Dynamic cooperative task scheduling method and system based on lightweight intelligence and storage medium

Abstract

The application relates to the technical field of computers, and provides a dynamic cooperative task scheduling method and system based on lightweight intelligence and a storage medium, wherein the method mines a skill co-occurrence mode from historical scheduling task records, constructs a skill co-occurrence graph, decomposes independent embedded vectors and associated implicit constraints, further generates task embedded representation fusing topological context information, and finally generates a cooperative task scheduling instruction sequence through adaptability calculation and scheduling strategy arrangement. The method can effectively utilize cooperative mode knowledge in historical data, improve the perception ability of scheduling decision to skill dependency, and thus improves the rationality of task scheduling and resource utilization efficiency.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a lightweight intelligent dynamic collaborative task scheduling method, system, and storage medium. Background Technology

[0002] In task scheduling of complex systems, scheduling decisions directly affect system operating efficiency. Existing scheduling methods mainly include rule-based methods and optimization model-based methods: rule-based methods rely on manually preset scheduling rules, which are simple to implement but their scheduling efficiency drops significantly when task requirements change dynamically; optimization model-based methods model the scheduling problem as a mathematical programming problem, and as the task size and the amount of resources increase, the solution space expands exponentially, the solution time increases sharply, and it is difficult to meet the real-time scheduling requirements.

[0003] In recent years, another set of existing technologies has improved scheduling efficiency by learning scheduling strategies from historical data. However, when dealing with the collaborative relationships between skills, the co-occurrence patterns of skills in historical data are not fully utilized. The skill requirements of the task and the collaborative structure between skills are handled separately, resulting in a weak ability of scheduling decisions to perceive skill dependencies and difficulty in ensuring the rationality of resource allocation. Summary of the Invention

[0004] This invention provides a lightweight, intelligent, dynamic collaborative task scheduling method, system, and storage medium that effectively utilizes collaborative pattern knowledge from historical data to improve the ability of scheduling decisions to perceive skill dependencies, thereby enhancing the rationality of task scheduling and the efficiency of resource utilization.

[0005] This invention provides a lightweight, intelligent dynamic collaborative task scheduling method, applied to a dynamic collaborative task scheduling system. The method includes: Obtain historical task records and a set of tasks to be scheduled. The historical task records contain historical task entries, and each historical task entry contains a combination of skill identifiers used by the historical task. The set of tasks to be scheduled contains the tasks to be scheduled and their corresponding sets of required skill identifiers. The skill identifier combinations in the historical scheduling task records are subjected to co-occurrence association analysis to generate a skill co-occurrence graph, which includes skill atomic nodes and associated edges connecting the skill atomic nodes; A decoupling transformation process is performed on the skill co-occurrence graph to decompose the independent embedding vector of each skill atomic node and the implicit constraints of association between the skill atomic nodes from the skill co-occurrence graph; Using the required skill identifier set for each scheduled task in the set of scheduled tasks and the independent embedding vector, an initial task representation for each scheduled task is generated. Based on the initial task representation and the associated implicit constraints, an activation mapping subgraph is determined on the skill co-occurrence graph. Based on the activation mapping subgraph, topological context information is extracted and fused with the initial task representation to generate a task embedding representation. Based on the task embedding representation and the current schedulable resource status, a scheduling strategy is orchestrated to generate a collaborative task scheduling instruction sequence. The collaborative task scheduling instruction sequence includes the mapping relationship between each task to be scheduled and the execution resource, as well as the execution timing.

[0006] This invention provides a dynamic collaborative task scheduling system, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method.

[0007] This invention provides a computer-readable storage medium including a computer program. When the computer program is run on a dynamic cooperative task scheduling system, the computer program is used to cause the dynamic cooperative task scheduling system to perform the steps of the above-described method.

[0008] In this invention, a skill co-occurrence graph is constructed by performing co-occurrence association analysis on skill identifier combinations in historical scheduling task records. Decoupling transformation is then used to separate the mixed node self-attribute information and inter-node interaction information in the graph into independent embedding vectors and associated implicit constraints, achieving a decoupled expression of inherent skill characteristics and collaborative relationship constraints. An initial task representation is generated by querying the independent embedding vectors using the required skill identifier set of the task to be scheduled. An activation mapping subgraph is then determined on the skill co-occurrence graph by combining the associated implicit constraints. Topological context information is extracted from this subgraph and fused with the initial task representation, ensuring that the generated task embedding representation simultaneously contains both inherent skill attributes and skill collaboration structure context. Scheduling strategy orchestration is performed based on the fit calculation between the task embedding representation and the resource capability embedding vectors, generating a collaborative task scheduling instruction sequence that includes task-resource mapping relationships and execution timing. Therefore, this method effectively improves the utilization rate of skill collaboration patterns by explicitly encoding historical collaborative knowledge into a graph structure and performing decoupling transformation. At the same time, it enhances the task representation's perception of the collaboration structure by activating the topological context fusion of the mapping subgraph, thereby improving the resource matching accuracy and the rationality of execution timing arrangement in multi-skill task scheduling. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a lightweight, intelligent, dynamic collaborative task scheduling method provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of a lightweight intelligent dynamic collaborative task scheduling method provided in an embodiment of the present invention.

[0011] Figure 3 This is a schematic diagram of the structure of a dynamic collaborative task scheduling system provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the technical solutions of the present invention, and not all embodiments. Based on the embodiments described in this document, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of the present invention.

[0013] See Figure 1 This is a lightweight intelligent dynamic collaborative task scheduling method provided in the embodiments of the present invention. The method can be applied to a dynamic collaborative task scheduling system. The specific process is as follows: steps 110-150.

[0014] The lightweight intelligent dynamic collaborative task scheduling method provided in this invention can be applied to multi-skill task scheduling scenarios.

[0015] Step 110: Obtain historical task records and a set of tasks to be scheduled. The historical task records contain historical task entries, and each historical task entry contains a combination of skill identifiers used by the historical task. The set of tasks to be scheduled contains the tasks to be scheduled and their corresponding sets of required skill identifiers.

[0016] When the dynamic collaborative task scheduling system initiates the scheduling process, the system reads pre-stored historical task records from the scheduling database or distributed storage component. Each historical task entry corresponds to a completed task execution process, recording the skill identifier combinations actually invoked during the execution of that historical task. Skill identifier combinations are stored in the form of an identifier list; for example, each skill identifier is a unique coded value representing a set type of operational capability or functional module. Historical task records can be arranged in reverse chronological order of task completion time, or grouped and stored according to the business domain to which the task belongs; the specific organization method depends on the actual configuration of the scheduling system.

[0017] Simultaneously, the system obtains a set of tasks to be scheduled within the current scheduling cycle. This set includes a unique identifier for each task and a set of skill identifiers required for its execution. The required skill identifier set describes the types of capabilities necessary to complete the task. Each task may correspond to multiple skill identifiers, and these identifiers have collaborative relationships. For example, if a task requires skill identifiers SK1, SK2, and SK3 to complete, then the required skill identifier set is {SK1, SK2, SK3}. The task set can be received from the upstream task generation module or retrieved in batches from the task queue. Its data format uses a structured object sequence, with each object containing a task identifier field and a skill identifier list field.

[0018] Step 120: Perform co-occurrence association analysis on the skill identifier combinations in the historical scheduling task records to generate a skill co-occurrence graph, which includes skill atomic nodes and associated edges connecting the skill atomic nodes.

[0019] To extract collaborative patterns between skills from historical data, the system performs co-occurrence association analysis on the historical scheduling task records obtained in step 110. The core of this analysis process is to identify which skill identifiers frequently co-occur in the same historical task and construct a graph structure based on co-occurrence frequency and coverage. The skill co-occurrence graph is an undirected graph where skill atomic nodes represent unique skill identifiers, and association edges connect skill atomic node pairs that co-occur in at least one historical task entry.

[0020] Step 121: Traverse each historical task entry in the historical scheduling task record, extract the skill identifier combination contained in the historical task entry, and pair the skill identifiers in each extracted skill identifier combination to generate candidate skill co-occurrence pairs.

[0021] The system first iteratively scans the historical task records. For each historical task entry read, the skill identifier combination field is parsed, and all skill identifiers contained in that combination are obtained. For example, if a historical task entry's skill identifier combination contains m skill identifiers, the system pairs these skill identifiers in an unordered manner, generating C(m, 2) candidate skill co-occurrence pairs. For instance, if the skill identifier combination is {SK1, SK2, SK3}, the generated candidate skill co-occurrence pairs include (SK1, SK2), (SK1, SK3), and (SK2, SK3). The pairing process does not consider the order of the skill identifiers; that is, (SK1, SK2) and (SK2, SK1) are considered the same candidate skill co-occurrence pair. Each historical task entry is processed independently, and the candidate skill co-occurrence pairs generated from different entries do not affect each other.

[0022] Step 122: Treat each candidate skill co-occurrence pair as a skill co-occurrence event and store the skill co-occurrence event in the skill co-occurrence event pool. The skill co-occurrence event pool is used to collect skill co-occurrence events generated by all historical task entries.

[0023] For each generated candidate skill co-occurrence pair, the system encapsulates it into a skill co-occurrence event. Each skill co-occurrence event contains at least the following information fields: a unique event identifier, the skill identifier pair involved, the identifier of the source historical task entry, and a timestamp. The system appends these skill co-occurrence events to the skill co-occurrence event aggregation pool. The skill co-occurrence event aggregation pool can be implemented using an in-memory database or a distributed message queue, supporting high-throughput write operations. As historical task entries are traversed, a large number of skill co-occurrence events gradually accumulate in the aggregation pool. The same skill identifier pair may appear repeatedly in different historical task entries, and each occurrence generates an independent skill co-occurrence event record.

[0024] Step 123: After traversing all historical task entries, extract all non-repeating skill identifiers from the skill co-occurrence event pool, establish a corresponding skill atomic node for each non-repeating skill identifier, and assign a node identifier to each skill atomic node. The node identifier corresponds one-to-one with the skill identifier.

[0025] After all historical task entries have been processed, the system performs a deduplication scan on the skill co-occurrence event pool. It extracts the unique skill identifiers involved in all skill co-occurrence events from the pool, forming a complete set of skill identifiers. For each skill identifier in the complete set, the system creates a corresponding skill atomic node. Each skill atomic node contains a node identifier field, an original skill identifier field, and attribute fields for storing subsequent statistical information. The node identifier uses either auto-incrementing integer encoding or hash encoding to ensure a one-to-one mapping with the original skill identifier.

[0026] In this way, each skill appearing in the historical scheduling task record is abstracted as an independent node in the graph structure.

[0027] Step 124: For each skill co-occurrence event in the skill co-occurrence event pool, determine the skill identifier pair involved in the skill co-occurrence event, locate the skill atomic node pair corresponding to the skill identifier pair, and establish an association edge between the skill atomic node pairs. If an association edge already exists between the skill atomic node pairs, add the skill co-occurrence event to the association event record list of the association edge. If no association edge exists between the skill atomic node pairs, create a new association edge and use the skill co-occurrence event as the starting record of the association event record list.

[0028] The system iterates through each skill co-occurrence event in the skill co-occurrence event pool again. For the current skill co-occurrence event, it parses out the two skill identifiers it contains and finds the corresponding skill atomic nodes through the mapping relationship established in steps 123. Then, it checks whether there is already an associated edge between the two skill atomic nodes. The way to determine whether there is an associated edge is to query the edge index table, which stores edge references with node pairs as keys.

[0029] If the corresponding record for the node pair exists in the edge index table, it means that an associated edge has been established before. The system will append the information of the current skill co-occurrence event to the associated event record list of the associated edge. The associated event record list is used to store references or copies of all skill co-occurrence events pointing to the edge.

[0030] If the corresponding record for the node pair does not exist in the edge index table, a new associated edge is created. The associated edge includes an edge identifier, a start node identifier, an end node identifier, and a newly created list of associated event records. Then, the current skill co-occurrence event is stored in this list as the first record, and the edge is added to the edge index table.

[0031] Step 125: Extract occurrence density features and coverage breadth features from the initial skill co-occurrence network obtained from skill co-occurrence event processing. Determine the skill co-occurrence map based on the occurrence density features, the coverage breadth features extracted, and the initial skill co-occurrence network.

[0032] Step 1251: After processing all skill co-occurrence events, an initial skill co-occurrence network is obtained, which includes all skill atomic nodes and all associated edges. Each associated edge in the initial skill co-occurrence network has a list of associated event records.

[0033] After processing all skill co-occurrence events in step 124, the system constructs an initial skill co-occurrence network. This network includes all skill atomic nodes generated in step 123 and all associated edges established in step 124. Each associated edge carries a list of associated event records, storing detailed information about all skill co-occurrence events pointing to that edge, including the historical task entry identifier of the event source and the event occurrence time. The initial skill co-occurrence network reflects the raw frequency information of skill co-occurrence in historical scheduled task records, but the significance of co-occurrence has not yet been quantitatively evaluated.

[0034] Step 1252: For each associated edge in the initial skill co-occurrence network, extract the occurrence density characteristics of co-occurrence events and the coverage breadth characteristics of historical task entries from the associated event record list of the associated edge.

[0035] The system iterates through each associated edge in the initial skill co-occurrence network and extracts features independently for each edge. The extraction method for dense occurrence features is as follows: The total number of events in the associated event record list is obtained and used as the basic quantification value for density. Simultaneously, the time interval distribution between adjacent events is calculated, and the dispersion of the time intervals determines whether the events exhibit a clustered or uniform distribution on the time axis. The extraction method for coverage breadth features is as follows: Source historical task entry identifiers are extracted from each event record in the associated event record list. The number of unique identifiers is counted and used as a dimension of coverage breadth. Simultaneously, the frequency distribution of the skill identifier pair in each source task entry is calculated to obtain a coverage breadth feature vector. This vector contains two components: the number of unique task entries and the degree of frequency difference between entries.

[0036] Step 1253: Generate a connection tendency descriptor for the associated edge based on the occurrence density feature and the coverage breadth feature. The connection tendency descriptor qualitatively describes the degree to which the skill atomic nodes at both ends of the associated edge tend to co-occur in historical scheduling tasks.

[0037] Step 12531: Obtain the historical task entry identifier from which each skill co-occurrence event originates in the associated event record list of the associated edge, and statistically analyze the occurrence distribution of different historical task entry identifiers to obtain the task entry distribution characteristics.

[0038] The system extracts the source historical task entry identifier field corresponding to each skill co-occurrence event from the associated event record list of associated edges. Frequency statistics are performed on these identifiers to calculate the number of times each different identifier appears, forming a mapping table from identifier to frequency. Based on this mapping table, task entry distribution characteristics are generated, which include three components: the total number of non-repeating task entries, the ratio of the maximum to the minimum number of occurrences of the associated edge in each task entry, and the percentage of task entries whose occurrence frequency exceeds a preset frequency threshold.

[0039] Step 12532: Generate a density feature based on the total number of skill co-occurrence events in the associated event record list. The density feature is used to reflect the frequency category of the co-occurrence of skill atomic nodes at both ends of the associated edge.

[0040] The system obtains the length of the list of associated event records, i.e., the total number of co-occurring events for the skill corresponding to the associated edge. This total number is compared with several preset density threshold intervals. Each threshold interval corresponds to a frequency category, such as a high-frequency dense category, a medium-frequency dense category, and a low-frequency sparse category. The frequency category corresponding to the threshold interval into which the total number falls is used as the value of the density feature. The density category is represented in discrete label form for subsequent qualitative analysis.

[0041] Step 12533: Determine the co-occurrence coverage feature of the skill atomic nodes at both ends of the associated edge based on the task item distribution feature. The co-occurrence coverage feature is used to reflect the degree of dispersion of co-occurrence events in the task space.

[0042] Based on the total number of non-repeating task entries in the task entry distribution characteristics obtained in step 12531, and the distribution of occurrence frequency among each task entry, the system determines the co-occurrence coverage characteristics. The coverage ratio is calculated by comparing the total number of non-repeating task entries with the total number of task entries in the historical scheduling task records. Simultaneously, the system combines this with the variance of the occurrence frequency distribution to comprehensively determine whether the skill co-occurrence pattern represented by the associated edge is widespread across multiple task types or concentrated only in a few specific tasks. Based on the combination of coverage ratio and distribution variance, the system quantifies the co-occurrence coverage characteristics into discrete dispersion categories, such as widely dispersed, moderately dispersed, or highly concentrated categories.

[0043] Step 12534: Combine the density feature with the co-occurrence coverage feature to form a correlation edge strength feature group, wherein the correlation edge strength feature group includes a qualitative pairing of frequency category and dispersion category.

[0044] The system concatenates the density features obtained in step 12532 and the co-occurrence coverage features obtained in step 12533 ​​into a composite feature group. This feature group is stored in binary form, such as (high-frequency dense category, widely dispersed category) or (low-frequency sparse category, highly concentrated category). Different qualitative combinations reflect different types of skill synergy relationships. For example, a high-frequency dense and widely dispersed combination indicates that the two skills frequently co-occur in many tasks, suggesting a stable and strong synergy relationship between them; while a high-frequency dense but highly concentrated combination indicates that the two skills frequently co-occur in only a few types of tasks, suggesting a strong synergy relationship between them in a specific scenario.

[0045] Step 12535: Invoke the preset connection tendency classification mapping rule to map the associated edge strength feature group to the set connection tendency category label. The connection tendency classification mapping rule is used to define the connection tendency category label corresponding to different combinations of frequency category and dispersion category. The connection tendency category label indicates that the cooperative tendency between the two skill atomic nodes is one of strong tendency cooperative, weak tendency cooperative or conditional tendency cooperative.

[0046] The system pre-configures a connection tendency classification mapping table, which defines the connection tendency category label corresponding to each qualitative combination in the associated edge strength feature group. For example, (high-frequency dense category, widely dispersed category) maps to a strong tendency collaboration label; (high-frequency dense category, highly concentrated category) maps to a conditional tendency collaboration label; (low-frequency sparse category, widely dispersed category) maps to a weak tendency collaboration label; and (low-frequency sparse category, highly concentrated category) maps to a weak tendency collaboration label. Based on the associated edge strength feature group obtained in step 12534, the system queries this mapping table to obtain the corresponding connection tendency category label. The connection tendency category label is represented by an enumeration type, including three values: strong tendency collaboration, weak tendency collaboration, and conditional tendency collaboration.

[0047] Step 12536: Convert the connection tendency category label into a vector form of connection tendency descriptor to generate a connection tendency descriptor. The connection tendency descriptor qualitatively describes the degree to which the skill atomic nodes at both ends of the associated edge tend to co-occur in historical scheduling tasks.

[0048] The system converts the connection tendency category labels obtained in step 12535 into dense vector representations. The conversion method employs embedding encoding, meaning each connection tendency category label is assigned a fixed-dimensional vector. For example, the strong tendency collaboration label corresponds to the vector `vecStrong`, the weak tendency collaboration label to `vecWeak`, and the conditional tendency collaboration label to `vecConditional`. The dimensions of these vectors are consistent with the input dimensions of subsequent processing modules, allowing them to directly participate in vector operations. The resulting vector is the connection tendency descriptor for that associated edge. The connection tendency descriptor not only contains the category information of collaboration tendency but also implicitly implies the relative degree of tendency through positional encoding in the vector space.

[0049] Step 1254: Combine all the associated edges carrying connection tendency descriptors and all the skill atomic nodes in the initial skill co-occurrence network to generate a skill co-occurrence graph. The skill co-occurrence graph includes skill atomic nodes and associated edges connecting skill atomic nodes. The skill atomic nodes correspond to the skill identifiers that appear in the historical scheduling task records. The associated edges connect skill atomic node pairs that appear together in the same historical task entry.

[0050] The system integrates the set of skill atomic nodes generated in step 123, the set of associated edges established in step 124, and the connection tendency descriptors generated for each associated edge in step 1253 to form a complete skill co-occurrence graph. In this graph, each skill atomic node maintains the mapping relationship between its node identifier and the original skill identifier, and each associated edge, in addition to connecting two skill atomic nodes, also carries a corresponding connection tendency descriptor vector. The skill co-occurrence graph can be stored in the form of an adjacency list or an adjacency matrix, where each entry in the adjacency list contains the target node identifier and edge attributes (i.e., connection tendency descriptors). This graph serves as the basic data structure for subsequent decoupling transformation processing and activation mapping subgraph determination, reflecting the distribution of collaborative patterns among skills in historical scheduling tasks.

[0051] Step 130: Perform decoupling transformation on the skill co-occurrence graph to decompose the independent embedding vector of each skill atomic node and the implicit constraints of association between the skill atomic nodes from the skill co-occurrence graph.

[0052] After obtaining the skill co-occurrence graph, the system needs to extract two types of complementary information from the graph structure: first, the inherent attribute representation of each skill atomic node itself; and second, the constraints of collaborative relationships formed between nodes based on historical co-occurrence patterns. The essence of the decoupling transformation process is to separate the node's own information and node interaction information mixed in the graph, and encode them respectively into independent embedding vectors and associated implicit constraints.

[0053] Step 131: Input the skill co-occurrence graph into a preset graph structure parsing component, which includes a neighbor information aggregation sublayer and a feature decoupling sublayer.

[0054] The system takes the skill co-occurrence graph generated in step 1254 as input data and passes it to a pre-trained graph structure parsing component. This component is a composite neural network module containing two functionally distinct sublayers. The neighbor information aggregation sublayer is responsible for collecting the local neighborhood information of each skill atomic node in the graph and integrating it into a context-aware node representation.

[0055] The feature decoupling sublayer is responsible for decomposing the hybrid representation into two mutually orthogonal subspaces, corresponding to the independent attributes and interaction attributes of the nodes, respectively. The parameters of the graph structure parsing component are pre-trained using historical scheduling task data, with the training objective being to minimize the reconstruction error of independent embedding vectors and the prediction error of associated implicit constraints.

[0056] Step 132: Traverse each skill atomic node in the skill co-occurrence graph through the neighbor information aggregation sub-layer to obtain the first-order neighbor node set of the skill atomic node, extract the connection tendency descriptor of the associated edge between each neighbor skill atomic node and the skill atomic node in the first-order neighbor node set, and perform weighted aggregation of the information of the neighbor skill atomic nodes according to the connection tendency descriptor to generate the neighbor aggregation representation of the skill atomic node.

[0057] The system calls the neighbor information aggregation sublayer to perform the following operations on each skill atomic node in the skill co-occurrence graph: First, query the first-order neighbor nodes of the current node from the adjacency structure of the graph, that is, other skill atomic nodes directly connected by the associated edges, to obtain the set of neighbor nodes. For each neighbor node in the set of neighbor nodes, obtain the connection tendency descriptor vector carried on the associated edge connecting the current node and that neighbor node.

[0058] Then, the system performs element-wise multiplication of the initial feature vectors of neighboring nodes with the connectivity tendency descriptor vectors to achieve information filtering based on the strength of cooperative tendency. Dimensions with larger values ​​in the connectivity tendency descriptor indicate a stronger cooperative correlation between the neighboring node and the current node in the corresponding feature space; information in this dimension is preserved after multiplication. Information corresponding to dimensions with smaller values ​​is suppressed. After performing the above weighting operation on all neighboring nodes, the resulting weighted vectors are summed element-wise to generate a neighbor aggregation representation. The dimension of the neighbor aggregation representation is the same as the dimension of the initial feature vectors of the nodes.

[0059] Step 133: The original identifier of the skill atomic node is mapped to a continuous vector space through the neighbor information aggregation sub-layer to obtain the skill atomic node's own representation. The skill atomic node's own representation is then concatenated with the neighbor aggregation representation to form the comprehensive initial embedding representation of the skill atomic node.

[0060] The system also processes the identification information of the skill atomic nodes themselves. Each skill atomic node carries a node identifier, which is a discrete integer code. The neighbor information aggregation sublayer contains a trainable embedding matrix. The number of rows in this matrix is ​​the same as the total number of nodes in the skill co-occurrence graph, and the number of columns is the preset embedding dimension. Based on the node identifier of the skill atomic node, the system looks up the corresponding row vector in the embedding matrix as the skill atomic node's own representation. The skill atomic node's own representation is a dense vector, and its dimension is the same as the number of columns in the embedding matrix.

[0061] Next, the system concatenates the neighbor aggregation representation obtained in step 132 with the representation of the skill's atomic node itself along the vector dimension, generating a comprehensive initial embedding representation whose dimension is the sum of the dimensions of the two. The comprehensive initial embedding representation integrates the node's own identity information and the context information provided by neighboring nodes.

[0062] Step 134: Based on the feature decoupling sub-layer, perform independent component enhancement transformation on the comprehensive initial embedding representation of each skill atom node, extract the inherent attribute components that do not depend on the information of other skill atom nodes from the comprehensive initial embedding representation, and map the inherent attribute components into independent embedding vectors. The independent embedding vectors are used to express the attribute features of the skill atom node itself.

[0063] The system inputs the integrated initial embedding representation of each skill atom node obtained in step 133 into the feature decoupling sublayer. The feature decoupling sublayer contains two parallel transformation paths, used to extract intrinsic attribute components and interactive attribute components, respectively. For independent component enhancement transformation processing, the feature decoupling sublayer first projects the integrated initial embedding representation to a low-dimensional subspace through a linear transformation matrix. The parameters of this linear transformation matrix are trained so that the projection result is as correlated as possible with the node's own static features, and as uncorrelated as possible with the features of neighboring nodes. Then, a nonlinear activation function is used to transform the projection result to obtain an intermediate representation.

[0064] Next, the system maps this intermediate representation back to the original embedding dimension through another linear transformation matrix to obtain the intrinsic attribute component vector. To ensure that this component truly does not depend on neighbor information, the feature decoupling sublayer also introduces an adversarial regularization term. This term attempts to predict the identifiers of neighboring nodes from the intrinsic attribute components through a discriminator network, while the training objective of the feature decoupling sublayer is to prevent the discriminator from making accurate predictions.

[0065] After the above processing, the system outputs the inherent attribute component vector as the independent embedding vector of the skill atomic node. The independent embedding vector expresses the attribute characteristics of the skill atomic node itself, such as the core function category and complexity level of the skill, which are essential attributes that do not change with the collaborative environment.

[0066] Step 135: Based on the feature decoupling sublayer, extract the interaction attribute components that have interaction relationships with other skill atomic nodes from the comprehensive initial embedding representation, and fuse the interaction attribute components with the connection tendency descriptor to obtain the initial association constraint components of the skill atomic node with respect to each associated neighbor.

[0067] In parallel with step 134, the feature decoupling sublayer simultaneously extracts interaction attribute components from the synthesized initial embedding representation. The extraction method employs another set of linear transformation matrices, the parameters of which are designed to preserve the interaction information between nodes in the projection result. The system projects the synthesized initial embedding representation onto a subspace orthogonal to the intrinsic attribute subspace using the interaction attribute extraction matrix, obtaining an interaction attribute component vector. These interaction attribute components reflect how the skill's atomic node interacts with other nodes in the collaborative task.

[0068] Then, for each neighbor node of the skill atomic node, the system obtains the connection tendency descriptor carried on the associated edge connecting the two nodes, and fuses the interaction attribute component vector with the connection tendency descriptor vector. The fusion operation can be performed by concatenating vectors and then applying linear mapping, or by adding element-wise and then normalizing. The result after fusion is the initial association constraint component of the skill atomic node with respect to that specific neighbor. The initial association constraint component is a vector, and its dimension is consistent with the dimension of the implicit association constraint.

[0069] Step 136: For any two directly connected skill atomic nodes in the skill co-occurrence graph, the initial association constraint components of the skill atomic nodes are combined in pairs based on the feature decoupling sublayer to generate pairwise implicit association constraints between the skill atomic node pairs. The pairwise implicit association constraints limit the co-occurrence of the skill atomic node pairs in the task requirements to the range allowed by the historical co-occurrence pattern.

[0070] The system traverses each association edge in the skill co-occurrence graph, obtaining the identifiers of the skill atomic nodes at both ends of the association edge. From the calculation results of step 135, it extracts the initial association constraint components of node A with respect to node B, and the initial association constraint components of node B with respect to node A. Then, it performs pairwise combination operations on these two components. The pairwise combination operation can be implemented using a bilinear transformation, combining two vectors through a trainable bilinear matrix to generate a scalar value or a low-dimensional vector. Another implementation is to concatenate the two vectors and map them to a preset constraint space through a fully connected layer. The result of the combination operation is the pairwise implicit association constraint between the skill atomic node pairs. The pairwise implicit association constraint can be expressed as a range, such as the co-occurrence strength of the node pair in the task requirements should be between the lower limit L and the upper limit U, or it can be expressed as a probability distribution, such as the probability density function parameters of the node pair's co-occurrence.

[0071] It is understandable that, regardless of the form used, the core function of pairwise implicit constraints is to limit the occurrence of these two skill atomic nodes to reasonable task requirements. The way they co-occur should conform to the range allowed by the historical co-occurrence pattern. Co-occurrence patterns that exceed this range are considered abnormal or infeasible.

[0072] Step 137: Collect all pairwise implicit constraints of skill atomic node pairs with directly related edges to obtain the set of implicit constraints.

[0073] The system collects the paired implicit constraints generated for each associated edge in step 136 to form a set of implicit constraints. This set can be stored using a mapping table structure, where the key is the identifier of the skill atomic node pair, and the value is the corresponding paired implicit constraint object. The set of implicit constraints is an important reference for subsequent generation of task embedding representations and scheduling strategy orchestration, as it encodes the cooperative relationship constraints formed between skills in historical scheduling tasks.

[0074] Step 138: Output the independent embedding vector of each skill atomic node and the set of associated implicit constraints as the result of the decoupling transformation process, so as to decompose the independent embedding vector and associated implicit constraints from the skill co-occurrence graph. The independent embedding vector represents the attribute characteristics of the skill atomic node itself, and the associated implicit constraints represent the cooperative relationship restrictions between different skill atomic nodes based on co-occurrence association.

[0075] The system uses the independent embedding vectors generated for each skill atomic node in step 134, and the set of associated implicit constraints generated in step 137, as the final output of the decoupling transformation process in step 130. The set of independent embedding vectors can be organized in matrix form, with rows corresponding to skill atomic nodes and columns corresponding to the dimensions of the embedding vectors. The set of associated implicit constraints is stored independently. Thus, the system has completed the task of separating the node's own attributes and the collaborative constraints between nodes from the skill co-occurrence graph, providing two different types of prior knowledge for subsequent task representation generation.

[0076] Step 140: Using the required skill identifier set for each scheduled task in the set of scheduled tasks and the independent embedding vector, generate an initial task representation for each scheduled task. Based on the initial task representation and the associated implicit constraints, determine an activation mapping subgraph on the skill co-occurrence graph. Extract topological context information based on the activation mapping subgraph and fuse it with the initial task representation to generate a task embedding representation.

[0077] After obtaining the independent embedding vectors of each skill atomic node and the implicit constraints of association between skills, the system needs to apply the aforementioned prior knowledge to the representation generation process of the tasks to be scheduled. Each task to be scheduled is defined by its required set of skill identifiers, but this set only provides the type information of the skills and does not reflect the collaborative structure between skills. The goal of step 140 is to convert the independent embedding vectors at the skill level into task-level representations and use the implicit constraints of association to locate the subgraph structure related to the task on the skill co-occurrence graph, thereby enriching the semantic information of the task.

[0078] Step 141: For each task to be scheduled in the set of tasks to be scheduled, extract each skill identifier from the set of required skill identifiers for the task to be scheduled, query the corresponding independent embedding vector using the correspondence between the skill identifier and the skill atomic node, and arrange all the queried independent embedding vectors to obtain a skill embedding vector sequence.

[0079] The system iterates through each task in the set of tasks to be scheduled. For the current task, it parses the required skill identifier set field to obtain a list of skill identifiers necessary for the task to execute. It iterates through each skill identifier in the list, finds the corresponding skill atomic node through the mapping relationship between skill identifiers and skill atomic nodes established in step 123, and then queries the independent embedding vector corresponding to the node from the set of independent embedding vectors output in step 138.

[0080] The retrieved independent embedding vectors are dense vectors of a fixed dimension, denoted as DimEmb. The system arranges the retrieved independent embedding vectors sequentially according to the original order in the list of required skill identifiers, forming a two-dimensional array structure, namely the skill embedding vector sequence. The length of this sequence is equal to the number of skills required for the current task to be scheduled, and each element is a DimEmb-dimensional vector.

[0081] Step 142: Perform a vector aggregation operation on the skill embedding vector sequence to compress multiple independent embedding vectors in the sequence into a single fixed-dimensional aggregated vector, and use the aggregated vector as the initial task representation of the task to be scheduled.

[0082] The system performs an aggregation operation on the skill embedding vector sequence obtained in step 141. This aggregation operation can be implemented in several ways. One approach is to calculate the element-wise average of all vectors in the sequence; that is, for each dimension, the values ​​of all vectors in the sequence at that dimension are summed, and then divided by the sequence length to obtain the average vector. Another approach is to use weighted summation, where the weights can be pre-set based on the importance of the skill in the task or dynamically calculated through an attention mechanism. Yet another approach is to use max pooling, where for each dimension, the maximum value of all vectors in the sequence at that dimension is taken.

[0083] It is understandable that, regardless of the method used, the result of the aggregation operation is a fixed-dimensional vector, whose dimension is consistent with the dimension DimEmb of the independent embedding vectors. This aggregated vector is the initial task representation of the task to be scheduled. The initial task representation incorporates the inherent attribute information of all the skills required for the task, but does not yet reflect the collaborative structural constraints between skills.

[0084] Step 143: Obtain the set of associated implicit constraints, and perform a compatibility evaluation on the initial task representation and each associated implicit constraint in the set of associated implicit constraints to determine whether the skill co-occurrence pattern implied by the initial task representation conforms to the collaborative relationship constraints defined by each associated implicit constraint.

[0085] The system loads the set of associated implicit constraints generated in step 137. For each pair of associated implicit constraints in the set, the system evaluates the compatibility between the initial task representation of the currently scheduled task and that constraint. The method of compatibility evaluation depends on the specific form of the associated implicit constraints.

[0086] If the constraint is expressed as a numerical range, the system first extracts the components related to the two skills from the initial task representation, calculates the estimated synergy strength of the two skills in the task, and then determines whether the estimated value falls within the range defined by the constraint. If the estimated value is within the range, it is considered compatible; if it is outside the range, it is considered a violation. If the constraint is expressed as a probability distribution, the system calculates the probability density value of the co-occurrence pattern implied in the initial task representation under that distribution. If the density value exceeds a preset threshold, it is considered compatible; otherwise, it is considered a violation. The system performs compatibility evaluation on all associated implicit constraints one by one and records the determination result for each constraint.

[0087] Step 144: Based on the results of the compatibility evaluation, select skill atomic nodes that are compatible with the initial task representation and do not violate the implicit association constraints, and generate a core node set by combining the skill atomic nodes directly corresponding to the required skill identifier set of the task to be scheduled.

[0088] Based on the compatibility evaluation results in step 143, the system selects two types of nodes from all nodes in the skill co-occurrence graph. The first type consists of atomic skill nodes that directly correspond to the set of required skill identifiers for the task to be scheduled. These nodes are the core skills necessary for task execution and must be included regardless of the compatibility evaluation results. The second type consists of other atomic skill nodes that, although not in the set of required skill identifiers, are compatible with the initial task representation and do not violate any implicit constraints.

[0089] The specific selection method is as follows: Traverse all atomic nodes of the skills in the skill co-occurrence graph. For each node, check whether there is a violation of the implicit pairwise association constraint between it and each skill in the required skill identifier set. If the implicit association constraint between the node and all required skill nodes is determined to be compatible, the node is selected. Merge the first type of nodes and the second type of nodes to form the core node set. The core node set contains the skill nodes necessary to perform the task, as well as auxiliary skill nodes compatible with the task in the historical co-occurrence pattern.

[0090] Step 145: Starting from the skill atomic nodes in the core node set, expand outward along the association edges on the skill co-occurrence graph. Incorporate neighboring skill atomic nodes in the graph that are connected to the core nodes through association edges and satisfy the implicit association constraints into the extended node set. Combine the core node set, the extended node set, and the association edges of the set to generate an activation mapping subgraph.

[0091] Step 1451: Initialize the extended node set to an empty set, and take the core node set as the current boundary node set. For each boundary node in the current boundary node set, traverse all first-order neighbor nodes of the boundary node on the skill co-occurrence graph. For each first-order neighbor node, obtain the connection tendency descriptor of the associated edge connecting the boundary node and the first-order neighbor node.

[0092] The system creates an empty set of expansion nodes to store nodes to be included in subsequent expansion processes. Simultaneously, it assigns the core node set to the current boundary node set, representing the starting node set for the current expansion round. For each boundary node in the current boundary node set, the system queries all its first-order neighbors from the adjacency structure of the skill co-occurrence graph—that is, other nodes directly connected via association edges. For each found first-order neighbor, the system obtains the connection tendency descriptor vector carried on the association edge connecting the boundary node and the neighbor node. This vector records the strength of the cooperative tendency between the two skills in historical scheduling tasks.

[0093] Step 1452: Based on the set of implicit association constraints, extract the pairwise implicit association constraints corresponding to the boundary node and the first-order neighbor node, and check whether the first-order neighbor node satisfies the cooperative conditions allowed by the pairwise implicit association constraints.

[0094] The system searches for corresponding pairwise implicit constraints from the set of implicit constraints based on the identifiers of the current boundary node and its first-order neighbor nodes. If a pairwise implicit constraint exists between the node pairs, the system extracts the constraint and checks whether the neighbor nodes satisfy the cooperation conditions defined by the constraint. The checking method is similar to step 143, but the object of the check is the cooperation pattern under the current task requirements. If there is no direct pairwise implicit constraint between the node pairs, the system assumes that the condition is met by default, but further judgment is required based on subsequent correlation evaluation. Only neighbor nodes that satisfy the cooperation conditions allowed by the pairwise implicit constraint can be included in the expanded node set.

[0095] Step 1453: Obtain the initial task representation of the task to be scheduled, and calculate the task correlation between the initial task representation and the independent embedding vector corresponding to the first-order neighbor node.

[0096] The system obtains the initial task representation vector of the task to be scheduled generated in step 142, and the independent embedding vector of the first-order neighbor node generated in step 134. The task relevance between the two can be calculated using cosine similarity, i.e., calculating the cosine of the angle between the two vectors; the closer the cosine value is to 1, the higher the relevance. Alternatively, the reciprocal of the Euclidean distance can be used; the smaller the distance, the higher the relevance. A learnable attention scoring function can also be used, concatenating the two vectors and inputting them into a feedforward neural network, outputting a scalar relevance score. The calculated relevance score is a value between 0 and 1, or falls within this range after normalization.

[0097] Step 1454: If the first-order neighbor node satisfies the pairwise association implicit constraint and the task relevance exceeds the preset activation threshold, the first-order neighbor node is included in the expanded node set and marked as a candidate boundary node for the next round of expansion.

[0098] The system sets an activation threshold, which can be a fixed value, such as 0.6, or dynamically adjusted according to the task type. If the judgment result of step 1452 satisfies the pairwise association implicit constraint, and the task relevance score calculated in step 1453 is greater than the activation threshold, then the system adds the first-order neighbor node to the expanded node set. Simultaneously, the node is marked as a candidate boundary node for the next round of expansion. If either of the two conditions is not met, the node is not included in the expanded node set.

[0099] Step 1455: After traversing all the boundary nodes in the current boundary node set, the candidate boundary nodes newly included in the expansion node set constitute the current boundary node set for the next round of expansion.

[0100] After the system completes traversing all boundary nodes in the current boundary node set, it collects all candidate boundary nodes marked in step 1454 and forms a new current boundary node set for the next round of expansion iteration.

[0101] Step 1456: Repeat the neighbor node traversal, constraint check, correlation calculation and candidate labeling steps until no new candidate boundary nodes are generated after one round of expansion, and stop the expansion process.

[0102] The system repeats steps 1451 to 1455, with each expansion exploring first-order neighbors based on the boundary nodes obtained in the previous expansion. The iteration terminates when the newly generated candidate boundary node set is empty after one expansion, meaning no new nodes satisfy the inclusion criteria. At this point, the expansion process stops, and the expanded node set contains all skill atomic nodes that can be reached through multi-hop expansion and satisfy the constraints and correlation conditions.

[0103] Step 1457: Take the union of the core node set and the extended node set as the activation node set, and construct an activation mapping subgraph based on the association edges between each node in the activation node set in the skill co-occurrence graph.

[0104] The system calculates the union of the core node set and the extended node set to obtain the active node set. Then, it extracts all associated edges from the skill co-occurrence graph where both ends of the graph belong to the active node set. These associated edges, along with all nodes in the active node set, form a subgraph structure, namely the activation mapping subgraph. The activation mapping subgraph is a subset of the original skill co-occurrence graph, focusing on the skill nodes related to the currently scheduled task and the connections between them.

[0105] Step 146: Perform topological context information extraction processing on the activated mapping subgraph to obtain the local connection mode, path connectivity structure between nodes, and community clustering features of each node in the activated mapping subgraph, and combine the local connection mode, path connectivity structure, and community clustering features into topological context information.

[0106] The system performs topological structure analysis on the activation mapping subgraph generated in step 1457, extracting three types of contextual information. The extraction method for local connectivity patterns is as follows: for each node in the subgraph, its degree value (i.e., the number of edges connected to it), the average degree of its neighboring nodes, and the mean of the connectivity tendency descriptor vectors of the edges connecting the node to its neighbors are calculated. This information is then combined to form the local connectivity pattern vector of that node. The extraction method for inter-node path connectivity is as follows: the shortest path length between any two nodes in the subgraph is calculated, forming a distance matrix. Simultaneously, the number of shortest paths passing through each node is counted to evaluate the betweenness centrality of that node in the subgraph.

[0107] The extraction method for community clustering features is as follows: Community detection algorithms, such as the Louvain algorithm or the label propagation algorithm, are used to divide the nodes in the subgraph into different communities. Each community corresponds to a group of closely connected skill nodes. The community detection results include the community identifier to which each node belongs and the list of nodes contained in each community. The system combines these three types of information—local connectivity patterns, path connectivity structure, and community clustering features—into a composite data structure called topological context information according to a preset format.

[0108] Step 147: The topological context information is vectorized and encoded to generate a topological context vector. The topological context vector is then fused with the initial task representation to obtain a task embedding representation that integrates skill collaboration context information.

[0109] The system converts the topological context information obtained in step 146 into a fixed-dimensional vector representation. This conversion can be achieved using a graph neural network encoder, such as a graph isomorphic network or a graph attention network. The node and edge features in the activation map subgraph are taken as input, and after multiple layers of graph convolution operations, a global pooling layer is used to obtain the representation vector of the entire subgraph. Alternatively, the structure-2vec method can be used to encode the adjacency matrix and node attributes of the subgraph. The generated topological context vector is a fixed-dimensional dense vector, denoted by DimTopo.

[0110] The system then fuses the initial task representation vector obtained in step 142 with the topological context vector. Fusion can be achieved through vector concatenation, where the two vectors are joined along their dimensional direction to generate a fused vector with dimensions DimEmb plus DimTopo. Alternatively, gated fusion can be used, where a learnable gated network is employed to calculate the fusion weights of the two vectors, followed by a weighted sum. The resulting fused vector is the task embedding representation of the task to be scheduled. This task embedding representation includes both the inherent attribute information of the skills required for the task and the collaborative structural context information formed by these skills in historical task scheduling.

[0111] Step 150: Based on the task embedding representation and the current schedulable resource status, perform scheduling strategy orchestration processing to generate a collaborative task scheduling instruction sequence. The collaborative task scheduling instruction sequence includes the mapping relationship between each scheduled task and the execution resource and the execution timing.

[0112] After obtaining the task embedding representation of each task to be scheduled, the system needs to assign these tasks to available execution resources and determine the execution order among the tasks. The goal of step 150 is to match the task-level representation information with the resource-level capability information to generate an executable sequence of scheduling instructions.

[0113] Step 151: Obtain the list of schedulable execution resources in the current scheduling period. The list of schedulable execution resources includes the resource identifier and resource capability description features of each execution resource.

[0114] The system retrieves a list of available execution resources from the resource management module within the current scheduling period. Each entry in the list of schedulable execution resources contains a resource identifier field and a resource capability description feature field. The resource identifier is a unique code for the execution resource. The resource capability description feature is structured data that describes the type of skill possessed by the resource and its proficiency or processing ability. For example, the resource capability description feature can be represented as a mapping from skill to capability value, where the skill identifier is taken from the complete set of skill identifiers, and the capability value can be a continuous numerical value or a discrete level.

[0115] Step 152: Perform capability embedding encoding on the resource capability description features of each execution resource to generate a resource capability embedding vector. The resource capability embedding vector is in the same vector space as the task embedding representation.

[0116] The system converts the resource capability description features of each execution resource into a vector representation. The conversion method can be embedding encoding, that is, for each skill identifier in the resource capability description features, query the independent embedding vector of the skill atomic node corresponding to the skill, and then aggregate all the queried independent embedding vectors according to the capability value.

[0117] Specifically, for each skill possessed by a resource, the independent embedding vector of that skill is multiplied by its corresponding capability value weight. Then, all weighted vectors are summed, and the result is mapped to a preset vector dimension via a linear transformation layer. Alternatively, a trainable resource encoder can be used directly, taking the original skill-capability mapping table as input and outputting fixed-dimensional resource capability embedding vectors. The dimension of the generated resource capability embedding vectors is consistent with the dimension of the task embedding representation generated in step 147 to ensure that they can be compared and matched in the same vector space.

[0118] Step 153: Perform a fitting degree calculation on the task embedding representation of each task to be scheduled in the set of tasks to be scheduled one by one with the resource capability embedding vector of each execution resource to generate a task resource fitting degree matrix. The task resource fitting degree matrix is ​​used to represent the capability matching status between each task to be scheduled and each execution resource.

[0119] The system constructs a two-dimensional matrix, where the number of rows equals the number of tasks to be scheduled, denoted as Ntask, and the number of columns equals the number of schedulable execution resources, denoted as Nres. For the i-th task to be scheduled and the j-th execution resource, the system obtains the task embedding representation vector T for that task. i and the resource capability embedding vector R of the resource j Calculate the fitness score between the two. Fitness can be calculated using the dot product operation, i.e., T i and R j The corresponding dimensions are multiplied and summed to obtain a scalar score. A larger dot product value indicates a higher degree of matching between the task and the resource. Another calculation method uses cosine similarity, dividing the dot product by the product of the magnitudes of the two vectors. Yet another method uses the reciprocal of the Euclidean distance; a smaller distance indicates a higher fit. The system fills the calculated fit score into the i-th row and j-th column of the matrix. After traversing all tasks and all resources, a complete task-resource fit matrix is ​​obtained.

[0120] Step 154: Based on the task resource adaptability matrix, use heuristic assignment logic to determine the candidate execution resource allocation scheme for each task to be scheduled and generate the initial scheduling assignment mapping.

[0121] The system allocates one or more candidate execution resources to each task to be scheduled based on the task-resource fit matrix. The heuristic assignment logic can employ a greedy strategy, selecting the execution resource with the highest fit score as its assignment target for each task. If multiple tasks compete for the same resource, they are sorted by fit score from highest to lowest, prioritizing tasks with higher scores. Alternatively, the Hungarian algorithm is used to solve the maximum weight matching problem in a bipartite graph. The task set and resource set constitute the two parts of the bipartite graph, with the fit score serving as the edge weight. The globally optimal assignment scheme is found using the Hungarian algorithm. The assignment result is a mapping relationship, associating each task to be scheduled with at least one execution resource, forming the initial scheduling assignment mapping.

[0122] Step 155: Based on the set of tasks allocated to each execution resource in the initial scheduling assignment mapping, check whether the task load of each execution resource exceeds the resource capacity limit, and perform task transfer adjustment on the scheduled tasks associated with the execution resources that exceed the resource capacity limit to obtain the adjusted scheduling assignment mapping.

[0123] The system iterates through each execution resource in the initial scheduling assignment map, counting the number of tasks assigned to that resource and the total skills required for those tasks. The results are compared to the execution resource's capacity limit. The capacity limit can be represented as the maximum number of concurrent tasks or the maximum processing capacity quota for each skill. If the total skills required by the tasks assigned to an execution resource exceed its capacity limit, the system triggers a task transfer adjustment.

[0124] The task transfer and adjustment strategy includes: selecting tasks with relatively low fit scores from the task set associated with overloaded resources, and reassigning them to other execution resources that still have idle capacity. The selection criteria are the fit scores of the task with other resources and the remaining capacity of the target resource. The system repeats the adjustment process until the load of all execution resources does not exceed their capacity limits, resulting in the adjusted scheduling and assignment mapping.

[0125] Step 156: Obtain the list of tasks to be scheduled for each execution resource in the adjusted scheduling assignment mapping. For each execution resource, determine the execution order of each task to be scheduled on that execution resource based on the skill dependency relationship between tasks and the implicit constraints of association, and generate a resource-level task execution sequence.

[0126] For each execution resource, the system retrieves all scheduled tasks assigned to that resource from the adjusted scheduling assignment map, forming a list of scheduled tasks for that resource. Then, the execution order is determined based on the skill dependencies between tasks. Skill dependencies manifest as some tasks potentially requiring the output of other tasks as input, which can be inferred from the inclusion relationships between the sets of required skill identifiers for each task.

[0127] Simultaneously, the system also needs to consider implicit constraints, namely, tasks corresponding to two skills with strong collaborative tendencies should be arranged as closely and consecutively as possible, while tasks corresponding to skills with weak or conditional collaborative tendencies should have sufficient intervals or be executed in a specific order. Based on these constraints, the system uses topological sorting or heuristic sorting algorithms to generate a linear execution order sequence of each task to be scheduled on the resource, i.e., a resource-level task execution sequence.

[0128] Step 157: Perform global timing alignment of the resource-level task execution sequences of all execution resources to eliminate timing conflicts between skill-dependent tasks on different execution resources, and obtain a globally consistent task execution timing scheme.

[0129] The system collects resource-level task execution sequences from all execution resources and performs global merging and alignment. The goal of global timing alignment is to ensure that if there is a skill dependency between tasks on two different resources, the task that creates the dependency must be completed before the task it depends on. The system first identifies all cross-resource task dependency edges and constructs a global directed acyclic graph, where nodes are tasks to be scheduled and directed edges represent dependencies.

[0130] Then, the directed acyclic graph is topologically sorted to obtain a global linear execution order. During the topological sorting process, if two tasks have no direct dependency, their relative order can be preserved based on the original order in the resource-level task execution sequence, or adjusted according to task priority. The final global task execution order is a globally consistent task execution timing scheme, which allocates a global time step or time window to each scheduled task.

[0131] Step 158: Based on the adjusted scheduling assignment mapping and the globally consistent task execution timing scheme, generate a collaborative task scheduling instruction sequence. Each instruction in the collaborative task scheduling instruction sequence includes a task identifier to be scheduled, an assigned execution resource identifier, and a task start execution timing marker.

[0132] The system combines the adjusted scheduling assignment mapping obtained in step 155 with the globally consistent task execution timing scheme obtained in step 157. For each task to be scheduled, the system obtains its assigned execution resource identifier from the scheduling assignment mapping and its start execution timing flag from the task execution timing scheme. The start execution timing flag can be in the form of an absolute timestamp, a relative time offset, or a timing position index. Following the order in the task execution timing scheme, the system combines the task identifier, assigned execution resource identifier, and start execution timing flag of each task into a scheduling instruction. All scheduling instructions are arranged in execution order to form a collaborative task scheduling instruction sequence, which can be directly issued to each execution resource to guide the actual task execution process.

[0133] After generating the cooperative task scheduling instruction sequence, to ensure the quality and robustness of the scheduling scheme, embodiments of the present invention may further include: Step 210: Obtain the task embedding representation of each task to be scheduled in the collaborative task scheduling instruction sequence and the resource capability embedding vector of the assigned execution resource. Perform vector concatenation processing on the task embedding representation and resource capability embedding vector with mapping relationship to generate a combination matching representation. Arrange all combination matching representations according to the execution time sequence in the collaborative task scheduling instruction sequence to form a scheduling execution snapshot feature sequence.

[0134] Based on the collaborative task scheduling instruction sequence generated in step 158, the system extracts the task embedding representation of the task to be scheduled and the resource capability embedding vector of the assigned execution resource for each instruction. These two vectors are concatenated along their dimensional direction to generate a combined fitting representation vector, the dimension of which is equal to the sum of the task embedding representation dimension and the resource capability embedding vector dimension. The system then arranges all combined fitting representation vectors sequentially according to the execution order in the scheduling instruction sequence, forming a two-dimensional array structure called the scheduling execution snapshot feature sequence. The length of this sequence is equal to the total number of tasks to be scheduled, and each element is a combined fitting representation vector.

[0135] Step 220: Input the scheduling execution snapshot feature sequence into the preset scheduling coherence analysis logic, and perform difference quantization processing on the combination adaptation representation of adjacent time sequence positions along the execution time sequence to obtain the scheduling coherence deviation distribution.

[0136] The system inputs the scheduling execution snapshot feature sequence generated in step 210 into the pre-configured scheduling coherence analysis logic. This analysis logic performs difference quantization on two paired matching representation vectors at adjacent time positions along the execution time sequence direction. Difference quantization can be performed using Euclidean distance, which is the square root of the sum of the squares of the differences between the two vectors in each dimension. Alternatively, cosine distance can be used, which is 1 minus the cosine similarity of the two vectors. For the t-th and t+1-th paired matching representations in the sequence, the calculated difference value is denoted as Dist_t. After performing difference quantization on all adjacent position pairs, a difference value sequence is obtained. Statistical analysis is performed on this sequence, such as calculating the mean, variance, and maximum value, to form a scheduling coherence deviation distribution. Larger difference values ​​in the deviation distribution indicate an anomaly in the connection between two adjacent tasks, that is, a drastic change in the capability matching pattern from task to resource.

[0137] Step 230: Based on the continuous segments in the scheduling coherence deviation distribution where the degree of deviation exceeds the preset deviation tolerance boundary, identify the task transfer points in the collaborative task scheduling instruction sequence that have an abnormal tendency in execution connection.

[0138] The system sets a deviation threshold, such as the mean of the difference values ​​plus twice the standard deviation, or a fixed empirical threshold. It iterates through the difference value sequence obtained in step 220, marking the locations where the difference values ​​exceed the deviation threshold. If multiple consecutive adjacent locations exceed the threshold, the continuous segment formed by these locations is identified as a region with a tendency for abnormal execution continuity. The transition point between the first task end position and the last task start position in this continuous segment is the task transfer point. The task transfer point indicates that the continuity of resource capabilities is disrupted when switching from one task mode to another.

[0139] Step 240: For the tasks to be scheduled involved in the task transfer point, extract the pairwise implicit constraints related to the required skill identifier set of the task to be scheduled from the set of associated implicit constraints, and rearrange the task assignment at the task transfer point according to the cooperative relationship restrictions defined by the pairwise implicit constraints to generate local task assignment adjustment instructions to update the cooperative task scheduling instruction sequence.

[0140] The system locates the two scheduled tasks involved in the task transfer point identified in step 230. Pairwise implicit constraints related to the required skill identifier sets of these two tasks are extracted from the set of associated implicit constraints. Specifically, all pairwise implicit constraints between the skill sets required by the first task and the skill sets required by the second task, as well as pairwise implicit constraints between skills within the two tasks, are extracted.

[0141] Then, based on the aforementioned constraints on collaborative relationships, the system determines whether the current task assignment violates these constraints or whether there are any assignment schemes that can be optimized to improve coherence. The system attempts to adjust the task assignment at the task transfer point, for example, by reassigning the second task to another execution resource, inserting a buffer task between the two tasks, or adjusting the execution order of the two tasks. After the reordering process is completed, a local task assignment adjustment instruction is generated to update the relevant positions in the collaborative task scheduling instruction sequence generated in step 158, resulting in an updated scheduling instruction sequence.

[0142] Furthermore, embodiments of the present invention may also include another optimization and adjustment step for achieving cross-resource load collaborative rebalancing: Step 310: Obtain the activation mapping subgraph of each task to be scheduled in the collaborative task scheduling instruction sequence, and perform topological structure compression processing on each activation mapping subgraph to obtain the subgraph structure feature code. The subgraph structure feature code represents the connection mode compression information between skill atomic nodes in the activation mapping subgraph.

[0143] The system extracts topological structure information from the activation map subgraph of each task to be scheduled generated in step 1457. Each activation map subgraph undergoes topological structure compression. Compression can be achieved using a graph hashing algorithm, mapping the adjacency matrix of the subgraph to a fixed-length binary or integer code using a defined hash function. Alternatively, the Weisfeiler-Lehman graph kernel method iteratively aggregates and sorts node labels to generate feature codes that can distinguish different graph structures. The subgraph structure feature codes reflect the connection patterns between skill atomic nodes in the activation map subgraph; subgraphs with the same or similar structure feature codes represent similar skill collaboration patterns.

[0144] Step 320: Interleave and combine the subgraph structure feature codes of all tasks to be scheduled belonging to the same execution resource to form a resource load topology profile of the execution resource. Perform structural similarity analysis on the resource load topology profiles of different execution resources, and divide the execution resources into different load coordination mode groups based on structural similarity.

[0145] Based on the collaborative task scheduling instruction sequence generated in step 158, the system associates each task to be scheduled with its assigned execution resource. For each execution resource, it collects the subgraph structure feature codes corresponding to all tasks assigned to that resource, sorts these feature codes according to task execution order or according to feature code value, and then concatenates them to form a composite feature sequence, which is the resource load topology profile of that execution resource. The resource load topology profile describes the overall distribution characteristics of the tasks undertaken by that execution resource in the skill collaboration structure.

[0146] The system then performs pairwise structural similarity analysis on the resource load topology profiles of all execution resources. Similarity can be calculated using sequence edit distance or dynamic time warping distance. Based on the calculated similarity matrix, hierarchical clustering or K-means clustering algorithms are used to group execution resources into different load collaboration pattern groups. Execution resources within the same group share similar skill collaboration structure load characteristics.

[0147] Step 330: For execution resources within the same load coordination mode group, perform internal difference quantification on their respective resource load topology profiles to obtain the load balance deviation degree within the group. When the load balance deviation degree within the group indicates that there is an imbalance tendency in the skill co-occurrence structure undertaken by different execution resources within the group, generate a resource coordination adjustment instruction based on the associated implicit constraint set. The resource coordination adjustment instruction is used to migrate some of the scheduled tasks on the execution resources with an imbalance in the skill co-occurrence structure to other execution resources within the same group, so as to achieve cross-resource load coordination rebalancing scheduling.

[0148] For each load coordination mode group, the system performs pairwise difference quantification on the resource load topology profiles of all execution resources within that group, calculating the average and maximum difference within the group. The degree of load balancing deviation within a group can be measured by calculating the difference between each resource profile within the group and the group center profile, which can be the average or median of all profiles within the group. If the deviation exceeds a preset balance tolerance threshold, it indicates an imbalance in the skill co-occurrence structure undertaken by different execution resources within the group, meaning that some resources are undertaking too many tasks in a certain coordination mode, while other resources are undertaking fewer. In this case, the system generates resource coordination adjustment instructions based on the associated implicit constraint set.

[0149] The adjustment instructions include: selecting a subset of tasks to be scheduled from overloaded resources, based on the degree of matching between their subgraph structure feature codes and the resource load topology profiles of less overloaded resources within the same group; then migrating these tasks to other resources in the same group with a higher degree of matching. After generating the resource coordination adjustment instructions, the system updates the coordination task scheduling instruction sequence accordingly, achieving cross-resource load coordination rebalancing scheduling.

[0150] Through the above steps, the lightweight intelligent dynamic collaborative task scheduling method provided by this invention mines skill co-occurrence patterns from historical task scheduling records, constructs a skill co-occurrence graph, decomposes it into independent embedding vectors and associated implicit constraints, and then generates a task embedding representation that integrates topological context information. Finally, it generates a collaborative task scheduling instruction sequence through fitness calculation and scheduling strategy orchestration. This method can effectively utilize the collaborative pattern knowledge in historical data, improve the ability of scheduling decisions to perceive skill dependencies, and thus improve the rationality of task scheduling and resource utilization efficiency.

[0151] Based on the above, the following is an exemplary description of the architecture, technical implementation, training process, and application of the graph structure parsing component involved in the embodiments of the present invention: The graph structure parsing component mentioned in step 131 of this embodiment adopts a graph attention network architecture, which includes two stacked graph attention encoding layers. Each graph attention encoding layer contains a multi-head attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by a layer normalization operation and a residual connection. The multi-head attention sublayer sets the number of attention heads to 4, and each attention head independently calculates the attention coefficients between nodes. The feedforward neural network sublayer adopts a two-layer fully connected structure. The first layer expands the input dimension to 4 times, and the second layer compresses it back to the original dimension, using Gaussian error linear units as activation functions in between. The feature decoupling sublayer adopts two parallel linear transformation paths. Each path contains a linear transformation matrix and a batch normalization layer. The outputs of the two paths are regularized by an orthogonality constraint loss function to ensure that the cosine similarity between independent embedding vectors and interactive attribute components approaches zero.

[0152] Regarding the training process of this component, the training dataset is derived from historical scheduling task records, containing 80,000 valid historical task entries. Skill co-occurrence graph samples are extracted from these entries. Each training sample consists of a skill co-occurrence graph and its corresponding node attribute annotations. The input format is the adjacency matrix and initial feature matrix of the skill co-occurrence graph. The adjacency matrix has a dimension of N×N, and the initial feature matrix has a dimension of N×F, where N is the number of skill atomic nodes in the graph, and F is the initial feature dimension of each node. The output format is the independent embedding vector of each node and the implicit constraint parameters between node pairs. The Adam optimizer is used during training, with an initial learning rate set to 0.001. The learning rate decay strategy uses cosine annealing scheduling, decaying to 0.9 times the current value every 10 training epochs. The batch size is set to 16, and the training epochs are 120. The loss function consists of three parts: the reconstruction loss of the independent embedding vectors uses mean squared error, the prediction loss of the implicit constraints uses binary cross-entropy, and the orthogonality constraint loss. The three loss components are weighted and summed at a ratio of 0.4:0.4:0.2 to obtain the total loss. The evaluation metrics are the cosine similarity preservation rate of independent embedding vectors and the prediction accuracy of the association implicit constraint. An early stopping mechanism is triggered when the average prediction accuracy on the validation set no longer improves for 5 consecutive rounds, and the model parameters with the highest accuracy on the validation set are saved.

[0153] When applying the model, the skill co-occurrence graph generated in step 1254 is first preprocessed. The preprocessing includes: converting the node identifiers of skill atomic nodes into one-hot encoded vectors, where the dimension of the one-hot encoded vector is equal to the size of the entire set of skill identifiers, with the dimension corresponding to the current node set to 1 and the other dimensions set to 0; stacking the one-hot encoded vectors of all nodes to form the initial node feature matrix. Simultaneously, an adjacency matrix is ​​extracted from the skill co-occurrence graph; if there is a connecting edge between two nodes, the value in the corresponding position of the adjacency matrix is ​​set to 1, otherwise it is set to 0. The initial node feature matrix and the adjacency matrix are then input into the graph structure parsing component. The forward propagation process of the component is as follows: First, two rounds of graph attention encoding are performed through the neighbor information aggregation sublayer. In each round of encoding, each node aggregates the features of its neighboring nodes. The aggregation weight is determined by the attention coefficient, which is calculated based on a weighted combination of node features and connectivity tendency descriptors. Then, the independent embedding vector and interaction attribute components are output through two parallel paths of the feature decoupling sublayer. Finally, the independent embedding vector of each node is output as the final independent embedding representation of the atomic node of that skill. The interaction attribute components between node pairs are combined in pairs and then output as the implicit constraint of association. The independent embedding vectors output by the model can be used directly without post-processing. The output implicit constraint of association is mapped to the 0-1 interval by the Sigmoid function and then used as the constraint strength value.

[0154] Based on the above content, combined with Figure 2 As shown in the embodiments of the present invention, the core concept of the dynamic collaborative task scheduling method based on lightweight intelligence is to extract the co-occurrence patterns between skills from historical task scheduling records, encode the patterns into graph structure knowledge, and then use the knowledge to enhance the representation of the task to be scheduled, and finally generate a reasonable task scheduling instruction sequence.

[0155] like Figure 2 As shown on the left, the system first acquires two types of input data: historical task records and a set of tasks to be scheduled. The historical task records contain a large number of completed task entries, each recording the skill identifier combination used in that task. These combinations reflect the collaborative relationships of different skills in a real-world scheduling scenario. The set of tasks to be scheduled contains the tasks that need to be processed in the current cycle and their corresponding sets of required skill identifiers. These two types of data constitute the knowledge source and processing object of the entire scheduling method.

[0156] like Figure 2As shown in the upper middle section, the system performs co-occurrence association analysis on skill identifier combinations in historical scheduling task records. This process traverses the skill identifier combinations in each historical task entry, pairing skill identifiers appearing in the same task to generate candidate skill co-occurrence pairs. Each co-occurrence pair is stored as a skill co-occurrence event in a pool. After processing, unique skill identifiers are extracted from the pool, a skill atomic node is created for each identifier, and association edges are established between skill atomic nodes that have co-occurred, forming an initial skill co-occurrence network. Subsequently, based on the occurrence density and coverage breadth characteristics of the co-occurrence events corresponding to each association edge, the system generates a connection tendency descriptor for each association edge. This descriptor qualitatively characterizes the degree to which two skills tend to co-occur in historical scheduling tasks. Combining all association edges carrying connection tendency descriptors with skill atomic nodes yields a complete skill co-occurrence graph, which contains skill atomic nodes and association edges connecting skill atomic nodes.

[0157] like Figure 2 As shown in the lower center, the system performs decoupling transformation on the skill co-occurrence graph. The core objective of the decoupling transformation is to decompose the mixed information in the graph into two complementary types of knowledge. The first type is the independent embedding vector of each skill atomic node. This vector is obtained by extracting inherent attribute components that do not depend on the information of other nodes from the node's comprehensive initial embedding representation, expressing the attribute characteristics of the skill atomic node itself. The second type is the implicit constraints on the associations between skill atomic nodes. This constraint is obtained by extracting the interaction attribute components between nodes and performing pairwise combination operations, limiting the range of historical co-occurrence patterns that different skills should conform to when they appear together in the task requirements. Through the decoupling transformation, the system separates two types of prior knowledge from the graph: the node's own attributes and the collaborative relationships between nodes.

[0158] like Figure 2As shown on the right, the system uses independent embedding vectors obtained from decoupling transformation and associated implicit constraints to enhance the representation of each task in the set of tasks to be scheduled. For each task to be scheduled, starting from the set of required skill identifiers, the system queries the corresponding independent embedding vectors and performs aggregation operations to obtain an initial task representation. Then, the initial task representation is evaluated for compatibility with the associated implicit constraints, and skill atomic nodes compatible with the task are selected. A core node set is generated by combining the nodes directly corresponding to the skills required by the task. Starting from the core node set, the system expands outward along the associated edges on the skill co-occurrence graph, and neighboring nodes that satisfy the associated implicit constraints and have a high correlation with the initial task representation are included in the extended node set. An activation map subgraph is generated by combining the core node set, the extended node set, and the associated edges between them. Next, topological context information, including the local connection patterns of nodes, the path connectivity structure between nodes, and community clustering features, is extracted from the activation map subgraph. This information is vectorized and encoded and then fused with the initial task representation to generate a task embedding representation. This representation integrates both the inherent attributes of the skill and the skill collaboration context.

[0159] like Figure 2 As shown at the bottom, the system performs scheduling strategy orchestration based on task embedding representations and the current schedulable resource status. First, it acquires the resource capability description features of schedulable execution resources and encodes them into resource capability embedding vectors in the same vector space as the task embedding representations. It calculates the fit between each task embedding representation and each resource capability embedding vector, generating a task-resource fit matrix. Based on this matrix, it uses heuristic assignment logic to determine the initial scheduling assignment mapping and adjusts the task transfer for execution resources exceeding their resource capability limits. After determining the mapping relationship between tasks and resources, it further determines the execution order of tasks on each execution resource based on skill dependencies and associated implicit constraints, and performs global temporal alignment of task sequences for different resources. Finally, it generates a collaborative task scheduling instruction sequence, which contains the mapping relationship between each scheduled task and execution resource, as well as the execution temporal sequence, and can be directly used to guide actual task scheduling execution.

[0160] The lightweight intelligent dynamic collaborative task scheduling method provided in this invention constructs a skill co-occurrence graph by performing co-occurrence association analysis on skill identifier combinations in historical scheduling task records. It then uses decoupling transformation to separate the mixed node self-attribute information and inter-node interaction information in the graph into independent embedding vectors and associated implicit constraints, achieving a decoupled expression of inherent skill characteristics and collaborative relationship constraints. The method uses the required skill identifier set of the task to be scheduled to query the independent embedding vectors to generate an initial task representation. Combined with the associated implicit constraints, it determines an activation mapping subgraph on the skill co-occurrence graph, extracts topological context information from this subgraph, and fuses it with the initial task representation. This results in a task embedding representation that simultaneously contains inherent skill attributes and skill collaboration structure context. Finally, based on the fit calculation between the task embedding representation and the resource capability embedding vector, a scheduling strategy is orchestrated to generate a collaborative task scheduling instruction sequence containing task-resource mapping relationships and execution timing. Therefore, this method effectively improves the utilization rate of skill collaboration patterns by explicitly encoding historical collaborative knowledge into a graph structure and performing decoupling transformation. At the same time, it enhances the task representation's perception of the collaboration structure by activating the topological context fusion of the mapping subgraph, thereby improving the resource matching accuracy and the rationality of execution timing arrangement in multi-skill task scheduling.

[0161] Based on the same inventive concept, embodiments of the present invention also provide a dynamic collaborative task scheduling system. See also... Figure 3 As shown, it is a schematic diagram of a possible dynamic collaborative task scheduling system provided in an embodiment of the present invention. Figure 3 The dynamic collaborative task scheduling system 200 includes a processor 210 and a memory 220. The processor 210 and the memory 220 are interconnected via a communication bus. The memory 220 stores computer programs that can be executed by the processor 210. By executing the instructions stored in the memory 220, the processor 210 can perform the steps of the aforementioned lightweight intelligent dynamic collaborative task scheduling method.

[0162] Based on the same inventive concept, embodiments of the present invention provide a computer-readable storage medium including a computer program. When the computer program runs on a dynamic collaborative task scheduling system, it causes the system to execute the steps of the aforementioned lightweight intelligent dynamic collaborative task scheduling method. In some possible implementations, various aspects of the lightweight intelligent dynamic collaborative task scheduling method provided by the present invention can also be implemented as a program product including a computer program. When the program product runs on a dynamic collaborative task scheduling system, it causes the system to execute the steps of the aforementioned lightweight intelligent dynamic collaborative task scheduling method. For example, the dynamic collaborative task scheduling system can execute... Figure 1The steps are shown in the diagram. The computer-readable storage medium includes volatile or non-volatile or a combination thereof, and may be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium.

[0163] Accordingly, embodiments of the present invention also provide a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above-described method embodiments. It should be understood that each step or combination of steps in the above-described method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above-described method embodiments.

[0164] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0165] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0166] Finally, it should be noted that the above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A lightweight, intelligent, dynamic collaborative task scheduling method, characterized in that, include: Obtain historical task records and a set of tasks to be scheduled. The historical task records contain historical task entries, and each historical task entry contains a combination of skill identifiers used by the historical task. The set of tasks to be scheduled contains the tasks to be scheduled and their corresponding sets of required skill identifiers. The skill identifier combinations in the historical scheduling task records are subjected to co-occurrence association analysis to generate a skill co-occurrence graph, which includes skill atomic nodes and associated edges connecting the skill atomic nodes; A decoupling transformation process is performed on the skill co-occurrence graph to decompose the independent embedding vector of each skill atomic node and the implicit constraints of association between the skill atomic nodes from the skill co-occurrence graph; Using the required skill identifier set for each scheduled task in the set of scheduled tasks and the independent embedding vector, an initial task representation for each scheduled task is generated. Based on the initial task representation and the associated implicit constraints, an activation mapping subgraph is determined on the skill co-occurrence graph. Based on the activation mapping subgraph, topological context information is extracted and fused with the initial task representation to generate a task embedding representation. Based on the task embedding representation and the current schedulable resource status, a scheduling strategy is orchestrated to generate a collaborative task scheduling instruction sequence. The collaborative task scheduling instruction sequence includes the mapping relationship between each task to be scheduled and the execution resource, as well as the execution timing.

2. The method according to claim 1, characterized in that, The step of performing co-occurrence association analysis on the skill identifier combinations in the historical scheduling task records to generate a skill co-occurrence map includes: Traverse each historical task entry in the historical scheduling task record, extract the skill identifier combination contained in the historical task entry, and pair the skill identifiers in the extracted skill identifier combination to generate candidate skill co-occurrence pairs. Each candidate skill co-occurrence pair is treated as a skill co-occurrence event, and the skill co-occurrence event is stored in a skill co-occurrence event pool. The skill co-occurrence event pool is used to collect skill co-occurrence events generated by all historical task entries. After traversing all historical task entries, extract all non-repeating skill identifiers from the skill co-occurrence event pool, establish a corresponding skill atomic node for each non-repeating skill identifier, and assign a node identifier to each skill atomic node. The node identifier corresponds one-to-one with the skill identifier. For each skill co-occurrence event in the skill co-occurrence event pool, the skill identifier pair involved in the skill co-occurrence event is determined, the skill atomic node pair corresponding to the skill identifier pair is located, and an association edge is established between the skill atomic node pairs. If an association edge already exists between the skill atomic node pairs, the skill co-occurrence event is added to the association event record list of the association edge. If no association edge exists between the skill atomic node pairs, a new association edge is created, and the skill co-occurrence event is used as the starting record of the association event record list. Based on the initial skill co-occurrence network obtained from skill co-occurrence event processing, occurrence density features and coverage breadth features are extracted. Based on the occurrence density features, the coverage breadth features extracted, and the initial skill co-occurrence network, the skill co-occurrence map is determined.

3. The method according to claim 2, characterized in that, The initial skill co-occurrence network obtained from skill co-occurrence event processing is used to extract occurrence density features and coverage breadth features. Based on the occurrence density features, the coverage breadth features, and the initial skill co-occurrence network, the skill co-occurrence map is determined, including: After processing all skill co-occurrence events, an initial skill co-occurrence network is obtained, which contains all skill atomic nodes and all associated edges. Each associated edge in the initial skill co-occurrence network has a list of associated event records. For each associated edge in the initial skill co-occurrence network, extract the occurrence density characteristics of co-occurrence events and the coverage breadth characteristics of historical task entries from which the co-occurrence events originate from the associated event record list of that associated edge; Based on the dense occurrence feature and the coverage breadth feature, a connection tendency descriptor is generated for the associated edge. The connection tendency descriptor qualitatively describes the degree to which the skill atomic nodes at both ends of the associated edge tend to co-occur in historical scheduling tasks. All associated edges carrying connection tendency descriptors and all skill atomic nodes in the initial skill co-occurrence network are combined to generate a skill co-occurrence graph. The skill co-occurrence graph includes skill atomic nodes and associated edges connecting skill atomic nodes. The skill atomic nodes correspond to skill identifiers appearing in the historical scheduling task records. The associated edges connect skill atomic node pairs that appear together in the same historical task entry.

4. The method according to claim 3, characterized in that, The connection tendency descriptor for the associated edge is generated based on the occurrence density feature and the coverage breadth feature. This connection tendency descriptor qualitatively describes the degree to which the skill atomic nodes at both ends of the associated edge tend to co-occur in historical scheduling tasks, including: Obtain the historical task entry identifier from which each skill co-occurrence event originates in the associated event record list of the associated edge, and statistically analyze the occurrence distribution of different historical task entry identifiers to obtain the task entry distribution characteristics; A density feature is generated based on the total number of skill co-occurrence events in the associated event record list. The density feature is used to reflect the frequency category of the co-occurrence of skill atomic nodes at both ends of the associated edge. Based on the task item distribution characteristics, the co-occurrence coverage characteristics of the skill atomic nodes at both ends of the associated edge are determined. The co-occurrence coverage characteristics are used to reflect the degree of dispersion of co-occurrence events in the task space. The density feature and the co-occurrence coverage feature are combined to form an association edge strength feature group, which includes a qualitative combination of frequency category and dispersion category; The preset connection tendency classification mapping rule is invoked to map the associated edge strength feature group to the set connection tendency category label. The connection tendency classification mapping rule is used to define the connection tendency category label corresponding to different combinations of frequency category and dispersion category. The connection tendency category label indicates that the cooperative tendency between the two skill atomic nodes is one of strong tendency cooperative, weak tendency cooperative, or conditional tendency cooperative. The connection tendency category label is converted into a vector form of connection tendency descriptor to generate a connection tendency descriptor. The connection tendency descriptor qualitatively describes the degree to which the skill atomic nodes at both ends of the associated edge tend to co-occur in historical scheduling tasks.

5. The method according to any one of claims 1-4, characterized in that, The decoupling transformation process performed on the skill co-occurrence graph, which decomposes the independent embedding vector of each skill atomic node and the implicit constraints of association between the skill atomic nodes, includes: The skill co-occurrence graph is input into a preset graph structure parsing component, which includes a neighbor information aggregation sublayer and a feature decoupling sublayer. By traversing each skill atomic node in the skill co-occurrence graph through the neighbor information aggregation sublayer, the first-order neighbor node set of the skill atomic node is obtained. The connection tendency descriptor of the associated edge between each neighbor skill atomic node and the skill atomic node in the first-order neighbor node set is extracted. The information of the neighbor skill atomic nodes is weighted and aggregated according to the connection tendency descriptor to generate the neighbor aggregation representation of the skill atomic node. The original identifier of the skill atomic node is mapped to a continuous vector space through the neighbor information aggregation sublayer to obtain the skill atomic node's own representation. The skill atomic node's own representation is then concatenated with the neighbor aggregation representation to form the comprehensive initial embedding representation of the skill atomic node. Based on the feature decoupling sublayer, the integrated initial embedding representation of each skill atom node is subjected to independent component enhancement transformation. Intrinsic attribute components that do not depend on the information of other skill atom nodes are extracted from the integrated initial embedding representation, and the intrinsic attribute components are mapped to independent embedding vectors. The independent embedding vectors are used to express the attribute features of the skill atom node itself. Based on the feature decoupling sublayer, interactive attribute components that interact with other skill atomic nodes are extracted from the comprehensive initial embedding representation, and the interactive attribute components are fused with the connection tendency descriptor to obtain the initial association constraint components of the skill atomic node with respect to each associated neighbor. For any two directly connected skill atomic nodes in the skill co-occurrence graph, the initial association constraint components of the skill atomic nodes are combined in pairs based on the feature decoupling sublayer to generate pairwise association implicit constraints between the skill atomic node pairs. The pairwise association implicit constraints limit the co-occurrence of the skill atomic node pairs in the task requirements to the range allowed by the historical co-occurrence pattern. Collect all pairwise implicit constraints of skill atomic node pairs with directly related edges to obtain the set of implicit constraints. The independent embedding vector of each skill atomic node and the set of associated implicit constraints are output as the result of the decoupling transformation process, so as to decompose the independent embedding vector and the associated implicit constraints from the skill co-occurrence graph. The independent embedding vector represents the attribute characteristics of the skill atomic node itself, and the associated implicit constraints represent the cooperative relationship restrictions between different skill atomic nodes based on co-occurrence association.

6. The method according to claim 1, characterized in that, The process involves using the required skill identifier set for each task in the set of tasks to be scheduled and the independent embedding vector to generate an initial task representation for each task. Based on the initial task representation and the associated implicit constraints, an activation mapping subgraph is determined on the skill co-occurrence graph. Topological context information is extracted from the activation mapping subgraph and fused with the initial task representation to generate a task embedding representation, including: For each task to be scheduled in the set of tasks to be scheduled, each skill identifier is taken out from the set of required skill identifiers for the task to be scheduled, and the corresponding independent embedding vector is queried using the correspondence between the skill identifier and the skill atomic node. All the queried independent embedding vectors are arranged to obtain a skill embedding vector sequence. Perform a vector aggregation operation on the skill embedding vector sequence to compress multiple independent embedding vectors in the sequence into a single fixed-dimensional aggregated vector, and use the aggregated vector as the initial task representation of the task to be scheduled; Obtain the set of associated implicit constraints, and perform a compatibility evaluation on the initial task representation and each associated implicit constraint in the set of associated implicit constraints to determine whether the skill co-occurrence pattern implied by the initial task representation conforms to the collaborative relationship constraints defined by each associated implicit constraint; Based on the results of the compatibility evaluation, skill atomic nodes that are compatible with the initial task representation and do not violate the implicit constraints of association are selected, and a core node set is generated by combining the skill atomic nodes directly corresponding to the required skill identifier set of the task to be scheduled. Starting from the skill atomic nodes in the core node set, expand outward along the association edges on the skill co-occurrence graph. Neighboring skill atomic nodes in the graph that are connected to the core nodes through association edges and satisfy the implicit association constraints are included in the extended node set. Combine the core node set, the extended node set, and the association edges of the set to generate an activation mapping subgraph. The activation mapping subgraph is processed to extract topological context information, and the local connection mode, path connectivity structure between nodes, and community clustering features of each node in the activation mapping subgraph are obtained. The local connection mode, path connectivity structure, and community clustering features are combined into topological context information. The topological context information is vectorized and encoded to generate a topological context vector. The topological context vector is then fused with the initial task representation to obtain a task embedding representation that integrates skill collaboration context information. The task embedding representation combines the inherent attributes of the task's skill requirements with the structural context of the skill in the historical collaboration graph.

7. The method according to claim 6, characterized in that, Starting from the skill atomic nodes in the core node set, the process expands outward along the association edges on the skill co-occurrence graph, including neighboring skill atomic nodes in the graph that are connected to the core nodes via association edges and satisfy implicit association constraints. Initialize the extended node set to an empty set, and take the core node set as the current boundary node set. For each boundary node in the current boundary node set, traverse all first-order neighbor nodes of the boundary node on the skill co-occurrence graph. For each first-order neighbor node, obtain the connection tendency descriptor of the associated edge connecting the boundary node and the first-order neighbor node. Based on the set of implicit association constraints, extract the pairwise implicit association constraints corresponding to the boundary node and the first-order neighbor node, and check whether the first-order neighbor node satisfies the cooperative conditions allowed by the pairwise implicit association constraints. Obtain the initial task representation of the task to be scheduled, and calculate the task correlation between the initial task representation and the independent embedding vector corresponding to the first-order neighbor node; If the first-order neighbor node satisfies the pairwise association implicit constraint and the task relevance exceeds the preset activation threshold, the first-order neighbor node is included in the expanded node set and marked as a candidate boundary node for the next round of expansion. After traversing all the boundary nodes in the current boundary node set, the candidate boundary nodes newly added to the expansion node set will form the current boundary node set for the next round of expansion. Repeat the neighbor node traversal, constraint check, correlation calculation and candidate labeling steps until no new candidate boundary nodes are generated after one round of expansion, then stop the expansion process; The union of the core node set and the extended node set is taken as the activation node set, and an activation mapping subgraph is constructed based on the association edges between each node in the activation node set in the skill co-occurrence graph.

8. The method according to claim 1, characterized in that, The scheduling strategy orchestration process based on the task embedding representation and the current schedulable resource status generates a cooperative task scheduling instruction sequence. This sequence includes the mapping relationship between each scheduled task and execution resource, as well as the execution timing. Obtain a list of schedulable execution resources for the current scheduling period, wherein the list of schedulable execution resources includes the resource identifier and resource capability description features of each execution resource; Capability embedding encoding is performed on the resource capability description features of each execution resource to generate a resource capability embedding vector. The resource capability embedding vector is in the same vector space as the task embedding representation. The task embedding representation of each task to be scheduled in the set of tasks to be scheduled is matched with the resource capability embedding vector of each execution resource to calculate the degree of fit, and a task resource fit matrix is ​​generated. The task resource fit matrix is ​​used to represent the capability matching status between each task to be scheduled and each execution resource. Based on the task resource suitability matrix, a heuristic assignment logic is used to determine the candidate execution resource allocation scheme for each task to be scheduled and to generate an initial scheduling assignment mapping. Based on the set of tasks allocated to each execution resource in the initial scheduling assignment mapping, check whether the task load of each execution resource exceeds the resource capacity limit, and perform task transfer adjustment on the scheduled tasks associated with the execution resources that exceed the resource capacity limit to obtain the adjusted scheduling assignment mapping. Obtain the list of tasks to be scheduled for each execution resource in the adjusted scheduling assignment mapping. For each execution resource, determine the execution order of each task to be scheduled on that execution resource based on the skill dependency relationship between tasks and the associated implicit constraints, and generate a resource-level task execution sequence. Globally align the execution sequences of resource-level tasks across all execution resources to eliminate timing conflicts between skill-dependent tasks on different execution resources, resulting in a globally consistent task execution timing scheme. Based on the adjusted scheduling assignment mapping and the globally consistent task execution timing scheme, a collaborative task scheduling instruction sequence is generated. Each instruction in the collaborative task scheduling instruction sequence includes a task identifier to be scheduled, an assigned execution resource identifier, and a task start execution timing marker.

9. A dynamic collaborative task scheduling system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the lightweight intelligent dynamic cooperative task scheduling method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program that, when run on the dynamic collaborative task scheduling system, causes the dynamic collaborative task scheduling system to perform the steps of any of the lightweight intelligent dynamic collaborative task scheduling methods described in claims 1 to 8.

Images