A method for constructing an enterprise labor employment efficiency evaluation model, medium and system

Abstract

The present application provides a kind of enterprise labor efficiency evaluation model construction method, medium and system, belong to artificial intelligence model technical field, the present application is by gathering staff behavior data to construct high-dimensional sparse feature matrix, after coding by hash embedding layer, gaussian mixture distribution modeling and double network cross screening noise correction are carried out to labeled label, employee collaboration relationship graph is constructed based on training sample set, after hierarchical random neighbor sampling and graph clustering presegmentation, employee performance evaluation vector is obtained by inputting causal inference enhancement double-flow comparative evaluation model, then the efficiency distribution is time series evolution by optimal transport gradient flow continuous flow algorithm, finally, according to the deviation of prediction result and equilibrium threshold, external incentive term is adjusted and intervention strategy is output, the technical problem that employee performance evaluation model cannot accurately estimate causal effect in organization network and realize group efficiency time series evolution prediction is solved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence model technology, and specifically relates to a method, medium and system for constructing an enterprise labor efficiency evaluation model. Background Technology

[0002] Evaluating employee performance is a core task of human resource management. Traditional methods involve collecting employee behavior data and constructing statistical regression or machine learning classification models to score or grade employee performance. With the rise of graph neural networks, some solutions have further incorporated employee collaboration graphs to model team collaboration characteristics, aggregating neighbor node information through graph convolutional networks to enhance individual performance representation. These solutions have been widely applied in performance management, talent inventory, and organizational diagnostics in medium and large enterprises.

[0003] However, in current employee performance evaluations, because employee behavior data is essentially observational data rather than randomized controlled trial data, there are numerous confounding variables between intervention variables such as working conditions, training intensity, and task allocation and performance labels. Traditional statistical regression models and graph neural network models can only capture the statistical correlation between features and labels, failing to distinguish between true causal effects and confounding biases. Furthermore, traditional performance diffusion modeling methods lack physical equation constraints, making them prone to extrapolation distortion in organizational structures with sparse training data, further leading to unreliable predictions of group performance evolution.

[0004] In other words, existing technologies have the technical problem that employee performance evaluation models cannot accurately estimate causal effects and predict the temporal evolution of group performance in organizational networks. Summary of the Invention

[0005] In view of this, the present invention provides a method, medium and system for constructing an enterprise labor efficiency evaluation model, which can solve the technical problem that existing employee efficiency evaluation models cannot accurately estimate causal effects in organizational networks and achieve prediction of the temporal evolution of group efficiency.

[0006] This invention is implemented as follows: This invention provides a method, medium, and system for constructing an enterprise labor efficiency evaluation model, including the following steps:

[0007] Employee behavior data is collected, a high-dimensional sparse feature matrix is ​​constructed, and it is stored in a hybrid format of compressed sparse row and compressed sparse column. After being encoded by a hash embedding feature mapping layer, the employee behavior feature vector set is obtained.

[0008] Noise detection is performed on the labeled tags in the set of employee behavior feature vectors. The label distribution is modeled using Gaussian mixture distribution, and noisy samples are identified through a dual-network cross-screening mechanism to obtain a noise-corrected training sample set.

[0009] An employee collaboration relationship graph is constructed based on the training sample set. Hierarchical random neighbor sampling and graph clustering pre-segmentation are used to divide the employee collaboration relationship graph into subgraphs. The causal inference-enhanced dual-flow comparison evaluation model is input to obtain the employee performance evaluation vector.

[0010] The efficiency diffusion propagation index is calculated based on the employee efficiency evaluation vector. The efficiency diffusion coefficient is adjusted according to the range of the efficiency diffusion propagation index. The efficiency diffusion propagation algorithm, which drives the partial differential operator of the heat conduction equation, is used to update the employee efficiency state vector and obtain the efficiency diffusion distribution.

[0011] The optimal transmission gradient flow continuous flow algorithm for efficiency diffusion distribution input is subjected to time-series evolution to obtain the group efficiency distribution prediction results;

[0012] Based on the deviation between the predicted group efficacy distribution and the efficacy equilibrium threshold, the parameters of the external incentive terms are adjusted, and an efficacy assessment report and intervention strategy recommendations are output.

[0013] The high-dimensional sparse feature matrix refers to a feature dimension that reaches... The employee behavior feature matrix, with a magnitude of less than 0.1% and a non-zero rate, is stored using a hybrid compressed sparse row format and a compressed sparse column format, and the computing kernel is optimized using a single instruction multiple data stream integer arithmetic instruction set for sparse data.

[0014] The hash embedding feature mapping layer refers to a network layer that maps a high-dimensional sparse feature matrix to a fixed-dimensional dense embedding space through a low-collision hash function. The collision rate is controlled below the collision rate threshold by combining multiple hash functions.

[0015] The Gaussian mixture distribution modeling of the label distribution refers to using a mixture model containing multiple Gaussian components to model the posterior probability distribution of the labeled labels in order to distinguish the components corresponding to clean samples and noisy samples. The number of Gaussian components is determined iteratively through the Bayesian information criterion.

[0016] The dual-network cross-selection mechanism refers to training two networks with the same structure but different initializations. In each iteration, each network updates its parameters using only samples that the other network determines to have low loss. The loss function uses symmetric cross-entropy instead of standard cross-entropy.

[0017] The hierarchical random neighbor sampling refers to randomly sampling a fixed number of neighbors from each node's neighbor set at each hop of the graph neural network message passing stage. The number of samples is determined by ablation experiments based on a comprehensive index of gradient variance and convergence speed.

[0018] The graph clustering pre-segmentation refers to using a multi-level graph partitioning algorithm to pre-segment the employee collaboration relationship graph into several subgraphs with low cross-partition edge ratios. The subgraph features are stored using half-precision floating-point numbers, and a gradient checkpoint recalculation strategy is combined to reduce peak memory usage.

[0019] The causal inference enhanced dual-stream contrastive evaluation model consists of eight components: a fact flow encoder, a counterfactual flow encoder, an interference budget sub-layer, a directed acyclic graph attention mask mechanism, a contrastive learning module, a bias score weighting layer, a counterfactual generation sub-network, and an effectiveness prediction main network. The fact flow encoder and the counterfactual flow encoder share the weight matrix of all convolutional and fully connected layers.

[0020] The directed acyclic graph attention masking mechanism uses the adjacency matrix of a predefined or automatically discovered causal directed acyclic graph as the causal mask of the attention matrix, forcing the attention weights to be passed along the causal direction and shielding non-causal related paths.

[0021] The propensity score weighting layer is composed of an independent four-layer fully connected lightweight multilayer perceptron that predicts the conditional probability of each sample receiving intervention, and the inverse probability weighting term is injected into the loss function of the performance prediction main network. The activation function of the multilayer perceptron is Sigmoid.

[0022] Among them, the effectiveness diffusion propagation index It is calculated from the node degree centrality, the mean edge weight, and the current component value of the employee performance state vector, using the following formula: The effectiveness diffusion coefficient is adjusted into three levels based on the range of the effectiveness diffusion propagation index.

[0023] The heat conduction equation partial differential operator discretization efficiency diffusion propagation algorithm treats the graph Laplacian matrix of the employee collaboration relationship graph as a discrete approximation of the spatial second-order partial differential operator, uses the Crank-Nicolson implicit difference scheme to discretize the time derivative, and solves the linear equation system formed by each time step using the conjugate gradient method.

[0024] The optimal transmission gradient flow continuous flow algorithm models the evolution of the employee group's efficiency distribution over time as a gradient flow on the Wasserstein-2 metric space. It adopts the Jordan-Kinderlehrer-Otto discretization scheme. The free energy functional updated at each step is composed of a linear combination of efficiency potential energy and entropy regularization term. The optimal transmission plan is solved in parallel on the GPU using the Sinkhorn algorithm.

[0025] The efficiency equilibrium threshold is determined by statistical analysis of historical efficiency distribution data. The upper limit of the threshold is the mean plus one standard deviation of the rate of change of the Wasserstein distance of the group efficiency distribution over several consecutive months, and the lower limit of the threshold is the mean minus one standard deviation.

[0026] The conflict rate threshold is 1%; the efficiency diffusion propagation index tier boundary values ​​are 0.3 and 0.7, and the efficiency diffusion coefficients for the three tiers range from 0.01 to 0.05, 0.05 to 0.15, and 0.15 to 0.30, respectively; the time step of the optimal transmission gradient flow continuous flow algorithm ranges from 0.01 to 0.10, and the entropy regularization weight ranges from 0.1 to 1.0; the efficiency equilibrium threshold statistical period is 12 to 36 months.

[0027] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for constructing an enterprise labor efficiency evaluation model.

[0028] A third aspect of the present invention provides a system for constructing an enterprise labor efficiency evaluation model, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0029] This invention constructs a causal inference-enhanced dual-stream comparative evaluation model, which combines the fact flow encoder and the counterfactual flow encoder with the synergistic effect of the directed acyclic graph attention masking mechanism and the bias score weighting layer. It simultaneously shields confounding biases from three levels: feature space, structural path, and loss function, making the performance evaluation vector causally interpretable, thereby overcoming the fundamental defect of traditional models that only model statistical correlations.

[0030] This invention further employs a discretized efficiency diffusion propagation algorithm using the partial differential operator of the heat conduction equation. It uses the graph Laplacian matrix as a discrete approximation of the second-order differential operator in space, constraining the propagation process of employee efficiency within the organizational network using physical equations, thus avoiding extrapolation distortion in sparse regions by purely data-driven models. Based on this, the optimal transport gradient flow continuous flow algorithm incorporates the evolution of the group efficiency distribution into a variational framework of the Wasserstein-2 metric space, making the distribution prediction path geometrically optimal and smooth.

[0031] In summary, this invention solves the technical problem mentioned in the background art that employee performance evaluation models cannot accurately estimate causal effects and achieve temporal evolution prediction of group performance in organizational networks. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method of the present invention.

[0033] Figure 2 A schematic diagram of the subgraph division of employee collaboration relationships and the distribution of performance evaluation vectors.

[0034] Figure 3 This diagram illustrates the steady-state convergence process of the efficiency diffusion propagation algorithm and the bottleneck node identification results.

[0035] Figure 4 This is a schematic diagram of the time-series evolution prediction results of the population effectiveness distribution. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0037] like Figure 1 The diagram shown is a flowchart of a method, medium, and system for constructing an enterprise labor efficiency evaluation model provided by this invention. The method includes the following steps:

[0038] S01. Collect employee behavior data, construct a high-dimensional sparse feature matrix, store it using a combination of compressed sparse row format and compressed sparse column format, and encode it through a hash embedding feature mapping layer to obtain the employee behavior feature vector set;

[0039] S02. Noise detection is performed on the labeled labels in the set of employee behavior feature vectors. The label distribution is modeled using Gaussian mixture distribution, and noise samples are identified through a dual-network cross-screening mechanism to obtain a noise-corrected training sample set.

[0040] S03. Construct an employee collaboration relationship graph based on the training sample set, and use hierarchical random neighbor sampling and graph clustering pre-segmentation to divide the employee collaboration relationship graph into subgraphs. Input the causal inference-enhanced dual-flow comparison evaluation model to obtain the employee performance evaluation vector.

[0041] S04. Calculate the efficiency diffusion propagation index based on the employee efficiency evaluation vector, adjust the efficiency diffusion coefficient according to the range of the efficiency diffusion propagation index, drive the partial differential operator discretization efficiency diffusion propagation algorithm of the heat conduction equation to update the employee efficiency state vector, and obtain the efficiency diffusion distribution.

[0042] S05. Perform time-series evolution on the optimal transmission gradient flow continuous flow algorithm for the efficiency diffusion distribution input to obtain the group efficiency distribution prediction results;

[0043] S06. Based on the deviation between the predicted results of group efficacy distribution and the efficacy equilibrium threshold, adjust the parameters of external incentive terms and output an efficacy assessment report and intervention strategy recommendations.

[0044] Among them, a high-dimensional sparse feature matrix refers to a feature dimension that reaches The high-dimensional sparse feature matrix contains employee behavior characteristics of magnitude less than 0.1% and a non-zero rate of less than 0.1%. This high-dimensional sparse feature matrix employs a hybrid storage method combining compressed sparse row and compressed sparse column formats to adapt to access patterns in different dimensions, and utilizes a single-instruction multiple-data-stream integer arithmetic instruction set optimized for sparse data to optimize the computational kernel.

[0045] The hash embedding feature mapping layer refers to a network layer that maps a high-dimensional sparse feature matrix to a fixed-dimensional dense embedding space using a low-collision hash function. The collision rate is controlled to below 1% through a combination of multiple hash functions. The optimal configuration of the number of hash functions and the embedding dimension is determined through comparative experiments with 10 to 20 different combinations, based on the trade-off curve between inference latency and memory usage.

[0046] Among them, Gaussian mixture distribution modeling of label distribution refers to using a mixture model containing 2 to 5 Gaussian components to model the posterior probability distribution of labeled samples, in order to distinguish the components corresponding to clean samples and noisy samples. The number of Gaussian components is determined iteratively in multiple experiments using the Bayesian information criterion.

[0047] The dual-network cross-selection mechanism refers to training two networks with the same structure but different initializations. In each iteration, each network only uses samples that the other network judges to have low loss for parameter updates. The loss function uses symmetric cross-entropy instead of standard cross-entropy to enhance noise robustness.

[0048] Hierarchical random neighbor sampling refers to randomly sampling a fixed number of neighbors for each node at each aggregation stage of the graph neural network message passing. A neighbor, The value range is 10 to 25. The value was determined by conducting ablation experiments in the range of 5 to 30 with a step size of 5, based on a comprehensive index of gradient variance and convergence speed.

[0049] In this context, graph clustering pre-segmentation refers to using a multi-level graph partitioning algorithm to pre-divide the employee collaboration graph into several subgraphs with low cross-partition edge ratios, thereby reducing the gradient estimation variance introduced by subgraph sampling. Subgraph features are stored using half-precision floating-point numbers, and a gradient checkpoint recalculation strategy is combined to reduce peak memory usage.

[0050] The causal inference-enhanced dual-stream contrastive evaluation model is based on deep learning. This model consists of eight components: a fact flow encoder, a counterfactual flow encoder, an interference budget sub-layer, a directed acyclic graph (DAG) attention masking mechanism, a contrastive learning module, a propensity score weighting layer, a counterfactual generation sub-network, and a performance prediction main network. The fact flow encoder and counterfactual flow encoder share the weight matrix of all convolutional and fully connected layers. The fact flow encoder receives real-world observations of employee working conditions and behavioral feature vectors from the training sample set, while the counterfactual flow encoder receives intervention sample feature vectors generated by the structural causal model. The interference budget sub-layer injects intervention operations into the intermediate representation of the counterfactual flow encoder through a differentiable perturbation operator; the perturbation magnitude is determined by the intervention variable assignment in the structural causal model. The DAG attention masking mechanism uses the adjacency matrix of a predefined or automatically discovered causal DAG as the causal mask for the attention matrix, forcing attention weights to propagate along the causal direction and shielding non-causal paths. The structure of the causal DAG is initialized with domain knowledge and then iteratively optimized using a differentiable structure learning algorithm. The contrastive learning module imposes causal invariance constraints on the factual and counterfactual embedding vectors of the same employee, minimizing their distance while simultaneously pushing the embedding vectors of different employees further apart, forming a causal invariant feature space. The propensity score weighting layer uses an independent four-layer fully connected lightweight multilayer perceptron to predict the conditional probability of each sample receiving intervention, and injects an inverse probability weighting term into the loss function of the efficacy prediction main network. The number of hidden neurons in the multilayer perceptron is 128, 64, 32, and 1 respectively, with the sigmoid activation function. The number of neurons was determined through grid search experiments increasing by a factor of two within the range of 64–256, using the log loss of propensity score estimation on the validation set as the evaluation metric. An alternating optimization iterative loop is used between the counterfactual generation subnetwork and the efficacy prediction main network. The counterfactual generation subnetwork generates high-quality intervention sample feature vectors that are difficult for the efficacy prediction main network to distinguish, while the efficacy prediction main network improves the accuracy of causal effect estimation under the pressure of stronger intervention sample feature vectors. In terms of CUDA stream allocation, the forward computation of the fact-fact encoder and the counterfact encoder is allocated to two independent CUDA streams for parallel execution. The kernel function of the interference budget sublayer is allocated to a third CUDA stream, and the multilayer perceptron in the propensity score weighted layer is allocated to a fourth CUDA stream. The four CUDA streams converge in the contrastive learning module through an event synchronization mechanism. Regarding memory allocation, the shared encoder weight matrix resides in GPU memory, the batch embedding vectors are stored using half-precision floating-point numbers, and the causal directed acyclic graph adjacency matrix resides in GPU memory in a sparse format. For memory allocation, training samples are batch-preloaded into paged memory and continuously filled by an asynchronous data prefetch thread. The connection weight sharing coefficients between neurons in each layer and adjacent layers are controlled within the range of 0.1 to 0.3 by sparse regularization constrained by the structural causal model. This range is determined through cross-validation experiments in the 0.05 to 0.5 interval with a step size of 0.05.The upper limit of the number of iterations in the data batch loop is set to 200-500 rounds, and the early stopping patience value is set to 20-50 rounds. The range is determined by multiple training experiments under different data scales based on the convergence curve of the causal effect estimation error of the validation set.

[0051] The steps for establishing the training dataset for the causal inference-enhanced dual-flow comparative evaluation model specifically include: collecting employees' historical work condition records, behavior logs, and performance evaluation labels, and generating corresponding intervention sample feature vectors for each record; the intervention values ​​of the intervention variables are randomly sampled by the structural causal model within a reasonable business range, with the intervention range being 10% to 50% of the original values, and the range is determined through business expert interviews and analysis of historical intervention experimental data; noise filtering is performed on the labeled labels, and samples with loss values ​​below the lower 20th percentile are retained as a clean training set; the training set, validation set, and test set are divided in a 7:1:2 ratio.

[0052] The specific steps for training the causal inference-enhanced dual-stream contrast evaluation model include: initializing the shared encoder weights using a normal distribution with a standard deviation of 0.01–0.05; initializing the multilayer perceptron in the propensity score weighting layer using Xavier; and using the adaptive moment estimation algorithm for the optimizer with an initial learning rate of [missing information]. ~ The range was determined through a learning rate warm-up scanning experiment; the loss function was a weighted sum of the performance prediction loss, the contrastive learning loss, and the propensity score estimation loss, and the weight coefficients were determined in 100 trials through Bayesian hyperparameter search; the counterfactual generation subnetwork and the performance prediction main network were optimized alternately, with the counterfactual generation subnetwork updating one step after the performance prediction main network updates one step in each round; the gradient pruning threshold was set to 1.0 to 5.0, and the threshold was determined by monitoring the gradient norm distribution of the first 10 rounds of training.

[0053] The causal inference-enhanced two-stream contrastive evaluation model brings the following technical benefits to the entire scheme: Traditional performance evaluation models only model the statistical correlation between features and performance labels, failing to distinguish between true causal effects and spurious correlations introduced by confounding variables. The causal inference-enhanced two-stream contrastive evaluation model, through a two-stream contrastive structure of fact-stream encoders and counterfactual-stream encoders, forces the learning of causal-invariant features in a causal-invariant feature space; a directed acyclic graph attention masking mechanism structurally shields non-causal information transmission paths; and a biased score weighting layer corrects sample selection bias at the loss function level. The synergy of these three elements enables the weights of performance influencing factors to possess causal interpretability, elevating performance intervention decisions from correlation-driven to causal-driven, effectively reducing the probability of erroneous intervention decisions.

[0054] The effectiveness diffusion and propagation index is a scalar index calculated by combining the node degree centrality, the mean edge weight, and the current component value of the employee effectiveness state vector in the employee collaboration relationship graph. The formula is as follows:

[0055] ;

[0056] in For nodes The degree, The maximum degree of a node in the employee collaboration graph. For nodes The average of all edge weights, Let the maximum mean edge weight be the value in the employee collaboration graph. For nodes Current employee performance state vector component values, This represents the maximum employee performance state vector component value in the employee collaboration graph. Based on the performance diffusion and propagation index. Range adjustment efficiency diffusion coefficient The rules are as follows: When hour, The value range is 0.01 to 0.05; when hour, The value range is 0.05 to 0.15; when hour, The value ranges from 0.15 to 0.30. The boundary values ​​of 0.3 and 0.7 were determined based on the principle of minimizing error after conducting a uniformly graded experiment on the entire domain of the efficiency diffusion and propagation index using the convergence error of the employee efficiency state vector as the evaluation index on 20 sets of collaborative graph data of enterprises of different sizes. The range of values ​​was determined by a combination of numerical stability condition analysis using the Crank-Nicolson scheme and comparative experiments with multiple time steps.

[0057] The efficiency diffusion propagation algorithm, which discretizes the partial differential operator of the heat conduction equation, treats the graph Laplace matrix of the employee collaboration graph as a discrete approximation of a second-order spatial partial differential operator. It uses the employee efficiency state vector as the discretized representation of the temperature field and employs a Crank-Nicolson implicit difference scheme to discretize the time derivative. Each time step advances to form a system of linear equations with the graph Laplace matrix as the coefficient matrix, which is solved using the conjugate gradient method. The computational complexity is linearly related to the number of edges in the employee collaboration graph. The efficiency diffusion coefficient is heterogeneously set according to edge type, and the external excitation term corresponds to external intervention input. The steady-state solution corresponds to the final equilibrium state of efficiency diffusion distribution in the organization, and the Fiedler vector component of the graph Laplace matrix is ​​used to identify bottleneck nodes in efficiency propagation.

[0058] The technical benefits of the discretized efficiency diffusion propagation algorithm based on the partial differential operator of the heat conduction equation are as follows: it models the propagation process of employee efficiency in the organizational network as a diffusion process driven by the heat conduction physical equation, enabling the temporal evolution of the employee efficiency state vector to have mathematical consistency with the constraints of the physical equation, and avoiding the extrapolation distortion problem of pure data-driven models in sparse training data regions. The second-order time precision of the Crank-Nicolson format makes the short-term prediction results of the employee efficiency state vector smoother, the heterogeneous efficiency diffusion coefficient enables the model to accurately reflect the differentiated impact of different collaboration strengths on efficiency propagation, the bottleneck node identification of the Fiedler vector component provides a structural basis for precise intervention, and the adjustability of the external incentive term enables the scheme to have the simulation evaluation capability of intervention strategies. Overall, the evaluation results of efficiency intervention strategies have stronger physical interpretability and generalization robustness.

[0059] The optimal transport gradient flow continuous flow algorithm models the evolution of employee group performance distribution over time as a gradient flow in the Wasserstein-2 metric space. The Jordan-Kinderlehrer-Otto discretization scheme transforms the continuous flow into a sequence of stepwise optimal transport problems. Each step updates and solves a minimization problem targeting the 2-Wasserstein distance and the free energy functional. The free energy functional is a linear combination of performance potential and entropy regularization, with the entropy regularization corresponding to the stochastic diffusion effect of employee behavior. The Sinkhorn algorithm solves the optimal transport plan for each step in parallel on a GPU. Momentum extrapolation accelerates the convergence of the Jordan-Kinderlehrer-Otto iterative algorithm, which outputs the predicted group performance distribution. (Time step...) The range is 0.01 to 0.10, and the entropy regularization weight ranges from 0.1 to 1.0. This range is determined by evaluating the convergence error of the Wasserstein distance of the group effectiveness distribution across five groups of employees of different sizes over a given time step. The results were determined by grid search experiments with entropy regularization weights.

[0060] The optimal transmission gradient flow continuous flow algorithm brings the following technical benefits to the solution: it incorporates the temporal evolution of employee group performance distribution into the optimal transmission geometry framework, making the distance metric between distributions geometrically optimal. The algorithm constrains the distribution evolution path at the analytical level through variational principles, resulting in smoother predictions of group performance distribution and lower sensitivity to initial distribution disturbances. The entropy regularization term gives the model built-in robustness to individual behavioral uncertainties. The GPU parallel implementation of the Sinkhorn algorithm makes the computation of large-scale employee group distribution evolution feasible in engineering, and the momentum extrapolation mechanism further reduces the number of iterations. Overall, it improves both the timeliness and accuracy of group performance distribution prediction, providing a quantitative basis for selecting the timing of performance interventions in large-scale organizations.

[0061] The efficiency equilibrium threshold is determined through statistical analysis of historical efficiency distribution data. The upper bound is the mean plus one standard deviation of the rate of change of Wasserstein distance in the group efficiency distribution over a continuous period of 12–36 months, and the lower bound is the mean minus one standard deviation. The monthly range is determined through a rolling window stability experiment on data from enterprises of different sizes, selecting the window length that minimizes the variance of the threshold estimate.

[0062] The parameters of the external incentive term are adjusted based on the deviation between the predicted group effectiveness distribution and the effectiveness equilibrium threshold, according to the following rules: When the mean of the predicted group effectiveness distribution is lower than the lower bound of the effectiveness equilibrium threshold, the weight of training intensity in the external incentive term is increased, with a value ranging from 0.6 to 1.0; when the mean of the predicted group effectiveness distribution is between the lower and upper bounds of the effectiveness equilibrium threshold, the external incentive term parameters remain unchanged; when the mean of the predicted group effectiveness distribution is higher than the upper bound of the effectiveness equilibrium threshold, the weight of incentive policy intensity in the external incentive term is decreased, with a value ranging from 0.1 to 0.4. The above weight ranges are determined by conducting uniform sampling experiments across the entire weight domain on multiple sets of historical intervention experimental data, using the absolute error between the actual change in effectiveness after intervention and the predicted change as an indicator, based on the principle of minimizing error.

[0063] Among them, the structural causal model refers to a mathematical model that describes the causal relationship between variables through a directed acyclic graph. It is used to generate feature vectors of intervention samples, support intervention operations on employee work condition variables, and derive the counterfactual distribution after intervention.

[0064] Here, the Fiedler vector refers to the eigenvector corresponding to the second smallest eigenvalue of the graph Laplacian matrix, and the nodes with smaller absolute values ​​of their components correspond to the bottleneck nodes of efficiency propagation in the employee collaboration relationship graph.

[0065] Employee behavior data refers to a set of quantitative data collected by the enterprise's business system, collaboration system, attendance system, project management system, performance management system, and training system within a preset statistical period, reflecting changes in employee work input, collaboration methods, task completion, and efficiency. The preset statistical period can be set by day, week, or month, with 30 days, 90 days, or 180 days being preferred as an observation window. Employee behavior data includes at least the following categories: task-related data, collaboration-related data, time input data, output-related data, quality data, learning and growth data, and organizational interaction data. Specifically, task-related data includes the number of tasks received, the number of tasks completed, the number of delayed tasks, the average processing time for tasks, and the number of rework tasks. Collaboration-related data includes the number of collaborating partners, the number of cross-departmental collaborations, message response time, meeting participation time, and the number of joint projects. Time input data includes effective working hours, overtime hours, the number of times late or early leave, and continuous working hours. Output-related data includes the number of deliverables, sales revenue, the number of customer response orders, the number of code submissions or work order processing volumes. Quality data includes the defect rate, the number of customer complaints, the review pass rate, and the rework rate. Learning and growth data includes the number of training sessions attended, exam scores, and the number of skills certifications. Organizational interaction data includes the number of knowledge sharing sessions, internal evaluation scores, and collaboration satisfaction scores. The above data are normalized, discretized, or binned to form numerical features. For example, message response time is quantified into minute-level values, task completion rate is quantified into a ratio between 0 and 1, and the number of collaborating objects is quantified into an integer count value, which serves as the input for subsequent performance evaluation models.

[0066] Enterprise labor efficiency refers to the comprehensive ability of an enterprise, under given human resource input, for individual employees, teams, or organizational units to generate business output, collaborative contributions, and quality results. It is not a single performance score, but a comprehensive performance indicator composed of employee output efficiency, task performance, collaborative contribution, work quality stability, and intervention response capability. It can usually be quantified as a continuous value between 0 and 1, or divided into discrete levels such as low efficiency, medium efficiency, and high efficiency.

[0067] The employee behavior feature matrix refers to a feature matrix formed by using employee or employee behavior records within a certain time window as rows and coded behavior indicators as columns. Each element in the matrix represents the value of an employee on a certain type of behavior indicator, such as the number of delayed tasks, average response time, number of collaborators, or training completion rate of an employee within 30 days. Because there are many types of actual corporate behavior indicators, but an individual employee will only trigger a small number of behavioral events, this matrix usually presents a high-dimensional and sparse form.

[0068] A high-dimensional sparse feature matrix refers to a matrix formed by expanding an employee behavior feature matrix, where the feature dimension reaches tens of thousands, but the proportion of non-zero elements is less than 0.1%. For example, if a company breaks down employee behavior into 12,000 behavioral feature dimensions, and a single employee corresponds to only 10 to 80 valid non-zero features within a statistical period, then most elements in the employee's row will be zero. This matrix is ​​used to accommodate a large number of fine-grained employee behavioral features while avoiding mistaking features with no observed behavior for valid contributions.

[0069] The hybrid storage method of compressed sparse row format and compressed sparse column format refers to the simultaneous retention of sparse index structures for both row-based and column-based access within the same employee behavior feature matrix. Row-based access is suitable for quickly reading all behavioral features of a single employee, while column-based access is suitable for quickly calculating the distribution of a specific behavioral feature among all employees. In this scheme, this terminology emphasizes bidirectional access optimization for the employee behavior matrix, rather than simply adopting a general sparse matrix storage format.

[0070] The employee collaboration graph refers to a graph structure data built with employees as nodes and their collaborative relationships as edges. The existence of an edge indicates that two employees have engaged in collaborative activities such as joint project development, task handover, approval processes, meeting collaboration, message interaction, or jointly serving customers within the statistical period. The edge weight can be calculated by comprehensively considering factors such as the number of collaborations, collaboration duration, number of shared tasks, response frequency, or collaboration quality score. This graph is used to describe the transmission path and scope of collaborative influence of employee effectiveness within the organizational network.

[0071] Graph clustering pre-segmentation refers to dividing the entire employee collaboration graph into multiple subgraphs based on the degree of collaboration before model training. This ensures that employee collaboration edges within the same subgraph are as dense as possible, while minimizing cross-partition collaboration edges between different subgraphs. This process reduces memory usage and gradient fluctuations during large-scale employee collaboration graph training and improves the stability of subsequent subgraph sampling.

[0072] The fact-flow encoder is the encoding branch in the causal inference-enhanced two-stream contrastive evaluation model that processes real-world employee observation data. Its input consists of the employee's working conditions, behavioral characteristics, and collaborative characteristics in the actual business environment, and its output is the employee's efficacy embedding representation under those real-world conditions. This encoder characterizes "what actually happened to the employee" and the corresponding efficacy characteristics of that actual state. The counterfactual flow encoder is the encoding branch in the causal inference-enhanced two-stream contrastive evaluation model that processes intervention samples. Its input is not actual employee data, but rather hypothetical intervention samples generated through a structural causal model, such as employee behavioral characteristics hypothesized after a 30% increase in training intensity, a 20% decrease in workload, or a 15% increase in collaboration frequency. This encoder is used to estimate "how employee efficacy might change if certain working conditions are changed."

[0073] The intervention sample feature vector refers to the simulated feature vector generated after applying a certain business intervention to the original behavioral characteristics of employees. The intervention may include increased training intensity, adjusted task allocation, increased collaborative resources, reduced workload, or changes in incentive policies, with the intervention magnitude set to 10% to 50% of the original variable values. This vector is used to support the model in learning the pattern of performance changes under intervention conditions.

[0074] The intervention layer refers to the differentiable computational layer that injects the influence of intervention variables into the counterfactual flow encoder. Based on the intervention variables and their values ​​given by the structural causal model, this layer perturbs the intermediate representation of the counterfactual flow, enabling the model to simulate the impact of different intervention measures on employee performance characteristics. Its function is to transform business actions such as "training, motivation, and task adjustment" into computable numerical changes within the model.

[0075] The counterfactual generation subnetwork is a sub-model used to generate high-quality feature vectors for intervention samples. Based on the original behavioral characteristics of employees and intervention variables, this subnetwork generates representations of behavioral characteristics that may occur under hypothetical intervention scenarios, enabling the main performance prediction network to be trained on samples that more closely resemble real business interventions.

[0076] The performance prediction main network is the predictive branch that ultimately outputs the employee performance evaluation results. Its inputs include fact flow, counterfactual flow, collaboration graph information, and propensity-weighted feature representations. The output can be an employee performance score, performance level, or employee performance evaluation vector. This network undertakes the core task of "providing performance evaluation results" in this solution.

[0077] In the efficacy diffusion and group efficacy evolution model, the external incentive term refers to the variable that represents the external driving force of proactive corporate interventions on employee efficacy. It can correspond to measures such as training investment, bonus incentives, job adjustments, management support, tool upgrades, and supplementary collaborative resources. This term is not a result of the natural evolution of employee behavior, but rather an external input of corporate management actions on efficacy.

[0078] Intervention responsiveness refers to the sensitivity of employees or teams to improvement or change in their performance status after receiving management interventions such as training, incentives, task adjustments, and collaboration optimization. This capability can be quantified by the magnitude of changes in employee performance evaluation vectors before and after intervention, the shift in group performance distribution, or the predicted performance gain, and is used to determine whether a certain type of intervention is worth implementing.

[0079] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for constructing an enterprise labor efficiency evaluation model.

[0080] A third aspect of the present invention provides a system for constructing an enterprise labor efficiency evaluation model, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0081] The specific implementation of step S01 is as follows: Employee behavior data is collected from the enterprise business system, collaboration system, attendance system, project management system, performance management system, and training system. The statistical period can be set to 30 days, 90 days, or 180 days, covering task-related data, collaboration data, time input data, output data, quality data, learning and growth data, and organizational interaction data. After collection, the data for each dimension is normalized, discretized, or binned. Message response time is quantified to minute-level values, task completion rate is quantified as a percentage value between 0 and 1, and the number of collaborating objects is quantified as an integer count. The employee behavior feature matrix is ​​constructed with employee or employee behavior records within a certain time window as rows and coded behavior indicators as columns, achieving a feature dimension of [number missing]. The matrix consists of elements of varying magnitudes, with non-zero elements accounting for less than 0.1%, forming a high-dimensional sparse feature matrix. This matrix employs a hybrid storage format of compressed sparse row and compressed sparse column formats to accommodate different access needs: quickly reading all features of a single employee in the row direction and quickly calculating the distribution of a specific behavioral feature across all data in the column direction. The computational kernel is optimized using a single-instruction multiple-data-stream integer arithmetic instruction set for sparse data. The high-dimensional sparse feature matrix is ​​encoded via a hash embedding feature mapping layer. Multiple hash functions are combined to map the high-dimensional sparse input to a fixed-dimensional dense embedding space, with the collision rate controlled below 1%. The optimal configuration of the number of hash functions and embedding dimension is determined through 10-20 sets of comparative experiments based on a trade-off curve between inference latency and memory usage. The final output is a set of employee behavioral feature vectors.

[0082] The specific implementation of step S02 is as follows: Noise detection is performed on the labeled tags in the employee behavior feature vector set. A mixture model with 2-5 Gaussian components is used to model the posterior probability distribution of the labeled tags. The number of Gaussian components is determined iteratively in multiple experiments using the Bayesian information criterion to distinguish between clean and noisy sample components. A dual-network cross-filtering mechanism trains two networks with identical structures but different initializations. In each iteration, each network updates its parameters using only samples deemed low-loss by the other network. The loss function uses symmetric cross-entropy instead of standard cross-entropy to enhance robustness to noisy labels. The two networks gradually filter noisy samples through cross-determination, ultimately retaining samples with loss values ​​below the lower 20th percentile as the clean training set. The training set, validation set, and test set are then divided in a 7:1:2 ratio to obtain the noise-corrected training sample set.

[0083] The specific implementation of step S03 is as follows: An employee collaboration graph is constructed based on the training sample set. Employees are used as nodes, and employee pairs that have engaged in collaborative activities such as project co-construction, task handover, approval processes, meeting collaboration, message interaction, or joint customer service within the statistical period are used as edges. The edge weight is calculated comprehensively based on the number of collaborations, collaboration duration, number of joint tasks, response frequency, or collaboration quality score. A hierarchical random neighbor sampling strategy is adopted. In each hop aggregation stage of the graph neural network message passing, a fixed number of neighbors are randomly sampled from the neighbor set of each node. The sampling number is determined through ablation experiments with a step size of 5, ranging from 5 to 30, with a default reference value of 10 to 25. Graph clustering pre-segmentation uses a multi-level graph partitioning algorithm to pre-segment the employee collaboration graph into several subgraphs with low cross-partition edge ratios. Subgraph features are stored using half-precision floating-point numbers, and a gradient checkpoint recalculation strategy is combined to reduce peak memory usage. The subgraph input causal inference-enhanced dual-flow contrastive evaluation model consists of eight components: a fact flow encoder, a counterfactual flow encoder, an intervention budget sublayer, a directed acyclic graph attention masking mechanism, a contrastive learning module, a propensity score weighting layer, a counterfactual generation subnetwork, and a performance prediction main network. The fact flow encoder and counterfactual flow encoder share all convolutional and fully connected layer weight matrices. During training, the fact flow encoder receives real observation features, while the counterfactual flow encoder receives intervention sample features generated by the structural causal model, with the intervention magnitude ranging from 10% to 50% of the original values. The directed acyclic graph attention masking mechanism uses the adjacency matrix of the causal directed acyclic graph as an attention mask to block non-causal paths. The contrastive learning module imposes causal invariance constraints, and the propensity score weighting layer injects inverse probability weights. The counterfactual generation subnetwork and the performance prediction main network are optimized alternately; after each round, the performance prediction main network updates by one step, followed by the counterfactual generation subnetwork, with an upper limit of 200–500 iterations and a tolerance value of 20–50 rounds for early stopping. The final output is an employee performance evaluation vector.

[0084] The specific implementation method of step S04 is: based on the employee performance evaluation vector, according to the formula... Calculate the efficiency diffusion propagation index for each node, where For node degree, The average edge weight of the node. This represents the current performance state vector component value. The performance diffusion coefficient is adjusted in three levels based on the performance diffusion propagation index range: when... When the value is taken as 0.01 to 0.05, When the value is taken as 0.05 to 0.15, The time range is 0.15 to 0.30. The efficiency diffusion propagation algorithm for the partial differential operator discretization of the heat conduction equation treats the graph Laplace matrix as a discrete approximation of the second-order partial differential operator in space, uses the employee efficiency state vector as the discretized representation of the temperature field, and uses the Crank-Nicolson implicit difference scheme to discretize the time derivative. Each time step advances to form a system of linear equations with the graph Laplace matrix as the coefficient matrix, which is solved using the conjugate gradient method. The computational cost is linearly related to the number of edges in the employee collaboration graph, and finally outputs the efficiency diffusion distribution.

[0085] The specific implementation of step S05 is as follows: The efficiency diffusion distribution is input into the optimal transport gradient flow continuous flow algorithm. The evolution of the population efficiency distribution over time is modeled as a gradient flow on the Wasserstein-2 metric space. The Jordan-Kinderlehrer-Otto discretization scheme is used to transform the continuous flow into a sequence of stepwise optimal transport problems. The free energy functional is composed of a linear combination of efficiency potential energy and entropy regularization term, with the entropy regularization weight ranging from 0.1 to 1.0. Each optimal transport plan is solved in parallel on the GPU using the Sinkhorn algorithm. Momentum extrapolation accelerates the convergence of the Jordan-Kinderlehrer-Otto iteration, with a time step ranging from 0.01 to 0.10. Finally, the predicted population efficiency distribution is output.

[0086] The specific implementation of step S06 is as follows: Calculate the deviation between the predicted mean of the group efficacy distribution and the efficacy equilibrium threshold. The efficacy equilibrium threshold is determined by adding or subtracting one standard deviation from the mean of the Wasserstein distance change rate of the historical group efficacy distribution over a continuous period of 12 to 36 months. When the predicted mean is lower than the lower threshold, increase the weight of training intensity in the external incentive item to 0.6–1.0; when the predicted mean is between the upper and lower thresholds, maintain the current parameters of the external incentive item; when the predicted mean is higher than the upper threshold, decrease the weight of incentive policy intensity in the external incentive item to 0.1–0.4. Finally, output an efficacy assessment report and intervention strategy recommendations.

[0087] It should be noted that the key technologies of this invention include: The causal inference-enhanced dual-flow comparative evaluation model, through a dual-encoder comparison structure of fact flow and counterfactual flow, forces the learning of causal invariant representations in the feature space; combined with a directed acyclic graph attention mask, it blocks non-causal paths at the structural level; and then, through a propensity score weighting layer, it corrects selection bias at the loss function level. The synergistic effect of these three elements transforms performance evaluation from statistical correlation-driven to causal-driven, thereby eliminating spurious correlations introduced by confounding variables. The heat conduction equation partial differential operator discretization performance diffusion propagation algorithm utilizes the semi-positive definiteness of the graph Laplacian matrix and the unconditional stability of the Crank-Nicolson scheme to introduce physical equation constraints into the performance diffusion process, avoiding extrapolation distortion in sparse regions by purely data-driven models. Fiedler vectors further provide a spectral structure basis for precise intervention. The optimal transport gradient flow continuous flow algorithm incorporates the temporal evolution of the population performance distribution into the Wasserstein-2 geometric framework, making the distribution evolution path optimal and smooth in a variational sense. The entropy regularization term has built-in robustness to the uncertainty of individual behavior. The three core technologies logically form a complete inference chain from individual causal effect estimation and organizational network diffusion modeling to group distribution time series prediction, which together solves the fundamental contradiction that a single technology module cannot simultaneously take into account causal identification, physical constraints of network diffusion and geometric optimality of distribution evolution.

[0088] It should be noted that in large-scale employee collaboration graphs, the message passing process of graph neural networks suffers from oversmoothing. As the number of aggregation hops increases, the embedding vectors of different employee nodes gradually converge, making it impossible to effectively distinguish the feature representations of high-performing and low-performing employees. This problem is particularly prominent in large organizations with thousands of employees and dense collaboration relationships. The root cause is that the low-pass filtering characteristic of the graph Laplacian operator causes high-frequency discriminative information to be attenuated during multi-hop propagation. Common solutions to this technical problem include limiting the number of aggregation hops or introducing residual connections. However, limiting the number of hops leads to the inability to effectively capture long-distance collaboration information, and residual connections can only alleviate gradient vanishing but cannot fundamentally prevent the low-pass filtering effect, still resulting in severe oversmoothing in dense graph structures. This invention effectively solves this technical problem. Firstly, graph clustering pre-segmentation pre-divides the entire employee collaboration graph into several subgraphs with low cross-region edge ratios according to the degree of collaboration, making the nodes within each subgraph more homogeneous and increasing the effective information density when the model aggregates within the subgraph. This structurally reduces the dilution effect of cross-community noise on discriminative features. Secondly, hierarchical random neighbor sampling samples a fixed number of nodes from the neighbor set in each hop, making the neighbor signals received by each node during aggregation random. This is equivalent to introducing random perturbation into the graph Laplace low-pass filtering process, preventing the embedding vectors from monotonically converging to the graph mean. Thirdly, the contrastive learning module in the causal inference-enhanced dual-stream contrastive evaluation model imposes exclusion constraints on the embedding vectors of different employees, directly preventing the embedding vectors of different employees from approaching each other at the loss function level, thus suppressing the oversmoothing trend from the optimization objective level. These three aspects synergistically suppress oversmoothing in three complementary dimensions: graph structure partitioning, sampling randomization, and loss function constraints, effectively preserving the performance discrimination features in large-scale dense employee collaboration graphs.

[0089] Specifically, the principle of this invention is:

[0090] The present invention can solve the above-mentioned technical problems, and the fundamental reason for this lies in the logical self-consistency of the following aspects.

[0091] First, traditional performance evaluation models directly use the statistical correlation between working condition characteristics and performance labels as a causal relationship, which is unreasonable in observational data because confounding variables simultaneously affect both intervention variables and performance outcomes, causing the model to learn spurious correlations rather than causal effects. This invention introduces a dual-stream structure: a fact-stream encoder processes real observational data, while a counterfactual stream encoder processes intervention samples generated by a structural causal model. Both share a weight matrix, and a contrastive learning module forces their embedding vectors to have the minimum distance in the causal-invariant feature space, ensuring the model retains only causal-invariant features. A directed acyclic graph attention masking mechanism blocks non-causal information transmission paths at the graph structure level, and a biased score weighting layer corrects sample selection bias at the loss function level. These three elements form an independent and complementary causal bias elimination mechanism, thus fundamentally solving the problem of confusion between statistical correlation and causal effects.

[0092] Secondly, traditional data-driven diffusion models lack prior constraints in sparse regions of employee collaboration graphs, making predictions prone to deviating from physically plausible ranges. This invention introduces the partial differential equation of heat conduction into efficiency diffusion modeling. The graph Laplace matrix naturally possesses positive semi-definiteness, ensuring that the efficiency state vector satisfies the energy dissipation principle during diffusion. The Crank-Nicolson implicit difference scheme guarantees unconditional stability over time. These two factors together ensure that efficiency diffusion propagation is mathematically consistent and physically reasonable. Heterogeneous efficiency diffusion coefficients accurately distinguish the contributions of collaboration edges of different intensities to efficiency propagation, while the bottleneck node identification using Fiedler vectors provides a graph-based theoretical basis for intervention location selection.

[0093] Finally, even if the single-step efficacy diffusion prediction is accurate, an unreasonable temporal evolution path of the group efficacy distribution can still lead to distortion of medium- and long-term prediction results. This invention employs an optimal transport gradient flow continuous flow algorithm, modeling the distribution evolution as a gradient flow on the Wasserstein-2 metric space. The Jordan-Kinderlehrer-Otto discretization scheme ensures the uniqueness of the path for each update in the geometrically optimal transport sense. The entropy regularization term provides built-in robustness to random perturbations in individual employee behavior, and the parallel implementation of the Sinkhorn algorithm on GPUs ensures computational feasibility in large-scale scenarios. These three core modules logically form a complete causal inference chain from individual causal effect estimation and organizational network diffusion modeling to temporal prediction of group distribution, thus systematically solving the aforementioned technical problems.

[0094] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0095] The specific implementation of step S01 is as follows: Collect employee behavior data, covering seven categories of quantitative indicators: task-related, collaboration-related, time input-related, output-related, quality-related, learning and growth-related, and organizational interaction-related. After normalization, discretization, or binning, construct a high-dimensional sparse feature matrix. ,in For the number of employees, For feature dimension and achieve The magnitude is on the order of magnitude, and the non-zero rate is less than 0.1%. This matrix employs a hybrid storage format of compressed sparse row and compressed sparse column formats to simultaneously support two access modes: reading all features of a single employee by row and statistically analyzing the distribution of a specific feature by column. The computational kernel is optimized using a single-instruction multiple-data stream integer arithmetic instruction set for sparse data. Subsequently, Input hash embedding feature map layer, through Low-collision hash functions Combining these methods maps high-dimensional sparse indexes to fixed-dimensional ones. With a dense embedding space and a conflict rate controlled below 1%, the employee behavior feature vector set is obtained. ,in For the first Dense behavioral feature vectors of employees. and The optimal configuration was determined by conducting 10 to 20 sets of comparative experiments based on the trade-off curve between inference latency and memory usage.

[0096] The specific implementation of step S02 is as follows: For the employee behavior feature vector set... label in Noise detection was performed, among which For the first Employee performance labels are assigned either continuous values ​​or discrete level codes between 0 and 1. (The last part, "using a combination of...", appears to be incomplete and possibly contains errors. It's best left as is.) Gaussian components ( Model the posterior probability distribution of the label using a Gaussian mixture distribution of 2 to 5. The probability density function of each component is expressed as follows:

[0097] ;

[0098] In the formula, For the sample The latent variables of the component. For the first The mean of each component, For the first The variance of each component, the mixing coefficient satisfy The overall mixed distribution is expressed as follows:

[0099] ;

[0100] In the formula, Given the complete set of parameters of a Gaussian mixture distribution, it is estimated iteratively using the expectation-maximization algorithm. Determined through the Bayesian information criterion, the criterion formula is as follows:

[0101] ;

[0102] In the formula, For containing The maximum likelihood value of the Gaussian mixture model for each component. The number of model parameters, For the total number of samples participating in the modeling, take... smallest The value is used as the optimal number of components. A dual-network cross-selection mechanism is employed, training two networks with identical structures but different initializations. and In each iteration of the network Use only For samples deemed to have low loss, update parameters using symmetric cross-entropy as the loss function. The statement is as follows:

[0103] ;

[0104] In the formula, For the network to the first The employee performance prediction value is a continuous value between 0 and 1. To prevent logarithmic underflow, the smoothing term is empirically set to... The training sample set is obtained after noise correction.

[0105] The specific implementation of step S03 is as follows: Construct an employee collaboration relationship graph based on the training sample set. ,in For a set of employee nodes, For a set of cooperative edges, This is the edge weight matrix. For nodes With nodes The weights of the edges between nodes are calculated based on factors such as the number of collaborations and the duration of collaboration. A hierarchical random neighbor sampling strategy is adopted, and the nodes are sampled at each hop of the aggregation phase. Random sampling of neighbors indivual( Take 10-25). Determined through ablation experiments within the range of 5–30 with a step size of 5. A multi-level graph partitioning algorithm was used to divide the graph. The pre-slicing process divides the image into subgraphs with low cross-partition edge ratios. Subgraph features are stored as half-precision floating-point numbers, and a gradient checkpoint recalculation strategy is used to reduce memory peaks. The subgraph feature vectors are input into a causal inference-enhanced dual-stream contrastive evaluation model, while the fact-stream encoder receives real employee feature vectors. ,in For the first The behavioral feature vectors of employees in real business environments, and the feature vectors of intervention samples generated by the counterfactual flow encoder receiving the structural causal model. ,in For the first The behavioral feature vectors of employees under the hypothetical intervention scenario are shared by the two encoders, encompassing all weight matrices of the convolutional and fully connected layers. The budget sublayer uses a differentiable perturbation operator to control the intervention variables. Injecting counterfactual stream intermediate representation, The intervention magnitude is determined by the assignment of intervention variables in the structural causal model, ranging from 10% to 50% of the original values. The directed acyclic graph attention masking mechanism uses the adjacency matrix of the causal directed acyclic graph... ( The number of feature nodes in the causal graph is used as the causal mask for the attention matrix, and the attention weights are... The calculation is expressed as follows:

[0106] ;

[0107] In the formula, For nodes For nodes The original attention score, for No. Line 1 For column elements, a value of 0 masks non-causal paths, while a value of 1 preserves causal paths. To find the sum index, iterate through all... Each feature node. The contrastive learning module embeds facts about the same employee. With counterfactual embedding Imposing a causal invariance constraint, where To embed the vector dimension, the loss is expressed as follows:

[0108] ;

[0109] In the formula, For comparison, the loss weight is empirically set to 1.0. To contrast the learning interval hyperparameter, an empirical value of 1.0 was used, determined through a grid search on the validation set using the convergence of the contrastive loss as an indicator. For traversal Index of all employees except those in the above categories. and All distances are Euclidean distances, dimensionless. The propensity score weighting layer consists of four fully connected multilayer perceptron layers (with hidden layer neuron numbers of 128, 64, 32, and 1 respectively, and activation functions of...). Predicted Samples Conditional probability of accepting intervention Inverse probability weighted term The performance prediction main network loss function is injected. The counterfactual generation subnetwork and the performance prediction main network are optimized alternately. After each round, the performance prediction main network is updated by 1 step, and then the counterfactual generation subnetwork is updated by 1 step. The overall training loss is expressed as follows:

[0110] ;

[0111] In the formula, To predict losses in effectiveness, To estimate the loss for the propensity score, The corresponding weighting coefficients were determined in 100 trials using a Bayesian hyperparameter search. All three loss terms are dimensionless scalars. It is also a dimensionless scalar. The weight sharing coefficient is controlled within the range of 0.1 to 0.3 through sparse regularization of the structural causal model constraints, and the final output is the employee performance evaluation vector. ,in For the output dimension of the performance evaluation vector, when the output is a continuous performance score... When the output is a multidimensional performance level probability distribution This represents the number of grade categories.

[0112] The specific implementation method of step S04 is: calculate the effectiveness diffusion and propagation index based on the employee effectiveness evaluation vector. The formula is expressed as follows:

[0113] ;

[0114] In the formula, For nodes The degree, The maximum node degree in the graph. For nodes Average weight of all edges It is the maximum value among the average edge weights of all nodes in the graph. For nodes At any moment The performance state scalar component value, This represents the maximum value of the scalar component of the performance state of all nodes in the graph. All three ratios are dimensionless. It is a dimensionless scalar. Based on... Adjusting the efficiency diffusion coefficient :when hour Take a value of 0.01 to 0.05; when hour Take a value of 0.05 to 0.15; when hour Use values ​​between 0.15 and 0.30. (Use the graphical Laplace matrix.) (in For degree matrix, , The edge weight matrix is ​​the same as that in step S03. (With the same meaning) The approximate second-order partial differential operator in space, using the Crank-Nicholson implicit difference scheme for discrete-time derivatives, advances step by step to form a system of linear equations, as follows:

[0115] ;

[0116] In the formula, for identity matrix This represents the discrete time step (dimensionless number of steps). For each node diagonal elements Heterogeneous diffusion coefficient diagonal matrix, For all employees at all times The efficiency state vector, components and In Same meaning Let be the external excitation term vector, with each component taking values ​​between 0 and 1, representing the dimensionless excitation intensity. Both sides of the equation are . The dimensionless efficiency state vector is solved using the conjugate gradient method to obtain the efficiency diffusion distribution. The definition of the graph Laplace matrix and the Fiedler vector is as follows:

[0117] ;

[0118] In the formula, for The second smallest eigenvalue (algebraic connectivity). The corresponding eigenvector is the Fiedler vector. Nodes with smaller absolute values ​​of the median components correspond to bottleneck nodes in performance propagation.

[0119] The specific implementation of step S05 is as follows: The efficiency diffusion distribution is input into the optimal transport gradient flow continuous flow algorithm, modeled as a gradient flow on the Wasserstein-2 metric space, and the Jordan-Kindlerey-Otto discretization scheme is adopted. The minimization problem is updated and solved in each step, as described below:

[0120] ;

[0121] In the formula, For the first Step group efficiency distribution (normalized probability density function, dimensionless). The time step is Jordan-Kindleray-Otto (dimensionless). It is the square of the 2-Wasserstein distance (dimensionless). The free energy functional (dimensionless) is expressed as follows:

[0122] ;

[0123] In the formula, Let be a continuous coordinate variable in the efficiency state space, representing the efficiency value points on the support domain of the group efficiency distribution, with values ​​ranging from 0 to 1. Let be a dimensionless potential energy function defined on the efficiency state space, obtained by fitting historical efficiency equilibrium states. For group effectiveness The normalized probability density at point satisfies The first term is the potential energy term, and the second term is the entropy regularization term. Here are the entropy regularization weights (dimensionless), and both are dimensionless. The equation is a dimensionless scalar, and both sides are dimensionless. The Sinkhorn algorithm is used to solve the optimal transport plan at each step in parallel on a graphics processing unit. Momentum extrapolation accelerates the Jordan-Kindleley-Otto iterative convergence, and the predicted population efficiency distribution is output. .

[0124] The specific implementation of step S06 is as follows: The performance equilibrium threshold is determined by statistical analysis of historical performance distribution data over 12 to 36 consecutive months, and the upper limit of the threshold is... and the lower world The statement is as follows:

[0125] ;

[0126] ;

[0127] In the formula, Let be the mean of the rate of change of Wassstein distance in the historical group effectiveness distribution. For the corresponding standard deviation, and All with All quantities are dimensionless rates of change. The mean of the predicted results is based on the population effectiveness distribution. Adjusting the external excitation term parameters with respect to the threshold deviation: when When, increase the weight of training intensity to 0.6–1.0; when When, the external stimulus term maintains its current parameters; when At that time, the weight of the incentive policy intensity is reduced to 0.1-0.4, and the final output is an effectiveness evaluation report and intervention strategy recommendations.

[0128] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2: Human resource data from a manufacturing company is selected as the test object. This company has 1200 employees, the statistical period is set to 90 days, and the behavioral feature dimensions are expanded to 12000 dimensions, forming a scale of... The high-dimensional sparse feature matrix has a non-zero element ratio of approximately 0.06%.

[0129] In step S01, technical personnel collected 90 days of employee behavior data from the business system, collaboration system, attendance system, project management system, performance management system, and training system, covering seven categories: task-related, collaboration-related, time-input-related, output-related, quality-related, learning and growth-related, and organizational interaction-related. The data for each dimension was normalized and binned before being encoded into numerical features; for example, message response time was quantified to minute-level values, and task completion rate was quantified as a percentage between 0 and 1. The high-dimensional sparse feature matrix was stored using a hybrid compressed sparse row format and compressed sparse column format. After being encoded by a hash embedding feature mapping layer, it was mapped to a 256-dimensional dense embedding space using a combination of three hash functions. The collision rate was experimentally verified to be less than 0.8%.

[0130] In step S02, technicians performed noise detection on the labels of 1200 employees. A mixture model with three Gaussian components was used to model the posterior probability distribution of the labels. The number of Gaussian components was determined to be 3 based on a comparative experiment with 2 to 5 components using the Bayesian information criterion. After 100 iterations, the dual-network cross-screening mechanism identified and removed approximately 8% of the noisy samples. Finally, samples with loss values ​​below the lower 20th percentile were retained as a clean training set. After being divided in a 7:1:2 ratio, the training set contained approximately 840 records, the validation set approximately 120 records, and the test set approximately 240 records. Some of the label noise detection results are shown in Table 1.

[0131] Table 1. Statistical Table of Noise Detection in Labels

[0132]

[0133] In step S03, technicians construct an employee collaboration graph based on the training sample set, identifying 1200 employee nodes and 14376 collaboration edges. Edge weights are calculated by combining the number of collaborations and response frequency. Graph clustering pre-segmentation divides the collaboration graph into 8 subgraphs, with the proportion of cross-partition edges controlled below 6%. The hierarchical random neighbor sampling hop count is set to 15, and ablation experiments verify that this setting minimizes gradient variance. The causal inference-enhanced dual-stream contrastive evaluation model uses an initial learning rate... The optimizer is an adaptive moment estimation algorithm, which triggers early stopping after 350 training epochs, with a patience value of 30 epochs. Intervention samples are generated by a structural causal model under the assumption of a 20% increase in training intensity and a 15% decrease in task load, with intervention magnitudes randomly sampled within the range of 10% to 50% of the original values. Figure 2 As shown, the subgraph partitioning results of the employee collaboration relationship graph and the distribution of performance evaluation vectors exhibit obvious community clustering characteristics, with high-performing employees tending to concentrate in the central region of subgraphs with dense collaboration edges. The employee performance evaluation vector output dimension is 64 dimensions, and the corresponding employee performance evaluation vector distribution results are shown in Table 2.

[0134] Table 2 Statistical Table of Employee Performance Level Distribution

[0135]

[0136] In step S04, technicians calculate the performance diffusion and propagation index of each node based on the employee performance evaluation vector, and the performance diffusion and propagation index among all employee nodes. There are a total of 47 high diffusion nodes. There are a total of 386 diffusion nodes in the middle. There are a total of 767 low-diffusion nodes. The diffusion coefficients corresponding to the three performance levels are set to 0.22, 0.09, and 0.03, respectively. The performance diffusion propagation algorithm for the partial differential operator discretization of the heat conduction equation adopts the Crank-Nicolson implicit difference scheme, with a time step set to 0.05. The conjugate gradient method converges in each iteration until the residual is below [a certain value]. The iteration terminates at a certain point, and after 50 iterations, the efficiency diffusion distribution tends to a steady state. Fiedler vector analysis of the graph Laplace matrix identified 12 bottleneck nodes, which are mainly distributed at the boundaries of different departments, corresponding to employees with low frequency of cross-departmental collaboration. For example... Figure 3 As shown, during the steady-state convergence process of the efficiency diffusion propagation algorithm at different efficiency diffusion coefficient levels, the convergence speed of the node efficiency state vector at the high diffusion coefficient level is significantly faster than that at the low diffusion coefficient level.

[0137] In step S05, the technician inputs the efficiency diffusion distribution into the optimal transport gradient flow continuous flow algorithm, with a time step of [missing information]. The entropy regularization weight was set to 0.05, the Sinkhorn algorithm was run in parallel on the GPU, the number of Sinkhorn iterations per step was set to 100, and the momentum extrapolation coefficient was set to 0.9. The algorithm evolved for 20 steps, corresponding to the predicted population performance distribution after 90 days. The Wasserstein distance convergence error of the population performance distribution was below 0.01 after the 15th step, and the prediction results tended to stabilize. The mean of the predicted population performance distribution was 0.61.

[0138] In step S06, technicians calculate the efficacy equilibrium threshold based on 24 consecutive months of historical efficacy distribution data. The mean Wasserstein distance change rate is 0.08, and the standard deviation is 0.03. The upper bound of the efficacy equilibrium threshold is set at 0.11, and the lower bound at 0.05. The estimated Wasserstein distance change rate for the current group efficacy distribution prediction is 0.04, which is lower than the lower bound of 0.05. Therefore, the weight of training intensity in the external incentive item is increased to 0.75, corresponding to the recommended intervention strategy of strengthening training investment. The efficacy evaluation report marks the intervention priorities of 12 bottleneck nodes, such as... Figure 3 As shown, the node number with the smallest absolute value of the Fiedler vector component is highly consistent with the intervention priority sorting.

[0139] like Figure 3 As shown, in this embodiment, the correspondence between the efficiency diffusion propagation process and the bottleneck node identification results of the Fiedler vector is clear. The efficiency diffusion propagation index of the bottleneck node is generally low, which verifies the structural interpretation ability of the graph Laplacian matrix in organizational network efficiency diffusion modeling.

[0140] Compared to traditional methods, this invention brings the following technological advancements. Traditional performance evaluation models only model statistical correlations, failing to distinguish between causal effects and confounding biases. This invention enhances the two-stream comparative evaluation model through causal inference, collaboratively eliminating confounding biases from three dimensions: feature space, structural path, and loss function, ensuring intervention decisions are based on causal estimation rather than spurious correlations. Traditional performance diffusion modeling lacks physical equation constraints, easily leading to extrapolation distortion in sparse data regions. This invention introduces a graph Laplace semidefinite constraint and the unconditional stability of the Crank-Nicolson scheme by discretizing the performance diffusion propagation algorithm using the partial differential operator of the heat conduction equation, making the performance state evolution mathematically consistent and physically interpretable. Traditional methods for predicting population performance distribution lack geometrically optimal transmission path constraints. This invention incorporates distribution evolution into the Wasserstein-2 variational framework through the optimal transmission gradient flow continuous flow algorithm, making the prediction results geometrically optimal and robust to initial perturbations.

[0141] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0142] Table 3. Variable Explanation Table (Part 1)

[0143]

[0144] Table 4. Variable Explanation Table (Part Two)

[0145]

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an enterprise labor efficiency evaluation model, characterized in that, Includes the following steps: Employee behavior data is collected, a high-dimensional sparse feature matrix is ​​constructed, and it is stored in a hybrid format of compressed sparse row and compressed sparse column. After being encoded by a hash embedding feature mapping layer, the employee behavior feature vector set is obtained. Noise detection is performed on the labeled tags in the set of employee behavior feature vectors. The label distribution is modeled using Gaussian mixture distribution, and noisy samples are identified through a dual-network cross-screening mechanism to obtain a noise-corrected training sample set. An employee collaboration relationship graph is constructed based on the training sample set. Hierarchical random neighbor sampling and graph clustering pre-segmentation are used to divide the employee collaboration relationship graph into subgraphs. The causal inference-enhanced dual-flow comparison evaluation model is input to obtain the employee performance evaluation vector. The efficiency diffusion propagation index is calculated based on the employee efficiency evaluation vector. The efficiency diffusion coefficient is adjusted according to the range of the efficiency diffusion propagation index. The efficiency diffusion propagation algorithm, which drives the partial differential operator of the heat conduction equation, is used to update the employee efficiency state vector and obtain the efficiency diffusion distribution. The optimal transmission gradient flow continuous flow algorithm for efficiency diffusion distribution input is subjected to time-series evolution to obtain the group efficiency distribution prediction results; Based on the deviation between the predicted group efficacy distribution and the efficacy equilibrium threshold, the parameters of the external incentive terms are adjusted, and an efficacy assessment report and intervention strategy recommendations are output.

2. The method for constructing an enterprise labor efficiency evaluation model according to claim 1, characterized in that, The high-dimensional sparse feature matrix refers to a feature dimension that reaches... The employee behavior feature matrix, with a magnitude of less than 0.1% and a non-zero rate, is stored using a hybrid compressed sparse row format and a compressed sparse column format, and the computing kernel is optimized using a single instruction multiple data stream integer arithmetic instruction set for sparse data.

3. The method for constructing an enterprise labor efficiency evaluation model according to claim 2, characterized in that, The hash embedding feature mapping layer refers to a network layer that maps a high-dimensional sparse feature matrix to a fixed-dimensional dense embedding space through a low-collision hash function. The collision rate is controlled below the collision rate threshold by combining multiple hash functions.

4. The method for constructing an enterprise labor efficiency evaluation model according to claim 3, characterized in that, The Gaussian mixture distribution modeling of label distribution refers to using a mixture model containing multiple Gaussian components to model the posterior probability distribution of labeled labels in order to distinguish the components corresponding to clean samples and noisy samples. The number of Gaussian components is determined iteratively through the Bayesian information criterion.

5. The method for constructing an enterprise labor efficiency evaluation model according to claim 4, characterized in that, The dual-network cross-selection mechanism refers to training two networks with the same structure but different initializations. In each iteration, each network only uses samples that the other network determines to have low loss for parameter updates, and the loss function uses symmetric cross-entropy instead of standard cross-entropy.

6. The method for constructing an enterprise labor efficiency evaluation model according to claim 5, characterized in that, The hierarchical random neighbor sampling refers to randomly sampling a fixed number of neighbors from each node's neighbor set at each hop aggregation stage of the graph neural network message passing. The number of samples is determined by ablation experiments based on a comprehensive index of gradient variance and convergence speed.

7. The method for constructing an enterprise labor efficiency evaluation model according to claim 6, characterized in that, The graph clustering pre-segmentation refers to using a multi-level graph partitioning algorithm to pre-segment the employee collaboration relationship graph into several subgraphs with low cross-partition edge ratios. The subgraph features are stored using half-precision floating-point numbers, and a gradient checkpoint recalculation strategy is combined to reduce peak memory usage.

8. The method for constructing an enterprise labor efficiency evaluation model according to claim 7, characterized in that, The causal inference-enhanced dual-stream contrastive evaluation model consists of eight components: a fact flow encoder, a counterfactual flow encoder, an interference budget sub-layer, a directed acyclic graph attention masking mechanism, a contrastive learning module, a bias score weighting layer, a counterfactual generation sub-network, and an effectiveness prediction main network. The fact flow encoder and the counterfactual flow encoder share the weight matrix of all convolutional and fully connected layers.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the enterprise labor efficiency evaluation model construction method according to any one of claims 1-8.

10. A system for constructing an enterprise labor efficiency evaluation model, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images