A project budget management method and device

Abstract

The application provides a project budget management method and device, relates to the technical field of budget management, and solves the technical problems of existing project budget management technology, such as static, budget response not timely, and lack of compliance traceability. The method specifically comprises the following steps: obtaining first information of a project; integrating project basic information into a project main node, integrating budget subject information into a budget subject node, integrating budget template information into a project edge between nodes, and constructing a knowledge graph; converting the project main node and the budget subject node into low-dimensional vectors through semantic embedding, generating a weight vector through relationship embedding of the project edge, and forming GNN model input data; iteratively training the GNN model input data by using a heterogeneous graph attention network architecture; predicting the knowledge graph with changed project basic information by using a trained model, outputting budget adjustment parameters, and updating the knowledge graph to complete project budget management. The application is used for project budget management.

Description

Technical Field

[0001] This application relates to the field of budget management technology, and in particular to a project budget management method and apparatus. Background Technology

[0002] Project budget management is crucial for project resource allocation, cost control, and supporting financial budget analysis and manpower hour assessment. However, existing technologies mostly employ a static initialization combined with offline adjustments, relying solely on experience to estimate the total budget at project initiation and registering it offline or at a basic online level, without establishing a linkage mechanism between budget and project execution information. Existing project budget management technologies suffer from data mismatch with management needs, delayed budget response to project changes, high compliance risks, and poor traceability. Therefore, existing project budget management technologies are characterized by static nature, untimely budget response, and lack of compliance traceability, making it difficult to meet the needs of dynamic project management. Summary of the Invention

[0003] This application provides a project budget management method and apparatus, which solves the technical problems of existing project budget management technologies, such as static nature, untimely budget response, and lack of compliance traceability.

[0004] To achieve the above objectives, this application adopts the following technical solution:

[0005] Firstly, a project budget management method is provided, comprising: acquiring the first information of the project, which includes basic project information, budget item information, and budget template information; budget model information is used to represent the relationships between various basic project information and / or budget item information; integrating the basic project information into project main nodes, integrating the budget item information into budget item nodes, and integrating the budget template information into project edges between the nodes; and constructing a knowledge graph based on the project main nodes, budget item nodes, and project edges; the project edges include a first project edge and a second project edge, wherein the first project edge is used to represent the logical relationships between the nodes, and the second project edge is used to represent the relationships between the nodes. The calculation relationship of project budgets is established; project main nodes and budget item nodes are transformed into low-dimensional vectors through semantic embedding, and project edges are generated into weight vectors through relational embedding. An adjacency matrix is ​​constructed by combining the graph topology structure to form the input data of the GNN model; a heterogeneous graph attention network architecture is used to iteratively train the GNN model input data to obtain a project budget-specific GNN training model; the project budget-specific GNN training model is used to predict changes in the knowledge graph of project basic information and output budget adjustment parameters, which are used to adjust the parameters of budget item nodes and project edges in the knowledge graph; the knowledge graph is updated based on the budget adjustment parameters to complete project budget management.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves integrating basic project information into project main nodes, budget item information into budget item nodes, and budget template information into project edges between nodes. A knowledge graph is then constructed based on the project main nodes, budget item nodes, and project edges. This includes: structuring the basic project information to form project main nodes, extracting budget item information, and associating parent / child node dependencies to form budget item nodes; using a semantic rule parser to extract the logical relationship between budget template information and budget item information to form a first project edge; extracting the computational relationship between budget template information and budget item information through parameter mapping to form a second project edge; and importing the project main nodes, budget item nodes, first project edges, and second project edges into a graph database, establishing topological connections through the mapping relationship between the first and last nodes of nodes and edges to form a complete project budget knowledge graph.

[0007] In conjunction with the first aspect mentioned above, one possible implementation involves transforming project main nodes and budget item nodes into low-dimensional vectors through semantic embedding, generating weight vectors for project edges through relation embedding, and constructing an adjacency matrix based on the graph topology. This includes: using a hybrid semantic embedding algorithm of TransE and BERT to encode and transform the text attributes of project main nodes and budget item nodes into fixed-dimensional low-dimensional vectors; using a hybrid relation embedding algorithm of RotatE and DistMult to parse attributes, convert features, and establish relations for the rules of project edges, generating weight vectors that distinguish relation types and influence strengths; and using the low-dimensional vectors as row / column bases, combined with the weight vectors, filling the adjacency matrix elements based on the knowledge graph topology to form an adjacency matrix containing node semantics, edge relation weights, and topological connections.

[0008] In conjunction with the first aspect mentioned above, one possible implementation involves iteratively training the GNN model input data using a heterogeneous graph attention network architecture. This includes: adapting the low-dimensional vectors and weight vectors in the GNN model input data using node type mapping matrices and edge relation mapping matrices, respectively; performing a linear transformation on the adapted low-dimensional vectors to unify the feature dimensions; calculating the attention coefficients between nodes based on the weight vectors; aggregating the features of neighboring nodes based on the attention coefficients to update the current node features; and using a dual-objective loss function that includes regression loss and classification loss to optimize and continuously update the node features to train the GNN model. The regression loss is used for budget adjustment prediction, and the classification loss is used for work hour allocation compliance verification.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the low-dimensional vectors and weight vectors in the GNN model input data are adapted using a type mapping matrix and a relation mapping matrix, respectively, to satisfy the following formula:

[0010]

[0011] in, It is a low-dimensional vector. A node type mapping matrix, The adapted low-dimensional vector, For the weight vector, This is the edge relation mapping matrix. This is the adapted weight vector.

[0012] In conjunction with the first aspect mentioned above, one possible implementation involves performing a linear transformation on the adapted low-dimensional vector to unify the feature dimension, calculating the attention coefficients between nodes using the weight vectors, and then aggregating the features of neighboring nodes based on these attention coefficients to update the features of the current node. This includes:

[0013] S1. Perform a linear transformation on the adapted low-dimensional vector to obtain a low-dimensional vector that satisfies the following formula:

[0014]

[0015] in, It is a linear transformation matrix. It is a low-dimensional vector after linear transformation.

[0016] S2. Calculate the attention coefficient, which satisfies the following formula:

[0017]

[0018] in, Attention coefficient (·) is a function that normalizes the intermediate results for all neighboring nodes u. (·) represents an activation function with leakage. For attention vectors, Let be the low-dimensional vector after the linear transformation of node u, and || be the vector concatenation operation. This is the normalization term.

[0019] S3. Aggregate the features of neighboring nodes to update the features of the current node. The updated node features satisfy the following formula:

[0020]

[0021] in, For the updated node features, (·) is the activation function. Let v be the set of neighboring nodes.

[0022] In conjunction with the first aspect mentioned above, in one possible implementation, the bi-objective loss function satisfies the following formula:

[0023]

[0024] in, To regress the loss, Adjusting the sample size for budget purposes Adjust parameters to forecast budget. To represent the actual adjustment parameters, i is the index of the budget adjustment sample. These are the weighting coefficients; The classification loss satisfies the following formula:

[0025]

[0026] in, Allocate compliant sample numbers to working hours. For authenticity and compliance labeling, Let j be the compliance probability predicted by the model, and j be the compliance sample index for work hours.

[0027] In conjunction with the first aspect mentioned above, one possible implementation involves using a project budget-specific GNN training model to predict changes in the knowledge graph of project fundamental information and outputting budget adjustment parameters. This includes: updating the corresponding project main nodes in the knowledge graph when project fundamental information changes; regenerating new GNN model input data based on the updated knowledge graph; inputting the new GNN model input data into the project budget-specific GNN training model to predict budget adjustment needs and outputting budget adjustment parameters.

[0028] In conjunction with the first aspect mentioned above, one possible implementation involves updating the knowledge graph based on budget adjustment parameters to complete project budget management. This includes: updating project nodes related to budget item information and project edges related to budget template information in the knowledge graph based on budget adjustment parameters; and generating a project budget execution snapshot based on the updated project nodes and project edges to complete project budget management.

[0029] Secondly, a project budget management device is provided, comprising: a communication unit and a processing unit; the communication unit is used to acquire first information of the project, the first information including basic project information, budget item information, and budget template information; budget model information is used to characterize the relationship between various basic project information and / or budget item information; the processing unit is used to integrate the basic project information into project main nodes, integrate the budget item information into budget item nodes, and integrate the budget template information into project edges between various nodes, and construct a knowledge graph based on the project main nodes, budget item nodes, and project edges; the project edges include a first project edge and a second project edge, the first project edge being used to characterize the logical relationship between various nodes. The second project edge is used to represent the calculation relationship of project budget between various nodes. The main project node and budget item node are transformed into low-dimensional vectors through semantic embedding, and the project edge is generated into a weight vector through relation embedding. Combined with the graph topology, an adjacency matrix is ​​constructed to form the input data of the GNN model. A heterogeneous graph attention network architecture is used to iteratively train the GNN model input data to obtain a project budget-specific GNN training model. The project budget-specific GNN training model is used to predict changes in the knowledge graph of basic project information and output budget adjustment parameters. The adjustment parameters are used to adjust the parameters of budget item nodes and project edges in the knowledge graph. The knowledge graph is updated based on the budget adjustment parameters to complete project budget management.

[0030] Thirdly, this application provides a project budget management apparatus, comprising: a processor and a storage medium; the storage medium includes instructions, and the processor is configured to execute the instructions to implement the methods described in the first aspect and any possible implementation thereof. This project budget management apparatus may be an electronic device or a chip within an electronic device.

[0031] Fourthly, this application provides a project budget management system, including: an API interface and a project budget management device; wherein the API interface is used to collect first information about the project, and the project budget management device is used to obtain the first information about the project; the basic project information is integrated into project main nodes, the budget item information is integrated into budget item nodes, and the budget template information is integrated into project edges between the nodes; a knowledge graph is constructed based on the project main nodes, budget item nodes, and project edges; the project edges include a first project edge and a second project edge, the first project edge being used to represent the logical relationship between the nodes, and the second project edge being used to represent the project between the nodes. The budget calculation process involves: transforming project main nodes and budget item nodes into low-dimensional vectors through semantic embedding; generating weight vectors from project edges through relational embedding; constructing an adjacency matrix based on the graph topology to form the input data for the GNN model; iteratively training the GNN model input data using a heterogeneous graph attention network architecture to obtain a project budget-specific GNN training model; using the project budget-specific GNN training model to predict changes in the knowledge graph of project basic information and outputting budget adjustment parameters, which are used to adjust the parameters of budget item nodes and project edges in the knowledge graph; and updating the knowledge graph based on the budget adjustment parameters to complete project budget management.

[0032] Fifthly, this application provides a computer-readable storage medium storing instructions that, when executed on a project budget management device, cause the project budget management device to perform the methods described in the first aspect and any possible implementation thereof.

[0033] In a sixth aspect, this application provides a computer program product containing instructions that, when run on a project budget management device, cause the project budget management device to perform the methods described in the first aspect and any possible implementation thereof.

[0034] This application provides a project budget management method and apparatus. By acquiring comprehensive first-dimensional information including basic project information, budget item information, and budget template information, it overcomes the limitation of existing technologies that only collect single total budget data. Then, it constructs a knowledge graph using project main nodes, budget item nodes, and project edges representing logical / computational relationships, breaking the static storage mode of budgets. Subsequently, it generates vectors through semantic / relational embedding and constructs an adjacency matrix, trains a GNN model using a heterogeneous graph attention network, accurately captures key influencing factors of the budget, and finally predicts budget adjustment parameters and updates the knowledge graph. This not only solves the problem of delayed budget response when projects change, but also achieves compliance traceability through the recording of knowledge graph nodes and edges, effectively achieving dynamic and refined budget management. It solves the technical problems of static budget management, untimely budget response, and lack of compliance traceability in existing project budget management technologies. Attached Figure Description

[0035] Figure 1 A system architecture diagram of a project budget management system provided in this application embodiment;

[0036] Figure 2 A flowchart illustrating a project budget management method provided in this application embodiment;

[0037] Figure 3 A flowchart illustrating another project budget management method provided in this application embodiment;

[0038] Figure 4 A flowchart illustrating another project budget management method provided in this application embodiment;

[0039] Figure 5 A flowchart illustrating another project budget management method provided in this application embodiment;

[0040] Figure 6 A schematic diagram of the structure of a project budget management device provided in this application embodiment;

[0041] Figure 7 This is a schematic diagram of the hardware structure of a project budget management device provided in an embodiment of this application. Detailed Implementation

[0042] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0043] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0044] The project budget management method provided in this application embodiment can be applied to, for example, Figure 1 The project budget management system shown includes an API interface 101 and a project budget management device 102.

[0045] The API interface 101 is used to collect the first information of the project, and the project budget management device 102 is used to obtain the first information of the project. The basic project information is integrated into project main nodes, the budget item information is integrated into budget item nodes, and the budget template information is integrated into project edges between each node. A knowledge graph is constructed based on the project main nodes, budget item nodes, and project edges. The project edges include a first project edge and a second project edge. The first project edge is used to represent the logical relationship between each node, and the second project edge is used to represent the calculation relationship of the project budget between each node. The project main nodes and budget item nodes are converted into low-dimensional vectors through semantic embedding, and the project edges are generated into weight vectors through relation embedding. An adjacency matrix is ​​constructed by combining the graph topology structure to form the input data of the GNN model. The heterogeneous graph attention network architecture is used to iteratively train the input data of the GNN model to obtain the project budget-specific GNN training model. The project budget-specific GNN training model is used to predict the changes in the knowledge graph of the basic project information and output budget adjustment parameters. The adjustment parameters are used to adjust the budget item nodes and project edges in the knowledge graph. The knowledge graph is updated based on the budget adjustment parameters to complete the project budget management.

[0046] To address the technical problems of static budget management, untimely budget response, and lack of compliance traceability in existing project budget management technologies, this application provides a project budget management method. This method obtains comprehensive first-dimensional information including basic project information, budget item information, and budget template information, overcoming the limitation of existing technologies that only collect single total budget data. It then constructs a knowledge graph using project main nodes, budget item nodes, and project edges representing logical / computational relationships, breaking the static storage mode of the budget. Subsequently, it generates vectors through semantic / relational embedding and constructs an adjacency matrix, trains a GNN model using a heterogeneous graph attention network, accurately captures key budget influencing factors, and finally predicts budget adjustment parameters and updates the knowledge graph. This not only solves the problem of delayed budget response when projects change but also achieves compliance traceability through the recording of knowledge graph nodes and edges, effectively achieving dynamic and refined budget management. Based on this, it solves the technical problems of static budget management, untimely budget response, and lack of compliance traceability in existing project budget management technologies.

[0047] Figure 2 A flowchart illustrating the project budget management method provided in this application embodiment is shown below. Figure 2 As shown, the method includes:

[0048] Step 201: The project budget management device obtains the first information about the project.

[0049] The first piece of information refers to the data set that supports the entire project budget management process, including basic project information (such as project name, duration, participants, project scope, etc.), budget item information (such as the definition and hierarchical relationship of different budget categories such as labor costs, material costs, operation and maintenance costs, etc.), and budget template information (data used to define the association rules between basic project information and budget item information).

[0050] In this embodiment, the project budget management device can automatically collect structured basic project information and budget item information through the API interface of the enterprise's internal project management system and financial system; it can also receive budget template information in Excel or CSV format manually uploaded by users through the front-end interface; and it can also call template information of similar projects from the historical project database as initial data. The method of obtaining the first information of the project is not specifically limited in this application.

[0051] As an example, when a software development project starts, the project budget management device automatically obtains basic information such as the project name "System Development," duration "6 months," and participants "10 people (including 3 backend developers, 2 frontend developers, 2 testers, and 3 maintenance personnel)" by connecting to the enterprise's Jira system (project management tool) API; it also obtains budget item information such as "labor costs," "server rental costs," and "software licensing costs" by connecting to the financial ERP system; and it receives the "software development budget template" uploaded by the finance personnel (which includes rules such as "server rental cost = number of servers × monthly rent × duration") to complete the acquisition of the first piece of information.

[0052] Based on the above steps, by integrating multi-dimensional and multi-source multimodal project information, budget estimation bias caused by incomplete data can be avoided.

[0053] Step 202: The project budget management device integrates basic project information into project main nodes, budget item information into budget item nodes, and budget template information into project edges between nodes, and constructs a knowledge graph based on project main nodes, budget item nodes, and project edges.

[0054] Among them, the project main node is the structured carrier of the project's basic information, with each node corresponding to a core piece of basic information; the budget item node is the structured carrier of budget item information, including attributes such as item name and hierarchical relationship; the project edge is the link connecting the nodes, including the first project edge and the second project edge. The first project edge is used to represent the logical relationship between each node, and the second project edge is used to represent the calculation relationship of the project budget between each node.

[0055] In this embodiment, for the main project node, different structured dimensions can be selected according to the project type; the hierarchical relationship of budget item nodes can be flexibly set according to enterprise management needs, supporting both two-level and multi-level hierarchies; in the integration of project edges, a semantic rule parser and natural language processing (NLP) are used to extract logical relationships, and parameter mapping and formula parsing are used to extract computational relationships. Finally, based on the graph database, a topological connection is established through the mapping relationship between node ID and the first and last nodes of the edge, forming a complete project budget knowledge graph.

[0056] As an example, in a certain construction project, the project budget management device integrates "Project Name: Office Building Construction", "Construction Period: 18 Months", and "Construction Team: Construction Team A (20 people)" into three main project nodes; it integrates "Labor Cost (Parent Node) - Construction Worker Cost (Child Node)", "Building Material Cost (Parent Node) - Steel Reinforcement Cost (Child Node)" and "Equipment Rental Cost (Parent Node) - Tower Crane Rental Cost (Child Node)" into budget item nodes; it extracts "The construction team node and the construction worker cost node have a hierarchical relationship" through a semantic rule parser to form the first project edge, and extracts "Tower Crane Rental Cost = Number of Tower Cranes × Monthly Rent × 18 Months" through parameter mapping to form the second project edge, and finally imports it into the Neo4j graph database to construct a knowledge graph.

[0057] Based on the above steps, the existing model of static budget storage is changed. A knowledge graph is used to establish a network linking project information and budget data, providing a structured data carrier for dynamic budget response when subsequent project changes occur.

[0058] Step 203: The project budget management device transforms the main project nodes and budget item nodes into low-dimensional vectors through semantic embedding, generates weight vectors for project edges through relational embedding, and constructs an adjacency matrix by combining the graph topology structure to form the input data for the GNN model.

[0059] Semantic embedding is the process of converting the textual attributes of nodes into low-dimensional, numerical vectors while preserving the semantic information of the nodes; relation embedding is the process of converting the association rules (logical relations, computational relations) of project edges into weighted vectors to reflect the type and intensity of the relations; the adjacency matrix is ​​a matrix that reflects the connection relationships between nodes in the graph topology, and the element values ​​contain the fusion information of node vectors and edge weight vectors.

[0060] In this embodiment, the vector dimension is determined based on the number of project nodes. A semantic embedding algorithm is used to encode the text attributes of the main project nodes and budget item nodes into low-dimensional vectors of the corresponding dimensions. Then, a relation embedding algorithm is used to parse the project edge rules and generate weight vectors according to the influence strength of the relations. Finally, based on the low-dimensional vectors, the row / column weights of the core node "total budget item node" are doubled, and matrix elements are filled to complete the adjacency matrix construction.

[0061] Based on the above steps, unstructured node text and edge relationships are transformed into machine-processable numerical data, while fully preserving node semantics and edge relationship features, providing high-quality input data for GNN models and avoiding model training bias caused by data format incompatibility.

[0062] Step 204: The project budget management device uses a heterogeneous graph attention network architecture to iteratively train the input data of the GNN model to obtain a project budget-specific GNN training model.

[0063] Among them, the heterogeneous graph attention network is a graph neural network architecture for heterogeneous graphs. It assigns different attention weights to different nodes and edges through the attention mechanism, thereby improving the model's ability to capture key relationships; the project budget-specific GNN training model refers to a GNN model adapted to the project budget scenario.

[0064] In this embodiment, a heterogeneous graph attention network architecture is used to learn the weight relationships between nodes and their neighbors. First, the low-dimensional vectors and weight vectors in the input data of the GNN model are adapted using a node type mapping matrix and an edge relationship mapping matrix, respectively, to provide an adapted data foundation for learning weight relationships. Then, the adapted low-dimensional vectors are linearly transformed to unify the feature dimensions. The attention coefficients between nodes are calculated in combination with the weight vectors, and the features of neighboring nodes are aggregated based on the attention coefficients to update the features of the current node. In this process, the learning of the weight relationships between nodes and their neighbors is completed. Finally, a dual-objective loss function including regression loss and classification loss is used to optimize and train the model, ensuring that the learned weight relationships between nodes and their neighbors can accurately support the predictive capabilities of the project budget-specific GNN training model.

[0065] Based on the above steps, the attention mechanism of the heterogeneous graph attention network is used to accurately capture the key influencing factors in the project budget. At the same time, the dual-objective loss function balances the accuracy and compliance of budget adjustment, significantly improving the practicality and reliability of the model.

[0066] Step 205: The project budget management device uses a project budget-specific GNN training model to predict changes in the knowledge graph of basic project information and outputs budget adjustment parameters.

[0067] Among them, the budget adjustment parameters are used to adjust the budget item nodes and project edges in the knowledge graph.

[0068] In this embodiment, when the basic information of the project changes, the corresponding project main node in the knowledge graph is updated first; based on the updated knowledge graph, the new GNN model input data is regenerated by selecting "real-time generation" or "periodic generation" as needed; the new GNN model input data is input into the project budget-specific GNN training model to predict budget adjustment needs, and "proportional parameter" or "absolute value parameter" is selected as the adjustment parameter type according to budget management needs.

[0069] It should be noted that model predictions are not the absolute basis for implementation. For major project changes, a manual review process is required to avoid budget overruns due to model bias. At the same time, the data sources and parameter settings during the prediction process should be recorded for subsequent compliance traceability.

[0070] As an example, due to increased construction difficulty, the project's basic information states that "the construction period has been extended from 18 months to 20 months." The project budget management device first updates the "project construction period nodes" in the knowledge graph (changing them from 18 months to 20 months), and then regenerates the GNN input data based on the updated knowledge graph (such as updating the low-dimensional vectors of the construction period nodes and adjusting the weight vectors of the relevant computational relationships). The new input data is then fed into the trained GNN model, and the model predicts and outputs budget adjustment parameters: "Tower crane rental costs increased by 11.1%" and "Construction worker costs increased by 11.1%".

[0071] Based on the above steps, the problem of budget response lag during project changes in existing technologies is solved. By predicting budget adjustment parameters in real time through the model, the project management personnel can be provided with fast and accurate decision support, while the adjustment risk is reduced by confidence level identification.

[0072] Step 206: The project budget management device updates the knowledge graph based on the budget adjustment parameters to complete the project budget management.

[0073] Among them, the project budget execution snapshot refers to a static data snapshot that records the project budget status after the knowledge graph is updated. It includes information such as updated node values, edge relationship rules, adjustment time, and adjustment basis, and is used for budget execution tracking and compliance auditing.

[0074] In this embodiment, the calculation relationship between the values ​​of budget item nodes and project edges in the knowledge graph is updated, and the "adjustment record attributes" (record adjustment time, value before adjustment, and source of adjustment parameters) of the nodes and the relationship version attributes of the edges are updated synchronously. The frequency of project budget execution snapshot generation is set according to management needs, and the snapshots are stored by selecting "associated storage with knowledge graph" or "independent archiving storage". The budget change trend can be viewed by comparing snapshot versions to complete project budget management.

[0075] It should be noted that the update records of the knowledge graph must be immutable. The authenticity and validity of the records can be ensured through technologies such as timestamps and hash values, so as to provide a complete adjustment traceability chain during compliance audits and avoid compliance risks caused by lost or tampered records.

[0076] As an example, based on the adjustment parameters of "tower crane rental cost increased by 11.1%" and "construction worker cost increased by 11.1%", the office building construction project updates the values ​​of the "tower crane rental cost node" (90,000 yuan → 100,000 yuan) and the "construction worker cost node" (2,880,000 yuan → 3,200,000 yuan) in the knowledge graph. At the same time, it updates the corresponding calculation relationship edges (changing the construction period parameter from 18 to 20) and adds adjustment records to the nodes. After the update is completed, a project budget execution snapshot is generated in real time. The snapshot contains all current node values, edge relationship rules and adjustment records, and is stored in the enterprise audit database, supporting the finance department to query and audit at any time.

[0077] Based on the above steps, dynamic updates to the project budget are achieved. Compliance traceability is accomplished through knowledge graph update records and execution snapshots, solving the problems of static technology and lack of compliance traceability in the existing technology, and achieving dynamic and refined management of the project budget.

[0078] Based on the above technical solution, this approach overcomes the shortcomings of existing technologies that only collect single total budget data by acquiring comprehensive first-dimensional information, including basic project information, budget item information, and budget template information. It then breaks the static storage mode of budget by constructing a knowledge graph using nodes and project edges, transforming the data into machine-processable GNN model input data through semantic / relational embedding. Subsequently, a specialized model is trained using a heterogeneous graph attention network to accurately capture key budget influencing factors and quickly predict budget adjustment parameters when basic project information changes. Finally, the knowledge graph is updated and an execution snapshot is generated, solving the problem of delayed budget response when projects change and enabling compliance traceability through node / edge records and snapshots. This achieves dynamic and refined management of project budgets, effectively addressing the technical problems of static budget management, untimely budget response, and lack of compliance traceability in existing project budget management technologies.

[0079] In one possible implementation, combining the above... Figure 2 ,like Figure 3As shown, in step 202 above, the project budget management device integrates basic project information into project main nodes, integrates budget item information into budget item nodes, and integrates budget template information into project edges between various nodes. The process of constructing a knowledge graph based on project main nodes, budget item nodes, and project edges can be specifically implemented through the following steps 301-303:

[0080] Step 301: The project budget management device performs structured processing on the basic project information to form the main project node, extracts the budget item information, and forms the budget item node by associating the parent / child node dependencies of the items.

[0081] Among them, structured processing is the process of converting non-standardized basic information into fixed field formats; the project main node is the structured carrier of the project's basic information, containing a unique identifier ID and attribute fields; the parent / child node dependency relationship of the subject is the hierarchical relationship between budget subjects, reflecting the subject inclusion logic.

[0082] In this embodiment, the project budget management device connects to the API interface of the enterprise project management system or financial system, standardizes the format of basic project information according to preset field mapping rules, extracts key content from manually entered unstructured information using NLP technology to complete the structured fields, and finally generates a project main node with a unique ID; at the same time, it extracts budget item information from the configuration database, and generates budget item nodes with hierarchical attributes by matching parent / child node dependencies through field matching according to preset item hierarchy rules.

[0083] It should be noted that structured processing must ensure the consistency of field mapping to avoid missing node attributes due to differences in data formats across different systems; the parent / child node dependency relationship of accounts must be aligned with the enterprise's budget management standards to ensure that the hierarchical logic complies with financial control requirements.

[0084] Based on the above steps, the scattered and non-standardized basic project information and budget item information are transformed into structured nodes, clarifying the hierarchical logic of the items, thus solving the problem of fragmented budget data and lack of hierarchical association in the existing technology.

[0085] Step 302: The project budget management device uses a semantic rule parser to extract the logical relationship between budget template information and budget item information to form the first project edge, and extracts the calculation relationship between budget template information and budget item information through parameter mapping to form the second project edge.

[0086] Among them, the semantic rule parser is a tool used to identify logical relationships in text-based rules; the first item edge is a link representing non-numerical relationships between nodes, such as logical relationships like "belonging" and "fitting"; parameter mapping is the process of decomposing numerical calculation rules in the template into field relationships; the second item edge is a link representing numerical calculation relationships between nodes, including calculation factors and formula logic.

[0087] In this embodiment, the project budget management device first calls the semantic rule parser to load the preset logical relationship recognition rules, performs text parsing on the budget template information, extracts the logical association between the template and the budget item (such as "Template 1 adapts to the labor cost item" and "Template 1 adapts to the server rental cost item"), and generates a first project edge with logical relationship labels; then, through the parameter mapping tool, the calculation rules in the template (such as "server rental cost = number of servers × monthly rent × construction period") are decomposed into field associations ("number of servers - from the main project node", "monthly rent - from the template parameter", "construction period - from the main project node"), determines the association between the calculation factor and the budget item, and generates a second project edge containing calculation rule attributes, and the second project edge needs to be labeled with the source node ID of the calculation factor.

[0088] As an example, for the "office building construction project template", the semantic rule parser analyzes the description "applicable subjects include construction worker cost and tower crane rental cost" in the template, extracts the logical relationship between "template 2 - adapt - construction worker cost subject" and "template 2 - adapt - tower crane rental cost subject", and forms the first project edge; through parameter mapping, it decomposes the rule "tower crane rental cost = number of tower cranes × 20,000 yuan / month × construction period" in the template, associates "number of tower cranes (project main node field)", "20,000 yuan / month (template parameter)", "construction period (project main node field)" with "tower crane rental cost subject", and generates the second project edge containing the calculation rule.

[0089] Based on the above steps, a two-way association (logic and calculation) is established between the budget template and the budget item, which improves the limitations of the existing technology where the template and item are not related and the calculation rules are isolated. This provides a link for subsequent dynamic budget updates and ensures that the budget calculation logic is traceable when the project changes.

[0090] Step 303: The project budget management device imports the project main node, budget item node, first project edge, and second project edge into the graph database, and establishes topological connections through the mapping relationship between the first and last nodes of the nodes and edges to form a complete project budget knowledge graph.

[0091] Among them, the graph database is a database used to store the relationship between nodes and edges; topological connection is the process of establishing a network of relationships between nodes by matching the IDs of the first and last nodes of the edge with the unique IDs of the nodes; the project budget knowledge graph is a visual network of relationships formed by integrating the main nodes of the project, the nodes of the budget items, and the two types of project edges, including node attributes, edge relationship attributes, and topological structure.

[0092] In this embodiment, the project budget management device first imports the project main nodes and budget item nodes in batches according to the graph database import format (such as JSON format, including node ID, node type, and attribute fields). Then, it imports the first project edge and the second project edge (including edge ID, relationship type, first and last node IDs, and edge attributes, such as the "adaptation" tag of the first project edge and the calculation formula of the second project edge). Subsequently, through the graph database topology construction tool, the association between nodes is automatically established based on the mapping relationship between the first and last node IDs of the edge and the unique ID of the node. At the same time, indexes are created for the nodes and edges to ensure the efficiency of subsequent queries and updates. Finally, a project budget knowledge graph with a complete topology structure is generated, and the graph supports visualization.

[0093] As an example, the following nodes of an office building construction project are imported into the Neo4j graph database: "Project Main Node (ID: P001, Attributes: Construction Period 20 Months, Number of Tower Cranes 2)", "Budget Subject Node (ID: C001, Attributes: Tower Crane Rental Cost)", "First Project Edge (ID: E101, Relationship: Template 2-Adaptation-C001, First and Last IDs: T002-C001)", and "Second Project Edge (ID: E201, Relationship: Calculation-Association, First and Last IDs: P001-C001, Attributes: Calculation Rule = Number of Tower Cranes × 20,000 Yuan / Month × Construction Period)". The system establishes topological connections between "P001-E201-C001" and "T002-E101-C001" through ID matching, ultimately forming a complete knowledge graph containing project foundations, budget subjects, and template associations.

[0094] Based on the above steps, a network of all elements of the project budget was constructed, transforming static node data into a dynamic and interconnected graph structure. This solves the problem of static storage and lack of connection of budget data in existing technologies, and provides a structured carrier for dynamic budget prediction and compliance traceability when the basic information of the project changes, ensuring the dynamism and traceability of budget management.

[0095] Based on the above technical solution, by structuring the basic project information and associating it with the hierarchical relationship of budget items, scattered and non-standardized project data is transformed into nodes with unique identifiers and clear attributes, solving the problems of hierarchical and chaotic budget data in existing technologies. By extracting the logical and computational relationships between budget templates and items to generate two types of project edges, the limitation of isolated templates and items is broken, providing a relational basis for dynamic budget calculation. Finally, the nodes and edges are imported into a graph database to construct a topological knowledge graph, transforming static data into a dynamic network that can be associated. This not only makes up for the defects of static storage and lack of association in existing technologies, but also constructs a carrier for the association of all elements of the project budget, effectively supporting the transformation of project budget from static to dynamic and refined management.

[0096] In one possible implementation, combining the above... Figure 2 ,like Figure 4 As shown, step 203 above transforms the main project nodes and budget item nodes into low-dimensional vectors through semantic embedding, generates weight vectors for project edges through relational embedding, and constructs an adjacency matrix in conjunction with the graph topology structure. This process can be specifically implemented through the following steps 401-403:

[0097] Step 401: The project budget management device uses a hybrid semantic embedding algorithm of TransE and BERT to encode and transform the text attributes of the project main node and budget item node to obtain a fixed-dimensional low-dimensional vector.

[0098] Among them, TransE is a classic algorithm for entity embedding in knowledge graphs, which models entity associations by mapping entities and relations to a low-dimensional space; BERT is a pre-trained language model that can extract deep semantic features of text; text attributes refer to the textual description information of the project main node and the budget subject node; low-dimensional vectors are numerical vectors that convert high-dimensional text attributes into fixed-dimensional ones, which are convenient for subsequent machine processing.

[0099] In this embodiment, the project budget management device first preprocesses the node text attributes, including removing meaningless stop words and segmenting by domain dictionary. The BERT model is then used to encode the preprocessed text to obtain an initial vector containing contextual semantics. Next, the TransE algorithm is introduced to fine-tune the initial vector based on the relationships between nodes in the knowledge graph, optimizing the distribution of nodes in the low-dimensional space and ensuring that nodes with similar semantics or close relationships are closer in the vector space. Finally, the vectors processed by the two algorithms are normalized to obtain a low-dimensional vector with a fixed dimension, and the low-dimensional vector dimension of all nodes is kept consistent to meet the input requirements of subsequent models.

[0100] As an example, for the "Project Main Node (text attributes: Project Name = Office Building Construction, Construction Period = 20 Months)" and the "Budget Item Node (text attributes: Item Name = Tower Crane Rental Cost, Description = Tower crane rental fee for project construction)", the text is first preprocessed (splitting "Office Building Construction" into "Office Building - Construction", removing stop words such as "for" and "fee"). Then, BERT-base is used to extract the initial vector, and the TransE algorithm is used for fine-tuning (with the relationship between "Project Main Node - Calculation Association - Tower Crane Rental Cost Node" as a constraint). Finally, a 128-dimensional low-dimensional vector is generated, in which the distance between the "Tower Crane Rental Cost" node vector and the "Equipment Rental Cost" node vector in space is significantly smaller than the distance between them and the "Labor Cost" node vector.

[0101] Based on the above steps, the unstructured text attributes of nodes are transformed into structured, machine-processable low-dimensional vectors. At the same time, by using a hybrid algorithm, both text semantics and node association relationships are taken into account, which solves the shortcomings of single embedding algorithms (such as BERT alone cannot utilize graph associations, and TransE alone is insufficient for semantic extraction).

[0102] Step 402: The project budget management device uses the RotatE and DistMult hybrid relation embedding algorithm to parse the attributes, convert them into features, and establish relations for the rules of the project edges, generating a weight vector that distinguishes the relation type and the intensity of influence.

[0103] Among them, RotatE is an algorithm for embedding relationships in knowledge graphs, which can model relationships as rotation operations in a low-dimensional space; DistMult is an embedding algorithm that models relationships through matrix multiplication, which can strengthen the influence of relationships on entity associations; the rules of item edges refer to the logical rules of the first item edge and the calculation rules of the second item edge; parsing attributes is to extract the type of item edge, the associated node ID, and the rule parameters; transforming features is to transform the parsed attributes into numerical features; the weight vector is a fixed-dimensional numerical vector that contains the relationship type identifier and the influence strength.

[0104] In this embodiment, the project budget management device extracts "relationship type = logical association", "association node = template node - subject node", and "rule description = template-adapted subject" for the first project edge; and extracts "relationship type = computational association", "association node = project main node - subject node", and "computation parameters = number of tower cranes, monthly rent (or construction period)" for the second project edge. The relationship type is converted into one-hot encoding, the computational parameter weights are converted into numerical features, and the associated node ID is mapped to a feature index. The RotatE algorithm is used to model the relationship as a rotation operation, handling complex relationship types such as "template-adapted subject" (symmetric relationship, A adapts to B, then B adapts to A) and "project main computational association subject" (asymmetric relationship, A computationally associates B ≠ B computationally associates A). The DistMult algorithm is used to strengthen the influence of the relationship on the node association through multiplication of the relationship matrix and the node vector. Finally, the results of the two algorithms are fused and normalized to generate a weight vector with the same dimension as the low-dimensional vector of the node.

[0105] Based on the above steps, the non-numerical rules of the project edge are transformed into weight vectors with relation type and influence intensity. The hybrid algorithm effectively processes complex relation types and accurately represents the degree of influence, solving the problem that a single relation embedding algorithm cannot simultaneously distinguish relation type and quantify influence intensity.

[0106] Step 403: The project budget management device uses a knowledge graph topology structure, with low-dimensional vectors as the row / column basis, combined with weight vectors, to fill the adjacency matrix elements, forming an adjacency matrix that includes node semantics, edge relationship weights, and topological connections.

[0107] Among them, the graph topology refers to the connection relationship between nodes and edges in the project budget knowledge graph constructed in step 303; the adjacency matrix is ​​a matrix that reflects the connection status and association strength between nodes, with rows and columns corresponding to all nodes in the knowledge graph, and the matrix element values ​​need to be integrated with the node low-dimensional vector and the edge weight vector.

[0108] In this embodiment, the project budget management device first determines the dimension of the adjacency matrix: the matrix dimension is equal to the total number of project main nodes and budget item nodes in the knowledge graph. The rows and columns are arranged in ascending order by node ID, and node ID indexes are added to the matrix rows / columns to facilitate subsequent node location. Next, all project edges in the knowledge graph are traversed, and the first and last node IDs of each edge are recorded. If there is an edge connection at the corresponding row and column position in the matrix, elements are filled. The filling method is to fuse the low-dimensional vector of the first node, the low-dimensional vector of the last node, and the edge weight vector according to a preset rule (such as first taking the mean of the node vectors, and then multiplying them element-wise with the weight vector, the formula is: element value = (first node vector + last node vector) / 2 × weight vector) to obtain the value of the matrix element; if there is no corresponding project edge between nodes, the element value at the corresponding position in the matrix is ​​set to 0. Finally, the filled matrix is ​​normalized to ensure numerical stability during subsequent GNN model training, ultimately forming an adjacency matrix containing node semantics, edge relationship weights, and topological connections.

[0109] Based on the above steps, the node semantics, edge relation weights and topological structure of the knowledge graph are integrated into a unified adjacency matrix, which is transformed into a data format that can be directly input into the GNN model. This solves the problem that the graph information (semantics, relations, topology) is scattered in the existing technology and cannot be directly used for model training.

[0110] In one possible implementation, combining the above... Figure 2 ,like Figure 4 As shown, the process of iteratively training the GNN model input data using a heterogeneous graph attention network architecture in step 204 can be specifically implemented through the following steps 404-406:

[0111] Step 404: The project budget management device adapts the low-dimensional vectors and weight vectors in the input data of the GNN model using the node type mapping matrix and the edge relationship mapping matrix, respectively.

[0112] Among them, the node type mapping matrix is ​​a matrix used to adapt low-dimensional vectors of different types of nodes, and its dimension is "target adaptation dimension × original low-dimensional vector dimension"; the edge relationship mapping matrix is ​​a matrix used to adapt edge weight vectors of different types of items, and its dimension matches that of the node type mapping matrix; the input data of the GNN model includes the low-dimensional vectors, weight vectors and adjacency matrix generated in steps 401-403.

[0113] In this embodiment, the project budget management device first constructs a node type mapping matrix and an edge relationship mapping matrix based on the node type and edge relationship type in the knowledge graph, respectively. The matrix parameters are initialized through pre-training using historical project data. Then, it traverses the nodes and edges in the adjacency matrix, determines the node type based on the node ID corresponding to the row / column of the adjacency matrix, calls the corresponding node type mapping matrix to perform a linear transformation on the low-dimensional vector, and simultaneously determines the edge relationship type based on the edge ID corresponding to the element of the adjacency matrix, calls the corresponding edge relationship mapping matrix to adapt the weight vector. During the adaptation process, the consistency of the node-edge association in the adjacency matrix needs to be verified synchronously to ensure that the adapted low-dimensional vector and weight vector can accurately correspond to the topological position of the adjacency matrix. Finally, it outputs an adapted vector set that perfectly matches the structure of the adjacency matrix.

[0114] As an example, the following formula is used to adapt the low-dimensional vectors and weight vectors in the input data of the GNN model using type mapping matrices and relation mapping matrices respectively:

[0115]

[0116] in, It is a low-dimensional vector. A node type mapping matrix, The adapted low-dimensional vector, For the weight vector, This is the edge relation mapping matrix. This is the adapted weight vector.

[0117] Based on the above steps, the feature adaptation problem of different types of nodes and edges in heterogeneous graphs is solved by using a typed mapping matrix. At the same time, it achieves accurate association between the adapted vector and the adjacency matrix, avoiding the chaos of model input data caused by differences in node / edge types.

[0118] Step 405: The project budget management device performs a linear transformation on the adapted low-dimensional vector to unify the feature dimension, calculates the attention coefficient between nodes in combination with the weight vector, and updates the current node feature by aggregating the features of neighboring nodes based on the attention coefficient.

[0119] Among them, linear transformation is an operation that transforms the low-dimensional vectors of different types of nodes after adaptation into feature vectors of the same dimension through a preset linear transformation matrix, ensuring the dimensionality consistency of subsequent attention calculations; attention coefficient is a value that characterizes the degree of influence of neighboring nodes on the features of the current node, and needs to be calculated in conjunction with the adapted weight vector; neighboring nodes refer to nodes that are connected to the current node by edges, determined by the non-zero positions of the adjacency matrix elements.

[0120] In this embodiment, a linear transformation is performed on the adapted low-dimensional vector to obtain a low-dimensional vector that satisfies the following formula:

[0121]

[0122] in, It is a linear transformation matrix. It is a low-dimensional vector after linear transformation;

[0123] The attention coefficient is calculated according to the following formula:

[0124]

[0125] in, Attention coefficient (·) is a function that normalizes the intermediate results for all neighboring nodes u. (·) represents an activation function with leakage. For attention vectors, Let be the low-dimensional vector after the linear transformation of node u, and || be the vector concatenation operation. For normalization terms;

[0126] Aggregate the features of neighboring nodes to update the features of the current node. The updated node features satisfy the following formula:

[0127]

[0128] in, For the updated node features, (·) is the activation function. Let v be the set of neighboring nodes.

[0129] Based on the above steps, the problem of incompatibility of heterogeneous node features was solved by unifying the feature dimension through linear transformation. The attention coefficient of the weight vector was combined to accurately capture the key associations between nodes. The node feature expression was updated by neighbor aggregation based on the adjacency matrix, which effectively improved the GNN model's ability to extract project budget-related features and provided high-quality node feature data for subsequent loss function optimization.

[0130] Step 406: The project budget management device uses a dual-objective loss function, which includes regression loss and classification loss, to optimize and continuously update the node features to train the GNN model.

[0131] The bi-objective loss function is a composite loss function that includes regression loss and classification loss. The regression loss is used for budget adjustment forecasting, and the classification loss is used for time allocation compliance verification.

[0132] In this embodiment, the project budget management device first extracts training samples from the adjacency matrix of the historical project knowledge graph. It then uses the budget adjustment data corresponding to the node features before and after the feature update in the adjacency matrix to calculate the regression loss. The device uses the work hour allocation data associated with the nodes in the adjacency matrix as work hour allocation compliance samples to calculate the classification loss. A total loss function is constructed based on the regression loss and classification loss. It then initiates iterative training. In each training round, steps 404-405 (adaptation, feature transformation, and update) are performed based on the adjacency matrix. Predicted values ​​are obtained through forward computation, substituted into the total loss function, and then the backpropagation algorithm is used to optimize parameters such as the linear transformation matrix, attention vector, and node / edge type mapping matrix using gradient descent, while simultaneously updating node features. The iteration is repeated until the total loss function is less than a preset threshold or the number of iterations reaches its upper limit, at which point training stops, resulting in a converged project budget-specific GNN training model.

[0133] As an example, the bi-objective loss function satisfies the following formula:

[0134]

[0135] in, To regress the loss, Adjusting the sample size for budget purposes Adjust parameters to forecast budget. To represent the actual adjustment parameters, i is the index of the budget adjustment sample. These are the weighting coefficients; The classification loss satisfies the following formula:

[0136]

[0137] in, Allocate compliant sample numbers to working hours. For authenticity and compliance labeling, Let j be the compliance probability predicted by the model, and j be the compliance sample index for work hours.

[0138] Based on the above technical solution, the text attributes and edge rules of project nodes are transformed into fixed-dimensional low-dimensional vectors and weight vectors through semantic embedding and relation embedding. Combined with the knowledge graph topology, an adjacency matrix containing node semantics, edge relation weights and topological connections is constructed, which solves the problem that unstructured data is difficult for machines to process. Then, based on the adjacency matrix, the node / edge type mapping matrix is ​​adapted to heterogeneous data, the node features are optimized by linear transformation and attention coefficient calculation, and the GNN model is iteratively trained by a dual-objective loss function. This ensures that the model accurately captures the project budget-related features and balances the accuracy of budget adjustment prediction with compliance verification capabilities.

[0139] In one possible implementation, combining the above... Figure 2 ,like Figure 5 As shown, step 205 above, which uses a project budget-specific GNN-trained model to predict changes in the knowledge graph of project basic information and outputs budget adjustment parameters, can be specifically implemented through the following steps 501-503:

[0140] Step 501: When the basic information of a project changes, the project budget management device updates the corresponding project main node in the knowledge graph.

[0141] In this embodiment, the project budget management device monitors the API interface of the enterprise project management system in real time to capture changes in basic project information, triggering a knowledge graph update process. It matches the corresponding project main node ID in the knowledge graph based on the field identifier of the changed information, then modifies the attribute value of the node and adds a change log attribute to the node. During the update process, it is necessary to verify the mapping relationship between the changed fields and the node attributes to ensure that there are no field mismatches, and simultaneously check whether the project edge associated with the node needs to be marked as pending update, in order to prepare for the generation of subsequent input data.

[0142] As an example, due to adjustments in the construction plan for an office building project, the "Construction Period" field in the project management system was changed from 18 months to 20 months. After the project budget management device listens for this change, it matches node P001 (the main project node with the attribute "Construction Period = 18 months") in the knowledge graph, updates its "Construction Period" attribute to 20 months, and adds a change log ("Change time, Zhang San, due to adjustments in the construction plan, the construction period has been changed from 18 months to 20 months"). At the same time, it marks the second project edge E201 associated with node P001 as "pending update".

[0143] Based on the above steps, real-time synchronization between changes in basic project information and the knowledge graph is achieved, avoiding the lag problem in existing technologies. At the same time, the change log provides the original basis for subsequent compliance traceability, ensuring that the knowledge graph can dynamically reflect the actual status of the project.

[0144] Step 502: The project budget management device regenerates new GNN model input data based on the updated knowledge graph.

[0145] In this embodiment, the project budget management device first re-executes the semantic embedding process of step 401 on the updated project main node, and generates a new low-dimensional vector based on the new attribute value ("construction period = 20 months"); then, it re-executes the relation embedding process of step 402 on the project edge marked as "to be updated", and adjusts the weight vector in combination with the updated association relationship of the node; finally, it reconstructs the adjacency matrix of step 403 based on the updated node low-dimensional vector and edge weight vector.

[0146] Based on the above steps, the consistency between the input data of the GNN model and the updated knowledge graph is ensured, avoiding model prediction bias due to data lag.

[0147] Step 503: The project budget management device inputs the new GNN model input data into the project budget-specific GNN training model to predict budget adjustment needs and outputs budget adjustment parameters.

[0148] In this embodiment, the project budget management device inputs the newly generated GNN model input data into the project budget-specific GNN training model, initiates the model forward computation process, and performs attention aggregation on the node features and edge relationship weights in the input data through the heterogeneous graph attention layer of the model, updating the node feature representation; then, through the output layer, it predicts the budget adjustment needs based on the aggregated features, the regression head outputs the adjustment parameters of each budget item node, and the classification head outputs the confidence level of the adjustment parameters; before output, the rationality of the adjustment parameters needs to be verified, and parameters with confidence levels lower than a preset threshold are marked as "awaiting manual confirmation"; finally, the structured budget adjustment parameters are output, including the adjustment object ID (node / edge), adjustment type (proportion / absolute value), adjustment value, confidence level, and prediction time.

[0149] Based on the above steps, the system enables rapid prediction of budget adjustment needs after changes in basic project information. The output adjustment parameters accurately correspond to the adjustment objects in the knowledge graph, solving the problem of budget response lag when project changes occur in existing technologies. At the same time, through confidence verification and manual review mechanisms, the system balances prediction efficiency and adjustment security.

[0150] In one possible implementation, combining the above... Figure 2 ,like Figure 5 As shown, the process of updating the knowledge graph based on budget adjustment parameters to complete project budget management in step 206 above can be specifically implemented through the following steps 504-505:

[0151] Step 504: The project budget management device updates the relevant project nodes of budget item information and the relevant project edges of budget template information in the knowledge graph based on the budget adjustment parameters.

[0152] In this embodiment, the project budget management device first parses the budget adjustment parameters to determine the budget item node ID and project edge ID that need to be updated. For budget item nodes, the corresponding attribute values ​​are modified according to the adjustment parameter type. If it is a proportional adjustment, the new amount is calculated as original amount × (1 + adjustment ratio), and the node's "current budget" attribute is updated. If it is an absolute value adjustment, the adjustment value is directly added to the original amount, and the node's "adjustment record" attribute is updated. For project edges, if the adjustment parameters involve changes in calculation rules, the "calculation formula" attribute of the second project edge is updated. If it involves changes in logical association, the "association status" attribute of the first project edge is updated.

[0153] Based on the above steps, dynamic updates of the knowledge graph are achieved, enabling the graph to reflect the adjustment status of the project budget in real time, thus solving the technical problem of static storage of budgets in existing technologies.

[0154] Step 505: The project budget management device generates a project budget execution snapshot based on the updated project nodes and project edges, thus completing project budget management.

[0155] Among them, the project budget execution snapshot is a static data set that records the complete state of the updated knowledge graph. It includes the current attributes of all project main nodes and budget item nodes, the current rules of the first / second project edges, the adjacency matrix topology, and the generation timestamp. The snapshot cannot be modified and can be used as the basis for budget execution tracking and compliance auditing.

[0156] In this embodiment, after the knowledge graph is updated, the project budget management device automatically triggers the project budget execution snapshot generation process. First, it extracts the complete attributes of all current nodes and the attributes of all project edges from the graph database. Then, it organizes the data according to a preset format and adds snapshot metadata. Next, it stores the snapshot data in the enterprise audit database, supporting two storage methods: associated storage with the graph database and independent archiving storage. Finally, it generates a snapshot query interface for the finance department, project management department, and audit department to call. It supports querying snapshot details by snapshot ID, project ID, and time range, and records the snapshot query log.

[0157] Based on the above steps, the generated project budget execution snapshot forms a "static audit node" for budget management, solving the problems of unrecorded budget adjustments and lack of compliance traceability. At the same time, through the snapshot query and audit functions, a closed-loop management of project budgets from dynamic adjustment to static archiving is realized, meeting the needs of corporate financial control and external audit, and ultimately achieving refined and compliant management of project budgets.

[0158] Based on the above technical solution, the knowledge graph's corresponding project main nodes are updated in real time when the project's basic information changes, solving the problem of asynchronous data between project changes and the knowledge graph. Input data for the GNN model is regenerated based on the updated knowledge graph, ensuring the accuracy of the model's predictions. A project-specific GNN is used to train the model, outputting accurate and confident budget adjustment parameters to quickly respond to budget adjustment needs. Budget item node attributes and project edge rules are updated synchronously based on the parameters, enabling dynamic matching of the knowledge graph with the budget status. Finally, a project budget execution snapshot containing complete node / edge information and tamper-proof verification values ​​is generated, achieving compliance traceability. This effectively solves the problems of static project budgets, delayed change response, and lack of compliance traceability in existing technologies, achieving the goals of dynamic, refined, and compliant project budget management.

[0159] The foregoing mainly describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, such as a project budget management device, includes at least one of the hardware structure and software module corresponding to the execution of each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in a hardware or computer software-driven hardware manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0160] This application embodiment can divide the project budget management device into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0161] When using integrated units, Figure 6 A possible structural schematic diagram of the project budget management device (referred to as project budget management device 60) involved in the above embodiments is shown. The project budget management device 60 includes a processing unit 601 and a communication unit 602, and may also include a storage unit 603. Figure 6 The structural diagram shown can be used to illustrate the structure of the project budget management device involved in the above embodiments.

[0162] when Figure 6The schematic diagram shown illustrates the structure of the project budget management device involved in the above embodiments. The processing unit 601 is used to control and manage the operation of the project budget management device, the communication unit 602 is used for the project budget management device to communicate with other devices, and the storage unit 603 is used to store the program code and data of the project budget management device.

[0163] For example, communication unit 602 is used to acquire the first information of the project, which includes basic project information, budget item information and budget template information; budget model information is used to characterize the relationship between the basic project information and / or budget item information.

[0164] Processing unit 601 integrates basic project information into project main nodes, budget item information into budget item nodes, and budget template information into project edges between nodes. It then constructs a knowledge graph based on the project main nodes, budget item nodes, and project edges. Each project edge includes a first project edge and a second project edge. The first project edge represents the logical relationship between nodes, and the second project edge represents the calculation relationship of the project budget between nodes. The processing unit transforms the project main nodes and budget item nodes into low-dimensional vectors through semantic embedding, and generates weight vectors for the project edges through relational embedding. It then constructs an adjacency matrix based on the graph topology to form the input data for the GNN model. A heterogeneous graph attention network architecture is used to iteratively train the GNN model input data to obtain a project budget-specific GNN training model. The project budget-specific GNN training model is used to predict changes in the knowledge graph of basic project information and outputs budget adjustment parameters. These adjustment parameters are used to adjust the budget item nodes and project edges in the knowledge graph. Finally, the knowledge graph is updated based on the budget adjustment parameters to complete project budget management.

[0165] In one possible implementation, the processing unit 601 is further configured to integrate basic project information into project main nodes, integrate budget item information into budget item nodes, and integrate budget template information into project edges between nodes. Based on the project main nodes, budget item nodes, and project edges, a knowledge graph is constructed, including: structuring the basic project information to form project main nodes, extracting budget item information, and associating parent / child node dependencies to form budget item nodes; using a semantic rule parser to extract the logical relationship between budget template information and budget item information to form a first project edge; extracting the computational relationship between budget template information and budget item information through parameter mapping to form a second project edge; importing the project main nodes, budget item nodes, first project edges, and second project edges into a graph database, and establishing topological connections through the mapping relationship between the first and last nodes of nodes and edges to form a complete project budget knowledge graph.

[0166] In one possible implementation, the processing unit 601 is further configured to convert the project main node and budget item node into low-dimensional vectors through semantic embedding, generate weight vectors for project edges through relation embedding, and construct an adjacency matrix in combination with the graph topology structure. This includes: using a hybrid semantic embedding algorithm of TransE and BERT to encode and convert the text attributes of the project main node and budget item node into fixed-dimensional low-dimensional vectors; using a hybrid relation embedding algorithm of RotatE and DistMult to parse the rules of project edges, convert them into features, and build relations to generate weight vectors that distinguish relation types and influence strengths; and based on the graph topology structure of the knowledge graph, using the low-dimensional vectors as the row / column basis, combining the weight vectors to fill the elements of the adjacency matrix to form an adjacency matrix containing node semantics, edge relation weights, and topological connections.

[0167] In one possible implementation, the processing unit 601 is further configured to iteratively train the input data of the GNN model using a heterogeneous graph attention network architecture, including: adapting the low-dimensional vectors and weight vectors in the input data of the GNN model through node type mapping matrix and edge relation mapping matrix respectively; performing linear transformation on the adapted low-dimensional vectors to unify the feature dimension, calculating the attention coefficients between nodes in combination with the weight vectors, and aggregating the features of neighboring nodes to update the features of the current node based on the attention coefficients; and using a dual-objective loss function including regression loss and classification loss to optimize and continuously update the node features to train the GNN model, where the regression loss is used for budget adjustment prediction and the classification loss is used for compliance verification of work hour allocation.

[0168] In one possible implementation, the low-dimensional vectors and weight vectors in the GNN model input data are adapted using a type mapping matrix and a relation mapping matrix, respectively, satisfying the following formula:

[0169]

[0170] in, It is a low-dimensional vector. A node type mapping matrix, The adapted low-dimensional vector, For the weight vector, This is the edge relation mapping matrix. This is the adapted weight vector.

[0171] In one possible implementation, the processing unit 601 is further configured to, in conjunction with the first aspect described above, perform a linear transformation on the adapted low-dimensional vector to unify the feature dimension, calculate the attention coefficients between nodes using the weight vectors, and aggregate the features of neighboring nodes based on the attention coefficients to update the features of the current node, including:

[0172] S1. Perform a linear transformation on the adapted low-dimensional vector to obtain a low-dimensional vector that satisfies the following formula:

[0173]

[0174] in, It is a linear transformation matrix. It is a low-dimensional vector after linear transformation.

[0175] S2. Calculate the attention coefficient, which satisfies the following formula:

[0176]

[0177] in, Attention coefficient (·) is a function that normalizes the intermediate results for all neighboring nodes u. (·) represents an activation function with leakage. For attention vectors, Let be the low-dimensional vector after the linear transformation of node u, and || be the vector concatenation operation. This is the normalization term.

[0178] S3. Aggregate the features of neighboring nodes to update the features of the current node. The updated node features satisfy the following formula:

[0179]

[0180] in, For the updated node features, (·) is the activation function. Let v be the set of neighboring nodes.

[0181] In one possible implementation, the bi-objective loss function satisfies the following formula:

[0182]

[0183] in, To regress the loss, Adjusting the sample size for budget purposes Adjust parameters to forecast budget. To represent the actual adjustment parameters, i is the index of the budget adjustment sample. These are the weighting coefficients; The classification loss satisfies the following formula:

[0184]

[0185] in, Allocate compliant sample numbers to working hours. For authenticity and compliance labeling, Let j be the compliance probability predicted by the model, and j be the compliance sample index for work hours.

[0186] In one possible implementation, the processing unit 601 is further configured to use a project budget-specific GNN training model to predict changes in the knowledge graph of project basic information and output budget adjustment parameters, including: updating the corresponding project main node in the knowledge graph when the project basic information changes; regenerating new GNN model input data based on the updated knowledge graph; inputting the new GNN model input data into the project budget-specific GNN training model to predict budget adjustment needs and output budget adjustment parameters.

[0187] In one possible implementation, the processing unit 601 is further configured to update the knowledge graph based on budget adjustment parameters to complete project budget management, including: updating project nodes related to budget item information and project edges related to budget template information in the knowledge graph based on budget adjustment parameters; generating a project budget execution snapshot based on the updated project nodes and project edges to complete project budget management.

[0188] The processing unit 601 can be a processor or a controller, and the communication unit 602 can be a communication interface, transceiver, transceiver circuit, transceiver device, etc. The term "communication interface" is a general term and may include one or more interfaces. The storage unit 603 can be a memory. When the project budget management device 60 is a chip, the processing unit 601 can be a processor or a controller, and the communication unit 602 can be an input interface and / or an output interface, pins, or circuits, etc. The storage unit 603 can be a storage unit within the chip (e.g., a register, cache, etc.) or a storage unit located outside the chip (e.g., read-only memory (ROM), random access memory (RAM, etc.).

[0189] The communication unit can also be called a transceiver unit. The antenna and control circuit with transceiver functions in the project budget management device 60 can be considered as the communication unit 602 of the project budget management device 60, and the processor with processing functions can be considered as the processing unit 601 of the project budget management device 60. Optionally, the device in the communication unit 602 that implements the receiving function can be considered as a communication unit. The communication unit is used to execute the receiving steps in the embodiments of this application, and the communication unit can be a receiver, a receiver circuit, etc. The device in the communication unit 602 that implements the transmitting function can be considered as a transmitting unit. The transmitting unit is used to execute the transmitting steps in the embodiments of this application, and the transmitting unit can be a transmitter, a transmitter, a transmitting circuit, etc.

[0190] Figure 6If the integrated units in the process are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. Storage media for storing computer software products include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0191] Figure 6 The units in the process can also be called modules; for example, a processing unit can be called a processing module.

[0192] This application embodiment also provides a hardware structure diagram of a project budget management device (denoted as project budget management device 70), see [link to diagram]. Figure 7 The project budget management device 70 includes a processor 701, and optionally, a memory 702 connected to the processor 701.

[0193] In the first possible implementation, see Figure 7 The project budget management device 70 also includes a transceiver 703. The processor 701, memory 702, and transceiver 703 are connected via a bus. The transceiver 703 is used to communicate with other devices or communication networks. Optionally, the transceiver 703 may include a transmitter and a receiver. The device in the transceiver 703 that implements the receiving function can be considered as a receiver, which is used to perform the receiving steps in the embodiments of this application. The device in the transceiver 703 that implements the transmitting function can be considered as a transmitter, which is used to perform the transmitting steps in the embodiments of this application.

[0194] Based on the first possible implementation method Figure 7 The structural diagram shown can be used to illustrate the structure of the project budget management device involved in the above embodiments.

[0195] in, Figure 7 This can also be illustrated by the system chip in the project budget management device. In this case, the actions performed by the aforementioned project budget management device can be implemented by this system chip; the specific actions performed can be found above and will not be repeated here.

[0196] In implementation, each step of the method provided in this embodiment can be completed by integrated logic circuits in the processor hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0197] The processor in this application may include, but is not limited to, at least one of the following: a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a microcontroller unit (MCU), or an artificial intelligence processor, etc., which are various computing devices that run software. Each computing device may include one or more cores for executing software instructions to perform calculations or processing. The processor may be a separate semiconductor chip or integrated with other circuits into a single semiconductor chip. For example, it may be integrated with other circuits (such as encoding / decoding circuits, hardware acceleration circuits, or various bus and interface circuits) to form a SoC (System-on-a-Chip), or it may be integrated as a built-in processor within an ASIC. The ASIC with the integrated processor may be packaged separately or together with other circuits. In addition to the cores for executing software instructions to perform calculations or processing, the processor may further include necessary hardware accelerators, such as field-programmable gate arrays (FPGAs), PLDs (programmable logic devices), or logic circuits that implement dedicated logic operations.

[0198] The memory in the embodiments of this application may include at least one of the following types: read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions; random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions; or electrically erasable programmable-only memory (EEPROM). In some scenarios, the memory may also be a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto.

[0199] This application also provides a computer-readable storage medium including instructions that, when run on a computer, cause the computer to perform any of the methods described above.

[0200] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform any of the methods described above.

[0201] This application also provides a chip including a processor and an interface circuit. The interface circuit is coupled to the processor, which is used to run computer programs or instructions to implement the above-described method. The interface circuit is used to communicate with other modules outside the chip.

[0202] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0203] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0204] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A project budget management method, characterized in that, include: Obtain the first information about the project, which includes basic project information, budget item information, and budget template information; The budget template information is used to characterize the relationship between the basic information and / or budget item information of each project; The basic project information is integrated into project main nodes, the budget item information is integrated into budget item nodes, and the budget template information is integrated into project edges between the nodes. A knowledge graph is constructed based on the project main nodes, budget item nodes, and project edges. The project edges include a first project edge and a second project edge. The first project edge is used to represent the logical relationship between the nodes, and the second project edge is used to represent the calculation relationship of the project budget between the nodes. The project main nodes and budget item nodes are transformed into low-dimensional vectors through semantic embedding, and the project edges are generated into weight vectors through relation embedding. Combined with the graph topology structure, an adjacency matrix is ​​constructed to form the input data of the GNN model. A heterogeneous graph attention network architecture is used to iteratively train the input data of the GNN model to obtain a project budget-specific GNN training model. The project budget-specific GNN training model is used to predict changes in the knowledge graph of the project's basic information and output budget adjustment parameters. These adjustment parameters are used to adjust the budget item nodes and project edges in the knowledge graph. Project budget management is completed by updating the knowledge graph based on the aforementioned budget adjustment parameters. The step of iteratively training the GNN model using a heterogeneous graph attention network architecture includes: The low-dimensional vector and the weight vector in the input data of the GNN model are adapted using a node type mapping matrix and an edge relationship mapping matrix, respectively. The adapted low-dimensional vector is linearly transformed to unify the feature dimension. The attention coefficient between nodes is calculated by combining the weight vector. The features of neighboring nodes are aggregated based on the attention coefficient to update the features of the current node. The GNN model is trained by optimizing the continuously updated node features using a dual-objective loss function that includes regression loss and classification loss. The regression loss is used for budget adjustment prediction, and the classification loss is used for compliance verification of work hour allocation.

2. The method according to claim 1, characterized in that, The process of integrating the project's basic information into project main nodes, integrating budget item information into budget item nodes, and integrating the budget template information into project edges between the nodes, and constructing a knowledge graph based on the project main nodes, budget item nodes, and project edges, includes: The basic information of the project is structured to form the main nodes of the project, and the budget item information is extracted. After associating the parent / child node dependencies of the items, budget item nodes are formed. A semantic rule parser is used to extract the logical relationship between the budget template information and the budget item information to form a first project edge, and a parameter mapping is used to extract the computational relationship between the budget template information and the budget item information to form a second project edge; The project main node, the budget item node, the first project edge, and the second project edge are imported into a graph database. A topological connection is established through the mapping relationship between the first and last nodes of the nodes and the edges to form a complete project budget knowledge graph.

3. The method according to claim 1, characterized in that, The process of converting the project main nodes and budget item nodes into low-dimensional vectors through semantic embedding, generating weight vectors from the project edges through relational embedding, and constructing an adjacency matrix by combining the graph topology structure includes: A hybrid semantic embedding algorithm combining TransE and BERT is used to encode and transform the text attributes of the project main node and budget item node to obtain a fixed-dimensional low-dimensional vector. The RotatE and DistMult hybrid relation embedding algorithm is used to parse the attributes, convert them into features, and establish relations for the rules of the project edges, generating weight vectors that distinguish relation types and influence strengths. Based on the knowledge graph topology, the low-dimensional vector is used as the row / column basis, and combined with the weight vector, the adjacency matrix elements are filled to form an adjacency matrix containing node semantics, edge relation weights, and topological connections.

4. The method according to claim 1, characterized in that, The adaptation process, which uses node type mapping matrix and edge relationship mapping matrix to match the low-dimensional vector and weight vector in the GNN model input data, satisfies the following formula: in, For the low-dimensional vector, This is the node type mapping matrix. The adapted low-dimensional vector, Let be the weight vector. Let be the edge relationship mapping matrix. This is the adapted weight vector.

5. The method according to claim 1, characterized in that, The process of performing a linear transformation on the adapted low-dimensional vector to unify the feature dimension, calculating the attention coefficient between nodes based on the weight vector, and updating the current node's features by aggregating the neighbor node features according to the attention coefficient includes: S1. Perform a linear transformation on the adapted low-dimensional vector to obtain a new low-dimensional vector, which satisfies the following formula: in, It is a linear transformation matrix. It is a low-dimensional vector after linear transformation; S2. Calculate the attention coefficient, which satisfies the following formula: in, The attention coefficient, (·) is a function that normalizes the intermediate results for all neighboring nodes u. (·) represents an activation function with leakage. For attention vectors, Let be the low-dimensional vector after the linear transformation of node u, and || be the vector concatenation operation. For normalization terms; S3. Aggregate the features of neighboring nodes to update the features of the current node. The updated node features satisfy the following formula: in, For the updated node features, (·) is the activation function. Let v be the set of neighboring nodes.

6. The method according to claim 1, characterized in that, The bi-objective loss function satisfies the following formula: in, For the regression loss, Adjusting the sample size for budget purposes Adjust parameters to forecast budget. To represent the actual adjustment parameters, i is the index of the budget adjustment sample. These are the weighting coefficients; The classification loss satisfies the following formula: in, Allocate compliant sample numbers to working hours. For authenticity and compliance labeling, Let j be the compliance probability predicted by the model, and j be the compliance sample index for work hours.

7. The method according to claim 1, characterized in that, The method of using the project budget-specific GNN-trained model to predict changes in the knowledge graph of project basic information and outputting budget adjustment parameters includes: When the basic information of a project changes, the corresponding main node of the project in the knowledge graph is updated. New GNN model input data is regenerated based on the updated knowledge graph; The new GNN model input data is input into the project budget-specific GNN training model to predict budget adjustment needs and output the budget adjustment parameters.

8. The method according to claim 1, characterized in that, The process of updating the knowledge graph based on the budget adjustment parameters to complete project budget management includes: Update the project nodes related to budget item information and the project edges related to budget template information in the knowledge graph based on the aforementioned budget adjustment parameters; Based on the updated knowledge graph, project nodes related to budget item information and project edges related to budget template information are used to generate project budget execution snapshots, thus completing project budget management.

9. A project budget management device, characterized in that, The device includes: a communication unit and a processing unit; The communication unit is used to acquire first information about the project, which includes basic project information, budget item information, and budget template information; the budget template information is used to characterize the relationship between the basic project information and / or budget item information. The processing unit is used to integrate the basic project information into project main nodes, integrate budget item information into budget item nodes, and integrate the budget template information into project edges between nodes. A knowledge graph is constructed based on the project main nodes, budget item nodes, and project edges. Each project edge includes a first project edge and a second project edge. The first project edge represents the logical relationship between nodes, and the second project edge represents the calculation relationship of the project budget between nodes. The project main nodes and budget item nodes are converted into low-dimensional vectors through semantic embedding. The project edges are generated into weight vectors through relational embedding. An adjacency matrix is ​​constructed using the graph topology to form the input data for the GNN model. A heterogeneous graph attention network architecture is used to iteratively train the GNN model input data to obtain a project budget-specific GNN training model, including: through nodes... The type mapping matrix and edge relationship mapping matrix are used to adapt the low-dimensional vector and the weight vector in the input data of the GNN model, respectively. A linear transformation is performed on the adapted low-dimensional vector to unify the feature dimension. The attention coefficient between nodes is calculated based on the weight vector, and the features of neighboring nodes are aggregated according to the attention coefficient to update the current node's features. A dual-objective loss function, including regression loss and classification loss, is used to optimize and continuously update the node features to train the GNN model. The regression loss is used for budget adjustment prediction, and the classification loss is used for work hour allocation compliance verification. The project budget-specific GNN training model is used to predict changes in the knowledge graph of project basic information and output budget adjustment parameters. These adjustment parameters are used to adjust the budget item nodes and project edges in the knowledge graph. The knowledge graph is updated based on the budget adjustment parameters to complete project budget management.

Images