Information service consulting fee management method and system

Abstract

The application discloses an information service consulting fee management method and system, the method comprising the following steps: in response to a consulting fee approval request, extracting a feature vector from a consulting project management system; inputting the feature vector into an approval path decision model, calculating a comprehensive decision score of each candidate path, and determining an execution path according to the calculation result of the comprehensive decision score; determining the identity of a current approver, and generating a dynamic abstract according to the identity of the approver, wherein the dynamic abstract comprises a semantic vector of the approval request, a budget execution deviation rate and a cost performance value; detecting the state of the approver, the budget execution deviation rate and the cost performance value, and triggering approval path re-routing to update the execution path when the state of the approver is abnormal, or the budget execution deviation rate is greater than a first threshold value, or the cost performance value is less than a second threshold value. The application guarantees the continuity and flexibility of the approval process, greatly improves the approval efficiency, and strengthens the fine management and control of the consulting fee.

Description

Technical Field

[0001] This invention relates to the field of information consulting technology, and in particular to a method and system for managing information service consulting fees. Background Technology

[0002] In the field of information service consulting fee management, existing technologies largely rely on fixed approval paths to control the fee approval process. Typically, the system pre-sets approvers at each level and their approval order, and consulting fee approval requests flow along this fixed path. Approver only receives basic information relevant to their own approval stage, such as the application amount and project name. However, existing technologies have significant drawbacks: First, fixed approval paths lack flexibility. When an approver is absent due to leave, resignation, or other reasons, the approval process easily stalls, severely impacting efficiency. Second, the information available to approvers is limited; relying solely on basic information makes it difficult to comprehensively assess the rationality of the approval request, and to promptly identify issues such as excessive budget execution deviations or poor cost performance. This can lead to unreasonable fees being approved, resulting in resource waste. Furthermore, even if such problems are discovered during the approval process, it is difficult to quickly adjust the approval path, forcing it to continue along the original path, further exacerbating management loopholes. Summary of the Invention

[0003] To address at least one of the aforementioned technical problems, this invention provides a method and system for managing information service consulting fees.

[0004] In a first aspect, the present invention provides a method for managing information service consulting fees, the method comprising:

[0005] In response to a consulting fee approval request, feature vectors are extracted from the consulting project management system.

[0006] The feature vectors are input into the approval path decision model to calculate the comprehensive decision score for each candidate path, and the execution path is determined based on the calculation result of the comprehensive decision score.

[0007] Determine the identity of the current approver and generate a dynamic summary based on the approver's identity. The dynamic summary includes the semantic vector of the approval request, the budget execution deviation rate, and the cost performance value.

[0008] The system monitors the approver status, budget execution deviation rate, and cost performance value. When the approver status is abnormal, the budget execution deviation rate is greater than the first threshold, or the cost performance value is less than the second threshold, the approval path rerouting is triggered to update the execution path.

[0009] Preferably, the step of inputting the feature vector into the approval path decision model and calculating the comprehensive decision score for each candidate path includes:

[0010] Determine the set of candidate paths, and calculate the ratio of the dimensions of the dynamic feature vector and the static feature vector of each candidate path based on the dynamic feature vector of the current consultation request and the static feature vector of each candidate path to obtain the dimension balance factor.

[0011] Three sets of trainable parameter matrices are established and multiplied by the dynamic feature vector, the static feature vector, and the dynamic-static feature cross matrix, respectively, to obtain the query, key, and value matrices of the attention mechanism; the cross-modal attention weight matrix is ​​calculated based on the query, key, and value matrices.

[0012] The path fit is calculated based on the cosine of the angle between the dynamic feature vector and the static feature vector. The comprehensive decision score for each candidate path is calculated based on the dimensional balance factor, the cross-modal attention weight matrix, and the path fit.

[0013] Preferably, the dynamic feature vector includes a vector composed of consulting fee amount, project urgency, client level, project complexity, historical approval record association information, and submission time attribute;

[0014] The static feature vector includes a vector composed of path nodes, node approval thresholds and permissions, expected processing time, path activation status, historical statistical information, and cost information attributes.

[0015] Preferably, determining the execution path based on the calculation result of the comprehensive decision score includes:

[0016] Dynamic threshold filtering is performed based on the comprehensive decision score calculated for each path to obtain the remaining candidate path set after filtering;

[0017] Based on the remaining set of candidate paths, the final target execution path is calculated according to the historical average approval time of each candidate path and the system's overall average approval time.

[0018] Preferably, the triggering of approval path rerouting to update the execution path includes:

[0019] Determine the number of joint approvers for the original path and the new path, and the path switching penalty coefficient. Multiply the number of joint approvers and the path switching penalty coefficient to obtain the path switching cost factor.

[0020] Calculate the comprehensive decision score for the new path, subtract the path switching cost factor from the comprehensive decision score calculated for each new path, and extract the maximum value of the result as the rerouting path.

[0021] Secondly, the present invention also provides an information service consulting fee management system, the system comprising:

[0022] The feature vector extraction unit is used to extract feature vectors from the consulting project management system in response to consulting fee approval requests.

[0023] The execution path decision unit is used to input feature vectors into the approval path decision model, calculate the comprehensive decision score of each candidate path, and determine the execution path based on the calculation result of the comprehensive decision score.

[0024] The dynamic summary generation unit is used to determine the identity of the current approver and generate a dynamic summary based on the approver's identity. The dynamic summary includes the semantic vector of the approval request, the budget execution deviation rate, and the cost performance value.

[0025] The path rerouting unit is used to detect the approver status, budget execution deviation rate, and cost performance value. When the approver status is abnormal, the budget execution deviation rate is greater than the first threshold, or the cost performance value is less than the second threshold, the approval path rerouting is triggered to update the execution path.

[0026] Preferably, the execution path decision unit is further configured to:

[0027] Determine the set of candidate paths, and calculate the ratio of the dimensions of the dynamic feature vector and the static feature vector of each candidate path based on the dynamic feature vector of the current consultation request and the static feature vector of each candidate path to obtain the dimension balance factor.

[0028] Three sets of trainable parameter matrices are established and multiplied by the dynamic feature vector, the static feature vector, and the dynamic-static feature cross matrix, respectively, to obtain the query, key, and value matrices of the attention mechanism; the cross-modal attention weight matrix is ​​calculated based on the query, key, and value matrices.

[0029] The path fit is calculated based on the cosine of the angle between the dynamic feature vector and the static feature vector. The comprehensive decision score for each candidate path is calculated based on the dimensional balance factor, the cross-modal attention weight matrix, and the path fit.

[0030] Preferably, the execution path decision unit is further configured to:

[0031] Dynamic threshold filtering is performed based on the comprehensive decision score calculated for each path to obtain the remaining candidate path set after filtering;

[0032] Based on the remaining set of candidate paths, the final target execution path is calculated according to the historical average approval time of each candidate path and the system's overall average approval time.

[0033] Thirdly, the present invention also provides an electronic device including a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs the method as described in the first aspect above and any possible implementation thereof.

[0034] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor of an electronic device, cause the processor to perform a method as described in the first aspect above and any possible implementation thereof.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] This invention utilizes the ability to dynamically determine execution paths, breaking the limitations of fixed paths. When the approver's status is abnormal, the path can be updated in a timely manner through path rerouting, avoiding process stagnation and solving the problems of rigid and inefficient approval paths. Secondly, it generates a dynamic summary for the approver, including semantic vectors, budget execution deviation rate, and cost performance value, enabling the approver to fully grasp the key information of the approval request and solving the problems of insufficient approval basis and difficulty in accurate evaluation. Finally, by real-time detection of the approver's status, budget execution deviation rate, and cost performance value and triggering path rerouting, dynamic control of the approval process is achieved, solving the deficiency of existing technologies that cannot adjust the path in a timely manner after problems are discovered. Therefore, this invention ensures the continuity and flexibility of the approval process through dynamic determination of execution paths and rerouting mechanisms, significantly improving approval efficiency; the dynamic summary provides the approver with comprehensive and accurate evaluation basis, helping the approver to make more reasonable decisions and reducing the risk of approving unreasonable expenses; real-time monitoring and path adjustment of budget execution deviation rate and cost performance value effectively strengthen the refined control of consulting fees, improve the accuracy of budget execution and cost performance level, and ultimately achieve efficient, accurate, and intelligent management of information service consulting fees.

[0037] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the background art, the accompanying drawings used in the embodiments of the present invention or the background art will be described below.

[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.

[0040] Figure 1 A flowchart illustrating an information service consulting fee management method provided in an embodiment of the present invention;

[0041] Figure 2 Provided for embodiments of the present invention Figure 1 A flowchart illustrating the sub-steps of S20;

[0042] Figure 3 Provided for yet another embodiment of the present invention Figure 1 A flowchart illustrating the sub-steps of S20;

[0043] Figure 4 This is a schematic diagram of the structure of an information service consulting fee management system provided in an embodiment of the present invention. Detailed Implementation

[0044] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] The terms "first," "second," etc., used in the technical solutions of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0046] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0047] Please see Figure 1 , Figure 1 This is a flowchart illustrating an information service consulting fee management method provided in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0048] S10. In response to the consulting fee approval request, extract the feature vector from the consulting project management system;

[0049] Existing technologies often only extract basic information such as application amount and project name as the basis for approval. This information lacks quantitative representation of key dimensions such as the nature of the expense, project risk, and budget execution, resulting in a lack of comprehensive data support for approval path decisions and difficulty in accurately matching the actual needs of different approval requests. Therefore, in this embodiment, the extracted feature vector is mainly a dynamic feature vector based on the current consultation request, including but not limited to a vector composed of consultation fee amount, project urgency, customer level, project complexity, historical approval record association information, and submission time attribute. The functions of each vector are as follows:

[0050] Representative object: The current specific consultation fee request.

[0051] Content: Contains real-time, attribute information specific to this request. This information is highly dependent on the specific content submitted by the user, the system environment, or the request context.

[0052] Types of dynamic feature vectors:

[0053] Fee Amount: The amount of money involved in this consultation.

[0054] Urgency level: The urgency level of a request as marked by the user or system.

[0055] Customer level: The importance or category of the customer who initiated the request.

[0056] Project complexity: The nature of the project associated with this consultation or the estimated difficulty of handling it.

[0057] Historical approval record information: past approval status of this user or similar requests (such as average approval time, common reasons for rejection, etc.).

[0058] Submission time: The time when the request was submitted (whether it was a weekday, holiday, etc.), current system load, etc.

[0059] In step S10, when the system receives a consulting fee approval request, it automatically triggers data interaction with the consulting project management system. Through a preset data interface, it extracts key data associated with the approval request from the system and transforms this data into feature vectors that can be used for model processing. Then, it extracts dynamic feature vectors that conform to the aforementioned attributes. It can be understood that each consulting fee approval request generates a unique dynamic feature vector, the value of which reflects the specific state at the time of request submission. The dynamic feature vectors of different requests differ significantly.

[0060] Regarding the amount of fees, a fee type quantification value can be calculated. This quantification value converts the type of consulting fees, such as research fees and service fees, into specific numerical values ​​through coding rules. This value can provide a category basis for subsequent approval path decisions. When there are differences in the amount, the logarithmic normalized value of the fee amount can be obtained. This allows for logarithmic transformation and normalization of the application amount of consulting fees, eliminating the impact of different amount levels and facilitating fair comparison and analysis of approval requests of different amount scales by the model.

[0061] Preferably, in addition to acquiring dynamic feature vectors such as project complexity and historical approval record association information, attributes such as project risk level score, current budget execution deviation rate, and contract urgency coefficient can also be considered. The project risk level score is a quantitative score derived from historical risk records, current risk assessment reports, and other data, according to preset scoring standards. The current budget execution deviation rate is obtained by calculating the ratio of the difference between the project's current actual expenditure and the budget to the budget. The contract urgency coefficient quantifies the urgency of contract execution based on information such as the service period and delivery nodes stipulated in the contract. The project risk level score directly reflects the project's risk status; high-risk projects may require more stringent approval processes. The current budget execution deviation rate and contract urgency coefficient reflect the project's budget execution status and time urgency, providing crucial references for assessing the rationality of approval requests.

[0062] Compared to the basic information extracted in existing technologies, such as application amount and project name, the feature vector extracted in this step covers multiple dimensions of key information, including the type and scale of consulting fees, project urgency and risk, customer level, project complexity, historical approval record association information and submission time attributes, budget execution and contract urgency. This provides rich and accurate input data for the subsequent approval path decision model, enabling the model to more scientifically calculate the comprehensive decision score of each candidate path based on these features, thereby determining the most suitable execution path. This solves the problems of insufficient basis for approval path decision and low matching degree in existing technologies.

[0063] S20. Input the feature vector into the approval path decision model, calculate the comprehensive decision score for each candidate path, and determine the execution path based on the calculation result of the comprehensive decision score.

[0064] See Figure 2 In one embodiment, the step of inputting feature vectors into the approval path decision model, calculating the comprehensive decision score for each candidate path, and determining the execution path based on the calculation result of the comprehensive decision score includes the following steps:

[0065] S201. Determine the candidate path set. Based on the dynamic feature vector of the current consultation request and the static feature vector of each candidate path, calculate the dimensional ratio of the dynamic feature vector and the static feature vector to obtain the dimensional balance factor.

[0066] S202. Establish three sets of trainable parameter matrices, and multiply them by the dynamic feature vector, static feature vector, and dynamic-static feature cross matrix respectively to obtain the query, key, and value matrix of the attention mechanism; calculate the cross-modal attention weight matrix based on the query, key, and value matrix.

[0067] S203. Calculate the path fit degree based on the cosine of the angle between the dynamic feature vector and the static feature vector, and calculate the comprehensive decision score of each candidate path based on the dimensional balance factor, the cross-modal attention weight matrix, and the path fit degree.

[0068] The dynamic feature vector has been obtained in step S10. To calculate the comprehensive decision score of the candidate paths, this step also requires obtaining the static feature vector for each candidate path. Preferably, the static feature vector includes a vector composed of path nodes, node approval thresholds and permissions, expected processing time, path activation status, historical statistical information, and cost information attributes. Static feature vectors typically describe the inherent, relatively stable attributes and rules of the path. This information defines the path's structure, constraints, and processing capabilities, and is independent of specific requests. For example:

[0069] Path node composition: which approval roles are included (e.g., specialist -> manager -> director -> finance).

[0070] Node approval threshold / permissions: The maximum amount that each node can approve, and the types of customers or projects that it can handle.

[0071] Expected processing time: The estimated time required to complete the approval process under normal circumstances.

[0072] Path availability status: Whether the path is currently available.

[0073] Historical statistics: the average processing time, success rate, and rejection rate of requests processed by this path in the past (as a reflection of the inherent performance of the path).

[0074] Cost information: Fixed or anticipated costs that may arise from using this route.

[0075] It is understandable that each candidate path has its own predefined static feature vector. These features typically do not change frequently with a single request, but are only updated when the path rules are modified by the administrator. Therefore, static feature vectors are inherent and are only used to describe the "configuration" or "attributes" of the path itself.

[0076] The reason for calculating the dimensionality ratio of the dynamic feature vector and the static feature vector in step S201 is to address the problems caused by differences in numerical scale. Dynamic features (such as amounts that may be tens or hundreds of thousands) and static features (such as the maximum approval amount for a node that may be millions, or the path cost that may be hundreds) typically have completely different orders of magnitude and value ranges. Directly comparing or combining them may cause the features with larger values ​​to dominate the entire calculation, drowning out the information of other important but smaller features.

[0077] To aid understanding, the detailed calculation formula for step S20 will be provided below:

[0078] In step S201, the candidate path set is determined: ;in, Representing the There are a total of candidate paths strip.

[0079] Calculate the dimensional balance factor:

[0080] ;

[0081] In the formula, The ratio of the dimensions of dynamic features to those of static features is called the dimensional balance factor. For the L1 norm of the input dynamic feature vector, The infinite norm of the path's static eigenvectors;

[0082] calculate This represents the sum of the absolute values ​​of all elements in the dynamic feature vector, which roughly represents a measure of the overall "scale" or "intensity" of the request (such as total amount, total urgency, etc.). The calculation... This is the maximum absolute value of all elements in the static feature vector, which represents the maximum "capacity" or "limit" of the path for a single feature (such as the maximum amount that can be approved at a certain approval node in the path). This ratio attempts to quantify "how large the overall size of this request is relative to the maximum single-point limit that the path can handle." A large ratio means the request is "robust" or the path has "limited capacity"; a small ratio means the opposite.

[0083] In step S202, three sets of trainable parameter matrices are established. , , Multiplying these by the dynamic feature vector, static feature vector, and dynamic-static feature cross matrix respectively, we obtain the query, key, and value matrix of the attention mechanism, i.e. The matrix is ​​then used to calculate the cross-modal attention weight matrix. :

[0084] ; ;

[0085] In the formula, For the attention mechanism, the query, key, and value matrix; For the extracted dynamic feature vector, For path static characteristics For path Dynamic-static feature cross matrix; , , It is a trainable parameter matrix; For transpose, For dynamic feature vectors The dimension; This is the cross-modal attention weight matrix.

[0086] In the above formula, the original space is... Mapped to a query, key, and value space suitable for calculating attention weights, through a learnable parameter matrix. , , The model's ability to dynamically learn which features or combinations of features are most important for calculating the correlation between dynamic features and static path features, as well as their intersection features.

[0087] Acting on dynamic feature vectors Transform it linearly into a query vector This represents "what kind of path features this request is looking for". It is randomly generated during model initialization and updated during training through backpropagation and the Adam optimizer. Its learning objective is to make the query vector... It can effectively express the path characteristics that the request needs to match. It determines which dynamic attributes of the request (such as amount, urgency, type) are emphasized for query path compatibility. Acting on static feature vectors Transform it linearly into a key vector, representing "the candidate path". "What features can be provided?" This parameter is also randomly generated during model initialization and determines which static attributes of the path (such as maximum amount limit, processing time, applicable type) are emphasized for matching the query request. Acting on dynamic-static feature cross matrix Transform it linearly into a value vector, and so on. , Both are randomly initialized and learned during training to determine which cross-feature information (such as "request amount") "Maximum amount limit for path", "Request type" Path applicable type rules, etc., are important and should be encoded into... The vector is used for attention weight aggregation, which is the core of cross-modal interaction.

[0088] In calculating the cross-modal attention weight matrix When the size is, This indicates that the query from the request is being calculated. and keys from the path The dot product of the two paths is the similarity score. The larger the dot product, the better the requested query matches the key provided by the path. Used to scale the dot product to prevent the dot product result from becoming too large when the dimension is large. Gradient vanishing. For dynamic feature vectors The dimension, i.e., the number of features. The value of directly reflects the feature complexity of the approval decision-making model, and typical features The composition is as follows: Numerical type: project budget amount, estimated time, number of related parties; Category type: project type; Time series type: historical approval rate, average time of similar projects in recent times.

[0089] The function then normalizes the scaled similarity scores to obtain the attention weight matrix. Each row of this weight matrix (corresponding to a query dimension of the request) represents the distribution of attention given to each key dimension of the path by that query dimension. Higher weights indicate a high correlation between the request feature and the corresponding path feature. Finally, the cross-modal attention weight matrix is ​​calculated. Attention weights are applied to the value vector. The cross-modal attention weight matrix is ​​obtained by weighted summation of information from both original and cross-features. It integrates request features, static path features, and their interaction features, and focuses on the most relevant parts through an attention mechanism. The cross-modal attention weight matrix... It is a key input for subsequent steps such as path scoring and selection.

[0090] Further, in step S203, the path fit is first calculated based on the cosine of the angle between the dynamic feature vector and the static feature vector, as follows:

[0091] ;

[0092] In the formula, For path adaptation, For cosine similarity, These are the weighting coefficients for the KL divergence. and Let be the feature mapping function, satisfying:

[0093] ;

[0094] ;

[0095] In the formula, , , , All of these are trainable parameters.

[0096] In the above formula, Used to calculate dynamic feature vectors With static feature vectors The direction is consistent, and the range is The closer to 1, the more similar the feature space directions are. This is used to convert cosine similarity to normalized similarity. The output is a value in radians, divided by The result is then scaled to [0,1] and subtracted by 1, so that the output is positively correlated with the similarity (the larger the value, the better the match). This is the KL divergence term, used to measure the dynamic eigenvectors. With static feature vectors The smaller the KL value, the closer the probability distributions are. The weighting of the KL divergence term in controlling the overall fitness requires training to balance the importance of geometric similarity and distributional similarity.

[0097] In calculating the feature mapping function and hour, , , , All of these are trainable parameters, among which, , These are the trainable projection matrix and bias vector of the dynamic features, respectively. Used to map the original features to a probability distribution space. Used to enhance the model's expressive power. Correspondingly, , These are the trainable projection matrix and bias vector for static features, used to map path features to a distribution space aligned with dynamic features and to enhance the model's expressive power, respectively. These four trainable parameters are also generated through random initialization and learned during training; for example, the Adam optimizer can be used. , , , and Optimization is achieved through gradient descent. and The output is converted into a probability distribution, which encodes the original features as a "feature importance weight distribution". Thus, This can be interpreted as the importance weight of each attribute in the request. This can be interpreted as the capacity allocation ratio of the path to each feature. Through the above formula, the model can simultaneously capture the explicit feature association (cosine term) and implicit distribution matching (KL term) between the request and the path, providing a more robust fit assessment for complex decision-making scenarios.

[0098] Furthermore, after obtaining the path fit, the comprehensive decision score for each candidate path is calculated based on the dimensional balance factor, the cross-modal attention weight matrix, and the path fit, as shown in the following formula:

[0099] ;

[0100] ;

[0101] In the formula, Candidate paths The overall decision score, For the Sigmoid function, For the slope parameter and center offset of the Sigmoid curve; This is the exponential amplification factor. .

[0102] In the above formula, Dimensional balance factor The sigmoid transformation function converts the dimensional balance factor. The Sigmoid function maps the path to the (0,1) interval, representing the path's capacity adaptability to the request. When At that time, the request strength was much lower than the path limit. Approaching 0, when At that time, the request intensity far exceeded the path limit. When the two approaches 0, the request matches the path capacity. Approaching 1. For the slope parameter and center offset of the Sigmoid curve, Used to control matching sensitivity, representing the system's tolerance for capacity deviations. The range of values ​​is It is usually initialized to a value of 1; This is the ideal dimensional ratio, which defines the optimal capacity ratio for business operations. For matrix norm, This indicates that attention weights are used to weight the cross features. It can capture key signals of dynamic-static feature interaction. The larger the value, the stronger the feature interaction between the path and the request, and the higher the matching degree. This is the exponential amplification factor, used to control the amplification intensity of the fit. At that time, the exponential amplification term of path fit It is always 1, and The larger the value, the more significant the score advantage of the highly fit path. Among these, Optimization through backpropagation aims to determine the accuracy of path selection, such as through cross-entropy loss. During the learning and training process, if capacity matching is prioritized in the data (e.g., in financial risk control), the model will increase in size. and reduce If feature matching is prioritized (e.g., in emergency medical care), the model will increase in size. And relax .

[0103] See Figure 3 In one embodiment, determining the execution path based on the calculation result of the comprehensive decision score includes:

[0104] S204. Perform dynamic threshold filtering based on the comprehensive decision score calculated for each path to obtain the remaining candidate path set after filtering.

[0105] S205. Based on the remaining set of candidate paths, calculate the final target execution path according to the historical average approval time of each candidate path and the system's overall average approval time.

[0106] Specifically, for each path calculate Filtering via dynamic thresholds:

[0107] ;

[0108] ;

[0109] In the formula, , The mean and standard deviation of the scores for the current batch. For adaptive coefficients, This is the set of paths remaining after filtering. This is the current timestamp. For the approval period;

[0110] ;

[0111] In the formula, For the final selected execution path, For path Historical average approval time This represents the average approval time across the entire system.

[0112] Comprehensive decision score calculation model It consists of three parts, using a geometrically weighted average to ensure the balance of indicators. Input feature vector This includes quantified values ​​for expense types, logarithmic normalized values ​​for amounts, project risk level scores, current budget execution deviation rates, and contract urgency coefficients. Path static features Such as approval level, historical approval rate, etc.; For dynamic-static feature cross matrix, satisfying , Represents the Kronecker product. Dynamic weighting function. The influence of feature magnitude differences can be dynamically adjusted using the Sigmoid function. These are learnable parameters. L1 norm of input dynamic feature vector Infinite norm of path static features The ratio serves to quantify the numerical scale difference between two types of features;

[0113] In the above formula, dynamic features are included. Static features Through the cross matrix Attention Matrix The fusion approach addresses the problem of feature fragmentation in traditional methods. Compared to traditional logistic regression or decision trees, which can only handle single-modal features, MAFS captures complex relationships through cross-modal interactions, such as the implicit relationship between "high-value projects + urgency" and specific approval paths. Traditional methods require manual feature standardization, while MAFS automatically adapts to changes in feature scale using the sigmoid function. Dynamically balances differences in feature dimensions to avoid manually setting fixed weights. Path adaptation metric function. MAFS combines cosine similarity and KL divergence to simultaneously measure the distance and probability distribution differences in vector spaces. Compared to traditional similarity calculations (such as Euclidean distance) that ignore feature distribution information, MAFS... and The mapping enhances semantic matching. Through Implement time-sensitive threshold relaxation to accelerate emergency processes; incorporate an efficiency factor into the final decision. This avoids choosing high-scoring but inefficient paths. Compared to traditional static thresholds, which cannot adapt to changes in business rhythm, it prioritizes efficient paths while ensuring quality.

[0114] Therefore, by combining decision scores and dynamic threshold filtering, approval time can be shortened, urgent processes can be accelerated, and approval efficiency can be greatly improved. At the same time, the accuracy of decision-making is enhanced, abnormal paths are avoided, and adaptive capabilities are significantly improved.

[0115] S30. Determine the identity of the current approver and generate a dynamic summary based on the approver's identity. The dynamic summary includes the semantic vector of the approval request, the budget execution deviation rate, and the cost performance value.

[0116] In this step, the system identifies the approver currently processing the approval request through user login information, authorization modules, or identity information such as employee ID and job code, clarifying their approval level, scope of responsibility, and authority. Subsequently, based on this identity information, a dynamic summary generation process is initiated. For the semantic vector of the approval request, the system first extracts triples related to the current approval request from the associated knowledge graph. During extraction, an attention mechanism is used to calculate the correlation between each entity and the approval request, filtering out associated entities with an attention weight exceeding 0.7. Then, combined with the project ID and the corresponding attribute values ​​of that entity, such as project stage, service type, and historical cooperation records, a triple containing "project ID-associated entity-attribute value" is constructed, which is then transformed into a semantic vector, presenting the core semantic information of the approval request in a structured form. Simultaneously, the system retrieves the budget execution deviation rate (the ratio of actual expenditure to budget) and cost performance value (such as the quantitative ratio of cost to output) corresponding to the approval request from the consulting project management system. These two types of data are integrated with the semantic vector to form a dynamic summary for the current approver.

[0117] This step generates customized dynamic summaries by identifying the approver, enabling approvers at different levels and with different permissions to accurately obtain key information relevant to their decisions. For example, senior approvers may focus more on cost-performance values ​​to control overall resource allocation, while business department approvers are more concerned with the details of project-related entities in the semantic vector. This avoids information overload or missing key information, improving the targeting and efficiency of approver decisions. In addition, the semantic vector is constructed based on knowledge graph triples and only includes highly related entities with an attention weight greater than 0.7. This ensures the accuracy and relevance of semantic information and allows approvers to quickly understand the background and core elements of the approval request through structured triples (project ID, related entity, attribute value), solving the problems of fragmented information and semantic ambiguity in traditional approvals. Finally, the inclusion of budget execution deviation rate and cost-performance values ​​provides approvers with quantitative financial indicators, enabling them to evaluate the rationality of approval requests from both business and financial dimensions by combining semantic information, reducing the subjectivity and blindness of decision-making.

[0118] S40. Detect the approver status, budget execution deviation rate, and cost performance value. When the approver status is abnormal, the budget execution deviation rate is greater than the first threshold, or the cost performance value is less than the second threshold, trigger approval path rerouting to update the execution path.

[0119] In one embodiment, triggering approval path rerouting to update the execution path includes:

[0120] Determine the number of joint approvers for the original path and the new path, and the path switching penalty coefficient. Multiply the number of joint approvers and the path switching penalty coefficient to obtain the path switching cost factor.

[0121] Calculate the comprehensive decision score for the new path, subtract the path switching cost factor from the comprehensive decision score calculated for each new path, and extract the maximum value of the result as the rerouting path.

[0122] During the approval process, three key indicators are continuously monitored: the approver's status (on-duty, off-duty, etc.) is obtained in real time through the identity authentication system, attendance management system, or work status monitoring module; the current budget execution deviation rate is synchronized periodically from the project management system; and the latest cost performance value is obtained from the cost accounting module. When an abnormal approver status is detected, such as being off-duty without designating a representative, the budget execution deviation rate exceeding the preset first threshold, or the cost performance value falling below the preset second threshold, the approval path rerouting mechanism is immediately triggered.

[0123] Specifically, the calculation model for rerouting paths is as follows:

[0124] ;

[0125] In the formula, Number the new candidate paths. This is the path switching penalty coefficient, with a value ranging from 0.1 to 0.5. This represents the number of joint approvers for both the original and new paths.

[0126] Typically, approval systems need to ensure at least 20%-50% continuity of approvers (as per industry standards in finance / healthcare, etc.). If there are too few co-approver individuals: business background needs to be re-examined, increasing time costs and the risk of approval errors due to decreased familiarity, and creating a communication burden for the applicant. This is particularly relevant in various industries, such as financial risk control. The value is typically between 0.3 and 0.5. Each change of approver results in an average delay of 6 hours, increasing the risk of misjudgment by 8%. For medical approval processes, timeliness is prioritized in emergency situations, but key approvers must be retained. A value between 0.1 and 0.3 is more appropriate. However, in the government procurement process, due to regulatory requirements for traceability of the approval chain and the inability to change the core approver, the value will be appropriately increased. The value range is, for example, between 0.4 and 0.5. In the formula above, This indicates the matching quality, i.e., maximizing path adaptability, while This indicates the switching cost, minimizes the risk of the approval chain breaking, and reflects the stability of the process.

[0127] In the path update phase, the system calculates the optimal rerouting path based on a multi-objective optimization algorithm. Specifically, it first filters all feasible paths from the candidate path set that match the current project type, budget range, and risk level; then, for each candidate path... Calculate its overall score Simultaneously, the original path is statistically analyzed. With new paths Number of joint approvers The larger this value, the lower the path switching cost; finally, substitute it into the formula " "By traversing all candidate paths, the path that maximizes the objective function value is selected as the final rerouting path." The path switching penalty coefficient is... The value can be dynamically adjusted within the range of 0.1-0.5 depending on the organization's emphasis on process stability; no specific limit is set here.

[0128] To facilitate understanding, the application process of the above embodiment will be illustrated below using a bank's corporate loan approval system as an example. Assume the current path is:

[0129] Original path = [Account Manager A → Risk Control Specialist B → Regional Manager C → Compliance Officer D → Head Office Approval E], due to an urgent customer need (loan amount exceeding regional authority), the system triggered a rerouting calculation. The candidate new path is:

[0130] z1 = [Account Manager A → Risk Control Director F → Head Office Approval E];

[0131] z2 = [Account Manager A → Risk Control Specialist B → Head Office Special Channel G];

[0132] z3 = [Account Manager A → Remote Risk Control H → Remote Compliance I → Head Office Approval E];

[0133] Of the three new paths mentioned above, Both are 2, meaning the number of joint approvers for the new path and the original path is 2. If... If the value is 0.3, then the calculation can be performed. And based on the aforementioned comprehensive decision score calculation model, the following were calculated respectively: , ,pass The final values ​​of the three paths can be calculated to be 4.2, 3.6, and 4.6. At this point, the path with the largest value, i.e., z3, is selected as the new path after rerouting.

[0134] Therefore, the above embodiments, by real-time monitoring of approver status, can automatically trigger path adjustments when personnel anomalies occur, avoiding process stagnation caused by single points of failure in traditional fixed paths. This ensures that approval efficiency is not affected by personnel changes, enhancing process robustness. Simultaneously, it improves budget control accuracy through dynamic monitoring of budget execution deviation rates and cost performance values, enabling the system to intervene promptly when financial indicators are abnormal. By introducing higher-level approvals or financial expert reviews through path rerouting, it effectively prevents over-budget expenditures and inefficient investments. Finally, the path switching strategy is optimized. The path switching penalty coefficient in the formula balances the path optimization effect with switching costs, avoiding communication losses caused by frequent approver changes while ensuring decisive path adjustments when necessary, achieving an optimal combination of process flexibility and stability. Compared to the crude approach of existing technologies that rely on either fixed paths or random adjustments, this solution's dynamic rerouting mechanism significantly improves the intelligence level of the approval process and resource allocation efficiency.

[0135] In summary, the method provided by this invention can dynamically determine the execution path, breaking the limitations of fixed paths. When the approver's status is abnormal, the path can be updated in a timely manner through path rerouting, avoiding process stagnation and solving the problems of rigid approval paths and low efficiency. Secondly, it generates a dynamic summary for the approver, including semantic vectors, budget execution deviation rate, and cost performance value, enabling the approver to fully grasp the key information of the approval request and solving the problems of insufficient approval basis and difficulty in accurate evaluation. Finally, by real-time detection of the approver's status, budget execution deviation rate, and cost performance value and triggering path rerouting, dynamic control of the approval process is achieved, solving the defect of existing technologies that cannot adjust the path in a timely manner after problems are discovered. This invention ensures the continuity and flexibility of the approval process through dynamic determination of execution paths and rerouting mechanisms, significantly improving approval efficiency. The dynamic summary provides the approver with comprehensive and accurate evaluation basis, helping the approver to make more reasonable decisions and reducing the risk of approving unreasonable expenses. Real-time monitoring and path adjustment of budget execution deviation rate and cost performance value effectively strengthen the refined control of consulting fees, improve the accuracy of budget execution and cost performance level, and ultimately achieve efficient, accurate, and intelligent management of information service consulting fees.

[0136] See Figure 4 In one embodiment, the present invention also provides an information service consultation fee management system, the system comprising:

[0137] Feature vector extraction unit 100 is used to extract feature vectors from the consulting project management system in response to a consulting fee approval request.

[0138] The execution path decision unit 200 is used to input the feature vector into the approval path decision model, calculate the comprehensive decision score of each candidate path, and determine the execution path based on the calculation result of the comprehensive decision score.

[0139] The dynamic summary generation unit 300 is used to determine the identity of the current approver and generate a dynamic summary based on the approver's identity. The dynamic summary includes the semantic vector of the approval request, the budget execution deviation rate, and the cost performance value.

[0140] The path rerouting unit 400 is used to detect the approver status, budget execution deviation rate, and cost performance value. When the approver status is abnormal, the budget execution deviation rate is greater than the first threshold, or the cost performance value is less than the second threshold, the approval path rerouting is triggered to update the execution path.

[0141] In one embodiment, the execution path decision unit 200 is further configured to:

[0142] Determine the set of candidate paths, and calculate the ratio of the dimensions of the dynamic feature vector and the static feature vector of each candidate path based on the dynamic feature vector of the current consultation request and the static feature vector of each candidate path to obtain the dimension balance factor.

[0143] Three sets of trainable parameter matrices are established and multiplied by the dynamic feature vector, the static feature vector, and the dynamic-static feature cross matrix, respectively, to obtain the query, key, and value matrices of the attention mechanism; the cross-modal attention weight matrix is ​​calculated based on the query, key, and value matrices.

[0144] The path fit is calculated based on the cosine of the angle between the dynamic feature vector and the static feature vector. The comprehensive decision score for each candidate path is calculated based on the dimensional balance factor, the cross-modal attention weight matrix, and the path fit.

[0145] In one embodiment, the execution path decision unit 200 is further configured to:

[0146] Dynamic threshold filtering is performed based on the comprehensive decision score calculated for each path to obtain the remaining candidate path set after filtering;

[0147] Based on the remaining set of candidate paths, the final target execution path is calculated according to the historical average approval time of each candidate path and the system's overall average approval time.

[0148] It is understood that the system provided in this embodiment has functions or includes modules that can be used to execute the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0149] The present invention also provides an electronic device including a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs a method as described in any of the above possible implementations.

[0150] The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0151] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software 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 implementations should not be considered beyond the scope of this invention.

[0152] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will also readily understand that the various embodiments of the present invention have different focuses, and for the sake of convenience and brevity, the same or similar parts may not be repeated in different embodiments. Therefore, parts not described or not described in detail in one embodiment can be referred to in other embodiments.

Claims

1. A method for managing information service consulting fees, characterized in that, The method includes: In response to a consulting fee approval request, feature vectors are extracted from the consulting project management system. The feature vectors are input into the approval path decision model to calculate the comprehensive decision score of each candidate path, and the execution path is determined based on the calculation result of the comprehensive decision score. The step of inputting the feature vector into the approval path decision model and calculating the comprehensive decision score for each candidate path includes: Determine the set of candidate paths: ; Establish a comprehensive decision-making score calculation model: ； ； ； ； ； In the formula, Candidate paths The overall decision score, The ratio of the dimensions of dynamic features to those of static features; For dynamic feature vectors Dimensions For dynamic feature vectors The L1 norm, Candidate paths static feature vectors, Static feature vectors The infinite norm, For the Sigmoid function, For the slope parameter and center offset of the Sigmoid curve; This is the exponential amplification factor; For the attention mechanism, the query, key, and value matrix; Candidate paths Dynamic-static feature cross matrix; , , It is a trainable parameter matrix; For transpose; This is the cross-modal attention weight matrix; For path adaptation, For cosine similarity, These are the weighting coefficients for the KL divergence. and Let be the feature mapping function, satisfying: ； ； In the formula, , , , All of these are trainable parameters; The process of determining the execution path based on the calculation result of the comprehensive decision score includes: For each candidate path calculate Filtering via dynamic thresholds: ； ； In the formula, , The mean and standard deviation of the scores for the current batch. For adaptive coefficients, This is the set of paths remaining after filtering. This is the current timestamp. For the approval period; ； In the formula, For the final selected execution path, Candidate paths Historical average approval time This represents the average approval time across the entire system. Determine the identity of the current approver and generate a dynamic summary based on the approver's identity. The dynamic summary includes the semantic vector of the approval request, the budget execution deviation rate, and the cost performance value. The system monitors the approver status, budget execution deviation rate, and cost performance value. When the approver status is abnormal, the budget execution deviation rate is greater than the first threshold, or the cost performance value is less than the second threshold, the approval path rerouting is triggered to update the execution path.

2. The information service consulting fee management method according to claim 1, characterized in that, The semantic vector of the approval request includes a triple extracted from the knowledge graph, which contains a project ID, an associated entity, and an attribute value; the associated entity is an entity with an attention weight greater than 0.

7.

3. The information service consulting fee management method according to claim 1, characterized in that, The dynamic feature vector includes the cost type quantification value, the logarithmic normalized value of the amount, the project risk level score, the current budget execution deviation rate, and the contract urgency coefficient.

4. The information service consulting fee management method according to claim 1, characterized in that, The triggering of approval path rerouting to update the execution path includes: ； In the formula, Number the new candidate paths. This is the path switching penalty coefficient, with a value ranging from 0.1 to 0.

5. This represents the number of joint approvers for both the original and new paths.

5. An information service consultation fee management system, characterized in that, The system includes: The feature vector extraction unit is used to extract feature vectors from the consulting project management system in response to consulting fee approval requests. The execution path decision unit is used to input feature vectors into the approval path decision model, calculate the comprehensive decision score of each candidate path, and determine the execution path based on the calculation result of the comprehensive decision score. The step of inputting the feature vector into the approval path decision model and calculating the comprehensive decision score for each candidate path includes: Determine the set of candidate paths: ; Establish a comprehensive decision-making score calculation model: ； ； ； ； ； In the formula, Candidate paths The overall decision score, The ratio of the dimensions of dynamic features to those of static features; For dynamic feature vectors Dimensions For dynamic feature vectors The L1 norm, Candidate paths static feature vectors, Static feature vectors The infinite norm, For the Sigmoid function, For the slope parameter and center offset of the Sigmoid curve; This is the exponential amplification factor; For the attention mechanism, the query, key, and value matrix; Candidate paths Dynamic-static feature cross matrix; , , It is a trainable parameter matrix; For transpose; This is the cross-modal attention weight matrix; For path adaptation, For cosine similarity, These are the weighting coefficients for the KL divergence. and Let be the feature mapping function, satisfying: ； ； In the formula, , , , All of these are trainable parameters; The process of determining the execution path based on the calculation result of the comprehensive decision score includes: For each candidate path calculate Filtering via dynamic thresholds: ； ； In the formula, , The mean and standard deviation of the scores for the current batch. For adaptive coefficients, This is the set of paths remaining after filtering. This is the current timestamp. For the approval period; ； In the formula, For the final selected execution path, Candidate paths Historical average approval time This represents the average approval time across the entire system. The dynamic summary generation unit is used to determine the identity of the current approver and generate a dynamic summary based on the approver's identity. The dynamic summary includes the semantic vector of the approval request, the budget execution deviation rate, and the cost performance value. The path rerouting unit is used to detect the approver status, budget execution deviation rate, and cost performance value. When the approver status is abnormal, the budget execution deviation rate is greater than the first threshold, or the cost performance value is less than the second threshold, the approval path rerouting is triggered to update the execution path.

6. An electronic device, characterized in that, include: A processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs the information service consulting fee management method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which includes program instructions that, when executed by a processor of an electronic device, cause the processor to perform the information service consulting fee management method according to any one of claims 1 to 4.

Images