavior lifecycle diagram, extract the initial responsibility node, summarize the event sequence from the initial responsibility node to the deviation result node to obtain the deviation causal chain set; and record the operation details, cross-departmental participation information and matching with policy objectives of each node on the path to form a complete responsibility path modeling result, and send the deviation causal chain set and the initial responsibility node information to the executable regulatory action output module.

[0049] The deviation behavior causal chain identification module is the set of deviation paths {r} identified in the semantic offset path identification and annotation module. k Based on this, and utilizing the global structural information of the lifecycle graph G, a causal chain C that accurately reflects the deviation propagation process is constructed. k And extract the initial responsible node v k1 .

[0050] Unlike general causal path analysis, the deviation behavior causal chain identification module not only considers the temporal and structural dependencies of events, but also introduces regulatory priority factors, cross-departmental interaction intensity correction terms, and policy correlation constraints specific to government scenarios, thereby forming a responsibility path model that is more in line with the actual operation characteristics of smart government. The result of this process directly determines whether the regulatory platform can provide a feasible basis for subsequent enforcement and accountability, and is therefore a crucial link in the entire patent solution.

[0051] First, on each deviation path r k The corresponding lifecycle subgraph G k A causal analysis subgraph is created within the graph. Node v i This indicates a specific government operation event, including the operator ID, timestamp, operation type, and target; edge e ij Indicates event v i With v j The direct dependencies between them. To reflect the special characteristics of government affairs scenarios, three specific modifications are introduced in the edge weight calculation: the regulatory priority factor φ. ij (Reflecting the importance of edge-connected nodes in regulations), cross-departmental interaction strength ψ ij (Reflecting the frequency and sensitivity of cross-departmental data exchange between events), and policy-related constraints ρ ij(reflecting whether the nodes connected by the edge are directly involved in the policy goal clause corresponding to the semantic deviation path). These factors can effectively filter out minor events with little policy impact, focusing on truly critical causal relationships.

[0052] In calculating the comprehensive causal score, the formula is as follows:

[0053] S ij = a w ij + b f ij + g p ij + h r ij - l R ij (4)

[0054] where w ij represents the basic dependency weight calculated based on the time interval and historical execution frequency in the life cycle graph; f ij is derived from the regulation level assessment of the nodes in the process definition file; p ij is derived from the cross-department data call frequency and sensitivity score recorded in the running log; r ij is determined according to the association analysis of the deviation label and the policy vector set P identified and labeled by the semantic deviation path identification and labeling module, and a higher value is given if the edge is directly related to the relevant policy clause; R ij is a regularization penalty term used to suppress redundant edges in the policy chain that have no substantial policy association, preventing the causal chain from being lengthened by noise events. a, b, g, h, and l are weight coefficients, which are obtained by historical case backtracking optimization.

[0055] In generating the causal chain, the node with the largest penalty term and the earliest time is selected from r k as the candidate starting point, and the S ij value is used to perform maximum comprehensive score path search in G k , ensuring that the path covers the main trigger event and meets the regulatory and policy constraints. To ensure the stability of the path, a path robustness coefficient s(C k ) is also introduced, which is calculated based on the structural similarity between multiple possible causal chains for the same deviation result, and is used as a regularization term for the final path score:

[0056]

[0057] where s(C k ) comes from the matching degree of different candidate chains in terms of node coverage and structural sequence, and m is an adjustment coefficient to prevent analysis bias caused by dependence on a single path.

[0058] The event sequence from the initial responsibility node to the deviation result node is summarized to obtain a deviation causal chain set, where the deviation causal chain set {C k} : the sequence of events from the initial responsibility node to the deviated result node, containing the integrated causal score of each edge and meta information.

[0059] In determining the final causal chain C k , its first node is marked as the initial responsibility node v k1 , and the operation details of each node on the path, cross-departmental participation information, and matching with policy objectives are recorded to form a complete responsibility path modeling result.

[0060] The initial responsibility node set is then obtained by summarizing the initial responsibility nodes, and the initial responsibility node set {v k1} : each node is explicitly identified as the first event triggering the deviation, and the operator, time, and policy association information are recorded.

[0061] The executable regulatory action output module is used to convert the set of deviated causal chains and the initial responsibility node set obtained by summarizing the initial responsibility nodes of each deviated path into directly executable structured abnormal event objects, introduce compliance feasibility masks, and generate regulatory actions accordingly;

[0062] The executable regulatory action output module converts the set of deviated causal chains {C k} calculated by the deviated behavior causal chain identification module and the initial responsibility node set {v k1} into directly executable structured abnormal event objects, and generates actionable regulatory actions accordingly. Subsequent modeling and training are no longer performed, and only event generation and action arrangement are performed: in a unified data structure, integrate causal chain scores, policy associations, and cross-departmental information, calculate risk priorities, and output executable disposal instructions under compliance and feasibility constraints. Considering the rigid process and hierarchical permissions of smart government, the executable regulatory action output module adjusts the risk by introducing a "compliance feasibility mask" to avoid producing unexecutable or unauthorized actions.

[0063] First, calculate the risk priority R k of each causal chain. R k is composed of three parts: the average causal strength within the chain (calculated directly from the S k of each edge in C ij ); the policy severity score L k (mapped from the policy target level associated with C k , which can produce a single value according to the "highest level first" or "appearance frequency weighted" rule and normalized); and the cross-departmental spread range score D k (calculated from the number of departments involved in C kThe ratio of the number of departments involved to the maximum number of departments allowed for the matter, both given in the link meta-information of the deviation behavior causal chain identification module. To ensure dimensional consistency, the above three items have been normalized before being synthesized. A fixed coefficient vector is used for linear synthesis to obtain R k :

[0064]

[0065] The edge score S k in C ij is calculated directly; L k is mapped and normalized from the policy target level table associated with C k ; D k is obtained from the ratio of the size of the department set in the meta-information of C k to the maximum number of departments that can be involved in the matter; w is a fixed coefficient vector calibrated offline. If the three quantities are not dimensionally consistent, normalization has been completed before entering the synthesis. Example: a construction permit deviation chain C k contains 6 edges, ∑S ij = 3.6, with an average of 0.60; the associated policy is high level, L k = 0.85; 2 departments are involved, with a maximum of 5, D k = 0.40; taking R k = 0.5 × 0.60 + 0.3 × 0.85 + 0.2 × 0.40 = 0.635.

[0066] R k is then adjusted for compliance and feasibility constraints to obtain the "executable risk". A compliance and feasibility mask m k ∈ 0,1] is constructed: when the node v k1 has process suspension authority, C k does not touch the non-interruptable link, and the platform capabilities required for the action exist, m k takes a higher value; if any of the conditions are not met, m k is reduced according to the rules. Condition determination comes from the node type, permission level, platform capability identification, etc. of C k and v k1 , without adding external inputs. The adjustment method is as follows:

[0067] R k ← R k × m k (7)

[0068] where m k is obtained from the three types of checks: permission matching, link interruptability, and platform capability availability, and is normalized to ensure dimensional consistency with Rk After multiplication, the dimensionless consistency is still maintained. Example: if the aforementioned R k = 0.635, and v k1 is located at a suspendable node but the link contains an interrupt-limited archival link (there is an alternative "read-only + manual review" capability), the rule gives m k = 0.8, then the adjusted R k = 0.508.

[0069] Based on the adjusted R k , a structured abnormal event object ε k is generated. The event object adopts a unified Schema, including: event identification (ID, time), policy association (policy number and abstract from C k ), behavior chain segment (node sequence and each S k from C ij ), responsibility information (account, department, node type from v k1 ), risk assessment (adjusted R k and level label), and suggested action (action identification and parameters). The suggested action is generated through deterministic rules: when R k falls into the highest level and v k1 node has process control capability, "suspend the current instance + transfer to manual review" is output, and the parameters are directly filled from the metadata of C k and v k1 (such as process instance ID, list of affected interfaces, notification object); when R k is medium level and the link contains a flow-limited data interface, "interface flow limiting + time limit review" is output; when R k is low level, "record and warning" is output. The rule also checks the permission and feasibility label, and if it is not satisfied, it is automatically downgraded to the next actionable action. Repetition instance: the R k of the cross-department unauthorized query chain C k is adjusted to 0.52, and the link contains a data sharing gateway node with flow limiting capability; the rule engine writes "limit flow for 24 hours + notify the data belonging department to review" in ε k , and the parameters (interface ID, time length, notification list) are all from the metadata fields of C k and v k1 .

[0070] The above describes the preferred embodiments of the present application. It should be noted that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements are also considered within the protection scope of the present application.

Claims

1. A big data operation supervision platform based on smart government, characterized in that, The application comprises: a data acquisition and processing module for constructing a policy target semantic vector set and a government affair behavior life cycle graph by acquiring policy text data, government affair operation logs and process definition files, and sending them to a semantic deviation path identification and labeling module and a deviation behavior causal chain identification module after processing; the semantic deviation path identification and labeling module for identifying and labeling deviation paths existing in platform operation according to the policy target semantic vector set and the government affair behavior life cycle graph, introducing compliance correction items, comprehensively determining the compliance of the paths to the applicable conditions of the legal processes and target groups, obtaining a deviation path set, and sending the deviation path set to the deviation behavior causal chain identification module; the deviation behavior causal chain identification module for constructing a causal chain reflecting the deviation propagation process according to the deviation path set and the global structure information of the government affair behavior life cycle graph, extracting an initial responsibility node, and obtaining a deviation causal chain set by summarizing the event sequence from the initial responsibility node to a deviation result node; recording the operation details, cross-department participation information and matching with policy targets of each node on the path to form a complete responsibility path modeling result, and sending the deviation causal chain set and the initial responsibility node information to an executable supervision action output module; the executable supervision action output module for converting the initial responsibility node set obtained by summarizing the deviation causal chain set and the initial responsibility node of each deviation path into a structured abnormal event object that can be directly executed, introducing a compliance feasibility mask and generating a supervision action accordingly.

2. The big data operation supervision platform based on smart government according to claim 1, characterized in that, In the construction process of the policy target semantic vector, the policy text data is obtained by sentence segmentation, word segmentation, word table mapping and context encoding, and the context vector of each policy clause is mapped to a fixed dimension.

3. The big data operation supervision platform based on smart government according to claim 1, characterized in that, In the construction process of the government affair behavior life cycle graph, the government affair operation logs and process definition files are aligned by event and process node and the aligned event node set and the edge set determined by time sequence and legal dependence are generated.

4. The big data operation supervision platform based on smart government affairs according to claim 1, characterized in that, In the semantic deviation path identification and labeling module, all candidate paths are enumerated in the government affair behavior life cycle graph according to start and end nodes, and the enumeration of the paths follows topological sorting to ensure that the node order is consistent with the time and legal process. The operation type and object field of the node sequence of each path are spliced into a path description text according to time, and input into the same government field Transformer encoder as the policy text data to obtain a path semantic vector.

5. The big data operation supervision platform based on smart government affairs according to claim 1, characterized in that, In the semantic deviation path identification and labeling module, the deviation degree calculation includes semantic similarity based on cosine similarity and structure consistency weight, the structure consistency weight includes node coverage rate, key node integrity and whether it meets the applicable conditions of a specific policy target, and a compliance correction item is introduced to amplify the impact of skipping mandatory nodes and unauthorized access to key violation behaviors.

6. The big data operation supervision platform based on smart government affairs according to claim 1, characterized in that, In the construction process of the causal chain, a causal analysis subgraph is established based on each deviation path and a corresponding government affair behavior life cycle graph, the causal analysis subgraph includes nodes and edges e ij , the nodes include specific government operation events, and the government operation events include an operator ID, a timestamp, an operation type and an object; the edges e ij are used to represent a direct dependency relationship between events v i and v j .

7. The big data operation supervision platform based on smart government according to claim 6, characterized in that, In the construction of the causal chain, the node with the maximum penalty term and the earliest time is selected from each deviated path as a candidate starting point, and a maximum comprehensive score path search is performed in the government behavior life cycle graph combined with the comprehensive causal score value, to ensure that the path covers the main trigger event and meets the regulatory and policy constraints; A path robustness coefficient is also introduced, which is calculated based on the structural similarity between multiple possible causal chains of the same deviated result, and the path robustness coefficient is also used as a regularization term for the final path score.

8. The big data operation supervision platform based on smart government according to claim 1, characterized in that, In the executable regulatory action output module, the risk priority of each causal chain is calculated, including the average causal strength in the chain, the policy severity score, and the cross-department propagation range score, and a structured abnormal event object is generated based on the risk priority.

9. The big data operation supervision platform based on smart government according to claim 8, characterized in that, The structured abnormal event object is configured with a unified Schema, including event identification, policy association, behavior chain segment, responsibility information, risk assessment, and recommended action.

10. The big data operation supervision platform based on smart government affairs according to claim 1, characterized in that, The compliance feasibility mask is used to make a constraint adjustment on the risk to avoid generating unexecutable or unauthorized actions. Specifically, the compliance feasibility mask takes a higher value when the node where the initial responsibility node is located has the process suspension authority, the deviated causal chain does not touch the non-interruptable link, and the platform capability required by the action exists. If any condition is not met, the compliance feasibility mask is reduced according to the rules. The value range of the compliance feasibility mask is [0, 1].

Images