File institution human resource configuration and risk prediction method based on graph neural network

Abstract

The invention discloses an archive institution human resource configuration and risk prediction method based on a graph neural network, and the method comprises the following steps: constructing an organization structure graph containing personnel nodes and post nodes, and fusing the static characteristics of personnel, task records and post information to form a graph edge; extracting personnel static features and task records, generating an adaptation degree change track, calculating an adaptation fatigue slope, and constructing a fatigue slope map; inputting the main graph and the auxiliary graph into an improved DCRNN model, and generating a personnel node initial embedding vector; collecting context information in a task period, constructing a context feature vector, and generating an updated personnel node embedding vector; generating post recommendation and adaptation scores based on the fusion result; and outputting a risk prediction result in combination with the adaptation degree change trajectory and the adaptation fatigue slope. According to the method, the personnel adaptation state can be dynamically described, the configuration risk is accurately evaluated, and the intelligent decision-making level of human resources is improved.

Description

Technical Field

[0001] This invention relates to the field of human resource management technology, and in particular to a method for human resource allocation and risk prediction in archival institutions based on graph neural networks. Background Technology

[0002] As the personnel structure of archival institutions becomes increasingly complex and the job task system dynamically evolves, human resource allocation and risk prediction based on historical configuration relationships and task execution data have become key issues in current personnel management. Existing methods typically rely on structured database statistical indicators or traditional machine learning models to conduct static assessments of job suitability, but these methods have significant limitations when dealing with situations involving multiple overlapping positions, strong temporal dependencies in tasks, and complex evolution patterns of job responsibilities.

[0003] Existing rule-based or shallow model-based configuration methods often fail to fully capture the multidimensional relationships between personnel and positions. They rely solely on current task metrics for suitability assessments, neglecting dynamic factors such as task frequency changes, job connection history, and adaptation evolution trends. Personnel allocation risk assessments are often based on thresholds for accumulated fatigue or job change frequency, lacking in-depth modeling of the hidden configuration path pressures and personnel status change trends within the job connection structure. Some solutions incorporating graph structure analysis only consider static topology or task distribution between positions, lacking the ability to dynamically adjust for embedding evolution directions and job relationship weights, and failing to characterize the impact of key nodes on configuration stability. In actual management scenarios, due to factors such as task category evolution, fatigue accumulation, and job structure reconfiguration during personnel allocation, traditional assessment methods struggle to achieve collaborative modeling of job recommendations and risk prediction. This results in a lack of foresight in job allocation suggestions and a lag in risk identification, making it difficult to support efficient and stable human resource allocation in archival institutions during high-load task cycles.

[0004] Therefore, how to provide a method for human resource allocation and risk prediction in archival institutions based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for human resource allocation and risk prediction in archival institutions based on graph neural networks. This invention fully integrates key steps such as organizational structure graph construction, node embedding vector generation, configuration feature score calculation, and job stability analysis. It describes in detail the entire process of constructing a candidate configuration pair set based on the embedding vectors of personnel nodes and job nodes, calculating the matching degree of job relationship features, the correlation of adaptation evolution trend, and the coupling degree of fatigue trend, and constructing job recommendation results and risk prediction results. It has the advantages of accurate job adaptation scoring, reasonable risk level classification, and high configuration optimization efficiency.

[0006] The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to embodiments of the present invention includes the following steps: Step 1: Collect static characteristic data, task execution records, and job-related information of personnel in the archives to construct an organizational structure graph; the organizational structure graph includes personnel nodes and job nodes, and the configuration relationship between personnel nodes and job nodes is the graph edge; Step 2: Based on the static feature data of personnel nodes and task execution records, construct an adaptation evolution function, calculate the adaptation degree value of the personnel nodes at different time points, and form an adaptation degree change trajectory; Step 3: Based on the time series change trend of the fitness value, calculate the fitness fatigue slope information of the personnel node and construct a fatigue slope map; Step 4: Input the tissue structure map and fatigue slope map as the main map and auxiliary map into the improved DCRNN model, respectively, initialize the main map path, auxiliary map path and fusion path, and generate the initial embedding vector of personnel nodes in the main map path and auxiliary map path; Step 5: Collect contextual information from the archival institution within the preset task period and construct a contextual feature vector; Step 6: Input the context feature vector and the initial embedding vector into the fusion path, perform gating adjustment operation, update the personnel nodes, and generate personnel node embedding vectors; Step 7: Based on the personnel node embedding vector, perform job configuration calculation, output job recommendation results and fit score results; and combine the fit change trajectory and fatigue slope information to perform job stability analysis and risk level prediction, output risk prediction results.

[0007] Preferably, step one specifically comprises: Set data collection cycles and data structure templates, and unify the coding rules for personnel identification, job identification, and task identification; The static feature data includes education category, professional title level, professional qualification, job level, years of service, department, skill tags, and historical adjustment records; the task execution record includes task number, task name, task execution frequency, task load intensity, task duration, task start and end time, and corresponding job information; the job association information includes job number, job name, job responsibility tags, department, job level information, and historical configuration information. Perform data cleaning operations; the data cleaning operations include field standardization, missing field filling, format unification, logical verification, and outlier removal; Construct a set of personnel nodes and a set of job nodes; the set of personnel nodes includes personnel nodes in the graph constructed according to personnel identifiers, and binding static feature data and task execution record vectors to the personnel nodes; the set of job nodes includes job nodes in the graph constructed according to job identifiers, and binding job association information vectors and responsibility tags to the job nodes. Graph edges are generated based on the configuration history relationship between personnel and positions. The graph edges record the configuration start time, configuration end time, configuration status, and cumulative configuration frequency. The personnel node set, position node set, graph edge set, node attribute set, and edge attribute set are combined to form an organizational structure graph.

[0008] Preferably, step two specifically includes: Collect static feature data and task execution records bound to the personnel nodes to construct an information sequence set containing time tags; Static feature data is encoded into static feature vectors. Task execution records are arranged in chronological order. Task frequency change values, load intensity change values ​​and job duration sequences are extracted to form dynamic feature vectors for tasks. The task dynamic feature vector is sliced ​​using a sliding window according to a preset time granularity, and an adaptation evaluation unit is constructed within the time window. In the fit evaluation unit, the matching score between the static feature vector and the corresponding task dynamic features in the window is calculated, and a single-cycle fit value is generated by combining the historical job level and skill tag weight. All single-cycle fit values ​​are connected in chronological order to construct a fit change trajectory; the fit change trajectory is a sequence reflecting the job fit ability of personnel during the time evolution of personnel nodes.

[0009] Preferably, step three specifically includes: A time series window is constructed based on the fitness change trajectory corresponding to the personnel node, and the fitness values ​​within the window are extracted. Perform linear fitting in each time series window to calculate the fitting slope of the fit trend; combine the positive and negative directions of the fitting slope with the magnitude of the absolute value of the slope to generate corresponding fatigue direction labels and fatigue level labels. Bind the corresponding slope value, fatigue direction label and fatigue level label to the personnel node to construct a fatigue slope node attribute set; Based on the personnel node identifiers in the organizational structure diagram, the fatigue slope node attribute set is mapped to the graph structure to construct a fatigue slope graph. The graph nodes in the fatigue slope graph are personnel nodes, and the graph node attributes include slope value, fatigue direction label, fatigue level label, and evolution direction information. The graph edge structure is consistent with the organizational structure diagram.

[0010] Preferably, step four specifically includes: Input the organizational structure map into the main graph path and perform graph structure initialization; input the fatigue slope map into the auxiliary graph path and perform auxiliary graph structure initialization; construct the adjacency matrix, node feature matrix and edge attribute matrix for the main graph path and the auxiliary graph path respectively, and generate the main graph structure tensor and the auxiliary graph structure tensor respectively. In the main graph path, the static feature vector, job relationship feature vector, and configuration frequency feature of the personnel nodes are extracted to construct the main graph input vector; the information propagation operation is performed on the main graph input vector through the diffusion convolution module to generate the initial embedding vector of the main graph. In the auxiliary graph path, the slope value, fatigue direction label, fatigue level label and evolution direction information of the personnel nodes are extracted to construct the auxiliary graph input vector; the feature extraction operation is performed on the auxiliary graph input vector through the diffusion convolution module to generate the initial embedding vector of the auxiliary graph. In the fusion path, the initial embedding vectors of the main graph and the initial embedding vectors of the auxiliary graph are received, and feature splicing operation is performed to construct the fusion input tensor. The fusion input tensor is input to the gated loop unit to extract the fusion state information and generate the initial embedding vector set of personnel nodes. The initial embedding vector set includes the structural association features, adaptation evolution features and fatigue trend features of the personnel nodes under the organizational structure map and fatigue slope map.

[0011] Preferably, the improved DCRNN model is as follows: The improved DCRNN model includes a main graph path, an auxiliary graph path, and a fusion path. Both the main graph path and the auxiliary graph path contain a diffusing convolution module and a gated recurrent unit structure, while the fusion path contains a feature concatenation structure and a fusion gate structure. In the main graph path, the diffusion convolution module receives the main graph structure tensor constructed from the organizational structure graph. Based on the adjacency matrix and node feature matrix of the main graph, it extracts the job relationship features and configuration frequency features of personnel nodes to generate the main graph diffusion representation. The gated recurrent unit performs temporal state modeling on the main graph diffusion representation and outputs the initial embedding vector of the main graph. In the auxiliary graph path, the diffusion convolution module receives the auxiliary graph structure tensor constructed from the fatigue slope map. Based on the adjacency matrix and node feature matrix of the auxiliary graph, it extracts the slope value, fatigue direction label, fatigue level label and evolution direction information of personnel nodes to generate the auxiliary graph diffusion representation. The gated recurrent unit performs temporal modeling on the auxiliary graph diffusion representation and outputs the initial embedding vector of the auxiliary graph. In the fusion path, the feature concatenation structure performs a concatenation operation on the initial embedding vector of the main image and the initial embedding vector of the auxiliary image to construct the fusion input tensor; the fusion gating structure jointly models the fusion input tensor and the context feature vector, performs a gating adjustment operation, and generates a fusion state representation. The fusion state represents the input to the final output layer, completing the generation of the embedding vector of the personnel node; the embedding vector includes job relationship features, adaptation evolution features and fatigue trend features.

[0012] Preferably, step five specifically includes: Set task cycle identification rules, perform time aggregation processing on task execution records, and divide them into continuous task cycle units according to the set cycle granularity; Within the task cycle unit, the number of tasks, the distribution of task categories, and the coverage of job positions are statistically analyzed to construct a set of task distribution features. The task category distribution includes the percentage of tasks categorized by task name and task responsibility tag. The job coverage includes the unique number of job positions associated with the task and the hierarchical distribution of the covered job positions. The task density index is calculated based on the task distribution feature set; the task density index is a weighted combination of the number of tasks and the number of job positions covered per unit time. The number of task-related positions marked as key positions in the task cycle unit is counted, and the key position percentage index is calculated; the key position percentage index is the ratio of the number of key positions to the total number of task-related positions. Generate cycle identifiers, task category distribution vectors, task intensity indicators, and key position percentage indicators, and construct a set of context feature vectors; each context feature vector in the set corresponds to a task cycle unit, containing cycle identification encoding and task feature encoding information; The context feature vector set is time-aligned with the adaptation change trajectory of personnel nodes to form a context input sequence; the context input sequence is synchronized with the main map path and the auxiliary map path.

[0013] Preferably, step six specifically includes: Construct the input tensor of the fusion path, and concatenate the context input sequence with the initial embedding vector of personnel nodes output by the main graph path and the fatigue trend vector of personnel nodes output by the auxiliary graph path in chronological order; Perform gating adjustment operations to construct a multi-dimensional gating factor set; the multi-dimensional gating factor set includes a job relationship gating factor, an adaptation evolution gating factor, and a fatigue trend gating factor; the job relationship gating factor adjusts the embedding weight distribution of the main graph path according to the task density index and job coverage characteristics in the context input sequence; the adaptation evolution gating factor adjusts the evolution direction attention weight of the adaptation evolution feature according to the task category distribution vector in the context input sequence; the fatigue trend gating factor adjusts the slope sensitivity distribution of the fatigue trend feature according to the key job proportion index in the context input sequence. Based on a multidimensional gating factor set, a fusion adjustment function is constructed to perform fusion processing on the main graph embedding vector, the auxiliary graph embedding vector, and the context input sequence to generate an updated personnel node embedding vector. The personnel node embedding vector includes a job relationship feature vector, an adaptation evolution feature vector, and a fatigue trend feature vector. The job relationship feature vector retains the time evolution pattern of the job connection structure, the adaptation evolution feature vector retains the adaptation degree change trajectory and evolution direction information, and the fatigue trend feature vector retains the fatigue slope and context association structure information.

[0014] Preferably, the step of performing job configuration calculation based on personnel node embedding vectors and outputting job recommendation results and adaptation scores specifically involves: Construct a job recommendation input tensor, and combine the personnel node embedding vector with the job association information vector in the job node set in pairs to form a candidate configuration pair set; each candidate configuration pair includes a personnel node embedding vector and a job node association information vector. Perform feature interaction operations on the candidate configuration pair set to calculate the job relationship feature matching degree, adaptation evolution trend correlation, and fatigue trend coupling degree, generating a configuration feature score set. The job relationship feature matching degree is calculated based on the structural similarity between the job relationship feature vector and the job responsibility label. The adaptation evolution trend correlation is calculated based on the evolution path similarity between the adaptation evolution feature vector and the historical configuration state of the job. The fatigue trend coupling degree is calculated based on the trend direction matching degree between the fatigue trend feature vector and the job task intensity index. An adaptation scoring function is constructed based on the configuration feature scoring set, and the adaptation scoring result of each candidate configuration pair is output. The adaptation scoring results are sorted and the job recommendation results are output. The job recommendation results are the set of job nodes with the highest scores corresponding to each personnel node. Record personnel node identifiers, recommended job identifiers, and suitability scores, and generate job recommendation results and suitability scores.

[0015] Preferably, the step of combining the fitness change trajectory and fatigue slope information to perform job stability analysis and risk level prediction, and outputting risk prediction results, specifically includes: The adaptation evolution feature vector and fatigue trend feature vector are extracted from the personnel node embedding vector to construct the risk analysis input tensor. The adaptation evolution feature vector contains information on the fitness change trajectory and evolution direction, while the fatigue trend feature vector contains information on fatigue slope and contextual relationship structure. Perform time series analysis on the fit change trajectory, calculate the consistency index of change amplitude, change frequency and direction, and generate a set of fit fluctuation characteristics; perform slope trend analysis on the fatigue slope, calculate the trend direction, fluctuation degree and correlation ratio with key positions, and generate a set of fatigue response characteristics. A job stability scoring function is constructed based on the set of adaptation fluctuation features and the set of fatigue response features, and the job stability scoring result is output. The job stability scoring result represents the degree of stability of the job adaptation status of personnel nodes in the current task cycle. Establish risk level prediction rules, and classify low-risk, medium-risk, and high-risk levels based on the job stability score and the combined threshold conditions of adaptation fluctuation and fatigue response indicators. Record personnel node identifiers, job identifiers, stability scores, and risk levels, and output risk prediction results.

[0016] The beneficial effects of this invention are: This invention addresses the challenges of complex job configuration relationships, scattered personnel suitability information, and difficulty in identifying risk trends in archival institutions by constructing an organizational structure graph and modeling the relationships between personnel and job nodes. It establishes a data collection cycle and structural template, constructs sets of personnel and job nodes, generates a graph structure, and integrates static feature data and task execution records to create a complete organizational structure graph. In the embedding vector generation stage, an improved DCRNN model is introduced to represent personnel and job nodes using vectors, preserving structural connectivity features and dynamic task execution features, and constructing a unified node embedding representation. In the job configuration recommendation stage, a set of candidate configuration pairs is generated based on the embedding vectors. Combining job association label features, task execution features, and configuration history features, the matching degree of job relationship features is calculated. Information on the trajectory and evolution direction of suitability changes is constructed to assess the correlation of suitability evolution trends. The coupling degree of fatigue trends is calculated by integrating job adjustment frequency and task load change features, and a set of configuration feature scores is output. Finally, combining the suitability change trajectory and fatigue slope information, job stability analysis and risk level prediction operations are performed to construct risk prediction results. This invention enables the reasonable generation of job recommendation results and the effective output of risk prediction results, ensuring the scientific and stable allocation of human resources in archival institutions. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 The flowchart shows the method for human resource allocation and risk prediction in archival institutions based on graph neural networks proposed in this invention. Figure 2 This is a schematic diagram of the improved DCRNN model proposed in this invention; Figure 3 This is a data flow diagram of the method for human resource allocation and risk prediction in archival institutions based on graph neural networks proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1-3 A graph neural network-based method for human resource allocation and risk prediction in archival institutions includes the following steps: Step 1: Collect static characteristic data, task execution records, and job-related information of personnel in the archives to construct an organizational structure graph; the organizational structure graph includes personnel nodes and job nodes, and the configuration relationship between personnel nodes and job nodes is graph edges; the task execution records include task execution frequency, task load intensity, and job duration; Step 2: Based on the static feature data of personnel nodes and task execution records, construct an adaptation evolution function, calculate the adaptation degree value of the personnel nodes at different time points, and form an adaptation degree change trajectory; Step 3: Based on the time series change trend of the fitness value, calculate the fitness fatigue slope information of the personnel node and construct a fatigue slope map; the fatigue slope map includes the slope value and evolution direction information of the personnel node; Step 4: Input the tissue structure map and fatigue slope map as the main map and auxiliary map into the improved DCRNN model, respectively, initialize the main map path, auxiliary map path and fusion path, and generate the initial embedding vector of personnel nodes in the main map path and auxiliary map path; Step 5: Collect contextual information of the archives within the preset task cycle and construct a contextual feature vector; the contextual information includes cycle identifier, task type distribution, task intensity index and key position ratio index; Step 6: Input the context feature vector and the initial embedding vector into the fusion path, perform gating adjustment operation, update personnel nodes, and generate personnel node embedding vectors; the personnel node embedding vectors include job relationship features, adaptation evolution features, and fatigue trend features; Step 7: Based on the personnel node embedding vector, perform job configuration calculation, output job recommendation results and fit score results; and combine the fit change trajectory and fatigue slope information to perform job stability analysis and risk level prediction, output risk prediction results.

[0021] This implementation method constructs an organizational structure graph by collecting static feature data, task execution records, and job-related information of personnel in archival institutions. This graph clarifies the configuration relationships between personnel nodes and job nodes, ensuring the integrity of the graph structure and the traceability of configuration relationships. Based on the static feature data of personnel nodes and task execution records, an adaptation evolution function is constructed to calculate the fitness degree value and form a fitness degree change trajectory, achieving time-series modeling of the dynamic changes in personnel job adaptation. Furthermore, based on the time-series change trend of the fitness degree value, fitness fatigue slope information is calculated, and a fatigue slope graph is constructed to characterize the fitness fatigue trend and evolution direction during the personnel job allocation process. The organizational structure graph and fatigue slope graph are input as the main graph and auxiliary graph, respectively, into the improved DCRNN model to initialize the main graph path, auxiliary graph path, and fusion path, generating initial embedding vectors for personnel nodes. Strengthen the joint expression capability of structural connections and trend evolution; further, collect contextual information of archival institutions within a preset task cycle, construct contextual feature vectors to reflect the external influencing factors of task background and configuration environment; simultaneously, input the contextual feature vectors and initial embedding vectors into the fusion path, perform gating adjustment operations, update personnel node status, generate personnel node embedding vectors, and integrate job relationship features, adaptation evolution features, and fatigue trend features to improve the comprehensive accuracy of personnel node expression; based on personnel node embedding vectors, perform job configuration calculation operations, output job recommendation results and adaptation score results, further combine adaptation change trajectory and fatigue slope information, perform job stability analysis and risk level prediction operations, output risk prediction results, and realize reasonable recommendation of job configuration schemes and forward-looking judgment of risk prediction results in archival institutions.

[0022] In this embodiment, step one specifically includes: Set data collection cycles and data structure templates, and unify the coding rules for personnel identifiers, job identifiers, and task identifiers to ensure the consistency of identifiers and the structural stability of data fusion during the map construction process; Collect static characteristic data of personnel in the archives, including education category, professional title level, professional qualification, job level, years of service, department, skill tags and historical adjustment records, covering information on changes in professional background and job ability of personnel in the long-term development process; The task execution records of the personnel are collected. The task execution records include task number, task name, task execution frequency, task load intensity, task duration, task start and end time and corresponding job information, which reflect the actual load and frequency of performance of personnel under different task conditions. Collect job-related information, including job number, job name, job responsibility tags, department to which the job belongs, job level information and historical configuration information, to clarify the responsibility structure and organizational level positioning of each job; Perform data cleaning operations, including field standardization, missing field filling, format unification, logical verification and outlier removal, to improve the quality of structured data and ensure the integrity of the basic data for map construction; Based on the cleaned personnel identification data, a set of personnel nodes is constructed. A corresponding node in the graph is created according to each personnel identification, and each personnel node is bound to its static feature data and task execution record vector to complete the initialization of personnel node attributes. Based on the cleaned job identifier data, a set of job nodes is constructed. Job nodes are established in the graph according to each type of job identifier, and job association information vectors and responsibility tags are bound to the job nodes to form a job node attribute structure. Based on historical task execution records and job configuration data, a graph edge relationship is constructed between personnel nodes and job nodes. The graph edge records the configuration start time, configuration end time, configuration status, and cumulative configuration frequency. The configuration status identifier indicates whether it is a current on-the-job relationship, and the configuration frequency records the historical job rotation situation. By combining the set of personnel nodes, the set of job nodes, the set of graph edges, the set of node attributes, and the set of edge attributes, an organizational structure graph is constructed. The organizational structure graph represents the relationship structure between personnel and jobs within the archives institution in the form of a heterogeneous graph.

[0023] In this embodiment, step two specifically includes: The static feature data bound to the personnel node and the task execution record are collected to construct an information sequence set containing time tags. The time tags are automatically generated based on the task start and end time and the sampling time granularity to ensure the consistency of the feature time sequence structure. Encoding operations are performed on static feature data to uniformly convert educational category, professional title level, professional qualification, job level, years of service, department, skill tags and historical adjustment records into static feature vectors. The static feature vectors maintain consistent dimensions and support subsequent fusion operations with dynamic features. The task execution records are sorted in chronological order, and the task frequency change value, load intensity change value and job duration data are extracted within each time segment. Standardization and normalization operations are used to construct a task dynamic feature vector, which reflects the task behavior change pattern of personnel in continuous time periods. The task dynamic feature vector is sliced ​​using a sliding window according to the set time granularity and sliding interval. Each window contains a fixed number of time steps and dynamic feature items, and an independent fitness evaluation unit is constructed in each time window. In the fit evaluation unit, the static feature vector is matched with the dynamic features in the current window. The weighted matching function is used to score the semantic similarity of professional qualifications and skill tags. Combined with the historical job level weight allocation strategy, a static and dynamic feature fusion mapping model is constructed. The single-cycle fit value is obtained by fitting the model. All single-cycle fit values ​​generated within the time window are concatenated in chronological order to form a complete fit change trajectory sequence. The fit change trajectory sequence uses time step as the horizontal axis and matching score as the vertical axis to represent the dynamic change in the fit ability between personnel nodes and their current positions during the evolution process.

[0024] This implementation constructs a time-series fit evaluation mechanism that combines static and dynamic features, effectively reflecting the matching trend of personnel nodes between different stages and job requirements, and possessing the technical effects of strong timeliness and accurate personalized expression.

[0025] In this embodiment, step three specifically includes: A fixed-length time series sliding window is constructed based on the fitness change trajectory corresponding to personnel nodes. The window length and sliding step size are set by system parameters. The fitness value sequence of continuous time steps is extracted in each time series window as the data input for fatigue trend analysis. A linear fitting operation is performed within each time series window. The least squares method is used to fit the trend of the fitness change sequence. The trend slope value within the current window is obtained by fitting a linear function. The slope value is used to characterize the rate of increase or decrease of fitness change. The trend type is determined by the positive or negative direction of the slope value. A positive value indicates an adaptation enhancement trend, while a negative value indicates an adaptation decline trend. Different trend intensity ranges are divided by combining the absolute value of the slope value, and fatigue direction labels and fatigue level labels are generated. The fatigue direction label marks the evolution direction of adaptation change, and the fatigue level label represents the adaptation decline rate level at the current stage. The slope value, fatigue direction label and fatigue level label are bound to the personnel node to construct a fatigue slope node attribute set. The node attribute set is indexed by the node identifier and contains trend slope and label information corresponding to multiple time windows. The fatigue slope node attribute set is mapped to the graph structure according to the personnel node identifier in the organizational structure graph, so that each personnel node has a trend evolution attribute, and a fatigue slope graph is constructed. The graph node is a personnel node, and the graph node attributes include trend slope value, fatigue direction label, fatigue level label and evolution direction information. The graph edge structure inherits the connection relationship in the organizational structure graph and maintains the consistency of the configuration path.

[0026] This implementation method realizes trend modeling based on fitness evolution sequence, constructs graph structure data integrating fatigue trend information, provides trend prior support for subsequent configuration risk prediction and scheduling strategy optimization based on graph neural network, and has the technical effects of accurate trend recognition and rich node expression.

[0027] In this embodiment, step four specifically includes: The organizational structure graph is input as the main graph structure into the main graph path. The graph structure initialization operation is performed, which includes constructing the adjacency relationship matrix between personnel nodes and job nodes, extracting node attributes to construct the node feature matrix and graph edge attribute matrix, and generating the main graph structure tensor. The fatigue slope map is input as an auxiliary graph structure into the auxiliary graph path. The auxiliary graph structure initialization operation is performed, including constructing an adjacency matrix based on the connection relationship between personnel nodes, extracting slope values ​​and fatigue labels to construct a node feature matrix, and generating an edge attribute matrix by combining evolution direction information to generate an auxiliary graph structure tensor. In the main map path, the static feature vector, job relationship feature vector and historical configuration frequency feature corresponding to the personnel node are extracted and concatenated to form the main map input vector set. The job relationship feature vector is obtained by encoding the node job level and responsibility label, and the configuration frequency feature is obtained by fitting the number of historical tasks and periodic distribution. The diffusion convolution module is invoked to perform information propagation operation on the main graph input vector set. The diffusion convolution module performs multi-hop neighbor information aggregation based on the adjacency matrix and edge attribute matrix in the graph, controls the length of the information diffusion path through the structure preservation mechanism, and outputs the initial embedding vector of the main graph, which represents the task history relationship and job feature expression of personnel nodes under the organizational structure. In the auxiliary graph path, the slope value, fatigue direction label, fatigue level label and evolution direction information of personnel nodes are extracted to construct an auxiliary graph input vector set. The fatigue label information is embedded into the feature vector in a one-hot encoding manner, and the evolution direction information is represented by a directional weight matrix to indicate the direction of change. The diffusion convolution module is invoked to perform feature extraction on the set of input vectors of the auxiliary graph. Based on the topological coupling relationship between fatigue trend labels, the structural features of nodes in the fatigue evolution path are extracted, and the initial embedding vector of the auxiliary graph is output, which represents the temporal feature distribution of personnel nodes in the fatigue trend evolution process. In the fusion path, the initial embedding vector of the main graph and the initial embedding vector of the auxiliary graph are received, a feature concatenation operation is performed, and a fusion input tensor is constructed. The concatenation operation retains the original information of the two graph paths and provides a basis for the combined representation. The fusion input tensor is input into the gated recurrent unit, and the multi-path features are fused using the memory state and forgetting mechanism. The fusion state information is extracted to generate a set of initial embedding vectors for personnel nodes that are structurally consistent and time-sensitive. The initial set of embedded vectors fully expresses the structural association features, adaptation evolution features, and fatigue trend features of personnel nodes under the organizational structure map and fatigue slope map.

[0028] This implementation method achieves multi-graph fusion and joint feature encoding operations, providing highly expressive node representations for downstream risk identification and dynamic configuration strategy generation, and possesses the technical effect of comprehensive integration of structural information and deep fusion of trend features.

[0029] In this embodiment, the improved DCRNN model is specifically as follows: The improved DCRNN model includes a main graph path, an auxiliary graph path, and a fusion path. Both the main graph path and the auxiliary graph path contain a diffusing convolution module and a gated recurrent unit structure. The fusion path contains a feature concatenation structure and a fusion gate structure. In the main graph path, the diffusing convolution module receives the main graph structure tensor constructed from the organizational structure graph. Based on the adjacency matrix and node feature matrix of the main graph, the job relationship features and configuration frequency features of personnel nodes are extracted to construct the main graph input vector. The diffusing convolution operation is performed on the main graph input vector to generate the main graph diffusing representation. The main graph diffusing representation is input to the gated recurrent unit to perform temporal state modeling operations and output the initial embedding vector of the main graph. In the auxiliary graph path, the diffusion convolution module receives the auxiliary graph structure tensor constructed from the fatigue slope map; based on the adjacency matrix and node feature matrix of the auxiliary graph, it extracts the slope value, fatigue direction label, fatigue level label and evolution direction information of personnel nodes to construct the auxiliary graph input vector; it performs a diffusion convolution operation on the auxiliary graph input vector to generate the auxiliary graph diffusion representation; it inputs the auxiliary graph diffusion representation into the gated recurrent unit to perform temporal modeling operation and outputs the initial embedding vector of the auxiliary graph. In the fusion path, the feature splicing structure performs a splicing operation on the initial embedding vectors of the main graph and the initial embedding vectors of the auxiliary graph to construct a fusion input tensor; the fusion gating structure receives the fusion input tensor and the context feature vector, performs gating adjustment operations, extracts key fusion features, and generates a fusion state representation; the fusion state representation is input to the final output layer to complete the generation of the embedding vectors of personnel nodes; the embedding vectors include job relationship features, adaptation evolution features, and fatigue trend features, and have the ability to express multi-dimensional structural information and time-series dynamic modeling capabilities, which helps to achieve accurate prediction of personnel allocation risks and effective capture of evolution trends, thereby improving the scientific and forward-looking nature of human resource allocation.

[0030] In this embodiment, step five specifically includes: Set task cycle identification rules; perform time aggregation processing based on task start and end time information in task execution records; divide the task time axis according to the set cycle granularity to form a continuous sequence of task cycle units; Within each task cycle unit, the number of tasks within the cycle is counted, task names and task responsibility tags are extracted, the percentage of tasks by category is calculated, and a task category distribution vector is generated. The associated job numbers of all tasks in each task cycle unit are extracted, the number of deduplicated jobs is counted, and the job coverage number is obtained. The hierarchical information of the covered jobs is statistically analyzed to generate a job coverage hierarchy distribution vector, and a job coverage feature is constructed. The task number, task category distribution vector, and job coverage feature are combined to generate a task distribution feature set. The task density index is calculated based on the task distribution feature set; the task density index is a weighted combination of the number of tasks per unit time and the number of job positions covered; the weighted combination coefficient is obtained by fitting the historical influence factors of tasks. In each task cycle unit, determine whether a position is a key position based on its job number and job level information; count the number of key positions and compare it with the ratio of the total number of positions associated with the total tasks in the cycle to calculate the key position percentage index. Generate a cycle identifier code based on the cycle number; combine the task category distribution vector, task density index and key position ratio index to form the context feature vector corresponding to each cycle unit; the context feature vector contains cycle identification code and task feature code information to generate a context feature vector set. Perform time alignment operation on the personnel node adaptability change trajectory; match the context feature vector set with the personnel adaptability change trajectory one by one according to the time boundary of the task cycle unit; construct the context input sequence; the context input sequence serves as the auxiliary input in the fusion path, maintaining temporal consistency and node alignment with the main graph path and the auxiliary graph path.

[0031] This implementation method can achieve synchronous modeling of task characteristics and personnel adaptation status under multiple cycles, enhance the model's ability to understand the background information of configuration evolution, and improve the temporal integrity of node embedding expression and the accuracy of configuration prediction.

[0032] In this embodiment, step six specifically includes: Construct the fusion path input tensor; align and concatenate the context input sequence with the initial embedding vector of personnel nodes output from the main graph path in terms of time dimension; concatenate the fatigue trend vector of personnel nodes output from the auxiliary graph path to the corresponding time segment in the same time order; complete the construction of the fusion path input tensor, which includes three substructures: main graph embedding structure, auxiliary graph embedding structure, and context structure; A gating adjustment operation is performed to construct a multi-dimensional gating factor set. This set includes a job relationship gating factor, an adaptation evolution gating factor, and a fatigue trend gating factor. The job relationship gating factor is calculated based on the task density index and job coverage characteristics of each period in the context input sequence. An attention mechanism is used to extract the connection strength and temporal change patterns between jobs, generating a job relationship adjustment weight distribution. The adaptation evolution gating factor is calculated based on the task category distribution vector. The importance distribution of configuration evolution direction features is identified through the changing trends of the task category dimension, generating adaptation evolution attention weights. The fatigue trend gating factor is calculated based on the key job percentage index, constructing a fatigue trend weight adjustment structure to improve the sensitivity of the key job percentage to changes in the fatigue slope. A fusion adjustment function is constructed. This function receives three types of input data: the main image embedding vector, the auxiliary image embedding vector, and the context input sequence. The fusion adjustment function adjusts the structural weights of the main image embedding vector according to the job relationship gating factor, adjusts the directional sensitive region of the adaptation evolution feature according to the adaptation evolution gating factor, and adjusts the gradient response of the fatigue trend feature according to the fatigue trend gating factor. The fusion adjustment function performs a dynamic weighting operation on the three types of inputs to generate an updated personnel node embedding vector. The updated personnel node embedding vector consists of a job relationship feature vector, an adaptation evolution feature vector, and a fatigue trend feature vector. The job relationship feature vector represents the evolution pattern of the connection structure and relationship strength between personnel and jobs under different task cycles. The adaptation evolution feature vector represents the trend continuation direction and evolution fluctuation structure of key nodes in the personnel adaptation change trajectory, fully preserving the evolution direction information. The fatigue trend feature vector represents the interaction structure between the cumulative slope change of the fatigue trend and contextual information, reflecting the adjustment path of the job load background on the personnel state.

[0033] This implementation method can achieve the joint expression of the main image embedding structure, the auxiliary image trend structure and the task context background. The fusion process has the ability to adjust for task dependence and enhance the sensitivity of state evolution, effectively improving the task matching accuracy of personnel embedding expression and the dynamic adaptability of risk prediction.

[0034] In this embodiment, the step of performing job configuration calculation based on personnel node embedding vectors and outputting job recommendation results and adaptation scores specifically includes: Construct a job recommendation input tensor; combine the updated personnel node embedding vector with the job association information vector in the job node set in pairs to form a candidate configuration pair set; each pair of data in the candidate configuration pair set includes a set of personnel node embedding vectors and a set of job node association information vectors; the job node association information vectors include job responsibility tags, job hierarchical structure, historical configuration frequency and task intensity distribution indicators; process them uniformly according to the embedding dimension and concatenate them to form the recommendation input tensor. Perform feature interaction operations on the candidate configuration pair set; construct a job relationship feature matching degree index based on the structural similarity between the job relationship feature vector and the job responsibility label, which represents the degree of structural fusion between the personnel's job experience embedded features and the current job responsibility structure; construct an adaptation evolution trend correlation index based on the evolution path similarity between the adaptation evolution feature vector and the job's historical configuration state sequence, which reflects the degree of consistency between the personnel's past adaptation trajectory and the job configuration change path; construct a fatigue trend coupling degree index based on the trend direction matching degree between the fatigue trend feature vector and the job task intensity index, which measures the degree of consistency between the fatigue slope change trend and the job task intensity time distribution. The matching degree of job relationship features, the correlation of adaptation evolution trend, and the coupling degree of fatigue trend are integrated to construct a configuration feature score set. The configuration feature score set is fitted using a score fusion model to construct an adaptation score function. The adaptation score function is generated through a multi-dimensional feature weighting mechanism, and each weight parameter is obtained by supervised training using historical configuration sample data. The adaptation score function outputs the adaptation score result for each candidate configuration pair. The score result reflects the degree of configuration adaptation between the current personnel node and the target job node in the current period. Sort the adaptation scores of all candidate configuration pairs; sort them from high to low scores, filter the set of job nodes with the highest scores for each personnel node, and output the job recommendation results; in the job recommendation results, each personnel node corresponds to only one set of job nodes with the best scores, which constitute the final recommendation pair; Record the personnel node identifier, recommended job identifier, and corresponding matching score for each recommendation pair; combine the ranking results to generate a job recommendation result table and a matching score result table. The recommendation result table provides the personnel-job correspondence, and the score result table provides the detailed score structure of the corresponding configuration pair.

[0035] This implementation method introduces a three-dimensional index system of job relationship feature matching degree, adaptation evolution trend correlation and fatigue trend coupling degree to achieve a deep-level assessment of the adaptability between personnel and jobs. The constructed adaptation scoring function can integrate structural similarity, evolution consistency and trend coupling to ensure that the recommendation results have contextual consistency and job matching rationality, thereby significantly improving recommendation accuracy and job allocation efficiency.

[0036] In this embodiment, the step of combining the fitness change trajectory and fatigue slope information to perform job stability analysis and risk level prediction, and outputting risk prediction results, specifically involves: Extract the adaptation evolution feature vector and fatigue trend feature vector from the personnel node embedding vector; the adaptation evolution feature vector contains the adaptation degree change trajectory, change amplitude sequence and evolution direction encoding information within the historical cycle; the fatigue trend feature vector contains the fatigue slope value sequence, fatigue trend fluctuation pattern and context task temporal structure information within the continuous cycle; concatenate the above two types of features to form the risk analysis input tensor; Perform time series analysis on the trajectory of fit change; calculate the amplitude index, frequency index, and direction consistency index of fit change within a continuous period based on a sliding window, which respectively represent the amplitude intensity of drastic changes in fit value, the number of fluctuations per unit time, and the degree of consistency with the trend direction of neighboring periods; construct a set of fit fluctuation characteristics to reflect the continuity and stability of personnel job fit status. Perform slope trend analysis on the fatigue slope value sequence; calculate the dominant direction index, fluctuation index and key post task intensity correlation ratio index of fatigue trend; the dominant direction index represents the overall trend direction of fatigue trend, the fluctuation index represents the standard deviation of fatigue trend slope change, and the key post task intensity correlation ratio represents the normalized proportion of current post task intensity in high-intensity posts; construct a fatigue response feature set to reflect the coupling risk between personnel state stability and post task intensity. A job stability scoring function is constructed based on the set of adaptation fluctuation features and the set of fatigue response features. The job stability scoring function is obtained by fitting a feature weighted fusion model. The fusion model uses historical stable configuration samples as the training set and adopts supervised learning to obtain weight parameters and nonlinear transformation paths. The job stability scoring results are used to quantify the continuity of job adaptation and the degree of exposure to state risks of personnel in the current task cycle. Establish risk level prediction rules; define the segmented range of job stability scores and the joint threshold of adaptive fluctuation range and fatigue response indicators; mark configurations with scores below the first threshold and high-frequency changes in fluctuation or fatigue characteristics as high-risk; mark configurations with scores in the middle range or only a single fluctuation risk as medium-risk; mark configurations with scores above the second threshold and stable characteristics as low-risk; construct a risk level prediction mapping function to realize the mapping operation from scores and indicator sets to risk level labels; Record the personnel node identifier, job identifier, job stability score and corresponding risk level label for each configuration; construct a risk prediction result table to display the potential stability risks and allocation priorities of personnel configuration in the organizational structure diagram.

[0037] This implementation method introduces multidimensional time series analysis and fatigue trend structure modeling, and combines job fit continuity and state fluctuation characteristics to construct a job stability scoring function and risk level prediction mechanism. This effectively realizes dynamic stability assessment of job matching status and identification of high-risk configurations, thereby improving the foresight and security of human resource allocation.

[0038] Example 1: To verify the feasibility of this invention in practice, it was applied to the human resource allocation system of a comprehensive archives management service center. This institution has long undertaken multiple high-density tasks, including digitizing paper archives, classifying electronic archives, compiling and researching documents, and providing information retrieval services. Due to the complex personnel structure and frequent changes in task assignments, there are prominent problems such as unreasonable job allocation, uneven personnel workload, and high job turnover, which affect overall work efficiency and management stability. The management department urgently needs an intelligent method that can dynamically identify the suitability of personnel and positions, provide early warnings of the risk of personnel-job mismatch, and improve the scientific nature of allocation.

[0039] In this scenario, the system first collects static characteristic data of all employees, including education level, professional title, length of service, department, professional qualifications, and job level. Combined with task execution records and job-related information from the past year, it constructs an organizational structure graph based on personnel and job nodes. The graph edges detail the start and end times, configuration frequency, and task load of each job configuration relationship, providing a high-quality structural foundation for subsequent dynamic modeling.

[0040] Based on this graph, the system introduces a time-series modeling mechanism to construct an adaptation trajectory based on changes in the task history of personnel nodes, and calculates fatigue slope information to dynamically quantify the long-term stability of each employee in their position. Subsequently, the structural graph and fatigue graph are simultaneously input into an improved DCRNN model. Structural configuration information is extracted from the main graph path, fatigue trends are modeled in the auxiliary graph path, and contextual factors within the task cycle, such as major tasks and pre-holiday peak periods, are integrated to jointly generate node embedding vectors reflecting personnel job adaptability.

[0041] In actual operation, the system performs job allocation recommendation and risk prediction tasks on a monthly cycle, and outputs a job recommendation list, suitability score, and job risk level results. To verify the improvement effect brought by the system, a three-month comparative experiment was conducted with the original manual allocation mechanism, focusing on four indicators: job suitability, personnel workload balance, job turnover rate, and high-risk job identification rate.

[0042] Table 1 Comparison of Personnel and Job Allocation Effects

[0043] As shown in Table 1, regarding job fit, the method of this invention, through the introduction of multi-source information modeling and dynamic graph analysis, significantly improves the matching degree between job recommendations and actual job capabilities. The average fit over three months increased from 0.73 to 0.87, an improvement of nearly 20%. The standard deviation of personnel workload decreased from 0.36 to 0.21, indicating a more balanced task load and avoiding the problem of some personnel being overworked while others have sparse tasks. In terms of job turnover, it decreased from 14.2% to 9.6%, effectively alleviating the problem of job instability. Meanwhile, the high-risk job identification rate increased significantly from 62.4% to 91.7%, demonstrating that the model has strong sensitivity and foresight in identifying personnel fatigue trends and job mismatches, helping management departments to intervene early and rationally adjust job allocation.

[0044] Further analysis was conducted on the effectiveness of the job risk level assessment, and the job categories were labeled and compared as A (high risk), B (medium risk), and C (low risk).

[0045] Table 2 Comparison of Risk Level Prediction Results

[0046] Table 2 shows that, across 80 job positions within a three-month data period, the accuracy of this invention's model in identifying high-risk positions significantly outperforms the original manual experience-based judgment mechanism. Particularly in Category A high-risk positions, manual judgment correctly identified 13 positions, while this invention's method correctly identified 22, increasing the accuracy from 52% to 88%. A steady improvement was also seen in identifying medium- and low-risk positions, with accuracy increasing by approximately 16% in each case, indicating that the improved DCRNN model has high reliability in modeling factors such as fatigue trends, job load, and contextual influences. This proactive judgment capability greatly enhances the timeliness and scientific rigor of job risk warnings, avoiding problems such as low personnel efficiency, job shortages, and task delays caused by mismatches between personnel and positions.

[0047] This embodiment fully demonstrates the practical application effectiveness of the present invention in the dynamic allocation of human resources. It not only improves the quality of job matching and personnel utilization efficiency, but also provides management with a scientific and reliable basis for risk identification. Through the deep integration of graph neural networks and organizational graphs, it provides a brand-new solution for intelligent human resource management in complex organizational systems.

[0048] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for human resource allocation and risk prediction in archival institutions based on graph neural networks, characterized in that, Includes the following steps: Step 1: Collect static characteristic data, task execution records, and job-related information of personnel in the archives to construct an organizational structure graph; the organizational structure graph includes personnel nodes and job nodes, and the configuration relationship between personnel nodes and job nodes is the graph edge; Step 2: Based on the static feature data of personnel nodes and task execution records, construct an adaptation evolution function, calculate the adaptation degree value of the personnel nodes at different time points, and form an adaptation degree change trajectory; Step 3: Based on the time series change trend of the fitness value, calculate the fitness fatigue slope information of the personnel node and construct a fatigue slope map; Step 4: Input the tissue structure map and fatigue slope map as the main map and auxiliary map into the improved DCRNN model, respectively, initialize the main map path, auxiliary map path and fusion path, and generate the initial embedding vector of personnel nodes in the main map path and auxiliary map path; Step 5: Collect contextual information from the archival institution within the preset task period and construct a contextual feature vector; Step 6: Input the context feature vector and the initial embedding vector into the fusion path, perform gating adjustment operation, update the personnel nodes, and generate personnel node embedding vectors; Step 7: Based on the personnel node embedding vector, perform job configuration calculation, output job recommendation results and fit score results; and combine the fit change trajectory and fatigue slope information to perform job stability analysis and risk level prediction, output risk prediction results.

2. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 1, characterized in that, Step one specifically involves: Set data collection cycles and data structure templates, and unify the coding rules for personnel identification, job identification, and task identification; The static feature data includes education category, professional title level, professional qualification, job level, years of service, department, skill tags, and historical adjustment records; the task execution record includes task number, task name, task execution frequency, task load intensity, task duration, task start and end time, and corresponding job information; the job association information includes job number, job name, job responsibility tags, department, job level information, and historical configuration information. Perform data cleaning operations; the data cleaning operations include field standardization, missing field filling, format unification, logical verification, and outlier removal; Construct a set of personnel nodes and a set of job nodes; the set of personnel nodes includes personnel nodes in the graph constructed according to personnel identifiers, and binding static feature data and task execution record vectors to the personnel nodes; the set of job nodes includes job nodes in the graph constructed according to job identifiers, and binding job association information vectors and responsibility tags to the job nodes. Graph edges are generated based on the configuration history relationship between personnel and positions. The graph edges record the configuration start time, configuration end time, configuration status, and cumulative configuration frequency. The personnel node set, position node set, graph edge set, node attribute set, and edge attribute set are combined to form an organizational structure graph.

3. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 2, characterized in that, Step two specifically involves: Collect static feature data and task execution records bound to the personnel nodes to construct an information sequence set containing time tags; Static feature data is encoded into static feature vectors. Task execution records are arranged in chronological order. Task frequency change values, load intensity change values ​​and job duration sequences are extracted to form dynamic feature vectors for tasks. The task dynamic feature vector is sliced ​​using a sliding window according to a preset time granularity, and an adaptation evaluation unit is constructed within the time window. In the fit evaluation unit, the matching score between the static feature vector and the corresponding task dynamic features in the window is calculated, and a single-cycle fit value is generated by combining the historical job level and skill tag weight. All single-cycle fit values ​​are connected in chronological order to construct a fit change trajectory; the fit change trajectory is a sequence reflecting the job fit ability of personnel during the time evolution of personnel nodes.

4. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 3, characterized in that, Step three specifically involves: A time series window is constructed based on the fitness change trajectory corresponding to the personnel node, and the fitness values ​​within the window are extracted. Perform linear fitting in each time series window to calculate the fitting slope of the fit trend; combine the positive and negative directions of the fitting slope with the magnitude of the absolute value of the slope to generate corresponding fatigue direction labels and fatigue level labels. Bind the corresponding slope value, fatigue direction label and fatigue level label to the personnel node to construct a fatigue slope node attribute set; Based on the personnel node identifiers in the organizational structure diagram, the fatigue slope node attribute set is mapped to the graph structure to construct a fatigue slope graph. The graph nodes in the fatigue slope graph are personnel nodes, and the graph node attributes include slope value, fatigue direction label, fatigue level label, and evolution direction information. The graph edge structure is consistent with the organizational structure diagram.

5. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 4, characterized in that, Step four specifically involves: Input the organizational structure map into the main graph path and perform graph structure initialization; input the fatigue slope map into the auxiliary graph path and perform auxiliary graph structure initialization; construct the adjacency matrix, node feature matrix and edge attribute matrix for the main graph path and the auxiliary graph path respectively, and generate the main graph structure tensor and the auxiliary graph structure tensor respectively. In the main graph path, the static feature vector, job relationship feature vector, and configuration frequency feature of the personnel nodes are extracted to construct the main graph input vector; the information propagation operation is performed on the main graph input vector through the diffusion convolution module to generate the initial embedding vector of the main graph. In the auxiliary graph path, the slope value, fatigue direction label, fatigue level label and evolution direction information of the personnel nodes are extracted to construct the auxiliary graph input vector; the feature extraction operation is performed on the auxiliary graph input vector through the diffusion convolution module to generate the initial embedding vector of the auxiliary graph. In the fusion path, the initial embedding vectors of the main graph and the initial embedding vectors of the auxiliary graph are received, and feature splicing operation is performed to construct the fusion input tensor. The fusion input tensor is input to the gated loop unit to extract the fusion state information and generate the initial embedding vector set of personnel nodes. The initial embedding vector set includes the structural association features, adaptation evolution features and fatigue trend features of the personnel nodes under the organizational structure map and fatigue slope map.

6. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 5, characterized in that, The improved DCRNN model is specifically as follows: The improved DCRNN model includes a main graph path, an auxiliary graph path, and a fusion path. Both the main graph path and the auxiliary graph path contain a diffusing convolution module and a gated recurrent unit structure, while the fusion path contains a feature concatenation structure and a fusion gate structure. In the main graph path, the diffusion convolution module receives the main graph structure tensor constructed from the organizational structure graph. Based on the adjacency matrix and node feature matrix of the main graph, it extracts the job relationship features and configuration frequency features of personnel nodes to generate the main graph diffusion representation. The gated recurrent unit performs temporal state modeling on the main graph diffusion representation and outputs the initial embedding vector of the main graph. In the auxiliary graph path, the diffusion convolution module receives the auxiliary graph structure tensor constructed from the fatigue slope map. Based on the adjacency matrix and node feature matrix of the auxiliary graph, it extracts the slope value, fatigue direction label, fatigue level label and evolution direction information of personnel nodes to generate the auxiliary graph diffusion representation. The gated recurrent unit performs temporal modeling on the auxiliary graph diffusion representation and outputs the initial embedding vector of the auxiliary graph. In the fusion path, the feature concatenation structure performs a concatenation operation on the initial embedding vector of the main image and the initial embedding vector of the auxiliary image to construct the fusion input tensor; the fusion gating structure jointly models the fusion input tensor and the context feature vector, performs a gating adjustment operation, and generates a fusion state representation. The fusion state represents the input to the final output layer, completing the generation of the embedding vector of the personnel node; the embedding vector includes job relationship features, adaptation evolution features and fatigue trend features.

7. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 6, characterized in that, Step five specifically involves: Set task cycle identification rules, perform time aggregation processing on task execution records, and divide them into continuous task cycle units according to the set cycle granularity; Within the task cycle unit, the number of tasks, the distribution of task categories, and the coverage of job positions are statistically analyzed to construct a set of task distribution features; Task category distribution includes the percentage of tasks categorized by task name and task responsibility tag; job coverage includes the unique number of job positions associated with each task and the hierarchical distribution of the covered positions; The task density index is calculated based on the task distribution feature set; the task density index is a weighted combination of the number of tasks and the number of job positions covered per unit time. The number of task-related positions marked as key positions in the task cycle unit is counted, and the key position percentage index is calculated; the key position percentage index is the ratio of the number of key positions to the total number of task-related positions. Generate cycle identifiers, task category distribution vectors, task intensity indicators, and key position percentage indicators, and construct a set of context feature vectors; Each context feature vector in the context feature vector set corresponds to a task cycle unit, which contains cycle recognition encoding and task feature encoding information; The context feature vector set is time-aligned with the adaptation change trajectory of personnel nodes to form a context input sequence; the context input sequence is synchronized with the main map path and the auxiliary map path.

8. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 7, characterized in that, Step six specifically involves: Construct the input tensor of the fusion path, and concatenate the context input sequence with the initial embedding vector of personnel nodes output by the main graph path and the fatigue trend vector of personnel nodes output by the auxiliary graph path in chronological order; Perform gating adjustment operations to construct a multidimensional set of gating factors; The multidimensional gating factor set includes job relationship gating factor, adaptation evolution gating factor, and fatigue trend gating factor; the job relationship gating factor adjusts the embedding weight distribution of the main graph path according to the task density index and job coverage characteristics in the context input sequence; the adaptation evolution gating factor adjusts the evolution direction attention weight of the adaptation evolution feature according to the task category distribution vector in the context input sequence. The fatigue trend gating factor adjusts the slope sensitivity distribution of the fatigue trend feature based on the key position percentage index in the context input sequence. Based on a multidimensional gating factor set, a fusion adjustment function is constructed to perform fusion processing on the main graph embedding vector, the auxiliary graph embedding vector, and the context input sequence to generate an updated personnel node embedding vector; the personnel node embedding vector includes a job relationship feature vector, an adaptation evolution feature vector, and a fatigue trend feature vector. The job relationship feature vector preserves the time evolution pattern of the job connection structure, the adaptation evolution feature vector preserves the adaptation degree change trajectory and evolution direction information, and the fatigue trend feature vector preserves the fatigue slope and contextual relationship structure information.

9. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 8, characterized in that, The process of performing job configuration calculation based on personnel node embedding vectors and outputting job recommendation results and adaptation scores is as follows: Construct a job recommendation input tensor, and combine the personnel node embedding vector with the job association information vector in the job node set in pairs to form a candidate configuration pair set; each candidate configuration pair includes a personnel node embedding vector and a job node association information vector. Perform feature interaction operations on the candidate configuration pair set, calculate the job relationship feature matching degree, the correlation of adaptation evolution trend and the coupling degree of fatigue trend, and generate a configuration feature score set; The job relationship feature matching degree is calculated based on the structural similarity between the job relationship feature vector and the job responsibility label; the adaptation evolution trend correlation is calculated based on the evolution path similarity between the adaptation evolution feature vector and the historical configuration state of the job; the fatigue trend coupling degree is calculated based on the trend direction matching degree between the fatigue trend feature vector and the job task intensity index. An adaptation scoring function is constructed based on the configuration feature scoring set, and the adaptation scoring result of each candidate configuration pair is output. The adaptation scoring results are sorted and the job recommendation results are output. The job recommendation results are the set of job nodes with the highest scores corresponding to each personnel node. Record personnel node identifiers, recommended job identifiers, and suitability scores, and generate job recommendation results and suitability scores.

10. The method for human resource allocation and risk prediction in archival institutions based on graph neural networks according to claim 9, characterized in that, The process of combining the fitness change trajectory and fatigue slope information to perform job stability analysis and risk level prediction, and outputting risk prediction results, specifically: The adaptation evolution feature vector and fatigue trend feature vector are extracted from the personnel node embedding vector to construct the risk analysis input tensor. The adaptation evolution feature vector contains information on the fitness change trajectory and evolution direction, while the fatigue trend feature vector contains information on fatigue slope and contextual relationship structure. Perform time series analysis on the adaptation change trajectory, calculate the consistency index of change amplitude, change frequency and direction, and generate a set of adaptation fluctuation characteristics; Perform slope trend analysis on the fatigue slope, calculate the trend direction, fluctuation degree and correlation ratio with key positions, and generate a fatigue response feature set; A job stability scoring function is constructed based on the set of adaptation fluctuation features and the set of fatigue response features, and the job stability scoring result is output. The job stability scoring result represents the degree of stability of the job adaptation status of personnel nodes in the current task cycle. Establish risk level prediction rules, and classify low-risk, medium-risk, and high-risk levels based on the job stability score and the combined threshold conditions of adaptation fluctuation and fatigue response indicators. Record personnel node identifiers, job identifiers, stability scores, and risk levels, and output risk prediction results.

Images