Enterprise business-oriented process node intelligent optimization and scheduling method and system

Abstract

This invention provides a method and system for intelligent optimization and scheduling of process nodes for enterprise business, relating to the field of business processing technology. The method includes obtaining a description of business objectives and a sequence of executed process steps, retrieving historical process nodes based on deep vector embedding, generating business intent features using a bidirectional long short-term memory network, constructing a context state representation for structural matching, and finally constructing a dynamic heterogeneous graph structure to generate a weighted recommendation sequence. This invention achieves accurate recommendation and intelligent scheduling of enterprise processes, improving process execution efficiency and the achievement rate of business objectives.

Description

Technical Field

[0001] This invention relates to business processing technology, and more particularly to a method and system for intelligent optimization and scheduling of process nodes for enterprise business. Background Technology

[0002] With the deepening of enterprise digital transformation, business process management and optimization have become core elements for improving operational efficiency. Traditional enterprise process management mainly relies on manual experience for design and adjustment, lacking in-depth mining and utilization of massive historical data. Enterprise knowledge graphs, as a structured knowledge representation method, can effectively organize and connect internal business process knowledge, providing a knowledge foundation for process optimization. In recent years, deep learning technology has made significant progress in natural language processing and graph structure data analysis, providing new technical paths for the intelligent optimization and scheduling of enterprise processes. Currently, there are some process recommendation systems based on rules or simple machine learning methods, but these systems are often designed for specific scenarios, with limited versatility and adaptability.

[0003] Existing process optimization methods mostly rely on predefined process templates, making it difficult to adjust in real time according to dynamically changing business objectives and contexts. They lack a deep understanding of business semantics and the ability to respond accurately, and thus cannot meet the complex and ever-changing business needs of enterprises.

[0004] Traditional process node recommendation technology fails to effectively integrate enterprise historical process knowledge with the current execution context. The recommendation results are often limited to surface similarity matching, ignoring the deep temporal dependencies and structural constraints between process nodes, which may result in recommended process nodes that do not conform to business logic or have low execution efficiency.

[0005] The existing system lacks a self-learning and continuous optimization mechanism, and cannot automatically adjust and optimize the node association patterns in the knowledge base based on process execution results and business feedback. As a result, the system performance is difficult to improve in sync with the development of enterprise business, and its long-term use value is limited. Summary of the Invention

[0006] This invention provides a method and system for intelligent optimization and scheduling of process nodes for enterprise business, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a method for intelligent optimization and scheduling of process nodes for enterprise business, comprising:

[0008] Obtain the business objective description and the sequence of executed process steps of the current process to be built, and retrieve the set of historical process nodes and their associated business semantic tags from the enterprise knowledge graph based on deep vector embedding technology;

[0009] The business objective description is semantically parsed using a bidirectional long short-term memory network to generate business intent features and constraints. The similarity between the business intent features and the business semantic tags is vectorized to obtain a set of candidate nodes for the objective.

[0010] Based on the sequence of executed process steps and the business intent features, the context state representation of the current process is constructed using the deep vector embedding technology. The context state representation is then structurally matched with the predecessor node pattern and successor node pattern of each candidate node in the candidate node set in the enterprise knowledge graph to filter out a subset of target candidate nodes that meet the business intent.

[0011] A dynamic heterogeneous graph structure is constructed based on the target candidate node subset, and the temporal correlation between nodes and the business intent features are uniformly mapped to the feature space to generate a recommendation sequence with adaptive weights. The node migration pattern in the enterprise knowledge graph is continuously optimized through the bidirectional long short-term memory network.

[0012] A bidirectional long short-term memory network is applied to the business objective description for semantic parsing to generate business intent features and constraints. The similarity between the business intent features and the business semantic tags is calculated by vectorization to obtain a set of target-oriented candidate nodes, including:

[0013] Based on a bidirectional long short-term memory network, bidirectional semantic modeling is performed on the description of the business objective. Through a skip-layer connection mechanism, semantic accumulation features are captured in the forward encoding path and the reverse encoding path respectively, and features are reorganized at each word position to generate a word-level semantic representation sequence.

[0014] Based on the skip-layer connection mechanism, the contribution of each word in the word-level semantic representation sequence to the expression of business intent is calculated. The semantic representations of words with contribution values ​​higher than the preset contribution threshold are used to construct a hierarchical feature pyramid to form business intent features. At the same time, the constraint type ontology in the enterprise knowledge graph is combined to perform structured representation and generate constraint conditions.

[0015] Identify business entities from the business objective description, retrieve the association paths between the business entities in the enterprise knowledge graph, and filter out semantically coherent association paths as semantic dependencies based on the bidirectional long short-term memory network.

[0016] The business intent features, constraints, and semantic dependencies are heterogeneously integrated. During the integration process, the hierarchical feature pyramid is used to establish a mapping between business intent features and constraints, as well as an enhanced mapping between business intent features and semantic dependencies, to generate a business target semantic vector.

[0017] Similarity is calculated between the semantic vector of the business objective and the business semantic tag vector of each node in the historical process node set. The connectivity of nodes is verified according to the semantic dependency relationship. Nodes that do not meet the connectivity or violate the constraints are filtered to obtain a candidate node set.

[0018] The business intent features, constraints, and semantic dependencies are heterogeneously integrated. During the integration process, a mapping between business intent features and constraints, as well as an enhanced mapping between business intent features and semantic dependencies, are established through the hierarchical feature pyramid to generate a business target semantic vector, including:

[0019] The business intent features are hierarchically analyzed based on a deep self-attention network. At the macro level, business domain features are captured, and at the micro level, the combined features of business actions and objects are extracted. A two-layer feature pyramid is constructed through a feature transfer model.

[0020] The constraints are semantically matched with the business intent features in the two-layer feature pyramid through a cross-layer progressive encoder. The limiting effect of the constraints on business actions and object combinations is identified at the micro level, and the limiting effect is transmitted to the macro level based on the deep self-attention network, thus establishing a cross-layer mapping relationship from constraints to business intent features.

[0021] The feature transfer model is used to perform semantic association modeling between the business entity association path in the semantic dependency relationship and the two-layer feature pyramid. Based on the cross-layer mapping relationship, the semantic dimension that needs to be enhanced is determined and the enhancement information is extracted to construct the cross-layer enhancement mapping of the semantic dependency relationship.

[0022] Based on the cross-layer mapping relationship and the cross-layer enhancement mapping, the two-layer feature pyramid is optimized and adjusted. The adjusted micro-level features and macro-level features are adaptively fused through the deep self-attention network to generate a business target semantic vector.

[0023] Based on the sequence of executed process steps and the business intent features, the context state representation of the current process is constructed using the deep vector embedding technique, including:

[0024] The deep temporal coding network is used to model each process step in the executed process step sequence, generate a representation vector for each step, and construct a dynamic propagation graph based on the temporal dependency relationship between the representation vectors to capture the execution influence of the preceding steps on the subsequent steps and generate a step sequence vector containing temporal dependency information.

[0025] The business intent feature is semantically aligned with the step sequence vector, and dynamic weights are assigned to the execution steps in the executed process step sequence that are similar to the business intent feature to generate an intent-related step vector.

[0026] The dynamic propagation graph is used to perform semantic matching analysis on the intent-related step vector and the business intent feature, and the step sequence vector and the business intent feature are reconstructed based on the deep temporal coding network to generate an initial context representation;

[0027] In the enterprise knowledge graph, a semantic propagation path between process steps and target business entities is constructed based on the initial context representation. The context information is dynamically enhanced along the semantic propagation path to generate a context state representation of the current process.

[0028] The process steps in the executed process step sequence are modeled using a deep temporal coding network to generate representation vectors for each step, and a dynamic propagation graph is constructed based on the temporal dependencies between these representation vectors, including:

[0029] Extract step attribute features and step execution time features from each process step in the executed process step sequence. Jointly encode the step attribute features and step execution time features through deep tensor decomposition and embed time dimension position information to generate an initial representation vector for the step.

[0030] Based on the recursive attention structure, the execution sequence of the executed process steps is analyzed for temporal dependency. The execution influence of the preceding process steps on the subsequent process steps is mapped to the initial representation vector of the step through the deep tensor decomposition, generating a step representation vector containing temporal dependency information.

[0031] The recursive attention structure is used to calculate the temporal correlation strength between the representation vectors of each step in the executed process step sequence, and an edge weight mapping is constructed for process step pairs that exceed a preset correlation threshold.

[0032] A dynamic propagation graph is constructed by taking each process step in the sequence of executed process steps as a node, taking the step representation vector as the node feature, and taking the edge weight mapping to determine the directed edge connection relationship.

[0033] The context state representation is structurally matched with the predecessor and successor node patterns of each candidate node in the candidate node set within the enterprise knowledge graph to filter out a subset of target candidate nodes that satisfy the business intent, including:

[0034] Based on the deep graph projection architecture, the context state representation is transformed into the node relationship expression of the enterprise knowledge graph, and the expected patterns of the predecessor node and the expected patterns of the successor node in the current process are constructed through dynamic feature propagation.

[0035] For each candidate node in the candidate node set, the deep graph projection architecture is used to construct a node relationship context in the enterprise knowledge graph, and the predecessor node pattern and successor node pattern of the candidate node are generated.

[0036] A graph structure comparison method is used to perform similarity analysis between the predecessor node pattern and the expected predecessor node pattern of each candidate node, and similarity analysis between the successor node pattern and the expected successor node pattern. The structure matching features of each candidate node are generated through dynamic feature propagation.

[0037] The graph structure comparison method is used to perform deep matching between the unmet intent information in the context state representation and the business semantic labels of each candidate node to generate intent satisfaction features; the structure matching features and the intent satisfaction features are fused through the deep graph projection architecture to filter out a subset of target candidate nodes.

[0038] For each candidate node in the candidate node set, the node relationship context is constructed in the enterprise knowledge graph using the deep graph projection architecture, generating the predecessor node pattern and successor node pattern of the candidate node, including:

[0039] For each candidate node in the candidate node set, trace the incoming edge relationship of the candidate node in the enterprise knowledge graph, extract the set of predecessor nodes with incoming edge relationships to the candidate node and their node types, and obtain the relationship type of the incoming edge relationship to construct a topological representation of the predecessor structure information.

[0040] The predecessor structure information is semantically mapped using the deep graph projection architecture. The node feature vectors and type information of the predecessor node set are uniformly projected into the semantic space, and the predecessor node pattern of the candidate node is generated by edge relationship aggregation.

[0041] The outgoing edge relationships of the candidate nodes are analyzed using the same tracing method. The set of successor nodes and their node types are extracted. Combined with the relationship types of the outgoing edge relationships, a topological representation of the successor structure information corresponding to the predecessor structure information is constructed.

[0042] The node feature vectors and type information of the successor structure information are projected onto the same semantic space as the predecessor node pattern using the deep graph projection architecture, and the successor node pattern of the candidate node is generated by edge relationship aggregation.

[0043] A second aspect of this invention provides a process node intelligent optimization and scheduling system for enterprise business, comprising:

[0044] The first unit is used to obtain the business objective description and the sequence of executed process steps of the current process to be built, and to retrieve the set of historical process nodes and their associated business semantic tags from the enterprise knowledge graph based on deep vector embedding technology;

[0045] The second unit is used to apply a bidirectional long short-term memory network to perform semantic parsing on the business objective description, generate business intent features and constraints, and perform vectorized similarity calculation on the business intent features and the business semantic tags to obtain a set of target-oriented candidate nodes.

[0046] The third unit is used to construct a context state representation of the current process based on the sequence of executed process steps and the business intent features, using the deep vector embedding technology, and to perform structural matching between the context state representation and the predecessor node pattern and successor node pattern of each candidate node in the candidate node set in the enterprise knowledge graph, thereby filtering out a subset of target candidate nodes that meet the business intent.

[0047] The fourth unit is used to construct a dynamic heterogeneous graph structure based on the target candidate node subset, uniformly map the temporal correlation between nodes and the business intent features to the feature space, generate a recommendation sequence with adaptive weights, and continuously optimize the node migration pattern in the enterprise knowledge graph through the bidirectional long short-term memory network.

[0048] A third aspect of the present invention provides an electronic device, comprising:

[0049] processor;

[0050] Memory used to store processor-executable instructions;

[0051] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] The beneficial effects of this application are as follows:

[0054] This invention combines deep vector embedding technology with bidirectional long short-term memory networks to achieve intelligent optimization and scheduling of enterprise process nodes, effectively improving the accuracy and efficiency of business process construction. Compared with existing technologies, this invention can automatically recommend optimal process nodes based on business objectives and historical execution experience, significantly reducing the time cost of manual process design and adjustment.

[0055] This invention constructs a dynamic heterogeneous graph structure and adaptive weighted recommendation sequences, enabling more precise matching of process node selection to specific business needs, thus enhancing the adaptability and flexibility of the process. Simultaneously, the system continuously learns and optimizes node migration patterns in the knowledge graph, allowing process recommendations to continuously improve as business evolves, providing enterprises with a sustainable process management solution.

[0056] This invention combines business semantic understanding with graph structure analysis, enabling intelligent recommendations based on contextual state representation while ensuring the correctness of business logic. This effectively solves the problems of blindness and subjectivity in traditional process design. This data-driven process optimization method not only improves the execution efficiency of business processes but also helps enterprises accumulate and solidify business knowledge, forming reusable process assets and providing strong support for enterprise digital transformation. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the intelligent optimization and scheduling method for enterprise business processes according to an embodiment of the present invention.

[0058] Figure 2 This is a flowchart of the candidate node selection process based on context depth feature projection in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0060] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0061] Figure 1 This is a flowchart illustrating the intelligent optimization and scheduling method for enterprise business process nodes according to an embodiment of the present invention, as shown below. Figure 1As shown, the method includes:

[0062] Obtain the business objective description and the sequence of executed process steps of the current process to be built, and retrieve the set of historical process nodes and their associated business semantic tags from the enterprise knowledge graph based on deep vector embedding technology;

[0063] The business objective description is semantically parsed using a bidirectional long short-term memory network to generate business intent features and constraints. The similarity between the business intent features and the business semantic tags is vectorized to obtain a set of candidate nodes for the objective.

[0064] Based on the sequence of executed process steps and the business intent features, the context state representation of the current process is constructed using the deep vector embedding technology. The context state representation is then structurally matched with the predecessor node pattern and successor node pattern of each candidate node in the candidate node set in the enterprise knowledge graph to filter out a subset of target candidate nodes that meet the business intent.

[0065] A dynamic heterogeneous graph structure is constructed based on the target candidate node subset, and the temporal correlation between nodes and the business intent features are uniformly mapped to the feature space to generate a recommendation sequence with adaptive weights. The node migration pattern in the enterprise knowledge graph is continuously optimized through the bidirectional long short-term memory network.

[0066] In one optional implementation, a bidirectional long short-term memory network is applied to the business objective description for semantic parsing to generate business intent features and constraints. The similarity between the business intent features and the business semantic tags is then calculated using vectorization to obtain a set of target-oriented candidate nodes, including:

[0067] Based on a bidirectional long short-term memory network, bidirectional semantic modeling is performed on the description of the business objective. Through a skip-layer connection mechanism, semantic accumulation features are captured in the forward encoding path and the reverse encoding path respectively, and features are reorganized at each word position to generate a word-level semantic representation sequence.

[0068] Based on the skip-layer connection mechanism, the contribution of each word in the word-level semantic representation sequence to the expression of business intent is calculated. The semantic representations of words with contribution values ​​higher than the preset contribution threshold are used to construct a hierarchical feature pyramid to form business intent features. At the same time, the constraint type ontology in the enterprise knowledge graph is combined to perform structured representation and generate constraint conditions.

[0069] Identify business entities from the business objective description, retrieve the association paths between the business entities in the enterprise knowledge graph, and filter out semantically coherent association paths as semantic dependencies based on the bidirectional long short-term memory network.

[0070] The business intent features, constraints, and semantic dependencies are heterogeneously integrated. During the integration process, the hierarchical feature pyramid is used to establish a mapping between business intent features and constraints, as well as an enhanced mapping between business intent features and semantic dependencies, to generate a business target semantic vector.

[0071] Similarity is calculated between the semantic vector of the business objective and the business semantic tag vector of each node in the historical process node set. The connectivity of nodes is verified according to the semantic dependency relationship. Nodes that do not meet the connectivity or violate the constraints are filtered to obtain a candidate node set.

[0072] When performing bidirectional semantic modeling of business objective descriptions using a bidirectional Long Short-Term Memory (LSTM) network, a parallel processing architecture of forward encoding and backward encoding paths is employed. The forward encoding path processes each word sequentially from the beginning of the text, while the backward encoding path processes each word in reverse order from the end of the text. The hidden layer dimension of both paths is set to 512. A skip-layer connection mechanism sets connection nodes in the second, fourth, and sixth layers of the network, performing feature concatenation operations between the output of the current layer and the outputs of the previous two layers. During forward encoding, the hidden state at each word position is calculated and generated using LSM units. The weight matrices of the forget gate, input gate, and output gate have dimensions of 512 x 812, and the bias vector dimension is 512. Backward encoding uses the same network parameter configuration, but processes the word sequence in the opposite direction.

[0073] Semantic feature accumulation is achieved through a skip-layer connection mechanism. When recombining features at each word position, the forward path hidden state, the reverse path hidden state, and the skip-layer connection features are concatenated dimensionally to form a comprehensive semantic representation vector for that word position. Each element of the word-level semantic representation sequence has 1536 dimensions, and the sequence length is consistent with the number of words in the input text. For long texts exceeding 128 words, a sliding window segmentation method is used, with the window overlap length set to 32 words to ensure that semantic continuity is not truncated.

[0074] The contribution of each word to the expression of business intent is calculated based on the word-level semantic representation sequence output by the skip-layer connection mechanism. Attention weights are obtained by the dot product of the word-level semantic representation and the learnable query vector. The query vector has a dimension of 1536 and is adaptively updated during training via backpropagation. The dot product result is normalized using the softmax function, generating an attention distribution with weights ranging from zero to one, where the sum of all word weights is always equal to one. A contribution threshold of 0.15 is set; words exceeding this threshold are marked as high-contribution words, and their corresponding semantic representation vectors are selected for subsequent feature pyramid construction.

[0075] The hierarchical feature pyramid adopts a three-layer pyramid structure. The bottom layer directly contains the original semantic representation vectors of all high-contribution words, maintaining a dimension of 1536. The middle layer performs local feature fusion on the semantic representations of adjacent words through one-dimensional convolution operations, with a kernel size of 3, a stride of 1, and 768 output channels, using the ReLU activation function. The top layer compresses the feature maps of the middle layer into a fixed-dimensional business intent feature vector through global average pooling, resulting in a final dimension of 768. The convolutional layers are configured with batch normalization, a momentum parameter of 0.9, an epsilon parameter of 1 x 10⁻⁵, and a dropout ratio of 0.2 to prevent overfitting.

[0076] The constraint generation process utilizes a structured representation based on the constraint type ontology within the enterprise knowledge graph. This ontology includes four main types: time constraints, resource constraints, permission constraints, and business rule constraints. Constraint entity recognition employs a named entity recognition method based on conditional random fields, with a tag set covering time entities, resource entities, permission entities, and rule entities. The recognition module extracts constraint-related keywords from the business objective description and performs semantic matching with the constraint type ontology in the knowledge graph. Constraint entities with a similarity exceeding 0.8 are retained. The structured representation uses a triple format, containing three fields: constraint type, constraint object, and constraint value. Time constraints use the ISO 8601 standard format, resource constraints are represented using numeric types, permission constraints use predefined enumerated values, and business rule constraints are described using Boolean expressions.

[0077] Business entity recognition extracts key business objects from business objective descriptions and processes them using a pre-trained named entity recognition model. This model is built on the BERT architecture and fine-tuned on a domain-specific corpus. Entity types cover six categories: organizations, personnel, products, services, processes, and systems. The entity recognition confidence threshold is set to 0.9; recognition results below this threshold are automatically filtered. Identified business entities undergo path retrieval within the enterprise knowledge graph. The retrieval algorithm is based on a depth-first search strategy using graph traversal, with a maximum search depth limited to five levels; paths exceeding this depth are truncated.

[0078] The semantic coherence screening of associated paths is evaluated using a bidirectional long short-term memory network. The evaluation network receives the semantic vector sequence of each node on the path as input and outputs a coherence score. The coherence score is represented by a continuous numerical value from zero to one; paths with a score higher than 0.7 are considered semantically coherent and retained. The screening process also verifies the consistency of path types, ensuring that the relationship types on the path conform to business logic constraints. For example, there must be a hierarchical relationship between personnel entities and organizational entities, and an inclusion or dependency relationship between product entities and service entities. Semantic dependencies are represented by a directed graph structure, where nodes represent business entities, directed edges represent dependencies, and edge weights are determined by the semantic coherence score of the associated paths, with weight values ​​ranging from 0.7 to 1.

[0079] The heterogeneous information integration process maps business intent features, constraints, and semantic dependencies to an 800-dimensional vector space. Business intent features are linearly transformed from 768 dimensions to 800 dimensions through a fully connected layer, with the weight matrix initialized using the Xavier method. Constraints are converted from structured triple representations into dense 800-dimensional vectors through a dedicated embedding layer; the size of the embedding dictionary is dynamically determined based on the number of constraint types. Semantic dependencies are encoded into 800-dimensional vector representations using a graph neural network architecture with two graph attention layers, each with eight attention heads and a 100-dimensional hidden layer. Finally, the readout layer aggregates node features to generate a graph-level representation.

[0080] When establishing the mapping relationship between business intent features and constraints in the hierarchical feature pyramid, a multilayer perceptron network is used to calculate the relevance matrix. The multilayer perceptron contains two hidden layers: the first hidden layer has a dimension of 512, and the second hidden layer has a dimension of 256. The ReLU activation function is used, and the dropout ratio is set to 0.3. The relevance matrix has a dimension of 768 multiplied by the number of constraints, and the matrix element values ​​represent the correlation strength between each dimension of the business intent feature and the constraints. Enhanced mapping with semantic dependencies is achieved through a residual connection mechanism, which weights the semantic dependency features and integrates them into the business intent features. The weighting coefficients are dynamically calculated using an attention mechanism; the attention query vector comes from the business intent features, and the key-value vector comes from the semantic dependency features.

[0081] The business target semantic vector generation employs a gated feature fusion network, which adaptively weights and fuses vector representations of three heterogeneous information sources. The gated network receives the concatenation of the three vector representations as input and outputs three fusion weight values. These weight values ​​are normalized using a softmax function and then used for weighted summation. The fused business target semantic vector has 800 dimensions and serves as the query representation for subsequent similarity calculations.

[0082] Similarity calculation compares the semantic vector of the business objective with the business semantic label vectors of each node in the historical process node set. The business semantic label vectors are pre-generated using the same embedding method, with a uniform dimensionality of 800. The similarity calculation employs the cosine similarity metric, calculating the dot product of two vectors divided by the product of their respective vector lengths; the similarity value ranges from -1 to +1. Nodes with a similarity higher than 0.6 are initially selected as candidate nodes.

[0083] Semantic dependency verification verifies the connectivity requirements of candidate nodes, checking whether each candidate node has a reachable path in the semantic dependency graph. Connectivity verification employs a breadth-first search algorithm, starting with entity nodes related to the business objective, with a maximum search depth of three layers. Access paths and hop counts are recorded during the search. Candidate nodes that do not meet the connectivity requirements are removed from the candidate set. Constraint verification checks whether candidate nodes violate identified structured constraints, including time window limits, resource capacity constraints, permission level requirements, and business rule compliance. Nodes violating any constraint are filtered out. Through this triple filtering mechanism of similarity screening, connectivity verification, and constraint checking, a final set of candidate nodes satisfying all conditions is obtained, typically ranging from five to twenty nodes.

[0084] In one optional implementation, the business intent features, the constraints, and the semantic dependencies are heterogeneously integrated. During the integration process, the hierarchical feature pyramid is used to establish a mapping between the business intent features and the constraints, as well as an enhanced mapping between the business intent features and the semantic dependencies, generating a business target semantic vector, including:

[0085] The business intent features are hierarchically analyzed based on a deep self-attention network. At the macro level, business domain features are captured, and at the micro level, the combined features of business actions and objects are extracted. A two-layer feature pyramid is constructed through a feature transfer model.

[0086] The constraints are semantically matched with the business intent features in the two-layer feature pyramid through a cross-layer progressive encoder. The limiting effect of the constraints on business actions and object combinations is identified at the micro level, and the limiting effect is transmitted to the macro level based on the deep self-attention network, thus establishing a cross-layer mapping relationship from constraints to business intent features.

[0087] The feature transfer model is used to perform semantic association modeling between the business entity association path in the semantic dependency relationship and the two-layer feature pyramid. Based on the cross-layer mapping relationship, the semantic dimension that needs to be enhanced is determined and the enhancement information is extracted to construct the cross-layer enhancement mapping of the semantic dependency relationship.

[0088] Based on the cross-layer mapping relationship and the cross-layer enhancement mapping, the two-layer feature pyramid is optimized and adjusted. The adjusted micro-level features and macro-level features are adaptively fused through the deep self-attention network to generate a business target semantic vector.

[0089] When performing hierarchical parsing of business intent features using a deep self-attention network, the deep self-attention network is constructed using a multi-head attention mechanism with sixteen attention heads, each with a 64-dimensional dimension, resulting in a total input dimension of 1024. The business intent feature input vector is transformed linearly to generate a query matrix, a key matrix, and a value matrix. The transformation weight matrix is ​​1024 x 1024 in dimension and initialized using a Xavier uniform distribution. Macro-level feature capture is achieved through global attention computation, performing fully connected attention operations on all word positions within the business intent feature. The attention weights are obtained by the dot product of the query vector and the key vector, and the dot product result is scaled by dividing by the square root of the dimension and then normalized using a softmax function. The macro-level output feature vector has a 1024-dimensional dimension. This feature vector captures the overall semantic information of the business domain, including high-level abstract features such as industry category, business scope, and strategic direction.

[0090] Micro-level feature extraction employs a local window attention mechanism, with a window size of eight word positions and a sliding step of four word positions to ensure overlapping areas between adjacent windows to maintain semantic continuity. Constrained attention calculations are performed on word positions within each local window, limiting the attention scope to the window itself to avoid interference from long-distance dependencies on local feature extraction. Business action recognition is performed through a predefined action dictionary containing 300 commonly used business action terms, covering action types such as creation, modification, deletion, query, approval, allocation, monitoring, and reporting. Business object extraction is based on a named entity recognition model trained on enterprise-specific domain corpora, capable of recognizing business object types such as personnel, departments, products, services, processes, systems, and resources. Combined action and object features are generated through a bilinear transformation layer with a transformation matrix dimension of 512 x 512 x 256, resulting in a 256-dimensional combined feature vector.

[0091] The feature propagation model constructs a two-layer feature pyramid using a combination of residual connections and skip connections. The bottom pyramid contains micro-level features of business actions and object combinations, with the number of feature vectors equal to the number of identified action-object combinations, and each feature vector has 256 dimensions. The top pyramid contains macro-level business domain features, with feature vectors having 1024 dimensions. Feature propagation is achieved through upsampling and downsampling operations. Upsampling extends micro-level features to the macro-level feature dimension through linear interpolation, while downsampling compresses macro-level features to the micro-level feature dimension through average pooling. Propagation weights are dynamically calculated using a learnable attention mechanism. Attention queries come from target-level features, and key values ​​come from source-level features. The attention output weights are used for weighted fusion of feature information from different levels.

[0092] When performing semantic matching between constraints and business intent features in a two-layer feature pyramid, the cross-layer progressive encoder employs a three-layer recurrent neural network structure, with each layer having a hidden state dimension of 512. Before being input into the encoder, constraints undergo structured preprocessing: time constraints are converted into timestamp sequences, resource constraints into numerical vectors, permission constraints into one-hot encoded vectors, and business rule constraints into symbolic sequences of logical expressions. The first layer of the encoder receives the original constraint representation and outputs a primary constraint feature vector. The second layer performs interactive encoding by combining micro-level business intent features, calculating the relevance score between the constraints and the business action object combination through an attention mechanism. The third layer passes the interactive encoding result to the macro-level for global semantic matching with business domain features.

[0093] Micro-level constraint identification is achieved through the dot product operation of the constraint conditions and the combined features of the business action objects. The dot product result represents the constraint strength of the constraint conditions on a specific combination of action objects. Combinations with a constraint strength exceeding 0.5 are marked as constrained combinations, combinations with a constraint strength between 0.2 and 0.5 are marked as partially constrained combinations, and combinations with a constraint strength below 0.2 are marked as unconstrained combinations. The feature vector of constrained combinations needs to be fused and updated with the corresponding constraint condition features. The fusion weight is controlled by a gating mechanism, and the gating value is calculated based on the importance and urgency of the constraint conditions. Partially constrained combinations use a weighted fusion method, with the fusion weight equal to the constraint strength value. Unconstrained combinations retain their original features unchanged.

[0094] Deep self-attention networks propagate constraints to the macro level using feature aggregation and redistribution mechanisms. Micro-level constraint information is aggregated into macro-constraint features through weighted average pooling, with weights calculated based on the importance scores of each micro-level combination. Macro-constraint features interact with macro-business domain features through attention, generating constraint-aware macro-feature representations. During this interaction, the query vector comes from business domain features, while the key and value vectors come from constraint features. The attention output is residually connected to the original business domain features, forming constraint-enhanced macro-features. Cross-layer mapping relationships are represented by a mapping matrix, where the matrix dimension is the number of micro-features multiplied by the macro-feature dimension, and the matrix element values ​​represent the weights of the micro-constraints on each dimension of the macro-features.

[0095] When the feature transfer model models the semantic association between business entity association paths in semantic dependencies and a two-layer feature pyramid, the association paths are first converted into path embedding vectors. Each business entity on the path is converted into a 256-dimensional vector through a pre-trained entity embedding layer, and the relationship type is converted into a 128-dimensional vector through a relationship embedding layer. Path embeddings are generated using a recurrent neural network encoder, which processes entity and relationship vectors sequentially according to the path traversal order, outputting a fixed-dimensional path representation vector. The path representation vector is then compared with micro-level features for similarity calculation, using cosine similarity as a metric. Paths with similarity scores higher than 0.6 are considered semantically associated with their corresponding micro-features.

[0096] The cross-layer mapping relationship determines the semantic dimensions that need enhancement through an importance scoring mechanism. The importance of each semantic dimension is calculated based on its activation frequency and activation strength during the constraint matching process. Dimensions with an activation frequency exceeding 20% ​​of the total frequency are marked as high-importance dimensions, and dimensions with an activation strength exceeding 0.7 are marked as strong-activation dimensions. Dimensions that meet both conditions are identified as semantic dimensions that need enhancement. Enhancement information is extracted from the associated semantic dependency paths, based on the attribute features of entities and the strength features of relationships along the path. Entity attribute features include entity type, entity level, entity activity, and entity connectivity. Relationship strength features include relationship type, relationship direction, relationship weight, and relationship credibility. Enhancement information is integrated into enhancement vectors through a feature fusion network. The fusion network adopts a multilayer perceptron structure with two hidden layers, with dimensions of 512 and 256 respectively. The activation function used is ReLU, and the output layer dimension is consistent with the number of semantic dimensions that need enhancement.

[0097] The cross-layer augmentation mapping for semantic dependencies employs an attention-guided feature propagation mechanism. At the micro-level, the augmentation vector is element-wise multiplied with the corresponding business action object's combined features; the result serves as the augmented micro-feature. This augmented micro-feature is then propagated to the macro-level through attention weights, calculated based on the magnitude of the augmentation vector and the importance of the micro-feature. The propagation process uses a weighted summation method, with weight normalization ensuring the stability of the macro-feature values ​​after propagation. The cross-layer augmentation mapping matrix records the contribution of the augmentation vector to features at different levels. The matrix dimension is the number of augmentation dimensions multiplied by the total number of feature dimensions. Matrix updates utilize gradient descent with a learning rate of 0.001.

[0098] The two-layer feature pyramid optimization is jointly guided by cross-layer mapping relationships and cross-layer enhancement mappings. The optimization objective function comprises three components: an original feature preservation term, a constraint consistency term, and a semantic enhancement term, with weights set to 0.5, 0.3, and 0.2, respectively. The original feature preservation term ensures that the similarity between the optimized features and the original features remains above 0.8; the constraint consistency term ensures that the optimized features meet all constraint requirements; and the semantic enhancement term ensures that the optimized features fully incorporate the enhanced information of semantic dependencies. The optimization algorithm employs an adaptive gradient descent method, with an initial learning rate set to 0.01, a momentum parameter set to 0.9, and a gradient clipping threshold set to 1.0.

[0099] Deep self-attention networks adaptively fuse adjusted micro-level features with macro-level features using a multi-scale attention mechanism. The fusion process consists of three stages: The first stage calculates the correlation matrix between micro- and macro-level features, obtained through the dot product of feature vectors. The second stage generates fusion weights based on the correlation matrix, using a softmax function to ensure the weights sum to one. The third stage performs a weighted average of the micro- and macro-level features according to the fusion weights to generate a fused feature vector. The adaptive mechanism is implemented through a gating network, which receives the concatenated vector of micro- and macro-level features as input and outputs a fusion gate value. This gate value controls the contribution ratio of micro- and macro-level features during the fusion process.

[0100] The business objective semantic vector generation employs feature compression and standardization. The fused feature vector is compressed to 800 dimensions through a linear transformation layer. The transformation matrix is ​​initialized using principal component analysis, retaining 95% of the original feature variation. The compressed feature vector undergoes L2 standardization to ensure a vector length of one, facilitating subsequent similarity calculations. The standardized vector serves as the final business objective semantic vector output, containing a complete business intent representation enhanced by constraints and semantic dependencies. The numerical values ​​of each dimension of the vector range from -1 to +1, and the vector sparsity is controlled below 30%, ensuring a balance between rich semantic information and computational efficiency.

[0101] In a specific data example, the input business objective is described as "assigning laptops to new employees and completing system permission configuration within three working days." Business intent feature extraction identifies the action words "assign" and "configure," and the object words "new employee," "laptop," and "system permission." The constraint is identified as the time constraint "within three working days," which is converted into a timestamp range. Semantic dependencies include the allocation relationship between employees and devices, and the configuration relationship between devices and permissions. After hierarchical parsing and cross-layer mapping, the generated business objective semantic vector has a value of 0.85 in the time sensitivity dimension, 0.92 in the resource allocation dimension, and 0.78 in the permission management dimension. The overall vector length is standardized to 1.0.

[0102] In one optional implementation, constructing a contextual state representation of the current process based on the executed process step sequence and the business intent features using the deep vector embedding technique includes:

[0103] The deep temporal coding network is used to model each process step in the executed process step sequence, generate a representation vector for each step, and construct a dynamic propagation graph based on the temporal dependency relationship between the representation vectors to capture the execution influence of the preceding steps on the subsequent steps and generate a step sequence vector containing temporal dependency information.

[0104] The business intent feature is semantically aligned with the step sequence vector, and dynamic weights are assigned to the execution steps in the executed process step sequence that are similar to the business intent feature to generate an intent-related step vector.

[0105] The dynamic propagation graph is used to perform semantic matching analysis on the intent-related step vector and the business intent feature, and the step sequence vector and the business intent feature are reconstructed based on the deep temporal coding network to generate an initial context representation;

[0106] In the enterprise knowledge graph, a semantic propagation path between process steps and target business entities is constructed based on the initial context representation. The context information is dynamically enhanced along the semantic propagation path to generate a context state representation of the current process.

[0107] When modeling each step in an executed process sequence, the deep temporal coding network employs a multi-layer coding architecture based on gated recurrent units. The coding network consists of three layers of gated recurrent units, each with a hidden state dimension of 512. The input gate weight matrix has a dimension of 512 x 768, the reset gate weight matrix has the same dimension, and the candidate state weight matrix has a dimension of 512 x 512. Before being input into the coding network, process steps need to be preprocessed into a standardized format, including six fields: step identifier, step type, execution timestamp, participating roles, processing object, and execution status. The step identifier uses one-hot encoding, and the encoding dimension is determined based on the total number of step types in the enterprise process library, typically set to 200 dimensions. The step type is converted into a 128-dimensional dense vector through a pre-trained embedding layer. The embedding layer dictionary size is 1000, covering common process step types. The execution timestamp is converted into a relative time code, and the time offset is calculated based on the process start time. The offset is encoded into a 64-dimensional time vector using sine and cosine functions.

[0108] The generation of representation vectors for each step employs a position-aware encoding method, concatenating step content features with position-encoded features to form a complete step representation. The step content features are mapped to a 400-dimensional vector through a fully connected layer, with the mapping layer weights initialized using a Gaussian distribution and a standard deviation of 0.02. The position encoding uses a learnable position embedding vector with 112 dimensions, which is concatenated with the step content features to form a 512-dimensional step representation vector. The representation vector undergoes layer normalization, with normalization parameters including a scaling factor and an offset factor. The scaling factor is initialized to one, and the offset factor to zero, adaptively updated during training. A dropout layer is placed after the normalization layer, with a dropout probability set to 0.1 to prevent model overfitting.

[0109] The dynamic propagation graph is constructed based on the temporal dependencies between representation vectors, and an attention mechanism is used to calculate the dependency strength between steps. Temporal dependencies are captured using a sliding window approach, with a window size of five steps and a sliding step size of one step, ensuring that each step establishes a dependency with its predecessors and successors. Dependency strength is calculated using scaled dot product attention, where the query vector, key vector, and value vector all originate from the step representation vectors. The number of attention heads is set to eight, and each head has a 64-dimensional dimension. Attention weights are temperature-scaled, with a temperature parameter set to 8.0, and the scaled weights are normalized using a softmax function. Step pairs with a dependency strength exceeding 0.2 are retained as directed edges in the graph, with edge weights equal to the corresponding dependency strength values.

[0110] The impact of preceding steps on subsequent steps is modeled using a graph neural network propagation mechanism, which employs a graph attention network architecture. The graph attention layer comprises two components: self-attention and neighbor attention. Self-attention updates the intrinsic representation of nodes, while neighbor attention aggregates information from neighboring nodes. Neighbor attention weights are calculated jointly based on node feature similarity and edge weights. Cosine similarity is used to measure similarity, and edge weights serve as adjustment factors for the attention weights. The graph neural network consists of two graph attention layers, each maintaining a 512-dimensional output feature dimension. Residual connections between layers prevent gradient vanishing. The activation function used is LeakyReLU, with the negative slope parameter set to 0.2.

[0111] The step sequence vector generation aggregates the node representations output by the graph neural network using global pooling. Global pooling employs an attention-weighted averaging method, with attention weights calculated based on the importance of nodes in the dynamic propagation graph. Node importance is determined by a weighted combination of degree centrality and proximity centrality, with a degree centrality weight set to 0.6 and a proximity centrality weight set to 0.4. The aggregated step sequence vector has a dimension of 512, containing temporal dependency information of the complete process execution history. The vector undergoes L2 normalization to ensure numerical stability and the effectiveness of subsequent calculations.

[0112] The semantic alignment of business intent features and step sequence vectors is achieved using a bidirectional attention mechanism. The semantic alignment process uses the business intent features as the query vector and the step sequence vectors as the key and value vectors, respectively, and calculates the semantic relevance between them. The business intent feature vector is 800-dimensional, and is mapped to 512 dimensions through a linear transformation layer to align with the step sequence vector. The transformation layer weight matrix has a dimension of 800 x 512 and uses the Xavier initialization method. The alignment calculation employs a multi-head attention mechanism with 16 heads, each with a dimension of 32 dimensions, ensuring the capture of semantic relationships at different granularities.

[0113] The identification of execution steps in the executed process sequence that are similar to the business intent features is achieved through similarity threshold filtering. Similarity calculation uses attention weight as the similarity measure. Steps with a weight value higher than 0.5 are marked as highly relevant steps, steps with a weight value between 0.2 and 0.5 are marked as moderately relevant steps, and steps with a weight value lower than 0.2 are marked as lowly relevant steps. Dynamic weight allocation is adjusted based on the similarity value and the criticality of the step in the process. Criticality is calculated through the step's execution frequency, execution time, and anomaly rate. The dynamic weight of highly relevant steps is set to an amplification factor of 1.5 times the original similarity; moderately relevant steps retain their original similarity unchanged; and lowly relevant steps have a reduction factor of 0.5 times the original similarity.

[0114] The generation of intent-related step vectors employs a weighted aggregation method, using dynamic weights as aggregation coefficients to sum the representation vectors of related steps. Before aggregation, the weights are normalized to ensure the sum of the weights is one. Weight normalization uses the softmax function with a temperature parameter set to 2.0 to control the sharpness of the weight distribution. The aggregated intent-related step vector has 512 dimensions, highlighting the historical execution information most relevant to the current business intent. The vector undergoes gradient clipping with a clipping threshold set to 1.0 to prevent gradient explosion.

[0115] The dynamic propagation graph performs semantic matching analysis between intent-related step vectors and business intent features using a graph matching network. The graph matching network receives two inputs: a subgraph constructed based on intent-related steps and a business intent feature vector. The subgraph is constructed by preserving the nodes corresponding to the intent-related steps and the edges connecting them; the subgraph typically contains five to fifteen nodes. The graph matching network employs a graph convolutional neural network architecture, containing three graph convolutional layers, each with a feature dimension of 256. Spectral domain convolution is used for the graph convolution operation, with a kernel size of 3 and ReLU activation. The matching analysis outputs a semantic matching score, ranging from 0 to 1. A score above 0.7 indicates a high match, a score between 0.4 and 0.7 indicates a medium match, and a score below 0.4 indicates a low match.

[0116] When reconstructing features from step sequence vectors and business intent features, deep temporal coding networks employ a variational autoencoder architecture. The encoder receives the concatenation of the step sequence vector and business intent features as input, with an input vector dimension of 1012. The encoder contains two hidden layers: the first hidden layer has a dimension of 768, and the second hidden layer has a dimension of 512, using tanh as the activation function. The encoder outputs a mean vector and a variance vector, both with a dimension of 256. Reparameterization is achieved by adding Gaussian noise, with the noise standard deviation calculated from the variance vector, and the noise mean being zero. The decoder receives the reparameterized latent vector as input and reconstructs the feature vector through two hidden layers with dimensions of 512 and 768, respectively. The output layer has a dimension of 1012, consistent with the input dimension.

[0117] The initial context representation generation further processes the reconstructed feature vector through a feature fusion network. The feature fusion network employs a multi-scale convolutional neural network architecture, containing three parallel one-dimensional convolutional branches with kernel sizes of three, five, and seven, respectively, and each branch has 128 output channels. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, with the batch normalization momentum parameter set to 0.9. The outputs of the three branches are concatenated to form a 384-dimensional feature vector. This concatenated feature vector is then compressed to 300 dimensions through a fully connected layer, serving as the initial context representation vector. The weights of the fully connected layer are initialized using an orthogonal method, and the bias term is initialized to zero.

[0118] In the enterprise knowledge graph, the semantic propagation path between process steps and target business entities, constructed based on the initial context representation, is implemented using the shortest path algorithm. The semantic propagation path search uses key entities identified in the initial context representation as starting nodes and the target business entity as the ending node, performing a breadth-first search within the knowledge graph. The search depth is limited to six hops; paths exceeding this depth are truncated. Path weights are calculated based on the semantic strength of edges and the importance of nodes. Semantic strength is represented by the confidence level of relationships between entities, ranging from zero to one. Node importance is calculated using the PageRank algorithm, with fifty iterations and a damping coefficient of 0.85. During the search process, the three paths with the highest weights are retained as candidate semantic propagation paths.

[0119] The dynamic enhancement of contextual information along the semantic propagation path employs a combination of path embedding and attention mechanisms. Each node on the path is transformed into a 256-dimensional vector through a pre-trained entity embedding layer, and each edge is transformed into a 128-dimensional vector through a relation embedding layer. Path embeddings are generated using a graph attention network, which receives node embeddings and edge embeddings as input and outputs a path-level representation vector. The attention mechanism calculates the correlation between the path representation vector and the initial context representation vector, and this correlation serves as the weight coefficient for dynamic enhancement. The enhancement weights are normalized using softmax to ensure that the sum of the weights for each path is one. The dynamic enhancement operation performs element-wise addition between the weighted path representation vector and the initial context representation vector to generate the enhanced context representation.

[0120] The current process's context state representation generation employs a multimodal fusion network to ultimately integrate the enhanced context information. This network comprises three input branches, processing step sequence information, business intent information, and knowledge graph enhancement information, respectively. Each branch utilizes an independent fully connected network for feature transformation, resulting in a unified 256-dimensional feature set. The outputs of the three branches are then fused using attention-weighted fusion. Attention weights are calculated based on the variance and mean of the branch features, with branches exhibiting higher variance receiving higher weights. The fused feature vector is then processed by a final fully connected layer and activation function, outputting a 300-dimensional context state representation vector for the current process.

[0121] In a specific data case, the executed process sequence contains five steps: application submission, preliminary review, secondary review, approval, and archiving. The business intent feature vector represents the semantic information of the equipment procurement application, with a dimension of 800. A deep temporal coding network generates a 512-dimensional representation vector for each step. The dynamic propagation graph identifies a strong dependency of 0.85 between application submission and preliminary review, and a dependency of 0.72 between preliminary review and secondary review. The semantic alignment process identifies the application submission and approval steps as having similarities of 0.89 and 0.76 with the business intent, respectively, and marks them as highly relevant steps. When generating the intent-related step vectors, the application submission step has a weight of 1.34, and the approval step has a weight of 1.14. The semantic matching analysis outputs a matching score of 0.81, indicating a high match. The initial context representation dimension after feature reconstruction is 300. The semantic propagation path constructed in the knowledge graph connects the procurement application entity and the equipment management entity, with a path length of four hops and a path weight of 0.73. The dynamically enhanced context state representation vector has a value of 0.91 in the device type dimension, 0.83 in the approval status dimension, and 0.67 in the time urgency dimension.

[0122] In one optional implementation, modeling each process step in the executed process step sequence using a deep temporal coding network to generate a representation vector for each step, and constructing a dynamic propagation graph based on the temporal dependencies between the representation vectors includes:

[0123] Extract step attribute features and step execution time features from each process step in the executed process step sequence. Jointly encode the step attribute features and step execution time features through deep tensor decomposition and embed time dimension position information to generate an initial representation vector for the step.

[0124] Based on the recursive attention structure, the execution sequence of the executed process steps is analyzed for temporal dependency. The execution influence of the preceding process steps on the subsequent process steps is mapped to the initial representation vector of the step through the deep tensor decomposition, generating a step representation vector containing temporal dependency information.

[0125] The recursive attention structure is used to calculate the temporal correlation strength between the representation vectors of each step in the executed process step sequence, and an edge weight mapping is constructed for process step pairs that exceed a preset correlation threshold.

[0126] A dynamic propagation graph is constructed by taking each process step in the sequence of executed process steps as a node, taking the step representation vector as the node feature, and taking the edge weight mapping to determine the directed edge connection relationship.

[0127] The feature extraction of each step in the executed process sequence is achieved using a multi-dimensional feature parser. The parser extracts six basic attributes from the process execution log: step identifier, step type, execution role, processing object, execution status, and exception marker. The step identifier is converted to a 64-bit fixed-length identifier using hash encoding. The step type is encoded using a predefined type dictionary containing 200 standard process step types, each mapped to a 16-dimensional hot vector. The execution role is converted to a 32-dimensional dense vector through a role embedding layer. The embedding layer dictionary size is 500, covering the enterprise's internal role system. The processing object uses an entity recognition model to extract key business objects, including five categories: documents, data, equipment, personnel, and resources, each represented by a 24-dimensional vector. The execution status uses status enumeration encoding, including five states: pending, processing, completed, paused, and abnormal, encoded as a 5-dimensional hot vector. The exception marker uses binary encoding, with zero for normal execution and one for abnormal execution.

[0128] The step execution time feature comprises two components: absolute time features and relative time features. The absolute time feature extracts six time components—year, month, day, hour, minute, and second—from the step execution timestamp. Each component is converted into a time vector through periodic encoding. The year component uses linear encoding, ranging from zero to ninety-nine, corresponding to the last one hundred years. The month and day components use sine and cosine encoding, with encoding dimensions of twelve and eight dimensions respectively, capturing the periodic patterns of months and days. The hour, minute, and second components use the same sine and cosine encoding method, each with a six-dimensional encoding dimension. The relative time feature calculates the time interval between steps and the execution time. The time interval is in minutes and is logarithmically scaled to prevent excessive values; the scaling formula is the time interval with a base of ten plus the logarithmic value. The execution time is processed using the same logarithmic scaling method to ensure the numerical stability of the time feature.

[0129] Deep tensor decomposition (DFD) jointly encodes step attribute features and step execution time features using a third-order tensor decomposition method. The attribute feature vector has a dimension of 141, and the time feature vector has a dimension of 59, together forming a 200-dimensional feature vector. The tensor decomposition employs the CP decomposition method, reshaping the 200-dimensional feature vector into a 10x10x2 third-order tensor structure. The rank of the CP decomposition is set to 50, resulting in three factor matrices with dimensions of 10x50, 10x50, and 2x50, respectively. The factor matrices are randomly initialized using a standard normal distribution, and then optimized using alternating least squares. The number of optimization iterations is set to 100, and the convergence threshold is set to 1 x 10^-6. Iteration stops when the reconstruction error change is less than the convergence threshold after ten consecutive iterations.

[0130] The temporal dimension location information embedding is implemented using a learnable location encoding method. The location encoding assigns a unique location vector to the position of each process step in the sequence, with the location vector dimension set to 64. The number of rows in the location encoding matrix equals the maximum sequence length, typically set to 200, and the number of columns equals the location vector dimension. The location encoding matrix is ​​randomly initialized using a normal distribution with a mean of zero and a standard deviation of 0.1, and adaptively learned during training using the backpropagation algorithm. The location information and the tensor decomposition result are added element-wise to generate a joint encoding vector containing location awareness. Before the addition operation, it is necessary to ensure that the dimensions of the two vectors are consistent. The tensor decomposition result is mapped to 64 dimensions through a linear transformation layer, and the transformation layer weights are initialized using the Xavier method.

[0131] The initial representation vector generation step further processes the joint encoded vector through a feature fusion network. The feature fusion network employs a multilayer perceptron architecture, containing two hidden layers: the first hidden layer has a dimension of 128, the second hidden layer has a dimension of 96, and the output layer has a dimension of 128. The activation function used is the GELU function, which exhibits smooth non-linearity near zero, facilitating gradient propagation. A batch normalization layer is placed after each hidden layer, with the momentum parameter set to 0.99 and the epsilon parameter set to 1 x 10⁻⁵. A dropout layer is placed after the batch normalization layer, with a dropout probability set to 0.2, randomly disabling some neurons during the training phase to prevent overfitting.

[0132] The recursive attention structure performs temporal dependency analysis on the sequence of executed process steps, employing a recursive variant of the self-attention mechanism. The recursive attention structure comprises multiple attention layers, each receiving the output of the previous layer as input, with a total of four layers. Each attention layer uses a multi-head attention mechanism with eight heads, each with a 16-dimensional dimension, ensuring a total dimension of 128, consistent with the initial representation vector dimension of the step. Attention calculation uses a scaled dot product, with a scaling factor equal to the square root of the head dimension, i.e., four. The query matrix, key matrix, and value matrix are generated from the input vector through linear transformations. The transformation weight matrix has a dimension of 128 by 128, and an orthogonal initialization method is used to ensure good matrix properties.

[0133] The influence of preceding steps on subsequent steps is mapped using a causal masking mechanism. This mechanism ensures that each step can only focus on its preceding steps and cannot access information from subsequent steps, thus preserving the causal relationship in the sequence. The mask matrix is ​​a lower triangular matrix, with elements on the diagonal and below set to zero, and elements in the upper triangular matrix set to negative infinity. Negative infinity values ​​result in near-zero weights in the softmax calculation, effectively masking information. The attention weights after masking are weighted and summed with the value vector to generate an intermediate representation that includes the influence of preceding steps. This intermediate representation is added to the input vector via a residual connection, which facilitates gradient propagation and feature preservation. Layer normalization is applied to the output of the residual connection, and the normalization parameters include learnable scaling and offset factors.

[0134] Deep tensor decomposition maps prior influences to the initial step representation vector using a tensor shrinking operation. Prior influence information is represented as a second-order tensor with a dimension equal to the sequence length multiplied by the feature dimension, i.e., a maximum of 200 multiplied by 128. The initial step representation vector is reshaped into a first-order tensor with a dimension of 128. Tensor shrinking is performed along the feature dimension, with each row of the second-order tensor performing a dot product with the first-order tensor to generate an influence weight vector of length equal to the sequence length. The influence weight vector is normalized using a softmax function to ensure that the weights sum to one. The normalized weights are then weighted and summed with the corresponding initial step representation vector to generate a step representation vector that incorporates prior influences.

[0135] The generation of step representation vectors containing temporal dependencies employs a gating fusion mechanism to control the degree of integration of influence information. The gating mechanism consists of two components: an update gate and a reset gate. The gating value is calculated using the sigmoid function, and the input is the concatenation of the initial step representation vector and the preceding influence vector. The update gate controls the proportion of new information integrated, while the reset gate controls the degree of retention of historical information. The gating value is element-wise multiplied with the corresponding vector. The output of the update gate is multiplied by the preceding influence vector, and the output of the reset gate is multiplied by the initial step representation vector. The two products are then weighted and summed to generate the final step representation vector, with weighting coefficients being the update gating value and one minus the update gating value, respectively. The weights of the gating network are randomly initialized, uniformly distributed, and range from -0.1 to +0.1.

[0136] The recursive attention structure calculates the temporal correlation strength between the representation vectors of each step in the executed process sequence using a pairwise similarity method. The temporal correlation strength is measured by the cosine similarity between the step representation vectors, calculated as the dot product of two vectors divided by the product of their lengths. Vector lengths are calculated using the L2 norm, which is the square root of the sum of the squares of the vector elements. Similarity values ​​range from -1 to +1, with positive values ​​indicating positive correlation and negative values ​​indicating negative correlation; the absolute value indicates the correlation strength. To improve computational efficiency, similarity calculation employs matrix operations, forming a feature matrix from all step representation vectors, and then calculating the similarity between all step pairs in a single matrix multiplication.

[0137] A preset association threshold is set to 0.4, determined through empirical analysis and cross-validation, effectively filtering out step pairs with significant temporal associations. Process step pairs with association strength exceeding the preset threshold are retained, while those below the threshold are filtered. The filtering process uses a mask matrix, comparing the association strength matrix with the threshold; values ​​greater than the threshold are set to 1, and values ​​less than or equal to the threshold are set to 0. Element-wise multiplication of the mask matrix and the association strength matrix retains association strength values ​​exceeding the threshold, setting other values ​​to zero. To ensure graph connectivity, for nodes with no outgoing edges, the edge with the highest association strength is retained to avoid creating isolated nodes.

[0138] The edge weight mapping is constructed using a nonlinear transformation of association strength. The nonlinear transformation function is the hyperbolic tangent function, which maps association strength values ​​to the range of zero to one. The transformation formula is: the hyperbolic tangent function takes the association strength minus a threshold as input, adds one, and divides by two. The transformed values ​​serve as edge weights; larger weight values ​​indicate stronger temporal associations between steps. The edge weight matrix is ​​an asymmetric matrix, reflecting the directionality of temporal relationships. The matrix elements correspond to the association relationships between step pairs. The row index represents the preceding step, the column index represents the subsequent step, and the matrix element values ​​represent the influence weight of the preceding step on the subsequent step. To control the complexity of the graph, each node retains a maximum of five outgoing edges, and the five edges with the highest weights are selected by weight sorting.

[0139] The dynamic propagation graph is constructed using each step in the executed process sequence as a node, with the number of nodes equal to the sequence length. Node features are represented by step representation vectors, with each node associated with a 128-dimensional feature vector. Directed edge connections are determined by edge weight mapping; steps with weights greater than zero are connected by directed edges. The direction of the directed edges points from the row index to the column index of the weight matrix, indicating the causal direction of the temporal relationship. Edge attributes include two fields: edge weight and edge type. The edge weight is the mapped association strength value, and the edge type is determined based on the combination of connected step types, including same-type connections, dissimilar-type connections, and cross-module connections.

[0140] The graph data structure uses an adjacency list, where each node maintains a list of outgoing edges. Each element in the list contains the target node index, edge weight, and edge type. The adjacency list supports efficient graph traversal and neighbor lookup operations, with a space complexity linear to the number of edges. The graph is stored in a compressed sparse row format, suitable for storing and computing sparse graphs. Cycle detection is performed during graph construction to avoid self-loops and multiple edges; this is implemented using a depth-first search algorithm. Graph updates are incremental; when a new step is added, only the relationship between the new step and existing steps needs to be calculated, avoiding the computational overhead of full graph reconstruction.

[0141] In the specific data case, the executed process sequence contains seven steps: requirement submission, requirement analysis, solution design, technical review, development implementation, testing and verification, and deployment. The feature vector extracted from the step attribute features of the requirement submission step includes a step type of "initiator," an execution role of "business user," and a processing object of "requirement document." The step execution time feature shows that the requirement submission was executed at 9:00 AM on Monday, with a relative time interval of zero. Depth tensor decomposition jointly encodes the attribute and time features into a 64-dimensional vector, and the position information embedding assigns a position code of zero to the requirement submission step. Recursive attention structure analysis shows that the impact weight of requirement submission on requirement analysis is 0.87, and its impact weight on solution design is 0.62. The temporal correlation strength calculation results show that the correlation strength between requirement submission and requirement analysis is 0.91, exceeding the preset correlation threshold of 0.4, and constructing a directed edge with a weight of 0.73. The dynamic propagation graph contains seven nodes and twelve directed edges, with an average degree of 3.43, a maximum out-degree of 5, and good connectivity, with all nodes reachable.

[0142] In one optional implementation, structural matching is performed between the context state representation and the predecessor and successor node patterns of each candidate node in the candidate node set within the enterprise knowledge graph to filter out a subset of target candidate nodes that satisfy the business intent, including:

[0143] Based on the deep graph projection architecture, the context state representation is transformed into the node relationship expression of the enterprise knowledge graph, and the expected patterns of the predecessor node and the expected patterns of the successor node in the current process are constructed through dynamic feature propagation.

[0144] For each candidate node in the candidate node set, the deep graph projection architecture is used to construct a node relationship context in the enterprise knowledge graph, and the predecessor node pattern and successor node pattern of the candidate node are generated.

[0145] A graph structure comparison method is used to perform similarity analysis between the predecessor node pattern and the expected predecessor node pattern of each candidate node, and similarity analysis between the successor node pattern and the expected successor node pattern. The structure matching features of each candidate node are generated through dynamic feature propagation.

[0146] The graph structure comparison method is used to perform deep matching between the unmet intent information in the context state representation and the business semantic labels of each candidate node to generate intent satisfaction features; the structure matching features and the intent satisfaction features are fused through the deep graph projection architecture to filter out a subset of target candidate nodes.

[0147] like Figure 2 As shown, the method includes:

[0148] The deep graph projection architecture transforms context state representations into node relationship representations of an enterprise knowledge graph using a multi-layer graph neural network projector. The projector comprises three graph convolutional layers, with input dimensions of 300, 256, and 128 dimensions for each layer, corresponding to a decreasing dimensionality transformation of the context state representations. The graph convolutional layers employ a graph attention network architecture with eight attention heads, the dimension of which is calculated by dividing the total dimension by the number of attention heads. The projection process uses the context state representations as initial node features, propagating these features through a virtual graph structure. The adjacency matrix of the virtual graph is constructed based on feature similarity, with a similarity threshold of 0.6. Node relationship representation generation utilizes relation-aware graph convolution operations, with convolution kernel parameters learned independently for different relationship types. The enterprise knowledge graph includes four main relationship types: hierarchical, dependency, collaboration, and constraint.

[0149] Dynamic feature propagation constructs the predecessor and successor node pattern expectations for the current process using a bidirectional graph traversal algorithm. The predecessor node pattern expectation is generated through reverse traversal, starting from the projected node relationship representation and performing a depth-first search along the incoming edges of the knowledge graph. The search depth is limited to three layers to ensure the capture of direct and indirect predecessor dependencies. During the traversal, the node type, relationship type, and path weight are recorded, with the weight calculated as the product of the edge confidence and node importance. The successor node pattern expectation is generated through forward traversal, with search parameters consistent with the predecessor pattern and a search depth also of three layers. Feature propagation employs a message-passing mechanism, where each node aggregates feature information from its neighbors. The aggregation function uses a weighted summation method, and the weights are dynamically calculated based on an attention mechanism.

[0150] The expected representation of node patterns uses a triple sequence format, with each triple containing three fields: subject node, relation type, and object node. Subject and object nodes are represented by node embedding vectors (128-dimensional), while relation types are represented by relation embedding vectors (64-dimensional). The length of the triple sequence is dynamically determined based on the traversal results, typically containing ten to thirty triples. The sequence is converted into a fixed-dimensional pattern vector by a sequence encoder using a Long Short-Term Memory (LSTM) network architecture. The hidden state dimension is 256, and the output vector dimension is 200. The pattern vector is then used to generate the final expected pattern representation through an attention pooling mechanism, with pooling weights calculated based on the importance scores of the triples.

[0151] The local subgraph extraction method is used to construct the node relationship context of each candidate node in the enterprise knowledge graph within the candidate node set. Local subgraph extraction expands outwards from the candidate node by two hops of its neighbors, forming a relationship context subgraph for the candidate node. The subgraph size is controlled within the range of fifty to one hundred nodes. If the size exceeds this limit, nodes are pruned based on importance ranking. Importance is calculated using a weighted combination of degree centrality and PageRank value, with a weight ratio of 7:3. Edges in the subgraph retain their original directionality and weight information, with edge weights normalized to the range of zero to one. The subgraph structure is standardized using a graph isomorphism detection algorithm to eliminate the influence of differences in node labels on structural comparisons.

[0152] The predecessor and successor node patterns of candidate nodes are generated using the same traversal strategy as the desired pattern. The deep graph projection architecture takes the local subgraph of the candidate node as input and generates node relationship representations through the same three-layer graph convolutional network. The predecessor node pattern is generated by traversing the in-neighbors of the candidate node in reverse, and the successor node pattern is generated by traversing the out-neighbors of the candidate node in forward. The traversal depth remains at three layers to ensure consistency with the desired pattern structure. The triple sequence length of the candidate node pattern is typically between five and twenty, and the sequence encoding uses the same long short-term memory network architecture, with an output pattern vector dimension of two hundred.

[0153] The graph structure comparison method performs similarity analysis on the predecessor node patterns of candidate nodes and their expected predecessor node patterns using graph kernel functions. The graph kernel function assesses structural similarity based on the quantity and quality of subgraph isomorphisms. Isomorphism detection is implemented using the VF2 algorithm, and the algorithm complexity is optimized through heuristic pruning strategies. Similarity calculation includes two dimensions: structural similarity and semantic similarity. Structural similarity is calculated by comparing the topological features of the graph, including node degree distribution, clustering coefficient, and path length distribution. Semantic similarity is calculated using the cosine similarity of the semantic vectors of nodes and edges, with a similarity threshold set to 0.7; matches below this threshold are filtered out. The overall similarity is calculated by a weighted average of structural and semantic similarity, with a weight ratio of 6:4.

[0154] The similarity analysis of successor node patterns employs the same graph structure comparison method, maintaining consistent computational parameters and thresholds. Dynamic feature propagation generates structural matching features for each candidate node through a feature fusion network. This network receives predecessor and successor pattern similarities as input and outputs a structural matching score. The fusion network uses a multilayer perceptron architecture with two hidden layers of 128 and 64 dimensions, respectively, using ReLU activation. The output layer employs a sigmoid activation function to ensure the matching score is within the range of zero to one. The structural matching features comprise three components: matching score, confidence, and stability. Confidence is calculated using variance during the matching process, and stability is evaluated through consistency assessment across multiple sampling iterations.

[0155] The identification of unmet intent information in the context state representation is achieved using an intent missing detection algorithm. This algorithm performs a difference analysis between the context state representation and the complete business intent representation, and calculates the missing intent vector through vector subtraction. The non-zero elements of the missing intent vector correspond to unmet intent dimensions, and the absolute value of each element represents the degree of missingness. The semantic interpretation of intent dimensions is provided through a predefined intent ontology library, which includes four main categories: time constraints, resource requirements, quality standards, and compliance requirements. Each category is further subdivided into several sub-dimensions. The priority of missing intent information is determined by ranking its importance, which is calculated based on the product of business impact and urgency.

[0156] The deep matching of business semantic labels and unmet intent information for each candidate node employs a semantic vector alignment method. Business semantic labels are pre-converted into vector representations using a domain-specific word embedding model with a 300-dimensional vector dimension. The embedding model is fine-tuned and trained on an internal enterprise document corpus. Deep matching calculates the semantic similarity between the label vector and the missing intent vector, using cosine similarity as a metric. The matching process considers the semantic hierarchy, expanding the label vocabulary with synonyms and hyponyms through the WordNet semantic network, with the expansion radius set to two layers of semantic relationships. A graph structure comparison method is used in deep matching to evaluate the structured relationships between labels; the label relationship graph is constructed through co-occurrence statistics and semantic associations.

[0157] Intent satisfaction features are generated through a satisfaction scoring function that comprehensively considers three factors: matching degree, coverage, and consistency. Matching degree is calculated using a weighted average of semantic similarity, with weights assigned based on the importance of the intent dimensions. Coverage is calculated by dividing the number of satisfied dimensions by the total number of dimensions. Consistency assesses the degree of semantic conflict between tags; conflict detection is achieved using an antonym dictionary and a contradiction rule base. The final intent satisfaction feature comprises three values: satisfaction score, coverage, and consistency index, each normalized to the range of zero to one.

[0158] The fusion of structural matching features and intent satisfaction features employs the feature fusion module within the deep graph projection architecture. This fusion module comprises two components: an attention mechanism and a gating mechanism. The attention mechanism calculates the importance weights of the two feature classes, while the gating mechanism controls the feature fusion ratio. Attention weights are calculated using multi-head attention, with four heads, each focusing on a different combination of feature dimensions. The gating value is calculated using the sigmoid function, with the input being the concatenated vector of the two feature classes and the output being a fusion coefficient between zero and one. The fusion operation uses a weighted summation method, with the weights determined by the product of the attention weights and the gating value. The fused feature vector has a dimension of 128, and numerical stability is ensured through layer normalization and dropout processing.

[0159] The selection of the target candidate node subset is based on a comprehensive score ranking using fused features. The comprehensive score is calculated through a linear combination and comprises three components: structural matching score, intent satisfaction score, and candidate node quality score, with weighting coefficients set to 0.4, 0.4, and 0.2, respectively. The candidate node quality is assessed based on its authority, activity, and credibility within the knowledge graph. Authority is calculated using in-degree, activity using update frequency, and credibility using historical verification accuracy. A selection threshold of 0.6 is set; candidate nodes with a comprehensive score exceeding this threshold are selected into the target candidate node subset. The subset size is controlled within the range of three to ten nodes; if the size exceeds this limit, the ten nodes with the highest scores are retained.

[0160] In a specific data case, the context state is represented as a 300-dimensional vector, indicating the current state of the procurement approval process, including information such as application type, amount range, urgency, and approval level. The candidate node set contains fifteen candidate process nodes, covering four categories: financial approval, technical evaluation, compliance check, and supplier evaluation. The deep graph projection architecture projects the context state representation into procurement topic nodes in the knowledge graph. The expected predecessor node pattern includes three preliminary steps: requirement confirmation, budget verification, and supplier qualification verification. The financial approval node in the candidate node has a predecessor node pattern in the knowledge graph that includes amount verification, budget check, and authorization verification, with a structural similarity of 0.82 to the expected pattern. Intent satisfaction analysis shows that this candidate node can satisfy 90% of the approval intent requirements, with a coverage of 0.90 and a consistency index of 0.95. The fusion feature calculation yields a comprehensive score of 0.87 for this candidate node, exceeding the screening threshold of 0.6, and it is selected into the target candidate node subset. Finally, six target candidate nodes are selected, with an average comprehensive score of 0.74, and the subset covers the complete approval process path.

[0161] In one optional implementation, for each candidate node in the candidate node set, a node relationship context is constructed in the enterprise knowledge graph using the deep graph projection architecture, generating the predecessor node pattern and successor node pattern of the candidate node, including:

[0162] For each candidate node in the candidate node set, trace the incoming edge relationship of the candidate node in the enterprise knowledge graph, extract the set of predecessor nodes with incoming edge relationships to the candidate node and their node types, and obtain the relationship type of the incoming edge relationship to construct a topological representation of the predecessor structure information.

[0163] The predecessor structure information is semantically mapped using the deep graph projection architecture. The node feature vectors and type information of the predecessor node set are uniformly projected into the semantic space, and the predecessor node pattern of the candidate node is generated by edge relationship aggregation.

[0164] The outgoing edge relationships of the candidate nodes are analyzed using the same tracing method. The set of successor nodes and their node types are extracted. Combined with the relationship types of the outgoing edge relationships, a topological representation of the successor structure information corresponding to the predecessor structure information is constructed.

[0165] The node feature vectors and type information of the successor structure information are projected onto the same semantic space as the predecessor node pattern using the deep graph projection architecture, and the successor node pattern of the candidate node is generated by edge relationship aggregation.

[0166] The inbound edge tracing of each candidate node in the candidate node set within the enterprise knowledge graph employs a depth-first search algorithm, with a search depth limited to two levels to ensure the capture of both direct and second-degree predecessor nodes. Inbound edge tracing starts from a candidate node and proceeds backward along the inbound edge directions of the graph, recording the node identifier, node type, and edge relationship type during the traversal. Node identifiers are globally unique 16-character strings. Node types are selected from a predefined type ontology, including four basic types: entity nodes, concept nodes, attribute nodes, and relationship nodes. Edge relationship types cover five main types: dependency, containment, inheritance, association, and constraint, each with specific semantic meaning and propagation weight.

[0167] The extraction of the predecessor node set employs a hybrid search method combining breadth-first and depth-first search strategies. The breadth-first part ensures the integrity of predecessor nodes at the same level, while the depth-first part captures multi-level dependency chains. The size of the predecessor node set is controlled within the range of ten to fifty nodes; if the upper limit is exceeded, pruning is performed based on node importance scoring. Node importance is calculated based on a weighted combination of in-degree, out-degree, and betweenness centrality, with weight coefficients set to 0.4, 0.3, and 0.3, respectively. Node type information is converted into a 128-dimensional vector representation through a type embedding layer. The embedding layer dictionary size is 1,000, covering all node types in the enterprise knowledge graph. The type vector is initialized randomly using a uniform distribution, with values ​​ranging from -0.1 to +0.1.

[0168] The relationship type acquisition for incoming edges utilizes a relationship extractor module. The extractor queries the graph database for all incoming edges from candidate nodes, returning a quadruple of information: source node, target node, relationship type, and relationship weight. Relationship types are standardized using a relationship ontology library containing 500 relationship type definitions, each with a semantic description, constraints, and propagation rules. Relationship weights represent the strength and credibility of the relationship, with values ​​ranging from zero to one. Weight calculation is based on the relationship's historical verification accuracy and usage frequency. Relationship types are converted into 64-dimensional embedding vectors. The embedding matrix is ​​obtained through pre-training on a knowledge graph using the TransE model, which undergoes 500 iterations on the enterprise knowledge graph.

[0169] The topological representation of the predecessor structure information is constructed using a combination of adjacency matrix and feature matrix. The adjacency matrix records the connection relationships within the predecessor node set, with a dimension equal to the square of the number of predecessor nodes. Matrix elements are binary values ​​of zero or one, where one indicates a connection and zero indicates no connection. The feature matrix contains node feature vectors and relation feature vectors; the node feature vectors are 256-dimensional, and the relation feature vectors are 64-dimensional. The topological representation generates fixed-dimensional structure vectors through a graph neural network encoder. The encoder employs a graph attention network architecture with two graph attention layers, each with eight attention heads and a hidden dimension of 128.

[0170] The deep graph projection architecture performs semantic mapping of predecessor structural information using a multi-scale graph convolutional network. The semantic mapping process projects the topological structure and feature information of the predecessor node set onto a 300-dimensional semantic space. The projection network contains three parallel graph convolutional branches with kernel sizes of one, three, and five, corresponding to different scale structural patterns. Each branch outputs a 100-dimensional vector, and the outputs of the three branches are concatenated to form a 300-dimensional feature vector. The graph convolution operation uses spectral domain convolution, the Laplacian matrix is ​​calculated from the graph's degree and adjacency matrices, and eigenvalue decomposition is implemented using the Lanczos algorithm with fifty iterations.

[0171] The unified projection of node feature vectors and type information from the predecessor node set employs an attention-weighted mechanism. The node feature vectors are derived from a pre-trained node embedding model with an embedding dimension of 256. The model is trained on an enterprise knowledge graph using the Node2Vec algorithm, with a walk length of 80, a walk count of 10, and a window size of 10. Type information projection is achieved through a type-aware transformation matrix, learned separately for different node types, with a matrix dimension of 128 x 300 to ensure that the dimensions of the projected vectors are consistent with the semantic space. The projection process utilizes residual connections and layer normalization techniques. Residual connections prevent information loss, and layer normalization ensures numerical stability.

[0172] Edge relation aggregation to generate predecessor node patterns employs relation-aware pooling. Relation aggregation groups edge relations of the same type, and edges within each group are represented using weighted average pooling to generate group-level representations. Weights are calculated based on the relation weights of the edges and the importance of the connecting nodes. Representation vectors from different groups are aggregated across groups using an attention mechanism, with attention weights calculated based on the semantic similarity of the relation types. Semantic similarity is measured using the cosine similarity of the relation embedding vectors, and the similarity matrix is ​​pre-calculated and cached to improve operational efficiency. The aggregated predecessor node pattern vector has a dimension of 300, and the vector undergoes L2 normalization to ensure the stability of the pattern vector's numerical range.

[0173] The outgoing edge relationship analysis of candidate nodes adopts the same tracing method as the incoming edge relationship analysis, with the search direction proceeding forward from the candidate node along the outgoing edges. The search depth for outgoing edge relationship tracing is also limited to two levels, and the same hybrid search strategy is used to ensure the completeness and representativeness of the successor node set. The size control parameters of the successor node set are consistent with those of the predecessor node set, and the pruning strategy is based on the same node importance scoring method. The relationship type acquisition for outgoing edge relationships uses the same relationship extractor module, and the extraction results include the directional information of the edges, distinguishing the semantic differences between incoming and outgoing edges.

[0174] The extraction process of successor node sets and their node types follows the same workflow as that of predecessor nodes. Node type information is processed through the same type embedding layer to ensure that the type representations of predecessor and successor nodes are within the same semantic space. Outgoing edge relation types are standardized using the same relation ontology library, and the relation weight calculation method remains consistent. The topological representation of successor structural information is constructed using the same adjacency matrix and feature matrix representation methods. The graph neural network encoder parameters are consistent with those used in predecessor structural processing to ensure the comparability of structural representations.

[0175] The topological representation of successor structure information uses the same representation format and encoding method as that of predecessor structure information. The adjacency matrix records the internal connectivity relationships of the successor node set, and the feature matrix contains the feature vectors of successor nodes and the feature vectors of outgoing edges. Topological encoding adopts the same graph attention network architecture, and the network parameters are shared in predecessor and successor structure processing to ensure semantic consistency of structure vectors. The encoder training adopts a joint optimization strategy, and the loss function includes two components: reconstruction error and contrastive learning error. The reconstruction error weight is set to 0.7, and the contrastive learning error weight is set to 0.3.

[0176] The deep graph projection architecture projects successor structure information into the same semantic space as the predecessor node pattern using a parameter-shared projection network. The weight parameters of the projection network remain consistent throughout the generation of predecessor and successor patterns, ensuring that both patterns are represented in the same semantic coordinate system. The node feature vectors of successor structure information and the projection of type information employ the same attention weighting mechanism and type-aware transformation matrix. The transformation matrix parameters are optimized through joint training, with training data containing paired samples of predecessor and successor structures. The optimization objective is to maximize the semantic similarity of the paired samples.

[0177] The generation of successor node patterns through edge relation aggregation employs the same relation-aware pooling operation as the predecessor node patterns. The grouping strategy, weight calculation method, and attention mechanism parameters for relation aggregation remain consistent, ensuring symmetry in the generation process of predecessor and successor patterns. The dimension and normalization method of the successor node pattern vector are consistent with those of the predecessor node pattern, facilitating subsequent pattern comparison and matching analysis. The quality assessment of pattern vectors uses two metrics: internal consistency and external discriminability. Internal consistency is calculated using the variance of the elements within the pattern vector, while external discriminability is calculated using the cosine distance between different pattern vectors.

[0178] In a specific data case, the candidate node is the financial audit node in the procurement approval process. This node has three inbound edges and four outbound edges in the enterprise knowledge graph. Inbound edge tracing identifies three predecessor nodes: a budget verification node, an amount verification node, and an applicant identity confirmation node. The relationship types are verification, dependency, and prerequisite, respectively. The topological representation of the predecessor node set shows an association between the budget verification node and the amount verification node, with node importance scores of 0.85, 0.72, and 0.68, respectively. The deep graph projection architecture maps the predecessor structure information into a 300-dimensional semantic vector, with a value of 0.91 in the budget dimension and 0.83 in the compliance dimension. The predecessor node pattern generated by edge relationship aggregation has a weight of 0.76 in verification-type relationships and 0.64 in dependency-type relationships. Outbound edge analysis identifies four successor nodes: a fund allocation node, an approval record node, a notification sending node, and a process jump node. The successor node pattern has a weight of 0.82 in execution-type relationships and 0.59 in record-type relationships. The semantic similarity between the predecessor and successor node patterns is 0.78, indicating that the candidate node has good structural consistency.

[0179] A second aspect of this invention provides a process node intelligent optimization and scheduling system for enterprise business, comprising:

[0180] The first unit is used to obtain the business objective description and the sequence of executed process steps of the current process to be built, and to retrieve the set of historical process nodes and their associated business semantic tags from the enterprise knowledge graph based on deep vector embedding technology;

[0181] The second unit is used to apply a bidirectional long short-term memory network to perform semantic parsing on the business objective description, generate business intent features and constraints, and perform vectorized similarity calculation on the business intent features and the business semantic tags to obtain a set of target-oriented candidate nodes.

[0182] The third unit is used to construct a context state representation of the current process based on the sequence of executed process steps and the business intent features, using the deep vector embedding technology, and to perform structural matching between the context state representation and the predecessor node pattern and successor node pattern of each candidate node in the candidate node set in the enterprise knowledge graph, thereby filtering out a subset of target candidate nodes that meet the business intent.

[0183] The fourth unit is used to construct a dynamic heterogeneous graph structure based on the target candidate node subset, uniformly map the temporal correlation between nodes and the business intent features to the feature space, generate a recommendation sequence with adaptive weights, and continuously optimize the node migration pattern in the enterprise knowledge graph through the bidirectional long short-term memory network.

[0184] A third aspect of the present invention provides an electronic device, comprising:

[0185] processor;

[0186] Memory used to store processor-executable instructions;

[0187] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0188] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0189] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent optimization and scheduling of process nodes for enterprise business, characterized in that, include: Obtain the business objective description and the sequence of executed process steps of the current process to be built, and retrieve the set of historical process nodes and their associated business semantic tags from the enterprise knowledge graph based on deep vector embedding technology; A bidirectional long short-term memory network is applied to the business objective description for semantic parsing to generate business intent features and constraints. The similarity between the business intent features and the business semantic tags is then calculated using vectorization to obtain a set of target-oriented candidate nodes, including: Based on a bidirectional long short-term memory network, bidirectional semantic modeling is performed on the description of the business objective. Through a skip-layer connection mechanism, semantic accumulation features are captured in the forward encoding path and the reverse encoding path respectively, and features are reorganized at each word position to generate a word-level semantic representation sequence. Based on the skip-layer connection mechanism, the contribution of each word in the word-level semantic representation sequence to the expression of business intent is calculated. The semantic representations of words with contribution values ​​higher than the preset contribution threshold are used to construct a hierarchical feature pyramid to form business intent features. At the same time, the constraint type ontology in the enterprise knowledge graph is combined to perform structured representation and generate constraint conditions. Identify business entities from the business objective description, retrieve the association paths between the business entities in the enterprise knowledge graph, and filter out semantically coherent association paths as semantic dependencies based on the bidirectional long short-term memory network. The business intent features, constraints, and semantic dependencies are heterogeneously integrated. During the integration process, the hierarchical feature pyramid is used to establish a mapping between business intent features and constraints, as well as an enhanced mapping between business intent features and semantic dependencies, to generate a business target semantic vector. Similarity is calculated between the semantic vector of the business objective and the business semantic tag vector of each node in the historical process node set. The connectivity of nodes is verified according to the semantic dependency relationship. Nodes that do not meet the connectivity or violate the constraints are filtered to obtain a candidate node set. Based on the sequence of executed process steps and the business intent features, the context state representation of the current process is constructed using the deep vector embedding technology. The context state representation is then structurally matched with the predecessor node pattern and successor node pattern of each candidate node in the candidate node set in the enterprise knowledge graph to filter out a subset of target candidate nodes that meet the business intent. A dynamic heterogeneous graph structure is constructed based on the target candidate node subset, and the temporal correlation between nodes and the business intent features are uniformly mapped to the feature space to generate a recommendation sequence with adaptive weights. The node migration pattern in the enterprise knowledge graph is continuously optimized through the bidirectional long short-term memory network.

2. The method according to claim 1, characterized in that, The business intent features, constraints, and semantic dependencies are heterogeneously integrated. During the integration process, a mapping between business intent features and constraints, as well as an enhanced mapping between business intent features and semantic dependencies, are established through the hierarchical feature pyramid to generate a business target semantic vector, including: The business intent features are hierarchically analyzed based on a deep self-attention network. At the macro level, business domain features are captured, and at the micro level, the combined features of business actions and objects are extracted. A two-layer feature pyramid is constructed through a feature transfer model. The constraints are semantically matched with the business intent features in the two-layer feature pyramid through a cross-layer progressive encoder. The limiting effect of the constraints on business actions and object combinations is identified at the micro level, and the limiting effect is transmitted to the macro level based on the deep self-attention network, thus establishing a cross-layer mapping relationship from constraints to business intent features. The feature transfer model is used to perform semantic association modeling between the business entity association path in the semantic dependency relationship and the two-layer feature pyramid. Based on the cross-layer mapping relationship, the semantic dimension that needs to be enhanced is determined and the enhancement information is extracted to construct the cross-layer enhancement mapping of the semantic dependency relationship. Based on the cross-layer mapping relationship and the cross-layer enhancement mapping, the two-layer feature pyramid is optimized and adjusted. The adjusted micro-level features and macro-level features are adaptively fused through the deep self-attention network to generate a business target semantic vector.

3. The method according to claim 1, characterized in that, Based on the sequence of executed process steps and the business intent features, the context state representation of the current process is constructed using the deep vector embedding technique, including: The deep temporal coding network is used to model each process step in the executed process step sequence, generate a representation vector for each step, and construct a dynamic propagation graph based on the temporal dependency relationship between the representation vectors to capture the execution influence of the preceding steps on the subsequent steps and generate a step sequence vector containing temporal dependency information. The business intent feature is semantically aligned with the step sequence vector, and dynamic weights are assigned to the execution steps in the executed process step sequence that are similar to the business intent feature to generate an intent-related step vector. The dynamic propagation graph is used to perform semantic matching analysis on the intent-related step vector and the business intent feature, and the step sequence vector and the business intent feature are reconstructed based on the deep temporal coding network to generate an initial context representation; In the enterprise knowledge graph, a semantic propagation path between process steps and target business entities is constructed based on the initial context representation. The context information is dynamically enhanced along the semantic propagation path to generate a context state representation of the current process.

4. The method according to claim 3, characterized in that, The process steps in the executed process step sequence are modeled using a deep temporal coding network to generate representation vectors for each step, and a dynamic propagation graph is constructed based on the temporal dependencies between these representation vectors, including: Extract step attribute features and step execution time features from each process step in the executed process step sequence. Jointly encode the step attribute features and step execution time features through deep tensor decomposition and embed time dimension position information to generate an initial representation vector for the step. Based on the recursive attention structure, the execution sequence of the executed process steps is analyzed for temporal dependency. The execution influence of the preceding process steps on the subsequent process steps is mapped to the initial representation vector of the step through the deep tensor decomposition, generating a step representation vector containing temporal dependency information. The recursive attention structure is used to calculate the temporal correlation strength between the representation vectors of each step in the executed process step sequence, and an edge weight mapping is constructed for process step pairs that exceed a preset correlation threshold. A dynamic propagation graph is constructed by taking each process step in the sequence of executed process steps as a node, taking the step representation vector as the node feature, and taking the edge weight mapping to determine the directed edge connection relationship.

5. The method according to claim 1, characterized in that, The context state representation is structurally matched with the predecessor and successor node patterns of each candidate node in the candidate node set within the enterprise knowledge graph to filter out a subset of target candidate nodes that satisfy the business intent, including: Based on the deep graph projection architecture, the context state representation is transformed into the node relationship expression of the enterprise knowledge graph, and the expected patterns of the predecessor node and the expected patterns of the successor node in the current process are constructed through dynamic feature propagation. For each candidate node in the candidate node set, the deep graph projection architecture is used to construct a node relationship context in the enterprise knowledge graph, and the predecessor node pattern and successor node pattern of the candidate node are generated. A graph structure comparison method is used to perform similarity analysis between the predecessor node pattern and the expected predecessor node pattern of each candidate node, and similarity analysis between the successor node pattern and the expected successor node pattern. The structure matching features of each candidate node are generated through dynamic feature propagation. The graph structure comparison method is used to perform deep matching between the unmet intent information in the context state representation and the business semantic labels of each candidate node to generate intent satisfaction features; the structure matching features and the intent satisfaction features are fused through the deep graph projection architecture to filter out a subset of target candidate nodes.

6. The method according to claim 5, characterized in that, For each candidate node in the candidate node set, the node relationship context is constructed in the enterprise knowledge graph using the deep graph projection architecture, generating the predecessor node pattern and successor node pattern of the candidate node, including: For each candidate node in the candidate node set, trace the incoming edge relationship of the candidate node in the enterprise knowledge graph, extract the set of predecessor nodes with incoming edge relationships to the candidate node and their node types, and obtain the relationship type of the incoming edge relationship to construct a topological representation of the predecessor structure information. The predecessor structure information is semantically mapped using the deep graph projection architecture. The node feature vectors and type information of the predecessor node set are uniformly projected into the semantic space, and the predecessor node pattern of the candidate node is generated by edge relationship aggregation. The outgoing edge relationships of the candidate nodes are analyzed using the same tracing method. The set of successor nodes and their node types are extracted. Combined with the relationship types of the outgoing edge relationships, a topological representation of the successor structure information corresponding to the predecessor structure information is constructed. The node feature vectors and type information of the successor structure information are projected onto the same semantic space as the predecessor node pattern using the deep graph projection architecture, and the successor node pattern of the candidate node is generated by edge relationship aggregation.

7. A process node intelligent optimization and scheduling system for enterprise business, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to obtain the business objective description and the sequence of executed process steps of the current process to be built, and to retrieve the set of historical process nodes and their associated business semantic tags from the enterprise knowledge graph based on deep vector embedding technology; The second unit is used to apply a bidirectional long short-term memory network to perform semantic parsing on the business objective description, generate business intent features and constraints, and perform vectorized similarity calculation on the business intent features and the business semantic tags to obtain a set of target-oriented candidate nodes. The third unit is used to construct a context state representation of the current process based on the sequence of executed process steps and the business intent features, using the deep vector embedding technology, and to perform structural matching between the context state representation and the predecessor node pattern and successor node pattern of each candidate node in the candidate node set in the enterprise knowledge graph, thereby filtering out a subset of target candidate nodes that meet the business intent. The fourth unit is used to construct a dynamic heterogeneous graph structure based on the target candidate node subset, uniformly map the temporal correlation between nodes and the business intent features to the feature space, generate a recommendation sequence with adaptive weights, and continuously optimize the node migration pattern in the enterprise knowledge graph through the bidirectional long short-term memory network.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images