A dynamic approval path generation method and device based on business status

Abstract

The application discloses a kind of dynamic approval path generation method and device based on business state, belong to data processing and intelligent decision-making technical field.Business state information of to-be-approved business is acquired, feature coding is generated state vector;Based on the business attribute in state vector, pre-defined approval rule is matched through rule engine, and candidate approval path set containing hard constraint is generated;State vector and candidate set are input into multiple AI models and parallel inference is carried out, and path selection strategy, node association relationship, approval person ability score, time consumption prediction and semantic feature are obtained;According to hard constraint and inference result, candidate path is fused and decided, and final approval path is generated.The application solves the problem that static approval process is difficult to adapt to dynamic change of business state, and realizes intelligent generation and efficient decision of approval path.

Description

Technical Field

[0001] This invention relates to the field of data processing and intelligent decision-making technology, and in particular to a method and apparatus for generating dynamic approval paths based on business status. Background Technology

[0002] In enterprise-level application systems, approval processes are a core component ensuring business compliance and operational efficiency, and the rationality of their path planning directly impacts task flow timeliness and resource allocation efficiency. Currently, the generation of approval paths primarily relies on the following four technical solutions: One approach is based on a workflow engine. This approach uses standardized modeling languages ​​such as BPMN to preload fixed process definitions during system startup. During operation, tasks are executed strictly according to a pre-defined directed graph structure, with each node and transition condition being statically configured.

[0003] Second, a predefined flowchart template solution. The system administrator configures multiple fixed approval templates in the backend. When a business is initiated, the user manually selects the corresponding template according to the business type, and subsequent approval steps proceed in the pre-set order of the template.

[0004] Thirdly, the rule engine solution. Approval rules are abstracted into "condition-action" production rules, and the next approval node is dynamically determined through rule matching, achieving a certain degree of conditional branching routing.

[0005] Fourth is the AI ​​recommendation system solution. This involves training a machine learning model on historical approval data to recommend possible approval paths based on current context features when a business request is initiated.

[0006] However, the aforementioned existing technologies all have significant drawbacks: workflow engines and predefined template solutions are rigid and unable to respond to real-time changes in business status; rule engines can only handle deterministic logic and lack the ability to generalize to complex nonlinear scenarios; while AI recommendation solutions possess adaptability, they lack explicit guarantees for hard constraints such as compliance and permissions, and there is a lack of collaborative decision-making mechanisms among multiple models, making it difficult to achieve global optimization. In summary, existing approval path generation technologies struggle to achieve dynamic, accurate, and compliant path decisions, necessitating a novel solution that integrates rule constraints and multi-dimensional AI reasoning. Summary of the Invention

[0007] The present invention aims to at least partially solve one of the technical problems in the related art.

[0008] To address this, this invention discloses a dynamic approval path generation method based on business status. The method involves acquiring the business status information of the business to be approved and generating a status vector through feature encoding. Based on the business attributes in the status vector, a rule engine is used to match predefined approval rules, generating a set of candidate approval paths containing hard constraints. The status vector and the candidate path set are then input into multiple AI models for parallel inference to obtain path selection strategies, node relationships, approver competence scores, time predictions, and semantic features. Based on the hard constraints and the results of multi-model inference, a fusion decision is made on the candidate approval paths to generate the final approval path. This invention overcomes the limitations of static processes and single rules in adapting to dynamic business changes and the lack of coordination in multi-dimensional decision-making, achieving efficient and accurate intelligent approval path generation.

[0009] Another objective of this invention is to provide a dynamic approval path generation device based on business status.

[0010] To achieve the above objectives, this invention proposes a method for generating dynamic approval paths based on business status, comprising: Obtain the business status information of the business to be approved, perform feature encoding on the business status information, and generate a status vector; Based on the business attributes in the state vector, the rule engine matches predefined approval rules to generate a set of candidate approval paths containing hard constraints. The state vector and the set of candidate approval paths are input into multiple AI models for parallel reasoning to obtain path selection strategies, node association relationships, approver ability scores, time consumption predictions and semantic features, respectively. Based on the hard constraints and the reasoning results of the AI ​​model, a fusion decision is made on the candidate approval paths to generate the final approval path.

[0011] The dynamic approval path generation method based on business status according to an embodiment of the present invention may also have the following additional technical features: In one embodiment of the present invention, the step of obtaining the business status information of the business to be approved, and performing feature encoding on the business status information to generate a status vector includes: Extract the business type field and amount field from the business status information, and perform a One process on the business type field. Hot encoding maps to dimensions 1 through 8 of the state vector; where dimensions 1 through 8 correspond to contract approval, procurement approval, personnel approval, expense reimbursement approval, budget approval, payment approval, loan approval, and other business types, respectively. By Amount Field Perform normalization processing and encode the current approval node status as "not started"; The historical approval statistics data for the corresponding business type and department are queried from the data layer, and the historical statistical dimensions of the state vector are populated to generate a 57-dimensional state vector.

[0012] In one embodiment of the present invention, the step of generating a set of candidate approval paths containing hard constraints by matching predefined approval rules through a rule engine based on business attributes in the state vector includes: Retrieve the business type and amount value from the state vector, and obtain the product, department, and risk level extended fields from the original request body. Iterate through the IF statements in the rule base. THEN rule chain matches the business type, amount value, and extended fields with the conditions of each rule; For a successfully matched rule, extract the set of required approval nodes and the order constraints between nodes from the corresponding THEN clause, merge the set of required approval nodes and the order constraints of all matched rules, and perform topological sorting on the order constraints. If a cycle exists, break it according to the approver level to generate a set of hard constraints. Enumerate all possible approval path topologies under the constraints of the hard constraint set, and generate a list containing... A set of candidate approval paths for each candidate path.

[0013] In one embodiment of the present invention, the step of performing a fusion decision on candidate approval paths based on hard constraints and the inference results of the AI ​​model to generate a final approval path includes: The candidate approval paths are verified for compliance using hard constraints to ensure that each candidate path meets the requirements for required approval nodes and order constraints. By applying the parallel proposals and association strength matrix output by the graph neural network mining model, serial nodes with association strength higher than the threshold and no sequential dependence are combined into parallel nodes to generate an optimized path topology. The optimized paths are sorted based on the Q-values ​​output by the reinforcement learning policy model, and the Top Paths are selected. K optimal candidate paths; Top The K optimal candidate paths are respectively used to calculate the expected total time consumption and timeliness fluctuation risk by calling the spatiotemporal characteristic prediction model; Calculate the overall score for each candidate path using the weighted scoring formula:

[0014] in The normalized result of the Q value, This is the normalized result of the expected total time. For hard constraint satisfaction, This is a semantic reasonableness penalty term output by the multimodal perception model. , , These are time-varying weighting coefficients. The final path is determined based on a comprehensive score with fixed weighting coefficients. Bind the approver recommended by the approver capability assessment model to each node of the final path to generate the final approval path.

[0015] In one embodiment of the present invention, the state vector and the set of candidate approval paths are input into a graph neural network mining model for inference, including: The state vector and the set of candidate approval paths are input into the graph neural network mining model. The graph neural network mining model reads the embedded representation of the approver node from the heterogeneous graph of approval history, performs message passing through a multi-layer graph attention network, calculates the correlation strength matrix between approver nodes, identifies node combinations with correlation strength higher than a preset threshold and no result dependence, and outputs parallel approval suggestions and correlation strength matrix.

[0016] In one embodiment of the present invention, inputting the state vector and the set of candidate approval paths into a reinforcement learning policy model for inference includes: The state vector and the set of candidate approval paths are input into the reinforcement learning policy model, which consists of a three-layer fully connected network. The first layer maps the 57-dimensional state vector to 256 dimensions, the second layer maps it to 128 dimensions, and the third layer outputs... 3D Q-value vector; in To preset the maximum number of candidate paths, the forward propagation calculates the Q-value of each candidate path. The Q-value integrates three dimensions: pass rate, timeliness, and rule satisfaction. The candidate path with the largest Q-value is selected as the recommended path.

[0017] In one embodiment of the present invention, inputting the state vector into the approver's competence assessment model for reasoning includes: For each approval node, the 42-dimensional features of the candidate approver are queried from the data layer. These 42-dimensional features include identity features, behavioral features, ability features, and status features. The comprehensive score of each candidate approver is calculated according to the formula:

[0018] in Rate the ability This is a time-normalized value. This is the load normalization value. For professional matching scoring, candidate approvers are ranked according to the comprehensive score, and a Top 100 is recommended for each approval node. 3. Approver.

[0019] In one embodiment of the present invention, inputting historical approval time series and current candidate path topology into a spatiotemporal characteristic prediction model, and inputting multimodal data into a multimodal perception model for inference, includes: The historical approval time series and the current candidate path topology are input into the spatiotemporal characteristic prediction model, which uses LSTM. The Transformer hybrid architecture predicts the expected time of each approval node and the overall completion time of the path, and assesses the risk of timeliness fluctuations. The text description, numerical features, and graph structure features in the state vector are encoded by BERT, fully connected layer, and graph attention network, respectively, and then a fused feature vector is generated through an attention fusion layer as semantic features.

[0020] In one embodiment of the present invention, it further includes: After the approval process is completed, the actual execution data is collected, the actual reward value is calculated, and the state vector, the selected final path, the actual reward value and the next state vector are stored in the experience replay buffer. When the buffer accumulates to the preset batch size, the parameter update of the deep Q network is triggered. The collaborative relationships of the approvers in this approval are incrementally updated in the heterogeneous graph of the approval history, and only the embedding vectors of the affected nodes are updated. Update the monthly approval rate and average processing time of approvers as dynamic statistical indicators; The deviation between the actual time taken and the expected time taken is recorded in the calibration log. When the accumulated deviation exceeds the preset threshold, the spatiotemporal characteristic prediction model parameters are fine-tuned and updated.

[0021] To achieve the above objectives, another aspect of the present invention proposes a dynamic approval path generation device based on business status, comprising: The state-aware encoding module is used to acquire the business state information of the business to be approved, perform feature encoding on the business state information, and generate a state vector. The rule matching and generation module is used to generate a set of candidate approval paths containing hard constraints by matching predefined approval rules with business attributes in the state vector through the rule engine. The intelligent parallel reasoning module is used to input the state vector and the set of candidate approval paths into multiple AI models for parallel reasoning, and obtain path selection strategy, node association relationship, approver ability score, time consumption prediction and semantic features respectively. The fusion decision output module is used to make fusion decisions on candidate approval paths based on hard constraints and the reasoning results of AI models, and generate the final approval path.

[0022] This invention discloses a method and apparatus for generating dynamic approval paths based on business status. It acquires business status information, encodes features to generate a state vector, and uses a rule engine to match and generate a candidate path set with hard constraints. The state vector and candidate set are then input into multiple AI models for parallel inference to obtain path selection strategies, node relationships, approver competence scores, time predictions, and semantic features. Finally, based on the hard constraints and the results of multi-model inference, a fusion decision is made to generate the approval path. This method effectively solves the limitations of existing technologies where static process templates and single rule engines are unable to adapt to dynamic changes in business status and lack collaborative integration of multi-dimensional decision elements. It achieves integrated solution across the entire process, from state perception, rule constraints, multi-model parallel inference to fusion decision-making, significantly improving the intelligence level and decision accuracy of approval path generation, and enhancing the method's adaptability and engineering practical value in complex enterprise scenarios.

[0023] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a dynamic approval path generation method based on business status according to an embodiment of the present invention; Figure 2 This is an architectural design diagram according to an embodiment of the present invention; Figure 3 This is a flowchart of the approval path generation according to an embodiment of the present invention; Figure 4 This is a flowchart of the rule engine matching process according to an embodiment of the present invention; Figure 5 This is a flowchart of AI model inference according to an embodiment of the present invention; Figure 6 This is a flowchart of path verification according to an embodiment of the present invention; Figure 7 This is a flowchart of real-time dynamic adjustment according to an embodiment of the present invention; Figure 8 This is a flowchart of the approver recommendation process according to an embodiment of the present invention; Figure 9 This is a schematic diagram of a dynamic approval path generation device based on business status according to an embodiment of the present invention. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0027] The following description, with reference to the accompanying drawings, describes a method and apparatus for generating dynamic approval paths based on business status, according to an embodiment of the present invention.

[0028] The core idea of ​​this invention is to address the problem of low-quality path generation in existing approval processes due to rigid processes, simplistic rules, and a lack of multi-dimensional collaborative decision-making. This is achieved by constructing a dynamic approval path generation method covering the entire process of "state awareness—rule constraint—parallel reasoning—fusion decision-making," thus realizing intelligent and adaptive generation of approval paths. First, the business state information of the business to be approved is acquired and its features are encoded to generate a state vector. Based on the business attributes in the state vector, a rule engine matches predefined approval rules to generate a set of candidate approval paths containing hard constraints. This candidate set, along with the state vector, is input into multiple AI models for parallel reasoning to obtain path selection strategies, node relationships, approver competence scores, time predictions, and semantic features. Finally, based on the hard constraints and the reasoning results of each model, a fusion decision is made on the candidate approval paths to generate the final approval path. By integrating the above-mentioned technical approach of "rule constraint - multi-modal reasoning - fusion decision-making", this invention transforms the traditional approval decision-making based on static templates or single rules into an intelligent path generation system that can adapt to dynamic changes in business status, fully mine multi-dimensional feature information, and realize multi-mechanism collaborative decision-making, significantly improving the accuracy, adaptability and engineering practical value of approval path generation.

[0029] The following description, with reference to the accompanying drawings, describes a method and apparatus for generating dynamic approval paths based on business status, according to an embodiment of the present invention.

[0030] like Figure 1 As shown, the dynamic approval path generation method based on business status of the present invention includes the following steps: S1, Obtain the business status information of the business to be approved, perform feature encoding on the business status information, and generate a status vector; S2, based on the business attributes in the state vector, generates a set of candidate approval paths containing hard constraints by matching predefined approval rules through the rule engine; S3, input the state vector and the set of candidate approval paths into multiple AI models for parallel reasoning to obtain path selection strategy, node association relationship, approver ability score, time consumption prediction and semantic features respectively; S4, based on the hard constraints and the reasoning results of the AI ​​model, performs a fusion decision on the candidate approval paths to generate the final approval path.

[0031] In general, the technical problems to be solved by this invention include: insufficient dynamic adjustment: the approval path cannot be adjusted in real time according to the business status, resulting in limited flexibility; limitations of preset rules: it only supports simple condition judgments (such as amount thresholds) and cannot handle complex business scenarios; insufficient utilization of historical data: it does not make full use of historical approval records to optimize the approval path and lacks continuous learning capabilities; non-intelligent approver selection: approver selection lacks a personalized recommendation mechanism and cannot be optimized based on historical preferences; and low approval efficiency: the approval path is lengthy, the approval node settings are unreasonable, and the overall approval timeliness is poor.

[0032] To achieve the above objectives, the present invention provides the following technical solution: like Figure 2 As shown, the overall system architecture of the present invention adopts a five-layer hierarchical design.

[0033] First layer: User interaction layer. Application submission: Users submit application information for services awaiting approval; Approval path display: Shows users the dynamic approval path generated by the system; Approval operation interface: Provides an interface for approvers to perform approval operations; Approval status query: Users can query the current approval progress and status.

[0034] The second layer: Business logic layer. Business status acquisition module: acquires complete status information of pending business transactions; Rule matching engine module: matches predefined approval rules based on business attributes; Path generation decision module: integrates rules and AI models to generate dynamic paths; Approver recommendation module: recommends the optimal approver based on capability assessment.

[0035] The third layer: AI model layer. This layer contains five core AI models that provide intelligent computing capabilities for the business logic layer: Reinforcement Learning Policy Model (DRL): learns the optimal path selection strategy based on Deep Q-Network (DQN); Graph Neural Network Mining Model (GNN): mines implicit relationships in the approval history; Approver Competency Assessment Model (ABE): evaluates and ranks approvers' competence from multiple dimensions; Spatiotemporal Feature Prediction Model (STL): predicts the time consumption of approval nodes and path change trends; Multimodal Perception Model (MMS): integrates multimodal features such as text, numerical values, and graph structures.

[0036] Fourth layer: Rule engine layer. Approval rule management: rule creation, deletion, modification, querying, and version management; rule condition configuration: supports multi-condition combination logic and nested conditions; rule execution engine: rule matching and execution engine; rule verification mechanism: rule conflict detection and priority handling.

[0037] Fifth layer: Data layer. Business data storage: business attributes and status data of business items pending approval; Approval history storage: historical approval records and approval behavior data; Rule configuration storage: definition and configuration data of approval rules; Model parameter storage: parameters and status data of AI models.

[0038] Specifically, the data flow at each layer is explained as follows: After a user submits an approval request through the interaction layer, the system completes the following end-to-end data flow sequentially: Step 1: Business status collection and feature encoding.

[0039] like Figure 3 As shown, after receiving the approval application (containing fields such as business type, amount, department, product, and risk level), the user interaction layer passes the original JSON request body to the business status acquisition module of the business logic layer. This module performs the following transformations: extracting business attribute fields and constructing a 57-dimensional business status vector S_initial; performing One-Hot encoding on the business type field (dimensions 1-8), and encoding the amount field... Normalization is performed (dimension 9), and the current approval node status is set to "not started" (dimension 10=1, the rest are 0); historical approval statistics for this business type and department are queried from the data layer, and dimensions 26-57 (historical number of approvals, average time taken, pass rate, rejection rate, etc.) are populated; output: a 57-dimensional state vector s_t ∈ R^57, which is passed to the rule engine layer and the AI ​​model layer as shared input.

[0040] Step 2: Rule engine layer matching and hard constraint generation.

[0041] like Figure 4As shown, the rule engine layer receives the business attribute portion of s_t (dimensions 1-8: business type, dimension 9: amount) and extended fields such as product, department, and risk level from the original JSON, and performs the following steps: Rule condition matching: Traverse the IF-THEN rule chain in the rule base, and compare the business type and amount values ​​in s_t with the conditions of each rule. Extended fields such as product, department, and risk level are directly obtained from the original JSON request body and participate in rule condition matching in parallel with the encoded dimensions in s_t. Example: Rule R_001 has the condition IF Business Type = Contract Approval AND Amount > 1 million AND Product = 5G Equipment. The business type is obtained from dimension 1 of s_t, the amount is obtained from dimension 9 of s_t through inverse normalization, and the product is obtained from the "Product" field in the original JSON. A successful match triggers its THEN clause. Hard constraint set generation: The THEN clause of a successfully matched rule must output the set of approval nodes M_must = {Department Manager, Product Manager, Deputy General Manager, General Manager}, along with the order constraints between nodes (Department Manager must precede Product Manager, and Deputy General Manager precedes General Manager). Candidate path enumeration: Enumerates all possible approval path topologies under hard constraints. For example, serial paths [Department Manager → Product Manager → Deputy General Manager → General Manager] and parallel paths [{Department Manager, Product Manager} → Deputy General Manager → General Manager] generate N = 10~50 candidate paths. Output: a candidate path set P_candidates = {p_1, p_2, ..., p_N}, where each candidate path p_i contains a list of nodes, node type (serial / parallel), and a hard constraint satisfaction flag. This is then passed to the path generation decision module in the business logic layer.

[0042] Step 3: Parallel inference and feature fusion at the AI ​​model layer.

[0043] like Figure 5As shown, after the AI ​​model layer receives s_t and P_candidates, five AI models perform inference in parallel: The GNN mining model reads the heterogeneous approval history graph (containing approximately 1000 approver nodes and 8 edge types) from the data layer and executes a 3-layer Graph Attention Network (GAT) message passing. It embeds the approver nodes in each candidate path, calculates the correlation strength matrix A_relation between nodes, and identifies node combinations that can be approved in parallel (e.g., if the collaboration frequency between department managers and product managers is > threshold 0.7, parallel processing is recommended). Output: Graph embedding vector h_graph ∈ R^64 and parallel suggestions; The DRL policy model inputs s_t and P_candidates into a trained DQN network (3 fully connected layers: 57→256→128→N), calculates the Q-value Q(s_t, p_i) for each candidate path through forward propagation, and selects the path with the largest Q-value as the DRL recommended path p_drl. The Q-value integrates three dimensions: pass rate, timeliness, and rule satisfaction. The ABE approver recommendation model: For each approval node, it queries the 42-dimensional features (identity, behavior, ability, and status) of candidate approvers from the data layer, and ranks them using a comprehensive scoring function Score(a) = 0.35×AbilityScore + 0.25×(1-TimelinessNorm) + 0.25×(1-WorkloadNorm) + 0.15×SpecialtyMatch, recommending the Top-3 approvers for each approval node. The STL spatiotemporal prediction model: Receives historical approval time series and the current candidate path topology, and uses an LSTM-Transformer hybrid architecture to predict the expected time t_i of each approval node and the overall path completion time T_total, while simultaneously assessing the timeliness fluctuation risk of each node in the path. The MMS multimodal perception model: Receives s_t The text description (reason for business application), numerical features (amount, quantity) and graph structure features are processed by BERT encoding text → 128 dimensions, FC encoding numerical values ​​→ 32 dimensions, and GAT encoding graph structure → 64 dimensions. Then, an attention fusion layer is used to generate a fused feature vector h_fused ∈ R^128, which is used to enhance the semantic understanding ability of subsequent decisions.

[0044] Step 4: Integration and output of the path generation decision module.

[0045] like Figure 6As shown, the path generation decision module of the business logic layer receives the output of all the above models and executes the following fusion decision process: Hard constraint verification: The candidate paths are verified for compliance using the hard constraint set M_must of the rule engine to confirm that all paths meet the requirements of the necessary approver (the candidate paths have been generated based on M_must in the rule engine layer, and this step is a cross-validation to ensure that subsequent AI model optimization does not introduce constraint violations); Parallel optimization and adjustment: The parallel suggestions of the GNN model and the association strength matrix A_relation are applied to combine serial nodes with high association strength and no order dependency into parallel nodes to generate the optimized path topology; Q-score ranking and Top-K screening: Based on the Q-score ranking of the DRL model, the Top-K (K=3~5) optimal candidate paths are selected; STL time consumption evaluation: The STL model is called for the Top-K paths to calculate the expected total time consumption and timeliness fluctuation risk; Comprehensive scoring and final decision: According to the weighted scoring formula: Scoring of Top-K paths; among which The normalized result of the Q value, This is the normalized result of the expected total time. For hard constraint satisfaction, This is a semantic reasonableness penalty term output by the multimodal perception model. , , These are time-varying weighting coefficients. Uses fixed weighting coefficients. ABE Approver Binding: Binds the Top-1 ABE-recommended approver to each node in the final path; Output: Final approval path JSON (including node list, approver, node type, and estimated time), returned to the user interaction layer for display.

[0046] Furthermore, such as Figure 7 As shown, the real-time dynamic adjustment process is as follows: Starting from the "Start" node, the system continuously executes the "Listen for Business Status Change Events" operation. When a real change in business status is detected, the system enters the judgment branch, triggering the "Identify Change Type" step to determine the specific attributes and scope of impact of the status change. Subsequently, based on the identification results, the system "triggers path regeneration" and executes the "Generate New Approval Path" operation to complete the construction of candidate paths. After generation, the system "notifies relevant approvers" of the path change information and "updates the original path" to synchronize the current approval process with the latest status. After the process is completed, it returns to the listening stage, maintaining continuous response to changes in business status and achieving dynamic adaptive adjustment of the approval path.

[0047] Furthermore, such as Figure 8As shown, the approver recommendation process is as follows: Starting from the "Start" node, the "Get Approval Node Information" operation is executed first to clarify the specific approval stage that requires an approver to be assigned. Then, the "Get Candidate Approver Information" operation extracts all available candidate approvers for that node from the system or data layer. Based on this, the process concurrently calls two intelligent analysis modules: "MMS Multimodal Perception" and "ABE Capability Assessment." The former performs semantic understanding of the multimodal data of the approval context, while the latter quantifies and scores the comprehensive capabilities of the candidate approvers. After obtaining the above analysis results, the "Multi-Dimensional Scoring" step is entered, integrating the indicators of each dimension. Then, a comprehensive score is generated for each candidate approver through "Weighted Fusion Scoring," and the order of merit is determined by "Sorting by Score." Finally, a "Top-N Recommendation List" is output as the approver recommendation result, and the process ends. The entire process achieves fully automated processing from information acquisition and multi-dimensional intelligent analysis to comprehensive ranking and recommendation.

[0048] The method of this invention effectively solves the prominent problems in the prior art, such as the inability of approval paths to be dynamically adjusted according to business status, the fact that preset rules only support simple condition judgments, insufficient utilization of historical data and personalized recommendations of approvers, and low approval efficiency due to lengthy paths. It realizes the integrated solution of the entire process from status perception, rule constraints, multi-model parallel reasoning to fusion decision-making, significantly improving the intelligence level, decision accuracy and adaptability of approval path generation, and enhancing the engineering practical value of the method in complex enterprise scenarios.

[0049] The application scenarios of the method in this embodiment of the invention also include: Complete process example.

[0050] Scenario: A user submits a contract worth 1.5 million yuan for approval.

[0051] Step 1: The user interaction layer receives and forwards the request.

[0052] Users fill out and submit a contract approval application through the system front end. The application includes the following business fields: { Business Type: Contract Approval Amount: 150 Department: Technology Department Product: 5G Equipment Risk Level: Medium Reason for application: "Purchase of 5G base station equipment for network expansion in South China" } The user interaction layer forwards this JSON request body to the business status collection module of the business logic layer.

[0053] Step 2: Business status collection and feature encoding.

[0054] After receiving the raw JSON, the business status collection module performs the following encoding: Business type One-Hot encoding: Dimension 1=1 (contract approval), the rest 0; Amount normalization: log(150+1) / log(1000+1)≈0.724; Current approval stage encoding: Dimension 10=1 (not started), the rest 0; Historical statistics query: Read the historical approval statistics data (historical number of approvals, average time consumption, pass rate, rejection rate, etc.) of this business type and department from the data layer, and populate dimensions 26-57.

[0055] Output a 57-dimensional state vector s_0 ∈ R^57, and pass it to the rule engine layer and the AI ​​model layer.

[0056] Step 3: Rule engine matching and candidate path generation.

[0057] After receiving s_0, the rule engine layer performs rule matching: iterates through the rule base and matches rule R_001 - "IF Amount > 1 million AND Product = 5G Device THEN Must = [Department Manager, Product Manager, Deputy General Manager, General Manager]".

[0058] The hard constraint set M_must contains four mandatory approvers and sequential constraints. Candidate paths are enumerated: a serial path [Department Manager → Product Manager → Deputy General Manager → General Manager], parallel path 1 [{Department Manager, Product Manager} → Deputy General Manager → General Manager], parallel path 2 [Department Manager → {Product Manager, Deputy General Manager} → General Manager], etc., for a total of 12 candidate paths. The output P_candidates is passed to the AI ​​model layer.

[0059] Step 4: Parallel inference at the AI ​​model layer.

[0060] Five models are executed in parallel: GNN: Reads the embedding representation of approver nodes from the heterogeneous graph and calculates the association strength of approvers in this business scenario. It finds that the department manager (approver A) and product manager (approver B) have a collaboration frequency higher than the threshold in the historical records, and recommends parallel approval; DRL: Inputs s_0 and 12 candidate paths into the DQN network and calculates the Q value. The path [parallel {approver A, approver B} → deputy general manager → general manager] has the highest Q value among the candidate paths because this path has a high historical pass rate and short total time; ABE: Scores each approval node: Department manager candidate - approver A (higher overall score), approver C (medium overall score); product manager - approver B (higher overall score); deputy general manager - approver D (higher overall score); general manager - approver E (highest overall score); STL: Predicts the time of each node and the overall completion time of the path, with a high probability of pass; MMS: Reads the application reason and extracts semantic features: medium to high urgency, high technical complexity, and high strategic importance. The fusion feature h_fused is recommended to strike a balance between timeliness and accuracy.

[0061] Step 5: Integration of path generation and decision-making modules.

[0062] The fusion decision module receives all model outputs: Hard constraint verification: All candidate paths contain 4 mandatory approvers and all pass; GNN parallel optimization: The department manager and product manager are merged into parallel nodes, and the path topology becomes [parallel {Approver A, Approver B} → Approver D → Approver E]; DRL Q-value ranking: The top-3 paths are ranked from high to low according to their Q-values; STL time consumption evaluation: The prediction time of the top-1 path is within an acceptable range; Comprehensive score: The top-K paths are scored according to the FinalScore formula, and the top-1 path has the highest comprehensive score.

[0063] This path was ultimately chosen as the optimal path.

[0064] Step 6: Path verification.

[0065] Perform triple verification on the selected path: Connectivity check: Node [parallel {Approver A, Approver B}] → Approver D → Approver E, connected and without broken links; Permission verification: Approver A is a department manager (Level 3), Approver B is a product manager (Level 3), Approver D is a deputy general manager (Level 2), and Approver E is a general manager (Level 1), with the level increment being valid; Rule compliance check: The path contains all 4 required approvers, and the order conforms to the hard constraint.

[0066] All three verifications passed.

[0067] Step 7: Approver binding and path output.

[0068] Assign approvers to each node: Parallel nodes are bound to approvers A and B; the approver D node is bound to approver D; the approver E node is bound to approver E. Generate the final approval path JSON (the following is sample data; actual values ​​will change dynamically based on the business environment): { "applicationId": "CON-2026-00158", "totalAmount": 1,500,000 "businessType": "Contract Approval", "product": "5G equipment", "approvalPath": { "nodes": [ { "id": "parallel1", "type": "parallel", "approvers": [ {"id": "A001", "role": "Department Manager"}, {"id": "B003", "role": "Product Manager"} ] }, { "id": "node2", "type": "serial", "approver": {"id": "D005", "role": "Deputy General Manager"} }, { "id": "node3", "type": "serial", "approver": {"id": "E001", "role": "General Manager"} } ] }, "estimatedTime": "approximately one day", "passProbability": "High", "decision summary": "GNN identifies parallel optimizations, DRL strategy selects the optimal path, STL prediction time is acceptable, and the overall score is optimal."} Step 8: The result is returned to the user interaction layer.

[0069] The path generation and decision-making module returns the aforementioned JSON to the user interaction layer. The user interface displays: a visualization of the approval path (parallel nodes are highlighted); information cards for approvers at each node; estimated total time and probability of approval; and a summary of the decision-making basis.

[0070] After the user confirms, the system will push the approval flow to the first approval node.

[0071] Step 9: Online learning feedback loop.

[0072] After the approval process is completed, the system collects actual execution data (actual total time, actual pass rate of each node), compares it with the predicted value, and triggers online learning: DQN experience replay: (s_0, p_selected, R_real, s_final) are stored in the experience replay buffer pool. When the buffer pool accumulates to batch_size (64 items), a DQN network parameter update is triggered; GNN graph update: the collaboration relationship of this approval is incrementally updated to the heterogeneous graph. After the new edge is added, only the embedding vector of the affected node is updated (local update strategy, no need for full graph retraining); ABE index refresh: the monthly pass rate and average time of each approver are updated as dynamic statistical indicators (pure statistical update, no model retraining involved); STL time series calibration: the deviation between the actual time and the predicted time is recorded in the calibration log. When the accumulated deviation exceeds the preset threshold (e.g., 20%), the LSTM parameter fine-tuning update is triggered.

[0073] The update frequency and triggering conditions of each model are set independently to avoid full retraining after each approval, thus forming an asynchronous, incremental continuous optimization loop.

[0074] The application scenarios of the method in this invention effectively solve prominent problems in the prior art, such as the inability of approval paths to dynamically adjust with business status, the fact that preset rules only support simple condition judgments, insufficient utilization of historical data and personalized recommendations from approvers, and low approval efficiency due to lengthy paths. It realizes integrated solution of the entire process from status perception, rule constraints, multi-model parallel reasoning to fusion decision-making, and is supplemented by an online learning feedback loop to achieve continuous optimization, significantly improving the intelligence level, decision accuracy and adaptability of approval path generation, and enhancing the engineering practical value of the method in complex enterprise scenarios.

[0075] Furthermore, the AI ​​model layer of this invention comprises five core models, each undertaking a different technical function, working together to intelligently generate the approval path. The following is a detailed description of each model: Cold start solution.

[0076] When the system is deployed for the first time and there is no historical approval data, the initialization strategies of each model are as follows: for dimensions 26-57 (historical statistical data) in the 57-dimensional state vector, industry default values or zero initialization are adopted; the GNN heterogeneous graph is initially constructed automatically based on the organizational structure, the features of approver nodes are preset by using organizational level and post information, and the initial edge weights are preset based on post responsibilities; the DQN model uses pre-training weights based on a simulation environment, and the simulation environment generates simulated approval trajectories according to enterprise approval rules; when the amount of historical data is less than T=10, the STL model adopts a zero-padding+mask mechanism, and outputs a default time consumption estimation based on the prior of organizational architecture levels.

[0077] As the system operates and accumulates real data, each model gradually transitions from a preset / simulation mode to a data-driven mode.

[0078] Model 1: Reinforcement learning policy model (DRL / DQN).

[0079] Application scenario: learning the optimal approval path selection strategy; balancing rule constraints and historical efficiency; intelligently selecting from candidate paths that satisfy the rules.

[0080] Function: modeling the approval path problem as a Markov Decision Process (MDP); learning the optimal policy through interaction with the environment; outputting the expected reward value of each candidate path to guide path selection.

[0081] Input parameter: state vector s_t (57-dimensional multi-dimensional service state encoding).

[0082] Output parameter: Q-value matrix Q(s, a): output dimension: N_max (N_max is the preset maximum number of candidate paths, which is 50 and covers all actual scenarios); value range: [-10, +10] (expected reward value); output meaning: long-term cumulative expected reward of each candidate path; usage: when the actual number of candidate paths N < N_max, only the first N dimensions are valid, and the remaining dimensions are masked and set to zero; argmax only selects the path with the maximum Q-value from the first N dimensions.

[0083] Architecture of the three-layer fully connected Q-value evaluation network: input layer: 57-dimensional state vector; hidden layer 1 (fully connected): 256 neurons, ReLU activation, Dropout=0.1; hidden layer 2 (fully connected): 128 neurons, ReLU activation, Dropout=0.1; output layer: N_max-dimensional Q-value vector (linear output), N_max=50 is a fixed value.

[0084] Training and optimization process: Phase 1: Environment definition; State space S: 57-dimensional business state vector; Action space A: a set of discrete actions for selecting execution paths; Reward function R: R = w1×pass rate + w2×timeliness + w3×rule satisfaction.

[0085] Phase 2: Hyperparameter settings. Learning rate α: 0.0001 (range: 1e-4 ~ 1e-3); Discount factor γ: 0.99 (range: 0.95 ~ 0.995); Exploration rate ε: decays from 1.0 to 0.05 (40,000 steps); Batch size: 64 (range: 32 ~ 128); Experience replay buffer: 100,000 (range: 50,000 ~ 500,000); Target network update cycle: 100 steps (range: 50 ~ 200).

[0086] Cold start strategy: When historical data is insufficient in the early stages of system deployment, a DQN model pre-trained in a simulation environment is used as the initial strategy. The simulation environment generates simulated approval trajectories based on the company's historical approval rules and typical approval patterns. The pre-trained model gradually adapts to the actual business distribution through incremental fine-tuning after the accumulation of real data. Training can begin when the experience replay buffer accumulates to 1000 records (using a smaller batch size of 16), and gradually transitions to standard training parameters as the data volume increases.

[0087] Phase 3: Training Loop. First, a batch of historical interaction samples is randomly sampled from the experience replay buffer. Each sample includes the current state, selected action, immediate reward, and next state. Then, the optimal action value for the next state is calculated using the target network, and this value is weighted and summed with the immediate reward using a discount factor to obtain the target Q-value. Simultaneously, the current state is input into the current Q-network for forward propagation to obtain the predicted Q-value. Next, the mean squared error between the predicted Q-value and the target Q-value is used as the loss function to measure the degree of deviation in the network output. Finally, gradient zeroing, loss backpropagation, and parameter update operations are performed in the optimizer to complete one iteration of network weight optimization, thereby gradually approaching the optimal action value function.

[0088] Phase 4: Model Evaluation. Test set accuracy: the matching rate between the selected path and the optimal path; Average reward: the cumulative reward in the test environment; Convergence: the curve of reward value changing with the number of training steps.

[0089] Model 2: Graph Neural Network Mining Model (GNN).

[0090] Application scenarios: Uncovering implicit relationships between approvers, business types, departments, and products; discovering indirect relationships that are difficult to capture using traditional statistical methods; supporting graph-based approver recommendation and cold start scenarios.

[0091] Functions: Construct a heterogeneous graph structure of approval history; learn node representations through information propagation in a multi-layer graph attention network (GAT); calculate the association strength between nodes and discover implicit relationships.

[0092] Input parameters: Heterogeneous graph structure data: approver nodes (approximately 1,000, initial 64-dimensional features); Business type nodes (approximately 50, initially 32-dimensional features); Department nodes (approximately 100, initially 32-dimensional features); Product nodes (approximately 200, initially 32-dimensional features).

[0093] Edge Types and Weights: Approver-Business Type Edge: Calculated based on historical approval frequency; Formula: weight = count(approver, bizType) / total. Meaning: Records the percentage of a particular approver's approvals for a specific business type. Approver-Department Edge: Affiliation relationship, fixed weight = 1.0. Meaning: Indicates that a particular approver is permanently assigned to a specific department. Approver-Product Edge: Approval capability weight, calculated using the formula: weight = success_rate(product) × log(count+1). Meaning: Records a particular approver's approval capability and proficiency with a specific type of product. Department-Business Type Edge: Business distribution weight; Formula: weight = count(dept, bizType) / total; Meaning: Records the degree of responsibility a department has for a specific business type.

[0094] Output parameters: Enhanced node representation and association matrix: Node embedding matrix: Approver embedding: Dimension [1,000,64]; Business type embedding: Dimension [50, 64]; Department / product embedding: Dimension [300, 64]; Node similarity matrix: Value range [0, 1].

[0095] GNN mining model architecture and training.

[0096] The graph neural network mining model adopts a two-layer graph attention network (GAT) architecture, with each layer having a hidden dimension of 64 dimensions. It is equipped with a 4-head attention mechanism, uses ReLU as the activation function, and finally outputs node embedding vectors.

[0097] The training process is divided into four stages: Phase 1: Graph Data Construction. Extract four types of nodes—approver, business type, department, and product—and their corresponding edge relationships from historical approval records. Calculate edge weights based on indicators such as approval frequency and success rate to complete node feature initialization.

[0098] Phase Two: Pre-training and Fine-tuning. First, link prediction is used as the self-supervised pre-training task, with a binary cross-entropy loss function. The general graph structure representation of approver nodes is learned by calculating the inner product similarity of node pairs. Then, based on the pre-training, a parallel approval recommendation task is introduced for fine-tuning. The loss function also uses binary cross-entropy, and the approval pairs actually executed in parallel in the past are used as labels to align the general graph representation to this specific downstream task of parallel approval recommendation.

[0099] Phase 3: Hyperparameter settings. The learning rate is 0.001, the hidden dimension is 64, the number of attention heads is 4, the Dropout is 0.2, and the number of training epochs is 200.

[0100] Phase Four: Model Application. Node embedding vectors are used for similarity calculation, supporting approver recommendation in cold start scenarios and can be migrated to new business scenarios. Result dependency verification is performed during parallel approval recommendations—if the approval result of node A affects whether node B needs approval, then this node pair cannot be parallelized; parallel approval is only possible when there is no order dependency and no result dependency between the two nodes.

[0101] Functional positioning of the ABE model.

[0102] The approver competence assessment model integrates features from four dimensions: identity, behavior, competence, and status. It outputs an interpretable competence score that dynamically reflects the approver's current status, provides a quantitative basis for approver selection, and supports real-time ranking and personalized recommendations.

[0103] Note: ABE and GNN employ different dimensional feature spaces for approvers, serving different technical objectives. GNN's 64-dimensional features focus on node embedding representations within the graph structure, used to uncover implicit relationships between approvers; ABE's 42-dimensional features focus on assessing the interpretability of individual approvers, used for approver ranking and recommendation. Both characterize approvers from different perspectives and are used collaboratively in the path generation and decision-making module.

[0104] Input parameters: Approver's multi-dimensional characteristics (42 dimensions): behavioral characteristics (16 dimensions); ability characteristics (12 dimensions); status characteristics (6 dimensions); Output parameters: Approver's overall score: Capability score: 0.0~1.0; Timeliness score: 0.0~1.0 (based on normalized average approval time); Workload score: 0.0~1.0 (based on normalized current workload); Professional matching score: 0.0~1.0 (based on the professional matching degree between the approver and the business type).

[0105] Final scoring formula: Score(a)=0.35×AbilityScore(a)+0.25×(1-TimelinessNorm(a))+0.25×(1-WorkloadNorm(a)) + 0.15×SpecialtyMatch(a).

[0106] Approver competence assessment model (ABE).

[0107] The ABE model takes 42 multidimensional features of the approver as input, including four dimensions: identity, behavior, ability, and status. The model architecture adopts a two-layer fully connected network: the first fully connected layer has 256 neurons (ReLU activation, Dropout=0.2), the second fully connected layer has 128 neurons (ReLU activation, Dropout=0.2), and the output layer is activated by Sigmoid to output a comprehensive score in the range of 0 to 1.

[0108] The training process is divided into three stages: Stage 1: Based on the actual performance scores of the approvers, labeled data is constructed, and the statistics and feature engineering of each dimension are completed; Stage 2: The model is trained using binary cross-entropy or mean squared error as the loss function and Adam as the optimizer (learning rate 0.001, training epochs 100); Stage 3: The accuracy is evaluated by the matching degree between the scores and the actual performance, and the contribution of each feature is analyzed by using the attention mechanism to ensure the interpretability of the model.

[0109] Spatiotemporal property prediction model (STL).

[0110] The STL model takes the historical approval time series of the 10 most recent approval nodes as input. Each time step contains 16-dimensional features, covering approval node ID encoding (4-dimensional), approval status (1-dimensional, including negative rejection values), normalized approval time (1-dimensional), day of the week encoding (7-dimensional), and time features such as working day identifier, the first 5 days of the month, and the last 5 days of the month.

[0111] The model architecture employs a two-layer LSTM (64 hidden units in the first layer and 32 hidden units in the second layer, both with tanh activation), followed by a fully connected output layer. It outputs two prediction results: approval time (in hours, ranging from 0 to 168) and approval success probability (ranging from 0 to 1). This model captures the regularity of approval node time consumption through temporal modeling, assisting in risk assessment and supporting approval path optimization and resource scheduling.

[0112] Multimodal sensing model (MMS).

[0113] The MMS model processes data of different modalities through three independent branches: the text modality branch takes business descriptions or application reasons as input, and after encoding by BERT-Base (12 layers, 768 dimensions), it is pooled into a 128-dimensional text vector; the numerical modality branch takes numerical values ​​such as amount, quantity, and deadline as input, and after normalization, it is encoded into a 32-dimensional numerical vector through a fully connected layer; the graph structure modality branch takes business relationship graphs as input, and after encoding by GCN or GraphSAGE, it is encoded into a 64-dimensional graph vector.

[0114] The output vectors of the three branches are concatenated into a 224-dimensional fusion vector, which is then deeply fused through a two-layer fully connected network combined with an attention mechanism, ultimately outputting a 128-dimensional fusion feature vector. This vector is used to support deep semantic understanding of complex business states and improve the accuracy of approval path generation.

[0115] Model collaboration relationship.

[0116] ABE focuses on the interpretable assessment of the individual approver's capabilities, STL focuses on the time dimension prediction of the approval process, and MMS focuses on the semantic fusion of multi-source heterogeneous information. Together with GNN and DRL, these three constitute a complete AI inference layer, supporting the intelligent generation of approval paths from different dimensions.

[0117] Hybrid Decision Engine: A Method for the Synergy and Integration of Rules and AI Models.

[0118] This invention proposes a three-layer hybrid decision-making architecture that prioritizes hard constraints, optimizes soft constraints, and integrates multiple models. This architecture resolves a long-standing technical contradiction in workflow systems: pure rule-based systems can guarantee 100% compliance but lack adaptability to dynamic business scenarios, while pure AI systems offer flexibility but cannot guarantee compliance. By using the hard constraints of the rule engine as a safety boundary, optimizing the soft constraints of the AI ​​model within the boundary, and integrating the outputs of multiple models under a unified scoring framework, a balance between compliance and flexibility is achieved.

[0119] In addition, the three-layer hybrid decision-making architecture of "hard constraint priority, soft constraint optimization, and multi-model fusion" proposed in this invention fundamentally solves the technical contradiction that pure rule systems lack flexibility and pure AI systems cannot guarantee compliance bottom lines.

[0120] I. Hard Constraint Priority Layer – Rule Engine

[0121] The rule engine uses a structured IF-THEN syntax to define a set of hard constraints. Each rule has the following format: IF<condition combination>THEN<list of required approvers>[WITH<order constraint>]. Taking rule R_001 as an example: when the business type is contract approval, the amount is greater than 1 million, and the product category is 5G equipment / core network / cloud computing, the required approvers are, in order, the department manager, product manager, deputy general manager, and general manager, with corresponding order constraints attached.

[0122] When multiple rules are matched, a "maximum set merging" strategy is used: M_final = ∪ M_must(r), which means taking the union of the sets of all required approvers for the matched rules. Sequential constraints are also topologically sorted after taking the union. If a cycle exists, it is broken according to the approver level (L1>L2>L3). The final output is the set of hard constraints H.

[0123] II. Soft Constraint Optimization Layer – Parallel Inference Algorithm for AI Models

[0124] DRL strategy optimization: The approval path selection is modeled as a Markov decision process, with the state as a 57-dimensional business vector and the action as a candidate path. The Q-value update follows the Bellman equation, and the reward function integrates three dimensions: pass rate, timeliness, and rule satisfaction, with weights of 0.4, 0.35, and 0.25, respectively.

[0125] GNN Parallel Optimization: Message passing is based on graph attention networks. A 4-head attention mechanism is used to calculate the association strength between nodes. When the association strength exceeds the threshold of 0.7 and there is no sequential dependency, parallel approval is recommended to shorten the approval path from a structural level.

[0126] ABE approver matching: The system scores approvals based on four dimensions: comprehensive ability, timeliness, workload, and professional matching, with weights of 0.35, 0.25, 0.25, and 0.15, respectively, to recommend the optimal approver for each approval stage.

[0127] III. Integrated Decision-Making Layer – Path Generation and Comprehensive Scoring

[0128] The comprehensive scoring formula is: FinalScore(p) = α(t)×Q_norm(p) + β(t)×(1 / T_norm(p)) + γ(t)×RuleScore(p) + δ×MMS_penalty(p). Where Q_norm is the normalized Q-value of the DRL output, T_norm is the normalized time consumption of the STL prediction, RuleScore is the hard constraint satisfaction, and MMS_penalty is the semantic reasonableness penalty. These four factors are organically integrated within a unified scoring framework; a higher score indicates better overall path quality.

[0129] IV. Online Learning Feedback Layer – Continuous Optimization of the Closed Loop.

[0130] Each approval process triggers six asynchronous update flows: DQN experience replay stores the quadruplets in a buffer pool and triggers parameter updates in batches; GNN graph updates use a local strategy to update only the embeddings of affected nodes; ABE metric refresh is a purely statistical calculation; STL time-series calibration fine-tunes the LSTM when the deviation exceeds the threshold; and MMS weight adjustment dynamically optimizes the time-varying coefficients. The update frequency of each model is set independently, forming an asynchronous incremental continuous optimization closed loop.

[0131] V. Architecture and Collaboration Mechanism.

[0132] The rules layer provides a safety boundary, ensuring 100% compliance; the AI ​​layer optimizes efficiency and quality within the boundaries; the fusion layer unifies scoring decisions; and the online learning layer drives continuous evolution—these four layers work together to achieve both compliance and flexibility, forming a complete technical closed loop of "rules as a safety net, AI optimization, and data-driven approach." This invention, through a hybrid decision architecture integrating five core modules—rule engine, GNN graph neural network, deep reinforcement learning (DRL), adaptive business encoder (ABE), spatiotemporal transfer learning (STL), and multimodal semantic calibration (MMS)—is expected to achieve significant improvements in approval efficiency, path quality, compliance, generalization ability, and inference efficiency.

[0133] To verify the technical advancement of this invention, three comparative schemes were set up: Scheme A is a pure rule engine, where the path is hard-coded in the business rules without AI participation; Scheme B is rules plus a single machine learning model, where candidate paths are generated by rules and then sorted by the single model; Scheme C is rules plus DRL, introducing DQN for path selection but lacking GNN parallel optimization and MMS semantic calibration. Based on the above schemes, this invention constructs a complete hybrid decision-making architecture combining rules and multiple AI models, with each module working synergistically and complementarily.

[0134] Dimension 1: Expected Improvement in Approval Efficiency. This invention utilizes Generative Neural Networks (GNNs) to uncover implicit parallel relationships between approval nodes, merging redundant nodes in traditional sequential paths into parallel approval groups, which is expected to significantly reduce the total approval time. Compared to the bottleneck of waiting time for approvers in pure rule engine solutions, this invention makes the workflow more compact and efficient, structurally addressing the business pain points of long approval cycles and slow response times. The graph structure induction capability of GNNs transforms the path that originally required sequential waiting of each node into a path where nodes satisfying dependencies can be processed synchronously and in parallel, thereby significantly reducing the overall workflow time while maintaining compliance.

[0135] Dimension Two: Optimization of Path Quality Expectations. This invention utilizes the MMS multimodal semantic calibration mechanism to achieve precise alignment between the approval path and business semantics, avoiding path bias caused by single-rule matching. Simultaneously, ABE performs deep semantic encoding of the business context, combined with the graph structure induction capabilities of GNN, which is expected to effectively identify and eliminate redundant nodes in the path, generating a more concise and accurate approval path and reducing resource waste caused by excessive approval. Multimodal fusion enables the system to understand the implicit semantics in the application reason text, ensuring that path selection matches the business substance, rather than relying solely on hard matching of explicit fields.

[0136] Dimension Three: Rule Compliance Assurance. This invention employs a triple verification mechanism—connectivity verification, permission compliance verification, and business rule conflict detection—through a hard constraint priority layer to ensure that the generated approval paths strictly comply with corporate management systems, compliance requirements, and organizational structure constraints. The AI ​​model's soft constraint optimization is performed under the premise of 100% compliance guaranteed by hard constraints, minimizing the risk of rule violations from an architectural design perspective and completely resolving the compliance and trust issues faced by pure AI solutions. The rule engine defines security boundaries first, and the AI ​​model only performs efficiency optimization within the compliant space, ensuring that no approval paths violating corporate regulations are ever generated.

[0137] Dimension Four: Cold Start Generalization Ability. This invention, through the STL mechanism, can transfer organizational structure knowledge and approval behavior patterns from richly learned tasks in scenarios with sparse data, such as new business types or new approvers. The graph structure inductive ability of GNNs enables the system to make effective inferences based on node structural features and multi-hop neighbor relationships, even when facing unseen approvers. It is expected to maintain high path recommendation accuracy in cold start scenarios, solving the performance degradation problem of traditional machine learning solutions in sparse data scenarios. The pre-training and fine-tuning training strategy allows the model to quickly adapt to new business scenarios, providing reasonable recommendations without accumulating large amounts of historical data.

[0138] Dimension Five: Model Inference Efficiency. This invention employs a multi-model parallel inference architecture, where five AI models execute in parallel on independent computing units without blocking each other. In real-world production environments, approval decisions are low-frequency scenarios, and the impact of single-inference latency (within seconds) on user experience is controllable. Each model can further reduce inference latency through lightweight techniques such as quantization and knowledge distillation, meeting the performance requirements of high-concurrency scenarios.

[0139] Summary of Expected Results: The rule-based multi-AI model hybrid decision-making architecture proposed in this invention ensures compliance baseline through a rule engine, shortens the approval path by mining implicit parallel relationships through GNN, optimizes long-term decision quality through DRL, achieves multimodal semantic calibration through MMS, and solves the cold start problem through STL. These modules form a complementary and enhanced technical closed loop. It is expected to significantly improve approval efficiency, path quality, rule compliance, and cold start generalization ability compared to existing single-technology solutions, fully demonstrating the technological advancement and practical value of this invention.

[0140] To achieve the above invention, such as Figure 9 As shown, this embodiment also provides a dynamic approval path generation device 10 based on business status, the device 10 including: The state-aware encoding module 100 is used to acquire the business state information of the business to be approved, perform feature encoding on the business state information, and generate a state vector.

[0141] The rule matching and generation module 200 is used to generate a set of candidate approval paths containing hard constraints by matching predefined approval rules with business attributes in the state vector through the rule engine.

[0142] The intelligent parallel reasoning module 300 is used to input the state vector and the set of candidate approval paths into multiple AI models for parallel reasoning, and obtain path selection strategy, node association relationship, approver ability score, time consumption prediction and semantic features respectively.

[0143] The fusion decision output module 400 is used to make fusion decisions on candidate approval paths based on hard constraints and the reasoning results of AI models, and generate the final approval path.

[0144] The apparatus of this invention effectively solves the problems in the prior art where approval paths cannot dynamically adapt to changes in business status and where rule constraints and AI optimization are disconnected. It achieves end-to-end integrated solution from business status perception, rule constraint generation, multi-model collaborative reasoning to fusion decision output, significantly improving the dynamic adaptability and decision accuracy of approval path generation, and enhancing the intelligence level and engineering practical value of the approval system in complex business scenarios.

[0145] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0146] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for generating dynamic approval paths based on business status, characterized in that, include: Obtain the business status information of the business to be approved, perform feature encoding on the business status information, and generate a status vector; Based on the business attributes in the state vector, the rule engine matches predefined approval rules to generate a set of candidate approval paths containing hard constraints. The state vector and the set of candidate approval paths are input into multiple AI models for parallel reasoning to obtain path selection strategies, node association relationships, approver ability scores, time consumption predictions and semantic features, respectively. Based on the hard constraints and the reasoning results of the AI ​​model, a fusion decision is made on the candidate approval paths to generate the final approval path.

2. The method as described in claim 1, characterized in that, The step of obtaining the business status information of the pending approval business, performing feature encoding on the business status information, and generating a status vector includes: Extract the business type field and amount field from the business status information, and perform a One process on the business type field. Hot encoding maps to dimensions 1 through 8 of the state vector; where dimensions 1 through 8 correspond to contract approval, procurement approval, personnel approval, expense reimbursement approval, budget approval, payment approval, loan approval, and other business types, respectively. By Amount Field Perform normalization processing and encode the current approval node status as "not started"; Query the historical approval statistics data of the corresponding business type and department from the data layer, populate the historical statistical dimensions of the state vector, and generate a 57-dimensional state vector.

3. The method according to claim 1, characterized in that, The process of generating a set of candidate approval paths containing hard constraints by matching predefined approval rules with business attributes in the state vector using a rule engine includes: Retrieve the business type and amount value from the state vector, and obtain the product, department, and risk level extended fields from the original request body. Iterate through the IF statements in the rule base. THEN rule chain matches the business type, amount value, and extended fields with the conditions of each rule; For a successfully matched rule, extract the set of required approval nodes and the order constraints between nodes from the corresponding THEN clause, merge the set of required approval nodes and the order constraints of all matched rules, and perform topological sorting on the order constraints. If a cycle exists, break it according to the approver level to generate a set of hard constraints. Enumerate all possible approval path topologies under the constraints of the hard constraint set, and generate a list containing... A set of candidate approval paths for each candidate path.

4. The method as described in claim 1, characterized in that, The step of fusing and deciding on candidate approval paths based on hard constraints and the inference results of the AI ​​model to generate the final approval path includes: The candidate approval paths are validated for compliance using hard constraints to ensure that each candidate path meets the requirements for required approval nodes and order constraints. By applying the parallel proposals and association strength matrix output by the graph neural network mining model, serial nodes with association strength higher than the threshold and no sequential dependence are combined into parallel nodes to generate an optimized path topology. The optimized paths are sorted based on the Q-values ​​output by the reinforcement learning policy model, and the Top Paths are selected. K optimal candidate paths; Top The K optimal candidate paths are respectively used to calculate the expected total time consumption and timeliness fluctuation risk by calling the spatiotemporal characteristic prediction model; Calculate the overall score for each candidate path using the weighted scoring formula: in The normalized result of the Q value, This is the normalized result of the expected total time. For hard constraint satisfaction, This is a semantic reasonableness penalty term output by the multimodal perception model. , , These are time-varying weighting coefficients. The final path is determined based on a comprehensive score with fixed weighting coefficients. Bind the approver recommended by the approver capability assessment model to each node of the final path to generate the final approval path.

5. The method as described in claim 1, characterized in that, The state vector and the set of candidate approval paths are input into the graph neural network mining model for inference, including: The state vector and the set of candidate approval paths are input into the graph neural network mining model. The graph neural network mining model reads the embedded representation of the approver node from the heterogeneous graph of approval history, performs message passing through a multi-layer graph attention network, calculates the correlation strength matrix between approver nodes, identifies node combinations with correlation strength higher than a preset threshold and no result dependence, and outputs parallel approval suggestions and correlation strength matrix.

6. The method as described in claim 1, characterized in that, The state vector and the set of candidate approval paths are input into the reinforcement learning policy model for inference, including: The state vector and the set of candidate approval paths are input into the reinforcement learning policy model, which consists of a three-layer fully connected network. The first layer maps the 57-dimensional state vector to 256 dimensions, the second layer maps it to 128 dimensions, and the third layer outputs... 3D Q-value vector; in To preset the maximum number of candidate paths, the forward propagation calculates the Q-value of each candidate path. The Q-value integrates three dimensions: pass rate, timeliness, and rule satisfaction. The candidate path with the largest Q-value is selected as the recommended path.

7. The method according to claim 1, characterized in that, The state vector is input into the approver's competence assessment model for reasoning, including: For each approval node, the 42-dimensional features of the candidate approver are queried from the data layer. These 42-dimensional features include identity features, behavioral features, ability features, and status features. The comprehensive score of each candidate approver is calculated according to the formula: in Rate the ability This is a time-normalized value. This is the load normalization value. For professional matching scoring, candidate approvers are ranked according to the comprehensive score, and a Top 100 is recommended for each approval node.

3. Approver.

8. The method according to claim 1, characterized in that, The historical approval time series and current candidate path topology are input into the spatiotemporal characteristic prediction model, and multimodal data is input into the multimodal perception model for inference, including: The historical approval time series and the current candidate path topology are input into the spatiotemporal characteristic prediction model, which uses LSTM. The Transformer hybrid architecture predicts the expected time of each approval node and the overall completion time of the path, and assesses the risk of timeliness fluctuations. The text description, numerical features, and graph structure features in the state vector are encoded by BERT, fully connected layer, and graph attention network, respectively, and then a fused feature vector is generated through an attention fusion layer as semantic features.

9. The method according to claim 1, characterized in that, Also includes: After the approval process is completed, the actual execution data is collected, the actual reward value is calculated, and the state vector, the selected final path, the actual reward value and the next state vector are stored in the experience replay buffer. When the buffer accumulates to the preset batch size, the parameter update of the deep Q network is triggered. The collaborative relationships of the approvers in this approval are incrementally updated in the heterogeneous graph of the approval history, and only the embedding vectors of the affected nodes are updated. Update the monthly approval rate and average processing time of approvers as dynamic statistical indicators; The deviation between the actual time taken and the expected time taken is recorded in the calibration log. When the accumulated deviation exceeds the preset threshold, the spatiotemporal characteristic prediction model parameters are fine-tuned and updated.

10. A dynamic approval path generation device based on business status, characterized in that, include: The state-aware encoding module is used to acquire the business state information of the business to be approved, perform feature encoding on the business state information, and generate a state vector. The rule matching and generation module is used to generate a set of candidate approval paths containing hard constraints by matching predefined approval rules with business attributes in the state vector through the rule engine. The intelligent parallel reasoning module is used to input the state vector and the set of candidate approval paths into multiple AI models for parallel reasoning, and obtain path selection strategy, node association relationship, approver ability score, time consumption prediction and semantic features respectively. The fusion decision output module is used to make fusion decisions on candidate approval paths based on hard constraints and the reasoning results of AI models, and generate the final approval path.

Images