Data processing method and system for human resource management based on cloud computing

Abstract

The invention discloses a data processing method and system for human resource management based on cloud computing, and the method comprises the steps: transmitting human resource data to a cloud platform, calculating a correlation coefficient matrix output coupling feature between features through a GCN graph convolution network, and calculating a personnel volatility index to obtain a fluctuation feature; inputting the coupling features and the fluctuation features into a feature coupling layer of a DyFPN dynamic feature fusion network, and generating a fusion feature vector through vector splicing and a nonlinear activation function; establishing a prediction model by using an LSTM (Long Short-Term Memory) network and an ATT attention mechanism, and capturing a nonlinear time sequence rule of the fusion feature vector by using the prediction model to obtain a personnel demand prediction index; and calculating the matching degree between the post and the personnel by using an FP-Growth association algorithm. And the human scheduling matching efficiency and the human resource management efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human resource management, in particular to a data processing method and system for human resource management based on cloud computing. BACKGROUND

[0002] The traditional human resource management method exposes problems such as response lag and low data utilization rate in aspects such as recruitment, employee performance evaluation, and resignation prediction. The existing method lacks effective analysis tools and methods. The degree of matching between employee skills and positions cannot be found through in-depth analysis of employee performance data, so that the human resource management plan cannot be made, and the efficiency of human resource management of enterprises is not high. SUMMARY

[0003] The purpose of the present application is to solve the above problems, and a data processing method and system for human resource management based on cloud computing are designed.

[0004] To achieve the above purpose, the technical scheme of the present application is as follows: further, in the data processing method for human resource management based on cloud computing, the data processing method for human resource management comprises the following steps: Collecting multi-source data, filling in missing data by using time series interpolation method, detecting abnormal values by using isolated forest algorithm and replacing them, and obtaining human resource data; Transmitting the human resource data to the cloud platform, calculating the correlation coefficient matrix between the features by the GCN graph convolution network, outputting the coupled features, and calculating the personnel volatility index to obtain the volatility features; Inputting the coupled features and the volatility features into the feature coupling layer of the DyFPN dynamic feature fusion network, generating a fusion feature vector through vector splicing and a nonlinear activation function; Establishing an LSTM-ATT prediction model by using an LSTM long short-term memory network and an ATT attention mechanism, capturing the nonlinear time sequence law of the fusion feature vector by using the LSTM-ATT prediction model, and obtaining a personnel demand prediction index; Based on the personnel demand prediction index, calculating the matching degree between the post and the personnel by using the FP-Growth association algorithm, generating a recommendation list in the cloud platform according to the calculation result, and sending the recommendation list to the management platform.

[0005] Further, in the data processing method for human resource management based on cloud computing, the multi-source data is collected, the missing data is filled in by using the time series interpolation method, the abnormal values are detected and replaced by using the isolated forest algorithm, and the human resource data is obtained, which comprises: Collecting employee change data, performance data, recruitment demand data, and market trend data to obtain multi-source data; The data formats of the unified multi-source data are unified, and the missing data is filled by using a time series interpolation method to obtain filled data. A plurality of random decision trees are constructed by the isolation forest algorithm to simulate the distribution law of the data, data points isolated in the trees are marked as abnormal and deleted to obtain the human resource data.

[0006] Further, in the data processing method for human resource management based on cloud computing, the human resource data is transmitted to the cloud platform, the correlation coefficient matrix between features is calculated by the GCN graph convolution network to output the coupling features, and the volatility features are obtained by calculating the personnel volatility index, including: The human resource data is encrypted by the AES-256 algorithm, the transmission link is encrypted by the SSL protocol, and is transmitted to the cloud platform. Taking an employee as a basic node, the correlation strength between nodes is quantified based on three dimensions of department, job level and performance, and a graph structure is established. The constructed graph structure is input into the GCN graph convolution network, and the deep correlation is mined by two layers of processing, the first layer focuses on local correlation and filters weak correlation, and the second layer extends to global correlation to generate coupling features.

[0007] Further, in the data processing method for human resource management based on cloud computing, the coupling features and the volatility features are input into the feature coupling layer of the DyFPN dynamic feature fusion network, the fusion feature vector is generated by vector splicing and a nonlinear activation function, including: The coupling features are compressed into low-dimensional vectors by feature importance sorting, and the volatility features are expanded by a sliding window to obtain processing feature data. The coupling feature vector and the volatility feature vector are connected head to tail after standardization to generate a joint feature vector. The joint feature vector is generated by using the Swish activation function, the feature value is compressed to the interval of 0-1 by using the Sigmoid activation function, and an enhanced feature vector is obtained. The compressed feature vector after nonlinear transformation is compressed by a fully connected layer to generate a fusion feature vector.

[0008] Further, in the data processing method for human resource management based on cloud computing, the LSTM-ATT prediction model is established by using the LSTM long short-term memory network and the ATT attention mechanism, the nonlinear time sequence law of the fusion feature vector is captured by using the LSTM-ATT prediction model, and a personnel demand prediction index is obtained, including: The first layer LSTM captures short-term dependence, the second layer LSTM focuses on long-term trends, and the hidden layer is set to 128 neurons. The memory unit optimizes the retention and forgetting of historical information by a gating mechanism.

[0009] Further, in the data processing method for human resource management based on cloud computing, the LSTM-ATT prediction model is established by using the LSTM long short-term memory network and the ATT attention mechanism, the non-linear time sequence law of the fusion feature vector is captured by using the LSTM-ATT prediction model, and a personnel demand prediction index is obtained, and the method further comprises: The similarity between the current hidden state and all historical hidden states is calculated by the attention mechanism, and the weight is multiplied by the LSTM output of the time step; The LSTM output weighted by the attention is mapped to a personnel demand prediction index for the next three months by a fully connected layer.

[0010] Further, in the data processing method for human resource management based on cloud computing, the matching degree of the post and the personnel is calculated based on the personnel demand prediction index by using the FP-Growth association algorithm, a recommendation list is generated in the cloud platform according to the calculation result, and the recommendation list is sent to the management platform, comprising: The post features and the personnel features are combined into transactions, and the transaction data set is analyzed by the FP-Growth algorithm; the association rules of the post features and the personnel features are mined from the frequent item set to obtain the matching degree of the post and the personnel.

[0011] Further, in the data processing system for human resource management based on cloud computing, the data processing system for human resource management comprises the following modules: The data acquisition and processing module is used for acquiring multi-source data, filling missing data by using time series interpolation method, detecting abnormal values by using isolated forest algorithm and replacing them to obtain human resource data; The cloud platform calculation module is used for transmitting the human resource data to the cloud platform, calculating the correlation coefficient matrix between the features by using the GCN graph convolution network to output the coupled features, and calculating the personnel volatility index to obtain the volatility features; The feature vector fusion module is used for inputting the coupled features and the volatility features into the feature coupling layer of the DyFPN dynamic feature fusion network, generating the fusion feature vector by vector splicing and a nonlinear activation function; The demand index prediction module is used for establishing the LSTM-ATT prediction model by using the LSTM long short-term memory network and the ATT attention mechanism, capturing the non-linear time sequence law of the fusion feature vector by using the LSTM-ATT prediction model, and obtaining a personnel demand prediction index; The recommendation list generation module is configured to calculate the matching degree between the post and the personnel by using the FP-Growth association algorithm based on the personnel demand prediction index, generate a recommendation list in the cloud platform according to the calculation result, and send the recommendation list to the management platform.

[0012] Further, in the data processing system for human resource management based on cloud computing, the demand index prediction module comprises the following sub-modules: The multiplication sub-module is configured to calculate the similarity between the current hidden state and all historical hidden states by using the attention mechanism, and multiply the weight and the LSTM output of the time step; The mapping sub-module is configured to map the attention-weighted LSTM output to the personnel demand prediction index for the next three months by using a fully connected layer.

[0013] Further, in the data processing system for human resource management based on cloud computing, the recommendation list generation module comprises the following sub-modules: The mining sub-module is configured to combine the post features and the personnel features into transactions, analyze the transaction dataset by using the FP-Growth algorithm, mine the association rules of the post features and the personnel features from the frequent item set, and obtain the matching degree between the post and the personnel.

[0014] The beneficial effects are as follows: 1. According to the personnel demand prediction index, the enterprise can make a talent reserve plan in advance to ensure that the required talents can be obtained in time when the business expands. At the same time, by analyzing the market trend data, the enterprise can adjust the talent structure to adapt to the industry development trend, improve the core competitiveness of the enterprise, and improve the efficiency of human scheduling and matching and the efficiency of human resource management. 2. According to the skills, performance and other characteristics of the employees, suitable posts and development directions are recommended to help the employees realize their personal value, improve the satisfaction and loyalty of the employees, and reduce the talent turnover rate. BRIEF DESCRIPTION OF DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not considered as limiting the application.

[0016] Figure 1 FIG. 1 is a schematic diagram of a first embodiment of the data processing method for human resource management based on cloud computing in the embodiments of the present application; Figure 2 FIG. 2 is a schematic diagram of a second embodiment of the data processing method for human resource management based on cloud computing in the embodiments of the present application; Figure 3 FIG. 3 is a schematic diagram of a first embodiment of the data processing system for human resource management based on cloud computing in the embodiments of the present application. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a cloud-based data processing method for human resource management includes the following steps: Step 101: Collect multi-source data, fill in missing data using time series interpolation, detect outliers and replace them using the Isolation Forest algorithm to obtain human resources data; Specifically, in this embodiment, employee change data, performance data, recruitment demand data, and market trend data are collected to obtain multi-source data; Unify the data format of multi-source data, and use time series interpolation to fill in the missing data to obtain filled data; By constructing multiple random decision trees using the isolated forest algorithm, the distribution pattern of the data is simulated. Data points that are isolated in the trees are marked as anomalies and deleted to obtain human resource data.

[0020] Specifically: 1.1 Data Acquisition Module; Data types and sources: Employee change data: including onboarding, offboarding, job transfer, promotion, etc., sourced from the company's HRM system, attendance system, and OA approval workflow.

[0021] Performance data includes monthly / quarterly / annual KPI scores (0-100 points), key project contributions, and 360-degree evaluation results, sourced from performance appraisal systems and project management tools.

[0022] Recruitment demand data includes job title, number of positions required, skill requirements, start date, and salary range, sourced from recruitment demand forms submitted by departments.

[0023] Market trend data includes industry turnover rate, competitor recruitment salaries, job supply-demand ratio, and skill popularity rankings, sourced from third-party data interfaces.

[0024] Data format standardization: The data format should be JSON or Parquet, and timestamps should be numerical data with two decimal places. Text data should be deduplicated.

[0025] 1.2 Missing data imputation; Applicable scenarios: Data is missing in continuous time dimensions.

[0026] Interpolation method selection: If the data trend is stable (frequency of job level changes): use linear interpolation (given the values ​​of t1 and t3, t2 = (t1 + t3) / 2). If the data exhibits periodicity (peak recruitment season demand): use spline interpolation (fitting curves using polynomials to preserve periodic characteristics). If the missing value rate is >30%: combine forward imputation (filling with the most recent non-missing values) + domain knowledge correction (manually adjusting during periods of declining recruitment demand). Execution flow: Group by time granularity, apply the corresponding interpolation method to each group of missing values, and output the complete time series.

[0027] 1.3 Outlier Detection and Replacement (Isolation Forest Algorithm); Outlier definition: Extreme values ​​that deviate from the normal distribution (e.g., an employee's performance score is 0 but there is no record of leaving the company, or the recruitment demand for a certain position is 100 people, far exceeding the historical average).

[0028] Step 102: Transmit human resources data to the cloud platform, calculate the correlation coefficient matrix between features through the GCN graph convolutional network to output coupled features, and calculate the personnel volatility index to obtain volatility features. Specifically, in this embodiment, human resources data is encrypted using the AES-256 algorithm, and the transmission link is encrypted using the SSL protocol before being transmitted to the cloud platform. A graph structure is established by using individual employees as basic nodes and quantifying the strength of relationships between nodes based on three dimensions: department, job level, and performance. The constructed graph structure is input into the GCN graph convolutional network, and deep associations are mined through two layers of processing. The first layer focuses on local associations and filters out weak associations, while the second layer extends to global associations and generates coupled features.

[0029] Specifically: Encryption and Access Control: The transmission process employs dual encryption: the data itself is encrypted using the AES-256 algorithm, and the transmission link is encrypted using the SSL / TLS protocol. Simultaneously, strict transmission permissions are set—only designated HR system administrators and cloud platform service accounts can initiate transmission requests, and each transmission requires verification of an identity token (dynamic verification code) to prevent unauthorized access.

[0030] Transmission method and rhythm; Real-time transmission: For high-frequency changing data, triggered transmission is used—data is automatically pushed to the cloud platform within 10 minutes of being generated. Batch transmission: For low-frequency static data, scheduled transmission is used—automatic batch uploads are performed at 2 AM daily (off-peak business hours) to avoid consuming daytime business bandwidth.

[0031] 2.1.3 Cloud platform data reception and storage; Receive verification: After receiving the data, the cloud platform automatically performs three checks: Integrity check: verifying whether the number of transmitted data entries matches the number of records in the preprocessing stage; Format check: checking whether the fields match the preset template; Timestamp check: ensuring that the data timestamp is within a reasonable range.

[0032] Graph structure construction: defining nodes and their relationships; Node Definition: Individual employees are the basic nodes, or departments are the aggregate nodes. Each node has the following basic attributes: Department (Technical Department), Job Level (P3 Engineer, M2 Manager), and average performance score over the past 3 months (85 points). Weight Calculation of Relationships (Edges): Based on three dimensions—department, job level, and performance—the strength of the relationship between nodes is quantified. Departmental dimension: Nodes within the same department are weighted at 0.4 (frequent daily collaboration), nodes across departments with overlapping projects are weighted at 0.2, and nodes across departments without overlap are weighted at 0.1 (basic organizational association); Job level dimension: Nodes with a job level difference ≤ 1 level are weighted at 0.3, and nodes with a difference > 1 level are weighted at 0.1; Performance dimension: Nodes with a performance score difference ≤ 10 points are weighted at 0.3, and nodes with a difference > 10 points are weighted at 0.1.

[0033] The final edge weight is the sum of the weights of the three dimensions (nodes in the same department + with a 1-level difference in job rank + with a 5-point difference in performance, weight = 0.4 + 0.3 + 0.3 = 1.0, the most closely related).

[0034] 2.2.2 GCN Network Processing: Extracting Coupling Features; Graph structure input and hierarchical processing: The constructed graph structure (nodes + weighted edges) is input into the GCN network, and deep connections are mined through two layers of processing: The first layer focuses on local connections (interactions between small teams within the same department), filters out weak connections (edges with weight < 0.2), and retains core collaborative relationships; The second layer expands to global connections and strengthens the influence of key nodes.

[0035] Coupling Feature Output: The final output is a node association strength matrix—each value in the matrix represents the degree of coupling between two nodes. This matrix directly reflects the collaborative network within the organization, providing association feature support for subsequent predictions.

[0036] 2.3 Extraction of fluctuation features (based on personnel volatility index); Core objective: To quantify the intensity of personnel changes within a certain period, generate fluctuation characteristics that reflect organizational stability, and capture dynamic changes such as peak turnover and frequent job transfers (increased turnover rate after year-end performance reviews and frequent job transfers in new business departments).

[0037] 2.3.1 Calculation dimensions of the volatility index; The volatility index is calculated using a monthly time granularity from three core change indicators: turnover rate: number of employees leaving in a month ÷ total number of employees at the beginning of the month; job transfer rate: number of employees transferred to a new job in a month ÷ total number of employees at the beginning of the month; promotion rate: number of employees promoted in a month ÷ total number of employees at the beginning of the month.

[0038] 2.3.2 Weight Allocation and Anomaly Correction; Weight setting logic: The weights are assigned based on the degree of business impact: turnover rate (0.5) > job transfer rate (0.3) > promotion rate (0.2). The reason is that turnover directly leads to job vacancies, which has the greatest impact on business continuity; job transfer affects team collaboration efficiency, but it is an internal transfer; promotion is usually a positive incentive and has the least impact on stability.

[0039] 2.3.3 Output of fluctuation characteristics; The final result is a monthly volatility index sequence with values ​​ranging from 0 to 100. This sequence directly reflects the trend of organizational stability and, when combined with coupling features, forms the input basis for subsequent prediction models.

[0040] Coupling characteristics and fluctuation characteristics are two key features used to characterize the state of an organization's human resources. Their definitions and expressions are as follows: The coupling feature focuses on the strength of collaborative relationships among members within an organization, which is quantitatively expressed through graph structure modeling and GCN network processing. Specifically, using individual employees or departments as nodes, a weighted association network is constructed based on three dimensions: department, job level, and performance. In the department dimension, the weight is 0.4 for members within the same department, 0.2 for members with overlapping projects across departments, and 0.1 for members without overlapping projects across departments. In the job level dimension, the weight is 0.3 for a job level difference ≤ 1 level and 0.1 for a difference > 1 level. In the performance dimension, the weight is 0.3 for a score difference ≤ 10 points and 0.1 for a score difference > 10 points. The weights of the three dimensions are then summed to form the final association strength of each edge (for example, a node in the same department + a job level difference of 1 level + a performance difference of 5 points has a weight of 0.4 + 0.3 + 0.3 = 1.0). This graph structure is then input into the GCN network for two layers of processing: the first layer focuses on local associations, filtering out weak association edges with weights < 0.2 and retaining core collaborative relationships; the second layer extends to global associations, strengthening the influence of key nodes. The final output is a node association strength matrix. Each element in the matrix represents the degree of coupling between two nodes, with a value range of 0-1 (e.g., 0.8 indicates strong coupling and 0.2 indicates weak coupling). This matrix can intuitively reflect the density of the collaborative network and the distribution of core nodes within an organization, providing association feature support for the prediction model.

[0041] Volatility characteristics are used to quantify the degree of personnel change within an organization over a specific period, dynamically expressed through a monthly volatility index sequence. Its calculation dimensions include turnover rate (number of departures in a given month ÷ total number of employees at the beginning of the month), job transfer rate (number of job transfers in a given month ÷ total number of employees at the beginning of the month), and promotion rate (number of promotions in a given month ÷ total number of employees at the beginning of the month), with weights assigned based on the degree of business impact: turnover rate (0.5) > job transfer rate (0.3) > promotion rate (0.2). For example, if the monthly turnover rate is 10%, the job transfer rate is 8%, and the promotion rate is 5%, then the volatility index = 10% × 0.5 + 8% × 0.3 + 5% × 0.2 = 8.4. This ultimately generates a monthly volatility index sequence from 0 to 100, with higher values ​​indicating lower organizational stability (e.g., 90 points indicates a coexistence of peak departures and frequent job transfers, while 30 points indicates stable personnel changes). This sequence can capture dynamic changes such as peak turnover and frequent job transfers (e.g., increased turnover rate after year-end performance evaluation and frequent job transfers in new business departments). When combined with coupling features, it forms the basis of the prediction model input, reflecting both the structural characteristics of the organization's collaborative network and the dynamic trend of personnel mobility, providing dual support for organizational stability analysis and risk warning.

[0042] Step 103: Input the coupling features and fluctuation features into the feature coupling layer of the DyFPN dynamic feature fusion network, and generate the fused feature vector through vector concatenation and nonlinear activation function; Specifically, in this embodiment, the coupling features are compressed into a low-dimensional vector by sorting the features by their importance, and the fluctuation features are expanded by using a sliding window to obtain the processed feature data; The standardized coupled feature vector is concatenated with the fluctuation feature vector to generate a joint feature vector. The intermediate feature vector of the joint feature vector is generated by using the Swish activation function, and the feature values ​​are compressed to the 0-1 interval by the Sigmoid activation function to obtain the enhanced feature vector. The compressed feature vector, after nonlinear transformation, is compressed through a fully connected layer to generate a fused feature vector.

[0043] Specifically: 3.1 Feature preprocessing: Dimension alignment and normalization; Core objective: To transform coupled features (correlation matrix) and fluctuation features (exponential sequence) into vectors of a unified dimension, eliminating the dimensional differences between features and preparing for fusion.

[0044] 3.1.1 Feature Dimensionality Reduction and Dimensionality Increase; Coupling feature processing: Dimensionality reduction of high-dimensional matrices: The association matrix output by GCN is compressed into a low-dimensional vector by sorting features by importance. For example, the relationship values ​​between each employee and the other employees with the strongest associations in the Top 5 are extracted (preserving the key collaborative network), reducing the dimensionality to 5000 dimensions.

[0045] Secondary dimensionality reduction: Principal component analysis (PCA) is applied to the 5000-dimensional vector, retaining the top 256 principal components with a cumulative contribution rate of 95%, and finally compressing it into a 256-dimensional vector.

[0046] Fluctuation characteristic processing: Time Series Expansion: The monthly volatility index series is expanded using a sliding window. For example, with a window size of 3, a new feature is generated: the current value of each month + the value of the previous month + the values ​​of the previous two months, expanding from 12 dimensions to 36 dimensions.

[0047] Dimensionality Upscaling: The 36-dimensional dimension is expanded to 256-dimensional dimension through a fully connected layer mapping, which is aligned with the dimension of the coupled feature.

[0048] 3.1.2 Standardization and Normalization; The 256-dimensional coupled feature vector after dimensionality reduction is standardized by Z-score (mean 0, standard deviation 1) to eliminate the dimensional differences between different departments / job levels.

[0049] The 256-dimensional fluctuation feature vector after dimensionality upgrade is normalized by Min-Max to ensure that the fluctuation index (0-100) and the correlation strength (0-1) are on the same scale.

[0050] 3.2 Feature Coupling Layer: Vector Concatenation and Nonlinear Fusion; Core objective: To deeply fuse the two types of preprocessed features and capture the complex interactive relationships between features through nonlinear transformation.

[0051] 3.2.1 Feature splicing strategy; Simple concatenation: Connect the standardized coupled feature vector (256-dimensional) and the fluctuation feature vector (256-dimensional) end to end to form a 512-dimensional joint feature vector.

[0052] Weighted concatenation (optional): Assign higher weight (0.6) to the fluctuation feature and a weight of 0.4 to the coupling feature (because the fluctuation feature more directly reflects changes in personnel demand). Scale the vectors according to the weights before concatenation.

[0053] 3.2.2 Nonlinear activation enhancement; Multi-layer transformation: The concatenated 512-dimensional vector is processed through two non-linear activation layers. First layer: Apply the Swish activation function (a smooth version of ReLU, enhancing gradient propagation) to generate intermediate feature vectors (preserving the non-linear relationships in the original features).

[0054] The second layer applies the Sigmoid activation function to compress the feature values ​​to the 0-1 range, highlighting key features (the cross-features of fluctuation periods and highly correlated teams).

[0055] Dynamic gating mechanism: Introducing feature gating to automatically learn the importance weights of two types of features. For example, during periods of business stability (when the weight of fluctuating features decreases), the model focuses more on coupling features (team collaboration networks); during periods of business adjustment (when the weight of fluctuating features increases), the model focuses more on personnel change trends.

[0056] 3.3 Output the fused feature vector; Dimensionality reduction output: The 512-dimensional vector that has undergone nonlinear transformation is compressed back to 256 dimensions through a fully connected layer (balancing expressive power and computational efficiency) to generate the final fused feature vector.

[0057] Feature verification: Check the reasonableness of the distribution of the fused vector (whether the mean is close to 0 and the standard deviation is close to 1), and ensure that there are no extreme values ​​or distribution biases (all values ​​in a certain dimension are 0, which may indicate that features are lost).

[0058] The fused feature vector is a composite feature representation generated by deep interaction and nonlinear transformation of coupled features and fluctuation features. Its core lies in capturing the synergistic effect of organizational collaboration networks and dynamic changes in personnel through multi-dimensional feature fusion. Specifically, the vector is constructed through the following steps: First, the original features are dimensionally aligned and standardized—coupled features (such as a 256-dimensional association strength vector) are Z-score standardized to eliminate dimensional differences, and fluctuation features (such as a 256-dimensional time-series extension vector) are compressed to the [0,1] interval through Min-Max normalization; then, the two types of feature vectors are merged into a 512-dimensional joint feature vector using a head-to-tail concatenation strategy, and the influence of key dimensions is strengthened by weight concatenation (such as a fluctuation feature weight of 0.6 and a coupled feature weight of 0.4).

[0059] In the feature coupling layer, the joint vector undergoes two layers of nonlinear activation enhancement: the first layer applies the Swish activation function (f(x)=x·sigmoid(x)) to preserve the nonlinear relationship in the original features and generate an intermediate feature vector; the second layer uses the Sigmoid activation function to compress the feature values ​​to the 0-1 range, highlighting the cross-features of high volatility periods and strongly correlated teams (such as the change in coupling strength of core collaboration nodes during peak resignation periods). Simultaneously, a dynamic gating mechanism is introduced to automatically adjust feature weights according to the business cycle—reducing the weight of volatile features (e.g., 0.3) during stable business periods to strengthen the characterization of the team collaboration network by the coupling features (0.7); and increasing the weight of volatile features (e.g., 0.7) during organizational adjustment periods to focus on the predictive value of personnel change trends.

[0060] Finally, a fully connected layer compresses the 512-dimensional vector to 256 dimensions, generating a fused feature vector that balances expressive power and computational efficiency. This vector is expressed numerically, with each dimension comprehensively reflecting the interaction between organizational collaboration strength and personnel dynamics (e.g., a value of 0.8 in a certain dimension might indicate that a highly correlated team is simultaneously experiencing a peak in personnel turnover). Its distribution must meet the validation criteria of a mean close to 0 and a standard deviation around 1 to ensure no feature loss or extreme value shift. The fused feature vector serves as input to subsequent prediction models, simultaneously capturing both the static collaboration patterns of the organizational structure and the dynamic trends of personnel mobility, providing more accurate decision-making basis for scenarios such as turnover prediction and personnel reassignment optimization.

[0061] Step 104: Build an LSTM-ATT prediction model using LSTM long short-term memory network and ATT attention mechanism. Use the LSTM-ATT prediction model to capture the nonlinear temporal pattern of the fused feature vector and obtain the personnel demand prediction index. Specifically, in this embodiment, the first LSTM layer captures short-term dependencies, the second LSTM layer focuses on long-term trends, and the hidden layer is set to 128 neurons. Memory unit optimization dynamically adjusts the retention and forgetting of historical information by memory units through a gating mechanism.

[0062] The similarity between the current hidden state and all historical hidden states is calculated using an attention mechanism, and the weights are multiplied by the LSTM output at each time step. The attention-weighted LSTM output is mapped to a personnel demand forecast index for the next three months through a fully connected layer.

[0063] Specifically: 4.1 LSTM layer: Capturing temporal dependencies; Core objective: To uncover long-term and short-term patterns in personnel demand over time based on historical sequences of fused feature vectors.

[0064] 4.1.2 Design of a dual-layer LSTM structure; The first LSTM layer focuses on capturing short-term dependencies (the chain reaction of monthly fluctuations), with 128 hidden neurons (matching the dimension of the input vector) and dropout=0.2 to prevent overfitting.

[0065] The second LSTM layer focuses on long-term trends (seasonal recruitment demand), with the input being the output of the first layer. The hidden layer also has 128 neurons, and recurrent_dropout=0.2 is added (to prevent overfitting from recurrent connections).

[0066] Memory unit optimization: The gating mechanism dynamically adjusts the retention and forgetting of historical information in memory units. For example, during the off-season (Q3), the retention of Q1 data is weakened, while the focus on Q2-Q3 trends is strengthened.

[0067] 4.2 Attention Mechanism (ATT): Focusing on key time steps; Core objective: To automatically identify the most valuable historical time points for predicting future demand (such as the recruitment peak before the launch of a major project), avoiding the misinterpretation of all historical data.

[0068] 4.2.1 Attention weight calculation; Context awareness: The importance weights for each time step are calculated using an additive attention model. Calculate the similarity between the current hidden state and all historical hidden states (through non-linear transformation).

[0069] Apply the softmax function to the similarity to normalize the weights to the 0-1 range.

[0070] Weight application: The normalized weights are multiplied by the LSTM output of the corresponding time step to enhance the influence of important time steps (the weight of the recruitment peak in Q4 of the previous year for a certain department is increased to 0.8, while the weight of regular months is 0.2).

[0071] 4.2.2 Multiscale attention; Simultaneously calculate short-term and long-term attention to capture key information at different time scales. For example: Short-term focus is on the impact of recent quantitative adjustments on hiring; Pay long-term attention to the cyclical fluctuations in the industry.

[0072] 4.3 Predicted Output and Post-processing; Prediction layer: The attention-weighted LSTM output is mapped to the personnel demand prediction index (range 0-100) for the next 3 months through a fully connected layer, and the activation function is ReLU (to ensure non-negative prediction values).

[0073] Confidence interval calculation: Simultaneously outputs the predicted confidence interval (the range of predicted values ​​at a 95% confidence level), reflecting the uncertainty of the prediction. For example: if the predicted demand index for July 2024 is 75, the confidence interval is [70, 80], indicating that there is a 95% certainty that the actual value will fall within this range.

[0074] Business logic correction: Lower limit constraint: If the predicted index is <20 (close to no demand), it will be automatically adjusted to 20 (considering the need for basic personnel replacement).

[0075] Upper limit constraint: If the predicted index is >90 (extreme demand), manual review will be triggered (to confirm whether there are any abnormal factors, such as major strategic adjustments).

[0076] Step 105: Based on the personnel demand prediction index, use the FP-Growth association algorithm to calculate the matching degree between positions and personnel, generate a recommendation list in the cloud platform according to the calculation results, and send the recommendation list to the management platform.

[0077] Specifically, in this embodiment, job features and personnel features are combined into transactions, and the transaction dataset is analyzed using the FP-Growth algorithm; association rules between job features and personnel features are mined from the frequent itemset to obtain the matching degree between job and personnel.

[0078] Specifically: 5.1 Data preparation and standardization for matching; Organize the forecast demand data: Extract the personnel demand forecast index output from step 4, classify it according to job type (technical position, management position) and urgency level (urgent and routine), and clarify the demand priority of each position (index ≥ 80 is urgent demand and should be matched first).

[0079] Analyze the job feature database: Collect detailed information on the jobs to be matched, including core skills (Java development and project management), job requirements (5 years of work experience and a bachelor's degree or above), job responsibilities (responsible for backend system development), work location, etc., and standardize the features (unify skills to industry-standard terms to avoid confusion with synonyms such as programming and software development).

[0080] Organize the personnel feature database: Extract employees' personal information, including existing skills (proficient in Python), work experience (3 years of e-commerce industry experience), education, current position, performance (performance rating of A in the past six months), mobility status (no intention to leave and can be internally transferred), etc., and also standardize them (experience is uniformly quantified by year, and performance is based on the corresponding score range of the rating).

[0081] Link historical matching data: retrieve successful job-person matching cases from the company's historical recruitment / transfer records (characteristics of the employees ultimately hired for a Java development position) as a reference for matching rules.

[0082] 5.2 Mining Association Rules Based on FP-Growth Algorithm; Construct a transaction dataset: Combine job characteristics and personnel characteristics into transactions, filter successfully matched transactions from historical data, and form a transaction dataset.

[0083] Generating frequent itemsets: Analyzing transaction datasets using the FP-Growth algorithm to identify frequently occurring feature combinations. For example, the combination of "Java developer + proficient Java skills + 3 years of experience" frequently appears in historical data and is marked as a frequent itemset.

[0084] Extract association rules: Mine association rules between job features and personnel features from frequent item sets (if the job requires 'Java development', then match personnel with 'proficient Java skills and 3 years of experience'), and filter reliable rules by confidence (confidence ≥70%, that is, the rule is true in more than 70% of cases).

[0085] 5.3 Job-person matching degree calculation; Feature mapping and rule matching: For each job position to be matched, suitable rules are selected from the association rules (the Python engineer position is matched with the rule of proficiency in Python and data analysis experience), and the job features and personnel features are mapped according to the rules (checking whether the employee has the skills, experience, etc. required by the rules).

[0086] Multi-dimensional weighting: Different characteristics are assigned weights based on job priority and feature importance. For example: core skills (40%), work experience (30%), performance (20%), and mobility (10%). For urgent positions, the weight of core skills can be increased (up to 50%).

[0087] Overall matching score: The degree to which an employee meets the job characteristics is scored according to weight (40 points for fully meeting the core skills, and 20 points for partially meeting them). The total score is the matching degree between the employee and the job (maximum score 100 points).

[0088] 5.4 Recommendation list generation and optimization; Preliminary sorting: Sort personnel from highest to lowest matching degree, and select the top 20 matching personnel for each position (balancing quantity and accuracy) to form a preliminary recommendation list.

[0089] Secondary screening and optimization: Exclude employees who do not meet the basic requirements (employees who have left the company or have explicitly refused to be transferred); The ranking will be adjusted based on the personnel demand forecast index (personnel with a matching score of ≥80 will be given priority for urgent positions, while the score can be relaxed to ≥60 for regular positions). Avoid duplicate recommendations (the same person should not appear in three consecutive recommendations for the same position to prevent waste of resources).

[0090] Additional recommended information: Supplement the list with auxiliary information, including the employee's current status (employed, available for transfer), matching advantages (core skills perfectly matched, excellent performance), and expected start date, to facilitate management platform decision-making.

[0091] 5.5 The recommended list is sent to the management platform; Format encapsulation: Convert the recommendation list into a management platform-compatible format (structured table, JSON format), including fields such as job title, recommender name, matching degree, and additional information, ensuring that the information is clear and readable.

[0092] Access control and targeted sending: The visibility of the list is assigned according to management permissions (HR administrators can view all job recommendations, while department managers can only view job recommendations within their own department). The list is sent to the management platform via an encrypted interface (HTTPS protocol) to ensure secure transmission.

[0093] Send confirmation and log recording: After sending, obtain the receipt confirmation from the management platform, and record logs such as sending time, recipient, and list content for subsequent traceability (to verify the adoption status of recommendation results).

[0094] Its beneficial effects are as follows: 1. Based on the personnel demand forecast index, companies can formulate talent reserve plans in advance to ensure that they can obtain the necessary talent in a timely manner when expanding their business. At the same time, through the analysis of market trend data, companies can adjust their talent structure to adapt to industry development trends, enhance their core competitiveness, and improve the efficiency of personnel scheduling and matching, as well as human resource management. 2. Based on employees' skills, performance, and other characteristics, suitable positions and development directions can be recommended to help employees realize their personal value, improve employee satisfaction and loyalty, and reduce talent turnover.

[0095] Please see Figure 2 In the data processing method for cloud-based human resource management, human resource data is transmitted to the cloud platform, and the correlation coefficient matrix between features is calculated through a GCN graph convolutional network to output coupled features. Simultaneously, the volatility index of personnel is calculated to obtain volatility features. The process includes the following steps: Step 201: Encrypt the human resources data using the AES-256 algorithm, and encrypt the transmission link using the SSL protocol before transmitting it to the cloud platform; Step 202: Using individual employees as basic nodes, quantify the strength of relationships between nodes based on three dimensions: department, job level, and performance, and establish a graph structure; Step 203: Input the constructed graph structure into the GCN graph convolutional network. The deep associations are mined through two layers of processing. The first layer focuses on local associations and filters out weak associations. The second layer extends to global associations and generates coupled features.

[0096] The above describes embodiments of the cloud computing-based human resource management data processing method of the present invention. Please refer to [link / reference]. Figure 3 In a cloud-based human resource management data processing system, the system includes the following modules: The data acquisition and processing module is used to collect multi-source data, fill in missing data using time series interpolation, detect outliers and replace them using the isolated forest algorithm to obtain human resources data. The cloud platform computing module is used to transmit human resources data to the cloud platform, calculate the correlation coefficient matrix between features through the GCN graph convolutional network to output coupled features, and calculate the personnel volatility index to obtain volatility features. The feature vector fusion module is used to input coupled features and fluctuation features into the feature coupling layer of the DyFPN dynamic feature fusion network, and generate fused feature vectors through vector concatenation and nonlinear activation functions. The demand index prediction module is used to build an LSTM-ATT prediction model using LSTM long short-term memory network and ATT attention mechanism. The LSTM-ATT prediction model is used to capture the nonlinear time series pattern of the fused feature vector to obtain the personnel demand prediction index. The recommendation list generation module is used to calculate the matching degree between job positions and personnel based on the personnel demand prediction index and the FP-Growth correlation algorithm. The module generates a recommendation list in the cloud platform based on the calculation results and sends the recommendation list to the management platform.

[0097] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data processing method for human resource management based on cloud computing, characterized in that, The data processing method for human resource management includes the following steps: Collect data from multiple sources, fill in missing data using time series interpolation, and detect and replace outliers using the isolated forest algorithm to obtain human resources data; The human resources data is transmitted to the cloud platform, and the correlation coefficient matrix between features is calculated through the GCN graph convolutional network to output the coupled features. At the same time, the personnel volatility index is calculated to obtain the volatility features. The coupling features and fluctuation features are input into the feature coupling layer of the DyFPN dynamic feature fusion network, and a fused feature vector is generated by vector concatenation and nonlinear activation function. An LSTM-ATT prediction model is established using an LSTM long short-term memory network and an ATT attention mechanism. The LSTM-ATT prediction model is then used to capture the nonlinear temporal patterns of the fused feature vectors to obtain the personnel demand prediction index. Based on the personnel demand prediction index, the matching degree between job positions and personnel is calculated using the FP-Growth association algorithm. A recommendation list is generated in the cloud platform based on the calculation results, and the recommendation list is sent to the management platform.

2. The data processing method for human resource management based on cloud computing as described in claim 1, characterized in that, The process involves collecting multi-source data, filling in missing data using time series interpolation, and detecting and replacing outliers using the Isolation Forest algorithm to obtain human resource data, including: Collect employee turnover data, performance data, recruitment demand data, and market trend data to obtain multi-source data; Unify the data format of multi-source data, and use time series interpolation to fill in the missing data to obtain filled data; By constructing multiple random decision trees using the isolated forest algorithm, the distribution pattern of the data is simulated. Data points that are isolated in the trees are marked as anomalies and deleted to obtain human resource data.

3. The data processing method for human resource management based on cloud computing as described in claim 1, characterized in that, The process of transmitting the human resources data to the cloud platform, calculating the correlation coefficient matrix between features using a GCN graph convolutional network to output coupled features, and simultaneously calculating the personnel volatility index to obtain volatility features includes: The human resources data is encrypted using the AES-256 algorithm, and the transmission link is encrypted using the SSL protocol before being transmitted to the cloud platform. A graph structure is established by using individual employees as basic nodes and quantifying the strength of relationships between nodes based on three dimensions: department, job level, and performance. The constructed graph structure is input into the GCN graph convolutional network, and deep associations are mined through two layers of processing. The first layer focuses on local associations and filters out weak associations, while the second layer extends to global associations and generates coupled features.

4. The data processing method for human resource management based on cloud computing as described in claim 1, characterized in that, The step of inputting the coupling features and fluctuation features into the feature coupling layer of the DyFPN dynamic feature fusion network, and generating a fused feature vector through vector concatenation and a nonlinear activation function, includes: The coupled features are compressed into a low-dimensional vector by sorting them by feature importance, and the fluctuating features are expanded by a sliding window to obtain the processed feature data. The standardized coupled feature vector is concatenated with the fluctuation feature vector to generate a joint feature vector. The intermediate feature vector of the joint feature vector is generated by using the Swish activation function, and the feature values ​​are compressed to the 0-1 interval by the Sigmoid activation function to obtain the enhanced feature vector. The compressed feature vector, after nonlinear transformation, is compressed through a fully connected layer to generate a fused feature vector.

5. The data processing method for human resource management based on cloud computing as described in claim 1, characterized in that, The LSTM-ATT prediction model is established using an LSTM (Long Short-Term Memory) network and an ATT (Attention Transaction) mechanism. This LSTM-ATT prediction model captures the nonlinear temporal patterns of the fused feature vectors to obtain a personnel demand prediction index, including: The first LSTM layer captures short-term dependencies, the second LSTM layer focuses on long-term trends, and the hidden layer is set to 128 neurons. Memory unit optimization dynamically adjusts the retention and forgetting of historical information by memory units through a gating mechanism.

6. The data processing method for human resource management based on cloud computing as described in claim 1, characterized in that, The method of establishing an LSTM-ATT prediction model using an LSTM long short-term memory network and an ATT attention mechanism, and using the LSTM-ATT prediction model to capture the nonlinear temporal patterns of the fused feature vectors to obtain the personnel demand prediction index, also includes: The similarity between the current hidden state and all historical hidden states is calculated using an attention mechanism, and the weights are multiplied by the LSTM output at each time step. The attention-weighted LSTM output is mapped to a personnel demand forecast index for the next three months through a fully connected layer.

7. The data processing method for human resource management based on cloud computing as described in claim 1, characterized in that, The process of calculating the matching degree between job positions and personnel based on the personnel demand prediction index and the FP-Growth correlation algorithm, generating a recommendation list in the cloud platform based on the calculation results, and sending the recommendation list to the management platform includes: Job characteristics and personnel characteristics are combined into transactions, and the transaction dataset is analyzed using the FP-Growth algorithm. Association rules between job characteristics and personnel characteristics are mined from the frequent itemsets to obtain the matching degree between jobs and personnel.

8. A cloud-based human resource management data processing system, characterized in that: The data processing system for human resource management includes the following modules: The data acquisition and processing module is used to collect multi-source data, fill in missing data using time series interpolation, detect outliers and replace them using the isolated forest algorithm to obtain human resources data. The cloud platform computing module is used to transmit the human resources data to the cloud platform, calculate the correlation coefficient matrix between features through the GCN graph convolutional network to output the coupling features, and calculate the personnel volatility index to obtain the volatility features. The feature vector fusion module is used to input the coupled features and fluctuation features into the feature coupling layer of the DyFPN dynamic feature fusion network, and generate a fused feature vector through vector concatenation and nonlinear activation function. The demand index prediction module is used to establish an LSTM-ATT prediction model using an LSTM long short-term memory network and an ATT attention mechanism. The LSTM-ATT prediction model is used to capture the nonlinear temporal patterns of the fused feature vectors to obtain the personnel demand prediction index. The recommendation list generation module is used to calculate the matching degree between job positions and personnel based on the personnel demand prediction index and the FP-Growth association algorithm. The module generates a recommendation list in the cloud platform based on the calculation results and sends the recommendation list to the management platform.

9. The cloud-based human resource management data processing system as described in claim 8, characterized in that, The demand index forecasting module includes the following sub-modules: The multiplication submodule is used to calculate the similarity between the current hidden state and all historical hidden states through the attention mechanism, and multiplies the weights with the LSTM output at each time step. The mapping submodule is used to map the attention-weighted LSTM output to a personnel demand forecast index for the next 3 months through a fully connected layer.

10. The cloud-based human resource management data processing system as described in claim 8, characterized in that, The recommendation list generation module includes the following sub-modules: The mining submodule is used to combine job features and personnel features into transactions. It analyzes the transaction dataset using the FP-Growth algorithm to mine association rules between job features and personnel features from the frequent itemsets, thereby obtaining the matching degree between jobs and personnel.

Images