Welfare plan generation method and system based on deep learning
Through deep learning technology, analyzing the correlation of welfare influencing factors, using attention mechanism neural network and optimization algorithm to generate enterprise welfare plans, solving the problems of low accuracy and poor timeliness of welfare plans in the existing technology, and realizing the welfare plan Automatic generation and real-time adjustments.
Patent Information
- Application Number
- CN202410904319.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-08
- Publication Date
- 2024-11-12
AI Technical Summary
The existing enterprise welfare plan generation method relies on simple linear statistics, resulting in low accuracy and inability to adjust in real time. It requires a lot of manual operation when personnel changes and standards are modified, which has a large amount of workload and poor timeliness.
Adopting a welfare plan generation method based on deep learning, analyzing the correlation of welfare factors, using attention mechanism neural network model to predict welfare plans, combining gray wolf optimization algorithm and genetic algorithm to optimize model parameters, to achieve automatic generation and update of welfare plans .
It improves the accuracy and timeliness of welfare plans, reduces manual operation costs, and achieves flexible planning adjustments and real-time response.
Smart Images

Figure CN118941261A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information processing technology, and in particular to a welfare plan generation method and system based on deep learning. Background Art
[0002] At present, the welfare plan data of enterprises mainly depends on the average data of each subsidiary over the years, combined with the implementation of the statistical indicator data, and the welfare plan for the next year is determined according to a simple linear statistical formula. If there are new standard inputs or data changes in the system, the system administrator needs to manually operate the system to make adjustments. The welfare plan generated by this method is less accurate, and when there are personnel changes and standard modifications, the plan cannot be adjusted in real time. The plan adjustment requires a lot of manpower, a large workload and poor timeliness. Summary of the invention
[0003] In order to overcome the problems of low accuracy in generating enterprise welfare plans, large workload in adjusting plans, and poor timeliness, the present invention provides a welfare plan generating method and system based on deep learning.
[0004] On the one hand, the present invention provides a method for generating a welfare plan based on deep learning, the method comprising:
[0005] Based on a plurality of welfare influencing factors set, determining the correlation between each of the welfare influencing factors and the welfare plan to be generated, wherein the plurality of welfare influencing factors include personnel characteristics, welfare types, and welfare payment characteristics;
[0006] Based on the correlation corresponding to each of the welfare influencing factors and the acquired historical welfare-related data corresponding to each of the welfare influencing factors, obtaining welfare-related data corresponding to each of the welfare influencing factors;
[0007] The welfare-related data, the correlations corresponding to each of the welfare influencing factors, and the historical welfare plan data are input into the trained welfare prediction model to obtain the prediction results of the welfare plan; the welfare prediction model is a neural network based on an attention mechanism, and the welfare prediction model determines the attention weight of the attention mechanism based on the correlations corresponding to each of the welfare influencing factors, so as to generate a welfare plan based on the attention weight.
[0008] Optionally, the welfare plan to be generated includes welfare expense data, and the determining of the correlation between each of the welfare influencing factors and the welfare plan to be generated includes:
[0009] The correlation coefficients between the personnel characteristics, the welfare type, the welfare payment characteristics and the welfare expense data are determined by factor analysis as the correlation corresponding to each of the welfare influencing factors.
[0010] Optionally, the historical welfare-related data includes historical personnel characteristic data, historical welfare type data, and historical welfare payment characteristic data, and the welfare-related data corresponding to each welfare influencing factor is obtained based on the correlation corresponding to each welfare influencing factor and the historical welfare-related data corresponding to each welfare influencing factor, including:
[0011] Using the correlation coefficient corresponding to the personnel characteristics as a weighting factor to weight and fuse the historical personnel characteristics data, so as to obtain the welfare-related data corresponding to the personnel characteristics;
[0012] Using the correlation coefficient corresponding to the welfare type as a weighting factor to perform weighted fusion on the historical welfare type data to obtain welfare-related data corresponding to the welfare type;
[0013] The correlation coefficient corresponding to the welfare payment feature is used as a weighting factor to perform weighted fusion on the historical welfare payment feature data to obtain welfare-related data corresponding to the welfare payment feature.
[0014] Optionally, the step of inputting the welfare-related data, the correlations corresponding to the welfare influencing factors, and the historical welfare plan data into a trained welfare prediction model to obtain a prediction result of the welfare plan includes:
[0015] Inputting the welfare-related data and the historical welfare plan data into the convolution unit of the welfare prediction model for multiple convolution and pooling processes to obtain spatial features;
[0016] The spatial features are subjected to global temporal dependency extraction by a gated recurrent unit of the welfare prediction model to obtain global temporal features;
[0017] Inputting the correlation between the global time series feature and each of the welfare influencing factors into the attention unit of the welfare prediction model for attention-based weighted processing to obtain extracted features;
[0018] The extracted features are fully connected to obtain the prediction results of the welfare plan.
[0019] Optionally, the training process of the welfare prediction model includes:
[0020] Acquire training data, and use a sliding window to acquire a corresponding spatiotemporal correlation matrix from the training data as a training sample;
[0021] Use the training samples to preliminarily train the initial welfare prediction model to obtain a primary welfare prediction model;
[0022] Optimizing the model parameters of the attention unit in the primary welfare prediction model based on the gray wolf optimization algorithm to obtain an optimized prediction model;
[0023] Based on a non-dominated sorting genetic algorithm with an elite strategy, the hyperparameters in the optimization prediction model are optimized to obtain a trained welfare prediction model.
[0024] Optionally, optimizing the model parameters of the attention unit in the primary welfare prediction model based on the gray wolf optimization algorithm to obtain the optimized prediction model includes:
[0025] The network parameters of the convolutional unit and the gated recurrent unit in the primary welfare prediction model are fixed, and the network parameters of the attention unit in the current primary welfare prediction model are used as the initial position of each gray wolf;
[0026] determining the fitness of each gray wolf based on the prediction performance index of the primary welfare prediction model;
[0027] The position of the gray wolf is updated based on the fitness and position of each gray wolf until the optimization is completed, the optimal network parameters of the attention unit are obtained, and then the optimized prediction model is obtained.
[0028] Optionally, the non-dominated sorting genetic algorithm with an elite strategy is used to optimize the hyperparameters in the optimization prediction model, including:
[0029] The hyperparameter combination in the optimized prediction model is used as a population, and the optimization is performed with the model evaluation index as the objective function.
[0030] Optionally, before inputting the welfare-related data, the correlations corresponding to the welfare-influencing factors, and the historical welfare plan data into the trained welfare prediction model, the method further includes:
[0031] The welfare-related data and the historical welfare plan data are preprocessed respectively, and the preprocessing includes at least one of outlier removal, missing value filling and normalization.
[0032] On the other hand, the present invention also provides a welfare plan generation system based on deep learning, comprising:
[0033] A correlation analysis module is used to determine the correlation between each of the welfare influencing factors and the welfare plan to be generated based on a plurality of welfare influencing factors set, wherein the plurality of welfare influencing factors include personnel characteristics, welfare types, and welfare payment characteristics; based on the correlation corresponding to each of the welfare influencing factors and the acquired historical welfare association data corresponding to each of the welfare influencing factors, obtain welfare association data corresponding to each of the welfare influencing factors;
[0034] A prediction module is used to input the welfare-related data, the correlations corresponding to each of the welfare influencing factors and the historical welfare plan data into a trained welfare prediction model to obtain a prediction result of the welfare plan; the welfare prediction model is a neural network based on an attention mechanism, and the welfare prediction model determines the attention weight of the attention mechanism based on the correlations corresponding to each of the welfare influencing factors, so as to generate a welfare plan based on the attention weight.
[0035] Optionally, the correlation analysis module includes:
[0036] The correlation determination submodule is used to determine the correlation coefficients between the personnel characteristics, the welfare type, the welfare payment characteristics and the welfare expense data respectively through factor analysis as the correlation corresponding to each of the welfare influencing factors.
[0037] Optionally, the historical welfare-related data includes historical personnel characteristic data, historical welfare type data, and historical welfare payment characteristic data, and the correlation analysis module includes:
[0038] A personnel feature weighting submodule, used to perform weighted fusion on the historical personnel feature data by using the correlation coefficient corresponding to the personnel feature as a weighting factor to obtain welfare-related data corresponding to the personnel feature;
[0039] A welfare type weighting submodule, used for weighting and fusing the historical welfare type data by using the correlation coefficient corresponding to the welfare type as a weighting factor, so as to obtain welfare-related data corresponding to the welfare type;
[0040] The benefit distribution feature weighting submodule is used to perform weighted fusion on the historical benefit distribution feature data by taking the correlation coefficient corresponding to the benefit distribution feature as a weighting factor to obtain the benefit association data corresponding to the benefit distribution feature.
[0041] Optionally, the prediction module includes:
[0042] A convolution submodule, used for inputting the welfare-related data and the historical welfare plan data into the convolution unit of the welfare prediction model for multiple convolution and pooling processes to obtain spatial features;
[0043] A temporal feature extraction submodule, used for extracting global temporal dependencies of the spatial features through the gated recurrent unit of the welfare prediction model to obtain global temporal features;
[0044] An attention processing submodule, used for inputting the correlation between the global time series feature and each of the welfare influencing factors into the attention unit of the welfare prediction model for attention-based weighted processing to obtain extracted features;
[0045] The fully connected submodule is used to perform fully connected processing on the extracted features to obtain the prediction results of the welfare plan.
[0046] Optionally, the system further includes a model training module, which includes:
[0047] A sample acquisition submodule is used to acquire training data and use a sliding window to acquire a corresponding spatiotemporal correlation matrix from the training data as a training sample;
[0048] A preliminary training submodule is used to perform preliminary training on the initial welfare prediction model using training samples to obtain a primary welfare prediction model;
[0049] A gray wolf optimization submodule, used for optimizing the model parameters of the attention unit in the primary welfare prediction model based on the gray wolf optimization algorithm to obtain an optimized prediction model;
[0050] The non-dominated sorting optimization submodule is used to optimize the hyperparameters in the optimization prediction model based on the non-dominated sorting genetic algorithm with an elite strategy to obtain a trained welfare prediction model.
[0051] Optionally, the gray wolf optimization submodule is specifically used for:
[0052] The network parameters of the convolutional unit and the gated recurrent unit in the primary welfare prediction model are fixed, and the network parameters of the attention unit in the current primary welfare prediction model are used as the initial position of each gray wolf;
[0053] determining the fitness of each gray wolf based on the prediction performance index of the primary welfare prediction model;
[0054] The position of the gray wolf is updated based on the fitness and position of each gray wolf until the optimization is completed, the optimal network parameters of the attention unit are obtained, and then the optimized prediction model is obtained.
[0055] Optionally, the non-dominated sorting optimization submodule is specifically used to: use the hyperparameter combination in the optimized prediction model as the population and optimize with the model evaluation index as the objective function.
[0056] Optionally, the system further includes:
[0057] A preprocessing module is used to preprocess the welfare-related data and the historical welfare plan data respectively, wherein the preprocessing includes at least one of outlier removal, missing value filling and normalization.
[0058] On the other hand, the present invention also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus;
[0059] The memory is used to store one or more programs;
[0060] When the one or more programs are executed by the at least one processor, any one of the above-mentioned methods for generating a benefit plan based on deep learning is implemented.
[0061] On the other hand, the present invention also provides a readable storage medium having an execution program stored thereon, which, when executed, implements any one of the above-mentioned deep learning-based welfare plan generation methods.
[0062] Compared with the prior art, the present invention has the following beneficial effects:
[0063] The present invention provides a method and system for generating a welfare plan based on deep learning. By analyzing the correlation between various welfare influencing factors and combining model prediction based on correlation, data association features are captured from different dimensions to improve the accuracy of the predicted welfare plan. Within the model, by combining the correlation between various welfare influencing factors and welfare plans with the attention mechanism, the model can accurately extract key information, improve the model's ability to identify key information, and thus improve the accuracy of prediction. By predicting and generating welfare plans through deep learning methods, the automatic generation and updating of welfare plans are realized, which is convenient for plan adjustment and has high timeliness. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] Figure 1 A flowchart of a method for generating a welfare plan based on deep learning according to the present invention;
[0065] Figure 2 It is a schematic diagram of the processing process of the welfare prediction model of the present invention;
[0066] Figure 3 It is a schematic diagram of the specific processing process of the CNN layer of the present invention;
[0067] Figure 4 A schematic diagram of the training process of the welfare prediction model of the present invention;
[0068] Figure 5 A schematic diagram of the process of optimizing attention weights using the Gray Wolf Optimization Algorithm of the present invention;
[0069] Figure 6 It is a schematic diagram of the optimization process of the genetic algorithm based on non-dominated sorting with elite strategy of the present invention;
[0070] Figure 7 It is a schematic structural diagram of the electronic device of the present invention. DETAILED DESCRIPTION
[0071] At present, the enterprise welfare plan data generation scheme mainly relies on the average data of various welfare types of subordinate units over the years, and combines the actual distribution statistical indicator data to predict and generate the indicator data of the planned amount of each welfare type for the next year according to a simple linear statistical formula. If there are new standards entered or modified in the system, the system administrator needs to manually operate the system again to adjust the planned amounts in batches. This scheme has inaccurate data for plan forecasting and is unable to cope with the actual situation of real-time data changes such as personnel changes and standard modifications. It is necessary to introduce a more flexible and accurate forecasting method to improve the system's forecasting effect and automation level to reduce the waste of manual operation costs. The current welfare plan generation method has the following problems:
[0072] (1) Lack of personalized forecasting: The current plan forecast generation can only provide a simple forecast result of the overall average level, and lacks the ability to personalize forecasts for different subordinate units and individual or type groups of employees in each unit. Since different types of employees have different welfare needs and backgrounds, a single linear statistical method cannot meet the needs of personalized forecasting.
[0073] (2) Data dimension limitation: Traditional methods for data processing and analysis are often limited to limited data dimensions and cannot fully explore the complex relationships and features between data. This results in the prediction results being inaccurate and incomplete and unable to fully reflect real needs.
[0074] (3) Unable to cope with nonlinear relationships: Traditional methods usually assume that the relationship between data is linear and cannot cope with complex nonlinear relationships. However, welfare plan forecasting often involves complex relationships between multiple influencing factors, which requires a more flexible and accurate forecasting model to cope with.
[0075] (4) Low prediction accuracy: Due to the limitations of traditional methods, prediction accuracy is often not high and is prone to deviations and errors, which brings certain uncertainties and risks to the planning and management of planning data.
[0076] (5) Unable to update forecast data in a timely manner: Traditional methods often require a long statistical process of data collection and batch calculation, resulting in the inability to update and optimize forecast plans in a timely manner. In the face of uncertain data such as rapid changes in personnel fluctuations in the company, this slow response speed may lead to inaccurate plan amounts and delays in the implementation of expenses, resulting in a waste of labor costs.
[0077] Based on the above analysis, the embodiment of the present invention provides a method and system for generating a welfare plan based on deep learning. The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings.
[0078] Embodiment 1:
[0079] The present invention provides a method for generating a welfare plan based on deep learning, the flow chart of which is as follows: Figure 1 As shown, including:
[0080] Step S110, based on a plurality of welfare influencing factors set, determining the correlation between each welfare influencing factor and the welfare plan to be generated, wherein the plurality of welfare influencing factors include personnel characteristics, welfare types, and welfare distribution characteristics;
[0081] Step S120, obtaining welfare-related data corresponding to each welfare influencing factor based on the correlation corresponding to each welfare influencing factor and the acquired historical welfare-related data corresponding to each welfare influencing factor;
[0082] Step S130, input the welfare-related data, the correlations corresponding to each welfare influencing factor and the historical welfare plan data into the trained welfare prediction model to obtain the prediction results of the welfare plan; the welfare prediction model is a neural network based on the attention mechanism, and the welfare prediction model determines the attention weight of the attention mechanism based on the correlations corresponding to each welfare influencing factor, so as to generate a welfare plan based on the attention weight.
[0083] In this embodiment, personnel characteristics include employee job status (on-the-job, retired, retired, rehired, seconded), job grade, skill grade, education grade, employee nature, job category, etc. Welfare types include heatstroke prevention and cooling, heating subsidies, funeral subsidies, pensions, physical examination fees, convalescent expenses, hardship subsidies, living subsidies, medical expenses, retirement activity funds, talent subsidies, etc. Welfare payment characteristics include payment coefficients, payment frequency, payment standard levels, etc. The welfare plan to be generated includes welfare expense data within the prediction time, such as welfare expense data for the next year. The present invention analyzes the correlation between various welfare influencing factors, combines model prediction based on correlation, captures data association features from different dimensions, and improves the accuracy of the predicted welfare plan. Within the model, by combining the correlation between various welfare influencing factors and welfare plans with the attention mechanism, the model can accurately extract key information, improve the model's ability to identify key information, and thus improve the prediction accuracy. The welfare plan is predicted and generated by a deep learning method, and the automatic generation and update of the welfare plan is realized, which is convenient for plan adjustment and has high timeliness.
[0084] In one implementation, determining the correlation between each welfare influencing factor and the welfare plan to be generated in the above step S110 includes the following process:
[0085] The factor analysis method is used to determine the correlation coefficients between personnel characteristics, welfare types, welfare payment characteristics and welfare cost data respectively, as the corresponding correlations of various welfare influencing factors.
[0086] In this implementation, the factor analysis method is a statistical analysis method that uses a statistical index system to analyze the degree of influence of each factor in the total change of a phenomenon. This example first determines the indicator to be analyzed (i.e., welfare cost data), then determines the welfare influencing factors (personnel characteristics, welfare types, welfare payment characteristics) that affect the indicator and their relationship with the indicator, and finally determines the degree of influence of each welfare influencing factor, and uses a correlation coefficient to represent the correlation corresponding to each welfare influencing factor, for example, using the Spearman correlation coefficient as the correlation corresponding to each welfare influencing factor.
[0087] In another implementation, the historical welfare-related data includes historical personnel characteristic data, historical welfare type data, and historical welfare payment characteristic data. The step S120 above, based on the correlation corresponding to each welfare influencing factor and the historical welfare-related data corresponding to each welfare influencing factor, obtains the welfare-related data corresponding to each welfare influencing factor, including the following process:
[0088] The correlation coefficient corresponding to the personnel characteristics is used as a weighting factor to perform weighted fusion on the historical personnel characteristics data to obtain the welfare-related data corresponding to the personnel characteristics;
[0089] The correlation coefficient corresponding to the welfare type is used as a weighting factor to weight and fuse the historical welfare type data to obtain the welfare-related data corresponding to the welfare type;
[0090] The correlation coefficient corresponding to the welfare payment characteristics is used as a weighting factor to perform weighted fusion on the historical welfare payment characteristic data to obtain the welfare-related data corresponding to the welfare payment characteristics.
[0091] In this embodiment, historical personnel characteristic data, historical welfare type data and historical welfare issuance characteristic data are welfare-related data within a preset historical period, for example, personnel characteristic data, welfare type data and welfare issuance characteristic data of the previous year. Historical personnel characteristic data include data such as employee job status (on-the-job, retired, retired, rehired, seconded), job grade, skill grade, education grade, employee nature, and job category of the previous year. Historical welfare type data include heatstroke prevention and cooling, heating subsidies, funeral subsidies, pensions, physical examination fees, convalescent expenses, hardship subsidies, living subsidies, medical expenses, retirement activity funds, talent subsidies, etc. of the previous year. Historical welfare issuance characteristic data include issuance coefficients, issuance frequencies, issuance standard levels, etc. of the previous year. The correlation coefficients of various welfare influencing factors are used as weighting factors of the corresponding historical welfare-related data (historical personnel characteristic data, historical welfare type data, and historical welfare issuance characteristic data) to perform weighted fusion processing on the corresponding historical welfare-related data to obtain welfare-related data of various welfare influencing factors. The association at the data level is achieved by weighted fusion processing of historical welfare-related data.
[0092] In another implementation, in the above step S130, the welfare-related data, the correlations corresponding to the welfare influencing factors, and the historical welfare plan data are input into the trained welfare prediction model to obtain the prediction results of the welfare plan, including:
[0093] The welfare-related data and historical welfare plan data are input into the convolution unit of the welfare prediction model for multiple convolution and pooling processes to obtain spatial features;
[0094] The spatial features are passed through the gated recurrent unit of the welfare prediction model to extract the global temporal dependency and obtain the global temporal features;
[0095] The global time series features and the correlation between each welfare influencing factor are input into the attention unit of the welfare prediction model for attention-based weighted processing to obtain the extracted features;
[0096] The extracted features are fully connected to obtain the prediction results of the welfare plan.
[0097] In this implementation, the welfare prediction model includes a convolution unit, a gated recurrent unit, and an attention unit. The convolution unit is used to extract features from input data, and the gated recurrent unit is used to perform time series prediction. Exemplarily, the welfare prediction model is a CNN-GRU-Attention (Convolutional Neural Network-Gate Recurrent Unit-Attention) time series model, such as Figure 2 As shown in the figure, the CNN-GRU-Attention time series model includes CNN layers, GRU layers, and Attention layers connected in sequence. The input sequence (such as X t0 , X t1 , X t2 ) CNN convolution extracts spatial features of different dimensions to obtain long sequence features L t0 , L t1 , L t2 , and then the GRU layer captures the global temporal features of the sequence data (H t0 , H t1 , H t2 ), the correlation between the global time series features and each welfare influencing factor is calculated by the Attention layer to calculate the importance of each time step in the sequence (the calculation results are such as a1, a2, a3), and the weighted GRU output is obtained by weighted summation; the GRU output is mapped to the feature space by the fully connected layer (Dense layer) to achieve dimensionality adjustment, and finally the prediction result of the welfare plan is output by the output layer. For example, Figure 3As shown in the figure, the specific processing process of the CNN layer is: extract the features of the input sequence data through the stacking of multiple CONV-1d convolutional layers and pooling layers. Each convolutional layer contains multiple convolution kernels to capture different features, and the pooling layer is used to reduce the dimension of the data. The convolution kernel size is 3, the step size is 1, and the ReLU activation function is used. Finally, the fully connected layer (FC) is used to output the long sequence features (such as L t0 ). This example can improve the forecasting process's ability to identify key information and enable the automatic generation and update of benefit plans.
[0098] Exemplarily, the processing formula involved in the above GRU layer is as follows:
[0099] GRU state update formula:
[0100]
[0101] Where:h t is the state at time t (current time step); h t-1 is the hidden state at time t-1; Z t is the output of the update gate, which represents the weight of the past state and the current input; ⊙ is the element-by-element multiplication operation; is the candidate state at time t.
[0102] The calculation formula of Update Gate is: t =σ(w z ·[h t-1 ,x t ])
[0103] The calculation formula of reset gate is: t =σ(w r ·[h t-1 ,x t ])
[0104] In the formula, r t is the output of the reset gate, w z and w r They represent the weight matrices corresponding to the update gate and the reset gate respectively; σ is the sigmoid activation function; [h t-1 ,x t ] is the hidden state h at time t-1 t-1 and the input x at time t t Vectors stitched together.
[0105] Candidate status update formula:
[0106] Among them, w cis the weight matrix corresponding to the candidate state; x t is the input at time t; b c is the bias term, and tanh is the hyperbolic tangent activation function.
[0107] In another implementation, the model needs to be trained before the welfare plan prediction. The training process of the welfare prediction model includes the following process:
[0108] Obtain training data, and use a sliding window to obtain the corresponding spatiotemporal correlation matrix from the training data as a training sample;
[0109] Use the training samples to preliminarily train the initial welfare prediction model to obtain a primary welfare prediction model;
[0110] Based on the gray wolf optimization algorithm, the model parameters of the attention unit in the primary welfare prediction model are optimized to obtain the optimized prediction model;
[0111] Based on the non-dominated sorting genetic algorithm with elite strategy, the hyperparameters in the optimization prediction model are optimized to obtain the trained welfare prediction model.
[0112] In this implementation, the training data includes the collected historical welfare plan and welfare cost implementation data, including the report csv (Comma Separate Values, character separator) file of welfare plan, welfare type, payment amount, payment time and other data. Figure 4 As shown, the training data is preprocessed and then input into the deep learning model for feature extraction and parameter adjustment. Based on the training termination conditions, it is determined whether the training is completed. If so, the test data is used to test the training model after the training is completed, otherwise the training continues. It is determined whether the test result meets the expectations. If so, the training ends and the final model is obtained, otherwise the training of the deep learning model continues.
[0113] In order to ensure the quality and integrity of the data, the training data is preprocessed, including outlier removal, missing value filling and normalization. Missing value filling uses mean filling or median filling to fill missing values in sequence data; normalization is used to scale data of different sizes to the same range. According to the relationship between the time when the data occurs and the various types of welfare expenses, the data is divided into training set and test set. Use a sliding window to slide in the training set and test set to generate a spatiotemporal correlation matrix as training samples and test samples. The training data is divided into training set and test set in a ratio of 8:2. The sliding window is used to slide in the training set and test set with a time step of k as the interval, that is, using k time periods, to predict the next time period or the next data, and create a k*s order spatiotemporal correlation matrix, where k is the time step and s is the welfare type X. s,tRepresents the implementation amount data of welfare type s at time t. During the training process, the model is first trained using the training set; then the gray wolf algorithm is used to optimize the weight of the attention mechanism to enhance the model's attention to key information. Finally, NSGA-II (Non-dominated Sorting Genetic Algorithm-II) is used to optimize the overall hyperparameters, find the optimal parameter combination, and obtain the final welfare prediction model.
[0114] Exemplarily, the model parameters of the attention unit in the primary welfare prediction model are optimized based on the gray wolf optimization algorithm to obtain an optimized prediction model, including:
[0115] The network parameters of the convolutional unit and the gated recurrent unit in the primary welfare prediction model are fixed, and the network parameters of the attention unit in the current primary welfare prediction model are used as the initial positions of each gray wolf;
[0116] The fitness of each gray wolf was determined based on the prediction performance index of the primary welfare prediction model;
[0117] The position of the gray wolf is updated based on the fitness and position of each gray wolf until the optimization is completed, the optimal network parameters of the attention unit are obtained, and then the optimized prediction model is obtained.
[0118] In this implementation, the network parameters of the convolutional unit and the gated recurrent unit in the primary welfare prediction model are fixed, and the process of optimizing the attention weights using the gray wolf optimization algorithm is as follows: Figure 5 As shown in FIG. 1 , the specific process includes the following steps: Initialize the gray wolf group: randomly generate a certain number of gray wolves and assign a random initial position to each gray wolf, for example, initialize the positions of gray wolves A, B, and C, and the position of each gray wolf represents the weight and bias of the neural network. Calculate the objective function value of the gray wolf group: calculate the fitness of each gray wolf and the current value of the gray wolf group. The fitness uses the performance indicators of the neural network, such as the mean square error. Update the position of the gray wolf and update the objective function value of the gray wolf group: according to the fitness of each gray wolf and the position of other gray wolves, update the position of each gray wolf (that is, update the weight and bias). This step simulates the search, siege, chase, and escape behaviors. The process of updating the position uses a standard optimization algorithm to calculate the average minimum value of the new position. Determine whether the termination condition is met. If so, output the optimal gray wolf position and the objective function value of the optimal gray wolf. Otherwise, continue to calculate the objective function value of the gray wolf group. The optimal gray wolf position (the position of the gray wolf with the best fitness) corresponds to the optimal weight and bias of the attention unit.
[0119] Exemplarily, the hyperparameters in the optimization prediction model are optimized based on a non-dominated sorting genetic algorithm with an elite strategy, including: taking the hyperparameter combination in the optimization prediction model as a population and optimizing with a model evaluation index as an objective function.
[0120] In this implementation, the hyperparameters optimized by the genetic algorithm based on non-dominated sorting with elite strategy include: hyperparameters of CNN layer, hyperparameters of GRU layer, hyperparameters of Attention layer, hyperparameters related to training, and hyperparameters of NSGA-II algorithm, wherein the hyperparameters of CNN layer include kernel size, i.e., the filter size of convolution operation, number of filters, i.e., the number of filters in each convolution layer, stride, i.e., the stride of convolution kernel movement, and padding, i.e., the padding method of convolution operation, such as 'valid' or 'same'. The hyperparameters of GRU layer include number of GRU units, i.e., the number of neurons in each GRU layer, activation function, i.e., the activation function used by GRU layer, and dropout rate, i.e., the dropout rate to prevent overfitting. The hyperparameters of Attention layer include number of attention heads, i.e., the number of heads parameter in multi-head attention mechanism. The training-related hyperparameters include the learning rate, which is the learning rate of the optimization algorithm, the batch size, which is the number of samples used each time the weight is updated, and the optimizer, which is the optimized Adam parameter. The hyperparameters of the NSGA-II algorithm include the population size, which is the number of individuals in each generation, the crossover rate, which is the probability of the crossover operation, and the mutation rate, which is the probability of the mutation operation.
[0121] The model evaluation indicators include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), goodness of fit R 2 (It is used to measure the degree to which the model explains the variability of the data. The value range is 0 to 1. The closer the value is to 1, the better the model fit is). The smaller the value of MAE, RMSE and MAPE, the better. The specific expressions of each model evaluation index are:
[0122]
[0123] In the formula, y i represents the actual value of the i-th sample, is the predicted value of the ith sample, and m is the number of samples.
[0124] For example, Figure 6 As shown in the figure, the optimization process of the genetic algorithm based on non-dominated sorting with elite strategy is implemented as follows:
[0125] (1) Initialize the population: Randomly generate the original population P0 (t = 0) and randomly distribute it throughout the space. The population is a collection of solutions, which is a combination of the above hyperparameters and output evaluation indicators.
[0126] (2) Fast non-dominated sorting: The population M is stratified according to the non-inferior solution level of the individual to obtain Fi, which is used to make the solution close to the Pareto optimal solution. This is a cyclic fitness value classification process. First, the non-dominated solution set in the group is found, denoted as F1, and all its individuals are assigned a non-dominated order i rank =1(where i rank is the non-dominated sequence value of individual i), and remove it from the entire population M, and then continue to find the non-dominated solution set in the remaining population, recorded as F2, and the individuals in F2 are assigned i rank =2, and so on until the entire population is stratified and the non-dominated order values in the Fi layers are the same.
[0127] (3) Selecting individuals from the current population (the parent population Pt) for selection, crossover, and mutation to generate a new population Qt (i.e., the daughter population), for example, by using a binary tournament method to select individuals from the population.
[0128] Selection: The operation imitates the law of "survival of the fittest" in nature. If the individual's fitness is high, it has a greater probability of being inherited to the next generation, otherwise the probability is small. Selection operations include roulette selection, sorting selection, optimal individual preservation, random league selection, etc.
[0129] Crossover: Simulate the crossover phenomenon of chromosomes in nature to generate new individuals. This step determines the global search capability of the algorithm.
[0130] Mutation: Simulates the genetic mutation of organisms. Like the crossover operation, it is used to generate new individuals.
[0131] (4) Combine the paternal population Pt and the daughter population Qt into population Rt.
[0132] (5) Perform a fast non-dominated sort on the population Rt. Classify the solutions into different fronts or levels, where the solutions in the first front are not dominated by any other solutions, while the solutions in the second front are only dominated by the solutions in the first front, and so on.
[0133] (6) Select N individuals through crowding calculation and elite retention strategy to form a new generation of population Pt+1, and record the number of population updates G;
[0134] (7) Determine whether G is greater than the maximum update number Gen. If so, take the parameters determined by the current population as the optimal set of target set parameters; otherwise, go to (3).
[0135] The present invention strategically trains and optimizes the prediction model, adopts the gray wolf optimization algorithm and the multi-objective NSGA-II genetic algorithm, and optimizes and trains the input parameters and the output error MA, the root mean square error RMSE, and the mean absolute percentage error MAPE in an iterative manner, thereby improving the accuracy of model prediction. In another implementation, before inputting the welfare-related data, the correlation corresponding to each welfare influencing factor, and the historical welfare plan data into the trained welfare prediction model, the method also includes: preprocessing the welfare-related data and the historical welfare plan data respectively, and the preprocessing includes at least one of outlier removal, missing value filling, and normalization.
[0136] In this implementation, the input data is preprocessed before being input into the welfare prediction model. Outlier removal is used to remove abnormal data, missing value filling is used to supplement missing values, and normalization is used to scale data of different sizes to the same range. Missing value filling includes mean value filling and median filling. Mean value filling means that the missing data is filled by averaging the two nearest valid values. The formula is as follows:
[0137]
[0138] In the formula, represents the missing value of the kth sequence at time t, represents the sampling value of the kth sequence at time t-1, Represents the sampling value of the kth sequence at time t+1.
[0139] Median filling means that when there is an obvious skewness in the adjacent data of the missing data, the adjacent data of the missing item cannot accurately express the data characteristics, so the median is used to fill in.
[0140] The Min-Max normalization method is used for normalization, and all eigenvalues are normalized to the interval [0,1]. The specific normalization formula is:
[0141]
[0142] In the formula, x i Represents the i-th value of the time series, min represents the minimum value of feature x in the time series, and max represents the maximum value of feature x in the time series.
[0143] The present invention is described below with a specific embodiment, which specifically includes the following steps:
[0144] Step 1: Based on the set multiple welfare influencing factors, determine the correlation between each welfare influencing factor and the welfare plan to be generated.
[0145] Step 2: Obtain historical welfare-related data and historical welfare plan data, and pre-process the acquired data to obtain welfare-related data based on the historical welfare-related data and correlation.
[0146] Step 3: Build a CNN-GRU-Attention time series model, including CNN convolution for feature extraction of input data, and neural network GRU for time series prediction and weighted Attention layer of attention mechanism.
[0147] Step 4: Use the Gray Wolf Optimization Algorithm and the multi-objective NSGA-II genetic algorithm to train the model, obtain the optimized CNN-GRU-Attention welfare prediction model and save it.
[0148] Step 5: Input the correlation of each welfare influencing factor, welfare-related data and historical welfare plan data into the optimized CNN-GRU-Attention welfare prediction model, and output the prediction results of the welfare plan.
[0149] The present invention is based on the correlation between different features and processes data based on the correlation, and predicts by inputting the correlation into the model and the action of the attention mechanism. The time series data of the welfare plan is modeled using a deep learning model to capture the time series characteristics and complex correlations in the data, and improve the prediction accuracy and generalization ability. By combining CNN, GRU and the attention mechanism, and introducing the Gray Wolf Optimization Algorithm and the NSGA-II Genetic Algorithm, the accuracy and robustness of the welfare plan prediction are significantly improved. By optimizing the weights and overall hyperparameters of the attention mechanism, the model performs better when processing complex time series data, can effectively capture the key features in the data, and is suitable for prediction scenarios of various welfare plans. Through the deep learning model algorithm, it is possible to provide more accurate welfare plan prediction results, greatly improving the prediction accuracy.
[0150] The present invention designs a real-time responsive prediction system that can timely adapt to changes in welfare plan data, provide a highly flexible prediction solution, support decision makers to timely adjust welfare plan strategies, and improve the real-time and flexibility of welfare plan generation. The present invention takes into account the characteristics of time series data and the complexity of welfare plans, and designs a comprehensive prediction method and system that can effectively respond to various welfare plan prediction needs; through the optimization of the model tuning algorithm, the model's ability to identify key information and processing efficiency are improved, and it has strong technical innovation.
[0151] The present invention can realize systematic welfare plan management. Based on the present invention, a complete welfare plan management system is generated, including data collection, preprocessing, model training, prediction generation, result optimization and other links, realizing a full-process closed loop of welfare plan prediction and management.
[0152] Embodiment 2:
[0153] Based on the same inventive concept, the present invention also provides a welfare plan generation system based on deep learning, comprising:
[0154] A correlation analysis module is used to determine the correlation between each welfare influencing factor and the welfare plan to be generated based on multiple welfare influencing factors that are set, wherein the multiple welfare influencing factors include personnel characteristics, welfare types, and welfare payment characteristics; based on the correlation corresponding to each welfare influencing factor and the historical welfare association data corresponding to each welfare influencing factor obtained, obtain the welfare association data corresponding to each welfare influencing factor;
[0155] The prediction module is used to input welfare-related data, the correlations corresponding to various welfare influencing factors and historical welfare plan data into the trained welfare prediction model to obtain the prediction results of the welfare plan; the welfare prediction model is a neural network based on the attention mechanism, and the welfare prediction model determines the attention weight of the attention mechanism based on the correlations corresponding to various welfare influencing factors, so as to generate a welfare plan based on the attention weight.
[0156] In a possible implementation, the correlation analysis module includes:
[0157] The correlation determination submodule is used to determine the correlation coefficients between personnel characteristics, welfare types, welfare payment characteristics and welfare cost data respectively through factor analysis as the corresponding correlation of each welfare influencing factor.
[0158] In a possible implementation, the historical welfare-related data includes historical personnel characteristic data, historical welfare type data, and historical welfare payment characteristic data, and the correlation analysis module includes:
[0159] The personnel characteristic weighting submodule is used to weight and fuse the historical personnel characteristic data by using the correlation coefficient corresponding to the personnel characteristic as the weighting factor to obtain the welfare-related data corresponding to the personnel characteristic;
[0160] The welfare type weighting submodule is used to weight and fuse the historical welfare type data using the correlation coefficient corresponding to the welfare type as a weighting factor to obtain the welfare-related data corresponding to the welfare type;
[0161] The distribution feature weighting submodule is used to use the correlation coefficient corresponding to the welfare distribution feature as a weighting factor to weightedly fuse the historical welfare distribution feature data to obtain the welfare-related data corresponding to the welfare distribution feature.
[0162] In a possible implementation, the prediction module includes:
[0163] The convolution submodule is used to input the welfare-related data and the historical welfare plan data into the convolution unit of the welfare prediction model for multiple convolution and pooling processes to obtain spatial features;
[0164] The temporal feature extraction submodule is used to extract the global temporal dependency of the spatial features through the gated recurrent unit of the welfare prediction model to obtain the global temporal features;
[0165] The attention processing submodule is used to input the correlation between the global time series features and each welfare influencing factor into the attention unit of the welfare prediction model for attention-based weighted processing to obtain the extracted features;
[0166] The fully connected submodule is used to perform fully connected processing on the extracted features to obtain the prediction results of the welfare plan.
[0167] In a possible implementation, the system further includes a model training module, and the model training module includes:
[0168] The sample acquisition submodule is used to acquire training data and use a sliding window to acquire the corresponding spatiotemporal correlation matrix from the training data as a training sample;
[0169] A preliminary training submodule is used to perform preliminary training on the initial welfare prediction model using training samples to obtain a primary welfare prediction model;
[0170] The gray wolf optimization submodule is used to optimize the model parameters of the attention unit in the primary welfare prediction model based on the gray wolf optimization algorithm to obtain an optimized prediction model;
[0171] The non-dominated sorting optimization submodule is used to optimize the hyperparameters in the optimization prediction model based on the non-dominated sorting genetic algorithm with elite strategy to obtain a trained welfare prediction model.
[0172] In a possible implementation, the gray wolf optimization submodule is specifically used for:
[0173] The network parameters of the convolutional unit and the gated recurrent unit in the primary welfare prediction model are fixed, and the network parameters of the attention unit in the current primary welfare prediction model are used as the initial positions of each gray wolf;
[0174] The fitness of each gray wolf was determined based on the prediction performance index of the primary welfare prediction model;
[0175] The position of the gray wolf is updated based on the fitness and position of each gray wolf until the optimization is completed, the optimal network parameters of the attention unit are obtained, and then the optimized prediction model is obtained.
[0176] In a possible implementation, the non-dominated sorting optimization submodule is specifically used to optimize the hyperparameter combination in the prediction model as a population and to optimize the model evaluation index as an objective function.
[0177] In a possible implementation, the system further includes:
[0178] The preprocessing module is used to preprocess the welfare-related data and the historical welfare plan data respectively, and the preprocessing includes at least one of outlier removal, missing value filling and normalization.
[0179] The system of the present invention includes a correlation analysis module and a prediction module, based on which a complete welfare plan generation system is constructed. The system has the characteristics of high automation and strong real-time performance, and can quickly respond to changes in welfare plans and provide accurate prediction results. In addition, the system of the present invention has a friendly interface, simple operation, and is user-friendly, which improves the efficiency of welfare plan management and reduces manual operation costs.
[0180] Embodiment 3:
[0181] like Figure 7 As shown, the present invention also provides an electronic device, which may be a computer device, a single-chip device, an intelligent mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected via a bus; the memory may be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory may also be used to store data, which may be called and / or modified when the instructions are executed.
[0182] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a storage medium to implement the corresponding method flow or corresponding functions, so as to implement the steps of a welfare plan generation method based on deep learning in the above embodiment.
[0183] Example 4
[0184] Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It can be understood that the storage medium here can include both built-in storage media in electronic devices and, of course, extended storage media supported by electronic devices. The storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions can be one or more execution programs (including program codes). It should be noted that the storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage. The processor loads and executes one or more instructions stored in the storage medium, which can implement the steps of a method for generating a benefit plan based on deep learning in the above embodiment.
[0185] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0186] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0187] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0188] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims to be approved.
Claims
1. A method for generating benefit plans based on deep learning, It is characterized in that include: Based on a plurality of welfare influencing factors set, determining the correlation between each of the welfare influencing factors and the welfare plan to be generated, wherein the plurality of welfare influencing factors include personnel characteristics, welfare types, and welfare payment characteristics; Based on the correlation corresponding to each of the welfare influencing factors and the acquired historical welfare-related data corresponding to each of the welfare influencing factors, obtaining welfare-related data corresponding to each of the welfare influencing factors; The welfare-related data, the correlations corresponding to each of the welfare influencing factors, and the historical welfare plan data are input into the trained welfare prediction model to obtain the prediction results of the welfare plan; the welfare prediction model is a neural network based on an attention mechanism, and the welfare prediction model determines the attention weight of the attention mechanism based on the correlations corresponding to each of the welfare influencing factors, so as to generate a welfare plan based on the attention weight.
2. The method according to claim 1, It is characterized in that The welfare plan to be generated includes welfare expense data, and the determining of the correlation between each of the welfare influencing factors and the welfare plan to be generated includes: The correlation coefficients between the personnel characteristics, the welfare type, the welfare payment characteristics and the welfare expense data are determined by factor analysis as the correlation corresponding to each of the welfare influencing factors.
3. The method according to claim 2, It is characterized in that The historical welfare-related data includes historical personnel characteristic data, historical welfare type data, and historical welfare payment characteristic data. The welfare-related data corresponding to each welfare influencing factor is obtained based on the correlation corresponding to each welfare influencing factor and the historical welfare-related data corresponding to each welfare influencing factor, including: Using the correlation coefficient corresponding to the personnel characteristics as a weighting factor to weight and fuse the historical personnel characteristics data, so as to obtain the welfare-related data corresponding to the personnel characteristics; Using the correlation coefficient corresponding to the welfare type as a weighting factor to perform weighted fusion on the historical welfare type data to obtain welfare-related data corresponding to the welfare type; The correlation coefficient corresponding to the welfare payment feature is used as a weighting factor to perform weighted fusion on the historical welfare payment feature data to obtain welfare-related data corresponding to the welfare payment feature.
4. The method according to claim 1, It is characterized in that The step of inputting the welfare-related data, the correlations corresponding to the welfare-influencing factors, and the historical welfare plan data into the trained welfare prediction model to obtain the prediction results of the welfare plan includes: Inputting the welfare-related data and the historical welfare plan data into the convolution unit of the welfare prediction model for multiple convolution and pooling processes to obtain spatial features; The spatial features are subjected to global temporal dependency extraction by a gated recurrent unit of the welfare prediction model to obtain global temporal features; Inputting the correlation between the global time series feature and each of the welfare influencing factors into the attention unit of the welfare prediction model for attention-based weighted processing to obtain extracted features; The extracted features are fully connected to obtain the prediction results of the welfare plan.
5. The method according to claim 4, It is characterized in that The training process of the welfare prediction model includes: Acquire training data, and use a sliding window to acquire a corresponding spatiotemporal correlation matrix from the training data as a training sample; Use the training samples to preliminarily train the initial welfare prediction model to obtain a primary welfare prediction model; Optimizing the model parameters of the attention unit in the primary welfare prediction model based on the gray wolf optimization algorithm to obtain an optimized prediction model; Based on a non-dominated sorting genetic algorithm with an elite strategy, the hyperparameters in the optimization prediction model are optimized to obtain a trained welfare prediction model.
6. The method according to claim 5, It is characterized in that The method optimizes the model parameters of the attention unit in the primary welfare prediction model based on the gray wolf optimization algorithm to obtain the optimized prediction model, including: The network parameters of the convolutional unit and the gated recurrent unit in the primary welfare prediction model are fixed, and the network parameters of the attention unit in the current primary welfare prediction model are used as the initial position of each gray wolf; determining the fitness of each gray wolf based on the prediction performance index of the primary welfare prediction model; The position of the gray wolf is updated based on the fitness and position of each gray wolf until the optimization is completed, the optimal network parameters of the attention unit are obtained, and then the optimized prediction model is obtained.
7. The method according to claim 5, It is characterized in that The non-dominated sorting genetic algorithm with an elite strategy is used to optimize the hyperparameters in the optimization prediction model, including: The hyperparameter combination in the optimized prediction model is used as a population, and the optimization is performed with the model evaluation index as the objective function.
8. The method according to any one of claims 1 to 7, It is characterized in that Before inputting the welfare-related data, the correlations corresponding to the welfare-influencing factors, and the historical welfare plan data into the trained welfare prediction model, the method further includes: The welfare-related data and the historical welfare plan data are preprocessed respectively, and the preprocessing includes at least one of outlier removal, missing value filling and normalization.
9. A welfare plan generation system based on deep learning, It is characterized in that include: A correlation analysis module is used to determine the correlation between each of the welfare influencing factors and the welfare plan to be generated based on a plurality of welfare influencing factors set, wherein the plurality of welfare influencing factors include personnel characteristics, welfare types, and welfare payment characteristics; based on the correlation corresponding to each of the welfare influencing factors and the acquired historical welfare association data corresponding to each of the welfare influencing factors, obtain welfare association data corresponding to each of the welfare influencing factors; A prediction module is used to input the welfare-related data, the correlations corresponding to each of the welfare influencing factors and the historical welfare plan data into a trained welfare prediction model to obtain a prediction result of the welfare plan; the welfare prediction model is a neural network based on an attention mechanism, and the welfare prediction model determines the attention weight of the attention mechanism based on the correlations corresponding to each of the welfare influencing factors, so as to generate a welfare plan based on the attention weight.
10. An electronic device, It is characterized in that include: at least one processor and memory; The memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the benefit plan generation method based on deep learning as described in any one of claims 1 to 8 is implemented.
Citation Information
Cited By
Field service meal supplement data processing method and program product
CN119623985A
