A knowledge graph-based fiscal fund project budget performance evaluation service system

Abstract

This invention relates to the field of fiscal big data processing and smart government supervision technology, specifically a knowledge graph-based fiscal fund project budget performance evaluation service system, comprising: an ideal benchmark construction module, which parses policy texts using natural language processing technology and constructs a full lifecycle ideal execution graph; a multi-dimensional data mapping module, used to acquire actual execution data; a knowledge-driven reverse simulation module, which calls preset violation risk parameters to generate a faulty simulation graph corresponding to different risk assumptions; and a dual-difference coupled decision module, which extracts the actual execution graph and, relative to the full lifecycle ideal execution graph, determines the violation evaluation result of the fiscal fund project based on the actual deviation manifold. This invention eliminates the contamination of evaluation standards by past violation inertia that may be contained in historical data, ensuring the objectivity and absoluteness of the performance evaluation benchmark.

Description

Technical Field

[0001] This invention relates to the field of fiscal big data processing and smart government supervision technology, specifically a fiscal fund project budget performance evaluation service system based on knowledge graphs. Background Technology

[0002] In the current regulatory environment for performance evaluation of fiscal fund project budgets, policies and regulations usually exist in the form of unstructured texts, while the actual implementation process of projects involves multi-source heterogeneous data such as payment records, invoices and acceptance reports, and the flow of funds is accompanied by complex time-series and cost logic interactions.

[0003] To evaluate the compliance and performance of these funding projects, existing solutions generally adopt rule-based linear comparison or simple quantitative threshold early warning modes, which directly calculate the numerical difference between actual execution data and budget indicators, relying on human experience or general statistical models to identify anomalies. Although such solutions can identify absolute numerical deviations in fund usage, they lack a unified theoretical reference system to remove human interference and inefficient factors, and are difficult to effectively integrate the heterogeneous dimensions of time and amount. This makes it impossible to distinguish, from topological perspective, passive anomalies caused by environmental noise and active anomalies caused by substantive violations. At the same time, traditional methods are difficult to identify hidden and complex violation patterns in the absence of large-scale historical violation annotation data, and can only output vague quantitative alarms without achieving accurate qualitative diagnosis of the nature of violations. Therefore, how to construct an objective and ideal execution benchmark, and achieve accurate qualitative identification and noise-resistant judgment of fund violations based on this benchmark, has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a knowledge graph-based fiscal fund project budget performance evaluation service system. Specifically, the technical solution of this invention includes:

[0005] The ideal benchmark construction module is used to parse policy texts based on natural language processing technology, extract execution logic triples of funding entities, and construct a full life cycle ideal execution graph. The execution logic triples include funding entities, task nodes, and output standards, and the full life cycle ideal execution graph includes ideal time nodes and ideal cost nodes set on the theoretically optimal path.

[0006] A multidimensional data mapping module is used to acquire real-world execution data, map the real-world execution data into graph nodes through entity alignment, and generate a real-world execution graph based on fund flow. A knowledge-driven reverse simulation module is used to call preset violation risk parameters and inject the violation risk parameters into the full life cycle ideal execution graph to generate a faulty simulation graph corresponding to different risk assumptions. A dual-difference coupled decision module is used to extract the real deviation manifold of the real-world execution graph relative to the full life cycle ideal execution graph, and the theoretical deviation manifold of the faulty simulation graph relative to the full life cycle ideal execution graph, and determine the violation evaluation result of the fiscal fund project based on the topological similarity between the real deviation manifold and the theoretical deviation manifold.

[0007] Preferably, the ideal benchmark construction module uses natural language processing technology to parse policy texts, extract execution logic triples of funding entities, and construct a full lifecycle ideal execution graph, including:

[0008] Read budget regulations, application guidelines, and industry quota documents; parse and extract the task breakdown structure and industry guidance unit price from the documents; based on first principles, assume the highest administrative efficiency and no rent-seeking behavior, construct the ideal execution graph for the entire life cycle, where the attribute vector of each node includes the theoretical planned time and the theoretical quota cost.

[0009] Preferably, the ideal benchmark construction module determines the theoretical quota cost, specifically including: calling the industry guidance unit price, the quantity of resources applied for in the project, and the administrative loss tolerance rate; calculating the sum of the products of the industry guidance unit price and the quantity of resources applied for in the project; applying a linear correction based on the administrative loss tolerance rate to the sum of the products, and defining the calculation result as the nodal potential function value of the ideal cost node.

[0010] Preferably, the knowledge-driven reverse simulation module calls preset violation risk parameters and injects these parameters into the full lifecycle ideal execution map to generate a faulty simulation map corresponding to different risk assumptions, including:

[0011] The ideal planning time and fiscal year end date in the full life cycle ideal execution graph are determined; the time compression rate and random offset risk factors representing the urgency of sudden spending are introduced; the time series compression perturbation operator is applied to simulate the exponential increase effect of the probability density of payment events as the ideal planning time approaches the fiscal year end date, thereby generating a simulated tampered timestamp after forced compression; the graph nodes are reconstructed based on the simulated tampered timestamp to generate the faulty simulation graph.

[0012] Preferably, the application of the time-series compression perturbation operator includes: constructing an inverse time potential field function, which contains a composite structure of logarithmic and exponential terms; when the ideal planning time is much earlier than the fiscal year deadline, controlling the function value to converge to the fiscal year deadline to simulate payment postponement behavior; when the ideal planning time is close to the fiscal year deadline, controlling the function value to be constrained by the fiscal year deadline to simulate hard constraint squeezing effect.

[0013] Preferably, the dual-difference coupled decision module extracts the actual deviation manifold of the actual execution graph relative to the ideal execution graph throughout the entire lifecycle, and the theoretical deviation manifold of the faulty simulation graph relative to the ideal execution graph throughout the entire lifecycle, including:

[0014] Calculate the first difference between the actual data and the ideal benchmark in terms of amount and time, and the second difference between the simulation data and the ideal benchmark in terms of amount and time. Introduce normalized weighting coefficients based on the average total project budget and the average total period. Use the normalized weighting coefficients to standardize the first difference and the second difference to construct node-level actual deviation vectors and simulation deviation vectors.

[0015] Preferably, the dual-difference coupled decision module determines the violation evaluation result of the fiscal fund project based on the topological similarity between the actual deviation manifold and the theoretical deviation manifold, including:

[0016] The real deviation vectors of all nodes in the entire graph are concatenated into a real high-dimensional hypervector; the simulated deviation vectors of all nodes in the entire graph are concatenated into a simulated high-dimensional hypervector; the cosine similarity value between the real high-dimensional hypervector and the simulated high-dimensional hypervector is calculated; the cosine similarity value is used as a quantitative indicator for judging the isomorphism of the violation pattern.

[0017] Preferably, the evaluation results for irregularities in fiscal fund projects include:

[0018] A similarity judgment threshold is set; if the cosine similarity value is greater than the similarity judgment threshold, the fiscal fund project is determined to be a specific violation corresponding to the current simulation hypothesis; if the cosine similarity value is less than or equal to the similarity judgment threshold, and the magnitude of the actual deviation vector is significant, the fiscal fund project is determined to have environmental noise or unknown anomalies.

[0019] Compared with existing technologies, this system has the following advantages:

[0020] 1. This system uses an ideal benchmark construction module to analyze policy regulations and industry quota documents based on natural language processing technology. Based on the first principle assumption of maximum administrative efficiency and no rent-seeking behavior, it constructs an ideal execution map for the entire life cycle. This map serves as a pure theoretical reference system for evaluation, eliminating the contamination of evaluation standards by past violations that may be contained in historical data, and ensuring the objectivity and absoluteness of the performance evaluation benchmark.

[0021] 2. This system uses a knowledge-driven reverse simulation module to inject preset risk parameters of violations, such as the urgency of rushing to spend money, into the ideal graph to generate a simulated graph with defects. The dual-difference coupled decision module is based on the real deviation manifold between the actual execution graph and the ideal graph, as well as the topological similarity between the simulated graph with defects and the theoretical deviation manifold between the ideal graph and the ideal graph. This enables it to make accurate qualitative diagnoses of abnormal situations in fiscal funds projects, rather than just giving vague quantitative alarms.

[0022] 3. This system uses a dual differential coupling decision mechanism to calculate the isomorphism between real-world deviations and theoretical simulation deviations, thereby effectively distinguishing between systemic anomalies caused by substantive violations and local or random anomalies caused by environmental noise. At the same time, by normalizing the difference vectors of monetary and time dimensions and concatenating them into a high-dimensional hypervector for topological similarity calculation, the system overcomes the limitations of single-dimensional deviation discrimination and significantly improves the robustness and accuracy of violation judgment.

[0023] 4. To address the difficulty in obtaining large-scale historical violation annotation data, this system's knowledge-driven reverse simulation module introduces time-series compression perturbation operators and other methods to dynamically generate simulation data fingerprints that conform to the dynamics of specific violation behaviors. This allows the system to identify and diagnose complex and hidden time-series violation patterns, such as sudden spending, without relying on a large amount of labeled historical data, through generative methods, effectively reducing the reliance on audit experience and manpower costs. Attached Figure Description

[0024] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0025] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0027] Example 1:

[0028] Please see Figure 1A knowledge graph-based fiscal fund project budget performance evaluation service system includes: an ideal benchmark construction module, which is used to parse policy texts based on natural language processing technology, extract execution logic triples of the funding entity, and construct a full life cycle ideal execution graph; wherein, the execution logic triples include the funding entity, task nodes, and output standards, and the full life cycle ideal execution graph includes ideal time nodes and ideal cost nodes set on the theoretical optimal path;

[0029] The multidimensional data mapping module is used to acquire real-world execution data, map the real-world execution data into graph nodes through entity alignment, and generate a real-world execution graph based on fund flow. The knowledge-driven reverse simulation module is used to call preset violation risk parameters and inject the violation risk parameters into the full life cycle ideal execution graph to generate a faulty simulation graph corresponding to different risk assumptions. The dual-difference coupled decision module is used to extract the real deviation manifold of the real-world execution graph relative to the full life cycle ideal execution graph, as well as the theoretical deviation manifold of the faulty simulation graph relative to the full life cycle ideal execution graph, and determine the violation evaluation result of the fiscal fund project based on the topological similarity between the real deviation manifold and the theoretical deviation manifold.

[0030] This embodiment provides a fiscal fund project budget performance evaluation service system based on knowledge graph; the system mainly includes an ideal benchmark construction module, a multi-dimensional data mapping module, a knowledge-driven reverse simulation module, and a dual differential coupling decision module;

[0031] Ideal Benchmark Construction Module: The core task of this module is to construct the zero point of evaluation; to establish a theoretical reference system that removes human interference and inefficient factors; in this embodiment, this module uses natural language processing technology to parse unstructured policy text and extract execution logic triples containing funding entities, task nodes, and output standards; based on this, the module constructs a full lifecycle ideal execution graph, denoted as... In this graph, the ideal time node located on the theoretically optimal path is defined as denoted as . And the ideal cost node is denoted as ; The edges and nodes in the diagram represent the trajectory of funds that should flow under the first-principles assumptions of full compliance, maximum administrative efficiency, and no rent-seeking behavior.

[0032] Multidimensional Data Mapping Module: This module is responsible for digitizing the execution traces of the physical world; transforming multi-source heterogeneous real-world data into a computable graph structure; in this embodiment, this module acquires real-world execution data including payment records, invoice OCR information, acceptance reports, etc., and maps the above real-world data into graph nodes using entity alignment technology; based on the actual flow of funds, it generates a real-world execution graph, denoted as... ;at this time, It includes compliant operations, illegal operations, and market environment noise;

[0033] Knowledge-driven reverse simulation module: This module is responsible for generating pathological slides for comparison; making implicit violation patterns explicit, and rehearsing what the graph would look like if a violation occurred; in this embodiment, this module calls preset violation risk parameters, such as the urgency of rushing to spend money, the threshold for splitting contracts, etc., and injects these parameters into the aforementioned ideal execution graph of the entire life cycle; by perturbing the ideal graph, it generates simulation graphs corresponding to different risks, for example, denoted as a faulty graph such as rushing to spend money or false reporting and fraudulent claims. ,in Representing the Such illegal assumptions;

[0034] Dual Differential Coupled Decision Module: This module is the core of the system's decision-making process; it identifies violations through topological comparison. In this embodiment, this module extracts two sets of differential features, resulting in the actual deviation manifold, i.e., the actual execution graph. Compared to the ideal execution map throughout the entire lifecycle Differences Theoretical deviation manifold: i.e., simulation graph with defects. Compared to the ideal execution map throughout the entire lifecycle Differences The module determines the evaluation results of violations in fiscal fund projects based on the topological similarity between the actual deviation manifold and the theoretical deviation manifold.

[0035] This system can accurately distinguish between anomalies caused by violations and those caused by environmental noise by calculating the isomorphism between actual deviations and theoretical simulation deviations. By using knowledge-driven reverse simulation, it can achieve qualitative diagnosis of specific violations, rather than just giving vague quantitative alarms, effectively solving the technical pain point of traditional methods in distinguishing the nature of violations.

[0036] Example 2:

[0037] The ideal benchmark construction module uses natural language processing technology to parse policy texts, extract execution logic triples of funding entities, and construct a full lifecycle ideal execution graph, including:

[0038] Read budget regulations, application guidelines, and industry quota documents; parse and extract the task breakdown structure and industry guidance unit price from the documents; based on the first principle assumption of maximum administrative efficiency and no rent-seeking behavior, construct an ideal execution graph for the entire life cycle, in which the attribute vector of each node contains the theoretical planned time and the theoretical quota cost.

[0039] This embodiment further defines the specific implementation logic of the ideal benchmark construction module, especially the construction process of the zero-entropy path;

[0040] The specific execution steps of the ideal benchmark construction module are as follows:

[0041] Document reading: The system automatically reads budget regulations documents, including the Budget Law, application guidelines for the project application stage, construction cost information, and industry quota documents;

[0042] Structured extraction: Using NLP technology, the document content is parsed, the task breakdown structure (WBS) of the project is extracted as the skeleton of the map, and the industry guidance unit price is extracted as the atomic data for cost calculation.

[0043] Specifically, the structured extraction employs a method combining rule parsing based on document structure and named entity recognition: Multi-level headings in the document are identified through document layout analysis, and a parent-child node relationship of a WBS tree-like skeleton is constructed based on the heading hierarchy; a pre-trained BERT-BiLSTM-CRF model is used to perform sequence annotation on the heading text, extracting action words such as "procurement" and "construction," and object words such as "server" and "main structure," which are then combined to generate standardized node names; regular expressions are used to match currency units, such as yuan / unit and yuan / square meter, and their preceding values ​​in the body paragraphs corresponding to the leaf nodes, extracting them as industry guidance unit prices.

[0044] Zero-entropy construction: Based on first principles, it assumes maximum administrative efficiency during project execution, i.e., no delays and no rent-seeking behavior, i.e., no false reporting or kickbacks; under this assumption, an ideal execution graph for the entire lifecycle is constructed; in this graph, the attribute vector of each node contains theoretical values ​​for two core dimensions:

[0045] Theoretical planned time: The optimal time point determined based on the task duration quota;

[0046] Theoretical cost quota: the optimal cost calculated based on industry standards;

[0047] By constructing a zero-entropy ideal map based on regulations and quota documents, this embodiment eliminates the contamination of evaluation criteria by past violations contained in historical data, ensuring the objectivity and absoluteness of the evaluation benchmark, and providing a pure coordinate origin for subsequent accurate calculation of deviations.

[0048] Example 3:

[0049] The ideal benchmark construction module determines the theoretical quota cost, specifically including: calling the industry guidance unit price, the quantity of resources applied for in the project, and the administrative loss tolerance rate; calculating the sum of the products of the industry guidance unit price and the quantity of resources applied for in the project; applying a linear correction based on the administrative loss tolerance rate to the sum of the products, and defining the calculation result as the nodal potential function value of the ideal cost node.

[0050] This embodiment details the calculation method of theoretical quota cost in the ideal benchmark construction module, which treats cost as potential energy in the graph network;

[0051] To accurately quantify the ideal cost, this embodiment defines the potential energy function at the ideal cost node, and its calculation formula is as follows:

[0052]

[0053] in, : Represents a subset of the resource list belonging to the task node, a set; Indicates belonging to a task node The first in the resource list subset Item resource index;

[0054] : The nodal potential energy function value of the ideal cost node, in unit yuan, which is derived from the calculation output of this step;

[0055] : Industry guidance unit price, in yuan / unit, which is derived from industry quota documents read in previous steps, such as official cost information database;

[0056] The number of resources and units involved in the project application are derived from the committed investment amount stated in the project application.

[0057] : Administrative loss tolerance rate, a dimensionless constant, which is derived from the unforeseen expense ratio or fault tolerance rate stipulated in the relevant fund management regulations of the Ministry of Finance. For example, it can be set to 0 or a minimum value.

[0058] The calculation logic and results are integrated: This embodiment calls upon the industry-guided unit price, the quantity of resources applied for in the project, and the administrative loss tolerance rate to calculate the sum of the products of the guided unit price and quantity of all resource items, which is the direct physical cost. A linear correction based on the administrative loss tolerance rate is then applied to this sum of products. This is achieved by introducing... In this embodiment, while adhering to strict quotas, on the one hand, compatibility with necessary administrative management losses is maintained, and on the other hand, the calculation results are defined as nodal potential energy, so that subsequent violation judgments can be compared with the anisotropic analysis of potential energy difference in physics, thereby more scientifically quantifying the degree of fund deviation.

[0059] Example 4:

[0060] The knowledge-driven reverse simulation module calls preset violation risk parameters and injects these parameters into the full lifecycle ideal execution graph to generate a faulty simulation graph corresponding to different risk assumptions, including:

[0061] The ideal planning time and fiscal year end date in the ideal execution graph of the entire life cycle are determined; the time compression rate and random offset risk factors representing the urgency of rushing to spend money are introduced; the time series compression perturbation operator is applied to simulate the exponential increase effect of the probability density of payment events as the ideal planning time approaches the fiscal year end date, thereby generating the simulated timestamp after forced compression; the graph nodes are reconstructed based on the simulated timestamp to generate the flawed simulation graph.

[0062] This embodiment details how the knowledge-driven reverse simulation module generates a simulated graph of a problem in a scenario of sudden spending;

[0063] To simulate the temporal characteristics of violations, this embodiment introduces a temporal compression perturbation operator; the specific execution of the knowledge-driven reverse simulation module includes:

[0064] Establishing a baseline: Determining the ideal planned time within the ideal execution map throughout the entire lifecycle. and the legally binding fiscal year end date For example, December 31st;

[0065] Introducing risk parameters: Retrieving risk factors representing the urgency of sudden spending from the risk knowledge base; time compression rate. And random offsets that simulate the randomness of artificial operations ;

[0066] Application Operator: The temporal compression perturbation operator is applied, which simulates the effect of an exponential increase in the probability density of payment events as the ideal planning time approaches the fiscal year deadline; the calculation results generate a simulated timestamp that has been forcibly compressed.

[0067] Reconstructing the graph: Based on the simulation's altered timestamps replacing the original ideal timestamps, the graph nodes are reconstructed to generate a faulty simulation graph;

[0068] By introducing risk parameters related to time urgency, this embodiment can dynamically generate simulation data that conforms to the dynamics of impulsive spending behavior; this eliminates the need for the system to rely on a large amount of historical violation annotation data, which is usually extremely difficult to obtain, and allows for the identification of complex temporal violation patterns through generative methods.

[0069] Example 5:

[0070] The application of a time-series compression perturbation operator includes: constructing an inverse time potential field function, which contains a composite structure of logarithmic and exponential terms; when the ideal planning time is much earlier than the fiscal year deadline, the control function value converges to the fiscal year deadline to simulate payment deferral behavior; when the ideal planning time is close to the fiscal year deadline, the control function value is constrained by the fiscal year deadline to simulate the hard constraint squeezing effect.

[0071] This embodiment is a further specification of embodiment 4, and describes in detail the mathematical mechanism of the timing compression perturbation operator;

[0072] To accurately simulate the time wall effect on the eve of the deadline, this embodiment constructs an inverse time potential field function containing a composite structure of logarithmic and exponential terms, mathematically expressed as:

[0073]

[0074] in, Nodes in the simulation diagram The simulation's timestamp was altered;

[0075] It should be noted that before substituting into the above formula for numerical calculation, all time variables include: , All of these are preprocessed and normalized to an integer number of days offset relative to the project initiation date, thereby transforming the time dimension into a dimensionless scalar real number to participate in exponential operations. The unit is date, and its source is the output of this step.

[0076] : is the base of the natural logarithm;

[0077] The fiscal year end date, the unit date, is derived from the budget regulations.

[0078] Ideal plan time, unit date, derived from the ideal baseline construction module;

[0079] Risk factor time compression rate, in units of 1 / day, is derived from historical violation statistics and represents the degree of urgency that violators face the deadline.

[0080] The specific calculation method for historical violation statistics fitting is as follows: collect the set of payment time points for historically confirmed violations of spending rules. The time interval between each point in time and the end date of the fiscal year is defined as follows: Based on the assumption that the probability of non-compliant payments increases exponentially with time, the maximum likelihood estimation (MLE) method is used to estimate the probability density function of the exponential distribution. Perform parameter solving and calculate the results. ;in This represents the total number of historical violation samples.

[0081] Random offset, in days, derived from a random number generator, used to simulate operational noise;

[0082] Function logic and effect integration: The physical significance of this operator lies in constructing an inverse potential field.

[0083] Simulated payment postponement behavior: when the ideal planned time Much earlier than the deadline When the exponential term in the function approaches 0, the function value converges to 0. This simulates the behavior of an executor who deliberately postpones payment of funds due in advance until the end of the year.

[0084] Simulating the effect of hard constraint compression: When the ideal planning time approaches the deadline, the function value is forcibly constrained by... The smoothing properties of the logarithmic function Softplus simulate the hard constraint squeeze effect of the deadline on payment behavior;

[0085] This mathematical model accurately reproduces the nonlinear compression characteristics of sudden spending on the time axis, namely, the early spending is delayed and the later spending is squeezed, thus providing a high-fidelity theoretical reference fingerprint for identifying such covert violations.

[0086] Example 6:

[0087] The dual-difference coupled decision module extracts the actual deviation manifold of the actual execution graph relative to the ideal execution graph throughout the entire lifecycle, and the theoretical deviation manifold of the faulty simulation graph relative to the ideal execution graph throughout the entire lifecycle, including:

[0088] The first difference between the actual data and the ideal benchmark in terms of amount and time is calculated, and the second difference between the simulation data and the ideal benchmark in terms of amount and time is calculated. Normalized weight coefficients based on the average total budget and the average total period of the project are introduced. The first and second differences are standardized using the normalized weight coefficients to construct node-level actual deviation vectors and simulation deviation vectors.

[0089] This embodiment details the method for constructing the deviation vector in the dual differential coupling decision module to solve the problem of inconsistency between the monetary and time dimensions;

[0090] To compare differences in a unified geometric space, this embodiment introduces normalization processing; the dual-difference coupled decision module executes as follows:

[0091] Calculate the original differences: Calculate the first difference between the actual data and the ideal benchmark in both the monetary and time dimensions. , And the second difference between the simulated data of defective patterns and the ideal benchmark in both monetary and time dimensions. , ;

[0092] Introducing weights: Calculating the average total project budget Compared with the total period average Based on this, normalized weight coefficients are constructed. and ;

[0093] Vector Construction: The first and second differences are standardized using normalized weight coefficients to construct a node-level reality deviation vector. Deviation vector from simulation The set of reality deviation vectors calculated for all nodes in the entire graph is defined as the reality deviation manifold. This manifold is mathematically represented as a point cloud distribution in a 3D feature space; similarly, a theoretically biased manifold can be constructed. For example, the reality deviation vector can be represented as:

[0094]

[0095] in, The unit is yuan , The unit is days. This ensures that both dimensions of the deviation vector are dimensionless.

[0096] By normalizing based on the overall project scale, this embodiment eliminates the huge dimensional difference between monetary amounts and time periods, allowing the deviations in the two dimensions to be geometrically calculated with equal weight in the same vector space, thus ensuring the mathematical rationality of subsequent similarity calculations.

[0097] Example 7:

[0098] The dual-difference coupled decision module determines the evaluation results of violations in fiscal fund projects based on the topological similarity between the actual deviation manifold and the theoretical deviation manifold, including:

[0099] The real deviation vectors of all nodes in the entire graph are concatenated into a real high-dimensional hypervector; the simulated deviation vectors of all nodes in the entire graph are concatenated into a simulated high-dimensional hypervector; the cosine similarity value between the real high-dimensional hypervector and the simulated high-dimensional hypervector is calculated; the cosine similarity value is used as a quantitative indicator to judge the isomorphism of the violation pattern.

[0100] This embodiment details how to utilize the characteristics of high-dimensional space to determine the violation evaluation result;

[0101] To capture violation patterns from a global perspective, this embodiment employs hypervector splicing technology; the specific execution steps of the dual-difference coupled decision module are as follows:

[0102] Preprocessing Alignment: Traversing the Ideal Execution Graph Throughout the Entire Lifecycle Each node in Using the Task Breakdown Structure (WBS) encoding as the primary key, a real-world execution graph is constructed. Searching for the corresponding real-world node ;

[0103] Among them, the actual execution map The nodes in the graph do not naturally carry WBS encoding and need to be associated in advance through semantic vector matching: the Sentence-BERT model is used to connect the nodes in the ideal graph. Convert the task name attribute into a target feature vector It also transforms real-world data, such as invoice details and summaries, into candidate feature vectors. ;calculate and The cosine similarity is used. If the similarity is greater than a preset semantic threshold, such as 0.85, then the real-world data node is marked as... And assign corresponding WBS codes; if there is a many-to-one mapping, then resolve conflicts according to the principle of the closest timestamp; after completing the above association, if the retrieval is successful, extract the attributes of the two to calculate the deviation; if the retrieval fails, that is, it was not executed or reimbursed in reality, then set the attribute value of the real node at that position to zero, and calculate the negative deviation vector relative to the ideal benchmark accordingly, so as to retain the difference characteristics caused by the task's absence; for It exists in but cannot be mapped to The free nodes are marked as independent anomalous subgraphs, do not participate in isomorphism calculations, and are directly output as additional violations; after alignment, it is ensured that the feature vector sequences of the two graphs have strictly consistent lengths;

[0104] Hypervector concatenation: Consider all nodes in the entire graph, denoted as the total number of nodes. The actual deviation vectors are concatenated in a fixed order to form a The simulation high-dimensional hypervector is obtained by concatenating the simulation deviation vectors of all nodes; similarly, the simulation high-dimensional hypervector is obtained by concatenating the simulation deviation vectors of all nodes.

[0105] Calculate similarity: Calculate the cosine similarity between the two high-dimensional hypervectors; this value reflects the degree of alignment between the two high-dimensional vectors in terms of direction;

[0106] Quantitative Indicator: The cosine similarity value is used as a quantitative indicator to judge the isomorphism of the violation pattern;

[0107] By constructing a high-dimensional hypervector at the full graph level, this embodiment is no longer limited to isolated deviations of a single node, but examines the distribution pattern of deviations throughout the entire project lifecycle, i.e., the shape of the deviations. This method can effectively identify systematic violations, such as systematic misappropriation of funds, while ignoring local and chaotic random errors, thereby significantly improving the robustness of the judgment.

[0108] Example 8:

[0109] Determine the results of the evaluation of irregularities in fiscal fund projects, including:

[0110] Set a similarity judgment threshold; if the cosine similarity value is greater than the similarity judgment threshold, the fiscal fund project is determined to be the specific violation corresponding to the current simulation hypothesis; if the cosine similarity value is less than or equal to the similarity judgment threshold, and the magnitude of the actual deviation vector is significant, the fiscal fund project is determined to have environmental noise or unknown anomalies.

[0111] This embodiment details the logic for judging violations;

[0112] To achieve tiered processing for accurate diagnosis and anomaly early warning, this embodiment sets the following decision rules:

[0113] Set threshold: Pre-set the similarity judgment threshold. For example, 0.85;

[0114] Diagnosis logic: If the calculated cosine similarity value is greater than the similarity judgment threshold, it indicates that the deviation pattern in reality is highly consistent with a certain simulated violation pattern, i.e., isomorphic. At this time, the system determines that the fiscal fund project is diagnosed as the specific violation behavior corresponding to the current simulation hypothesis, such as being diagnosed as rushing to spend money.

[0115] Noise elimination logic: If the cosine similarity value is less than or equal to the similarity judgment threshold, but at the same time the magnitude of the actual deviation vector is significant, that is, the deviation amplitude is large, it indicates that although there is a problem with the project, it is an unknown or non-systematic problem; at this time, the system determines that the fiscal fund project has environmental noise or unknown anomalies, and manual intervention is required.

[0116] This judgment logic achieves the effect of not only knowing what happened but also why it happened; high similarity directly corresponds to specific violations, providing explainable audit clues; low similarity and high deviation indicate new violations or external force majeure, guiding auditors to focus on key areas; this binary judgment mechanism greatly improves audit efficiency and reduces false alarm rate.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A knowledge graph-based fiscal fund project budget performance evaluation service system, characterized in that, The method comprises the following steps: An ideal benchmark modeling module is used to parse policy texts based on natural language processing technology, extract execution logic triples of fund subjects, and construct a full-life-cycle ideal execution graph; wherein the execution logic triples comprise fund subjects, task nodes, and output standards, and the full-life-cycle ideal execution graph comprises ideal time nodes and ideal cost nodes set on a theoretically optimal path; A multi-dimensional data mapping module is used to obtain real execution data, map the real execution data into graph nodes through entity alignment, and generate a real execution graph based on fund flow directions; a knowledge-driven reverse simulation module is used to call preset irregular behavior risk parameters, inject the irregular behavior risk parameters into the full-life-cycle ideal execution graph, and generate a simulation graph with irregularities corresponding to different risk assumptions; a double-difference coupling judgment module is used to extract real deviation manifolds of the real execution graph relative to the full-life-cycle ideal execution graph, and theoretical deviation manifolds of the simulation graph with irregularities relative to the full-life-cycle ideal execution graph, and determine irregular evaluation results of a fiscal fund project based on topological similarity of the real deviation manifolds and the theoretical deviation manifolds. 2.The knowledge graph-based financial fund project budget performance evaluation service system according to claim 1, characterized in that, The ideal benchmark modeling module is used to parse policy texts based on natural language processing technology, extract execution logic triples of fund subjects, and construct a full-life-cycle ideal execution graph, comprising the following steps: Budget regulation documents, declaration guide documents, and industry quota documents are read; task decomposition structures and industry guide unit prices in the documents are parsed and extracted; a full-life-cycle ideal execution graph is constructed according to the first principle assumption that administrative efficiency is the highest and there is no rent-seeking behavior, wherein an attribute vector of each node comprises theoretical planned time and theoretical quota cost. 3.The knowledge graph-based financial fund project budget performance evaluation service system according to claim 2, characterized in that, The ideal benchmark modeling module determines the theoretical quota cost, specifically comprising the following steps: calling the industry guide unit price, the number of project declaration resources, and the administrative loss tolerance rate; calculating a sum of products of the industry guide unit price and the number of project declaration resources; applying a linear correction based on the administrative loss tolerance rate to the sum of products, and defining a calculation result as a node potential energy function value of the ideal cost node. 4.The knowledge graph-based financial fund project budget performance evaluation service system according to claim 1, characterized in that, The knowledge-driven reverse simulation module calls preset irregular behavior risk parameters, injects the irregular behavior risk parameters into the full-life-cycle ideal execution graph, and generates a simulation graph with irregularities corresponding to different risk assumptions, comprising the following steps: The ideal planned time and the fiscal year deadline in the full-life-cycle ideal execution graph are determined; a risk factor time compression rate representing the degree of urgent spending in a sudden attack and a random offset are introduced; a time sequence compression disturbance operator is applied to simulate an exponential rising effect of a payment event probability density as the ideal planned time approaches the fiscal year deadline, thereby generating a simulation tampered timestamp under forced compression; the simulation graph with irregularities is generated by reconstructing graph nodes based on the simulation tampered timestamp. 5.The knowledge graph-based financial fund project budget performance evaluation service system according to claim 4, characterized in that, The application of the time-series compression perturbation operator includes: constructing an inverse time potential field function, which contains a composite structure of logarithmic and exponential terms; when the ideal planning time is much earlier than the fiscal year deadline, controlling the function value to converge to the fiscal year deadline to simulate payment postponement behavior; when the ideal planning time is close to the fiscal year deadline, controlling the function value to be constrained by the fiscal year deadline to simulate hard constraint squeezing effect. 6.The knowledge graph-based financial fund project budget performance evaluation service system according to claim 1, characterized in that, The dual differential coupling decision module extracts the actual deviation manifold of the actual execution graph relative to the ideal execution graph throughout the entire lifecycle, and the theoretical deviation manifold of the faulty simulation graph relative to the ideal execution graph throughout the entire lifecycle, including: Calculate the first difference between the actual data and the ideal benchmark in terms of amount and time, and the second difference between the simulation data and the ideal benchmark in terms of amount and time. Introduce normalized weighting coefficients based on the average total project budget and the average total period. Use the normalized weighting coefficients to standardize the first difference and the second difference to construct node-level actual deviation vectors and simulation deviation vectors. 7.The knowledge graph-based financial fund project budget performance evaluation service system according to claim 6, characterized in that, The dual-difference coupled decision module determines the violation evaluation result of the fiscal fund project based on the topological similarity between the actual deviation manifold and the theoretical deviation manifold, including: The real deviation vectors of all nodes in the entire graph are concatenated into a real high-dimensional hypervector; the simulated deviation vectors of all nodes in the entire graph are concatenated into a simulated high-dimensional hypervector; the cosine similarity value between the real high-dimensional hypervector and the simulated high-dimensional hypervector is calculated; the cosine similarity value is used as a quantitative indicator for judging the isomorphism of the violation pattern. 8.The knowledge graph-based financial fund project budget performance evaluation service system according to claim 7, characterized in that, The determination of the evaluation results for violations of fiscal fund projects includes: A similarity judgment threshold is set; if the cosine similarity value is greater than the similarity judgment threshold, the fiscal fund project is determined to be a specific violation corresponding to the current simulation hypothesis; if the cosine similarity value is less than or equal to the similarity judgment threshold, and the magnitude of the actual deviation vector is significant, the fiscal fund project is determined to have environmental noise or unknown anomalies.

Images