Personnel multi-dimensional score prediction method
By building an encoder and decoder model, combining non-uniform sampling and attention modules, the black box problem of employee score prediction method is solved, the interpretability and generalization ability of the model are realized, and the prediction accuracy and adaptability are improved.
Patent Information
- Application Number
- CN202510363598.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-26
- Publication Date
- 2025-07-11
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing employee score prediction methods have black box problems, which is difficult to resolve the solution process, and the generalization ability is limited, making it difficult to adapt to the needs of different corporate environments and jobs.
The multi-dimensional scoring prediction method is adopted to build an encoder and decoder model, combining the non-uniform sampling mechanism, self-attention module and cross-attention module, and using the Huber loss function for training to achieve the interpretability and generalization performance of the model.
It improves the prediction accuracy and adaptability of the model, can adapt to different enterprise environments and job needs, while retaining the interpretability of the model.
Smart Images
Figure CN120297792A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of human resource data processing, and specifically to a method for predicting multi-dimensional scores of personnel. Background Art
[0002] Currently, in human resource management, predicting scores of employees is crucial. It can objectively and fairly evaluate employees' performance, provide a basis for enterprise decision-making such as promotions and salary adjustments, and also motivate employees to improve themselves and enhance their work motivation. At present, related research has been continuously deepening, shifting from traditional single evaluation to multi-dimensional evaluation, and with the help of technologies such as big data and artificial intelligence, striving to make score prediction more scientific, accurate, and efficient. However, the existing prediction methods still have the following problems: the prediction algorithms adopted have black box problems, making it difficult to understand their decision-making processes and prone to biases. In addition, the generalization ability of the prediction model is limited and it is difficult to adapt to different enterprise environments and job requirements. Summary of the Invention
[0003] The technical problem to be solved by the present invention is to provide a method for predicting multi-dimensional scores of personnel, which predicts based on multi-dimensional time value data, while the generalization performance of the model is strong, the interpretability of the model is retained, the prediction results are accurate, and it can adapt to different enterprise environments and job requirements.
[0004] The technical solution of the present invention is as follows:
[0005] A method for predicting multi-dimensional scores of personnel specifically includes the following steps:
[0006] (1) Collect a training data set: Collect the historical data set of multi-dimensional scores of personnel and perform data cleaning operations to obtain initial data, and then obtain multi-dimensional time value data from the initial data according to the non-uniform sampling mechanism to form a training data set;
[0007] (2) Construct a model: Construct a multi-dimensional score prediction model for personnel, including an encoder and a decoder;
[0008] (3) Model training: Input the training data set into the multi-dimensional score prediction model for personnel for training to obtain a trained multi-dimensional score prediction model for personnel;
[0009] (4) Model fitting and training: Fit the basis function of the trained multi-dimensional score prediction model for personnel, and then input the time point to be predicted into the basis function to obtain the prediction result.
[0010] The distribution formula of the multi-dimensional time value data sampled under the non-uniform sampling mechanism is as follows (1):
[0011]
[0012] In Equation (1), T represents the actual time value, and R represents the sample resolution. Divide the actual time by the sample resolution to convert data with different time resolutions into a unified time step x that is convenient for processing in non-uniform sampling.
[0013] The processing process of the non-uniform sampling mechanism is specifically shown in Equation (2) below:
[0014]
[0015] In Equation (2), p non-usm (x) represents the value of the probability density function at which the time step x is sampled, and the normalization constant σ is the width of non-uniform sampling;
[0016] The non-uniform sampling mechanism specifically samples x and constructs (t, v) tuples from it, where v is the value corresponding to time t. For the input That is, an input sequence of M variables of length L, after padding, the input sequence is divided into N sequence segments of size P. The step size during the division process is S, and the step size S also serves as the length of the non-overlapping region between two consecutive sequence segments. Then, the N sequence segments of size P are embedded into a D-dimensional vector, and the embedded vector is used as the training dataset;
[0017] Send the training dataset to a linear differential equation solver to initialize the basis coefficients θ0. The initialized basis coefficients have corresponding frequency information and are embedded into basis coefficient tokens, while the training dataset is embedded into time value tokens.
[0018] The processing process of the described multi-dimensional scoring prediction model for personnel is specifically as follows:
[0019] S21. The time value token is sent as input to the self-attention module for preliminary feature extraction. The extracted features and the original time value token are sent to the token fusion module for feature fusion to reduce redundant features. After fusion, the fused time value token and the fused attention features are obtained. After adding the two and performing layer normalization, they are input to the feed-forward neural network. The features output by the feed-forward neural network are added to the fused attention features and then layer normalization is performed to update the time value token, which is then used as the time value token for the next encoder layer as input for repeated operations;
[0020] S22. The decoder layer receives the output of the corresponding encoder layer as the input to its cross-attention module's K and V. The input to the cross-attention module's Q comes from the base coefficient tokens. The decoder layer adds the output of the cross-attention module to the original base coefficient tokens, performs layer normalization, and then feeds the result into a feed-forward neural network. The output features of the feed-forward neural network are processed by another layer normalization and feed-forward neural network, and finally, after a layer normalization, the base coefficient tokens are updated and output to the next decoder layer;
[0021] S23. Both the encoder layer and the decoder layer have four layers, that is, the operations in steps S21 and S22 are repeated four times to output the updated base coefficient tokens.
[0022] The self-attention module consists of a pure convolutional structure. After the input is reshaped, it undergoes a Conv_step1 convolution operation followed by two Conv_step2 convolution operations, that is, convolutional feature extraction is performed in the feature dimension. Then the output vector is reshaped and flipped, and two more Conv_step2 convolution operations are carried out, that is, new feature representations are obtained on the target values. Finally, after reshaping and flipping again, it is processed with the input through residual processing to obtain the output of the self-attention module; For the Conv_step1 convolution operation on a vector with an input dimension of 1×(M×D)×N, there is one filter in the Conv_step1 convolution operation, and the filter contains M convolutional kernels, each with a size of 3×3; For the Conv_step2 convolution operation on a vector with an input dimension of 1×(M×D)×N, there are M filters in the Conv_step2 convolution operation, each filter contains M convolutional kernels, and each convolutional kernel has a size of 1×1.
[0023] The cross-attention module receives three inputs Q, K, and V, where Q, K, and V represent Query, Key, and Value respectively. Specifically, for the input sequence X, there are three learnable weight matrices W Q 、W K 、W V , and Q, K, and V are obtained through matrix multiplication, that is, Q = XW Q , K = XW K , V = XW V ; Then, the similarity matrix between Q and K is obtained through dot product operation, and then after normalization, it is multiplied by V to infer the common information P between Q and V QV , as shown in the following formula (3):
[0024]
[0025] In formula (3), d k is the scaling factor, and softmax is the softmax activation function;
[0026] Subsequently, by removing the common information, the difference information between Q and V is obtained. The specific processing process is shown in the following formula (4):
[0027] P = MLP(LN(Linear(V - P QV ) + Q)) + (Linear(V - P QV ) + Q) (4);
[0028] In formula (4), P represents the output of the cross - attention module, Linear represents the linear layer, LN represents layer normalization, and MLP represents the multi - layer perceptron.
[0029] The described multi - dimensional scoring prediction model for personnel is trained using the Huber loss function. The Huber loss function is specifically shown in the following formula (5):
[0030]
[0031] In formula (5), y is the true value, is the predicted value, δ is the hyper - parameter of the Huber loss function. The constructed multi - dimensional scoring prediction model for personnel is trained using the training data set, and the model parameters are updated to minimize the Huber loss function until the Huber loss function converges or reaches the upper limit of the number of training times, and a trained multi - dimensional scoring prediction model for personnel is obtained.
[0032] The basis function of the described trained multi - dimensional scoring prediction model for personnel is f(t, θ, v), which is specifically shown in the following formula (6):
[0033]
[0034] In formula (6), t represents the time, θ is the output vector containing the basis coefficients, v ∈ V is the frequency, and are the basis coefficients of the sine and cosine functions with frequency v respectively; IQR represents the inter - quartile range, MED is the abbreviation of Median, and is calculated based on each data point in the multi - dimensional scoring data of personnel obtained by non - uniform sampling; both a and b are affine adjustment coefficients.
[0035] Advantages of the present invention:
[0036] (1) The multi - dimensional scoring prediction model for personnel of the present invention adopts an encoder - decoder structure. The encoder makes the coefficients have a more comprehensive perspective in the process of learning and training by decomposing the continuous function in terms of frequency. While improving the generalization performance of the model, the interpretability of the model is retained.
[0037] (2) In the present invention, a self-attention module based on a pure convolutional structure is introduced into the encoder. Different from traditional convolutional methods, the self-attention module segments traditional convolutional operations, enabling it to capture cross-temporal information and effectively enhancing the model's ability to establish long-term dependencies.
[0038] (3) In the present invention, a cross-attention module is introduced into the decoder to further extract differential features, establish long-term dependencies, and lay the foundation for accurate model prediction. Description of the Drawings
[0039] Figure 1 is a flowchart of the present invention.
[0040] Figure 2 is a structural diagram of the multi-dimensional scoring prediction model for personnel of the present invention.
[0041] Figure 3 is a structural diagram of the self-attention module of the present invention.
[0042] Figure 4 is a structural diagram of the cross-attention module of the present invention. Detailed Embodiments
[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0044] See Figure 1 , a multi-dimensional scoring prediction method for personnel, specifically including the following steps:
[0045] (1) Collect the training data set: Collect the historical data set of multi-dimensional scoring for personnel and perform data cleaning operations to obtain the initial data. Then, according to the non-uniform sampling mechanism, multi-dimensional time value data is obtained from the initial data to form the training data set;
[0046] The distribution formula of the multi-dimensional time value data sampled under the non-uniform sampling mechanism is as follows in (1):
[0047]
[0048] In formula (1), T represents the actual time value, and R represents the sample resolution. Divide the actual time by the sample resolution to convert data with different time resolutions into a unified time step x that is convenient for processing in non-uniform sampling;
[0049] The processing process of the non-uniform sampling mechanism is specifically as follows in formula (2):
[0050]
[0051] In formula (2), p non-usm (x) represents the value of the probability density function sampled at time step x, and the normalization constant σ is the width of non-uniform sampling;
[0052] The non-uniform sampling mechanism specifically samples x and constructs (t, v) tuples from it, where v is the value corresponding to time t. For the input That is, for an input sequence of M variables of length L, after padding, the input sequence is divided into N sequence segments of size P. The step size during the division process is S, and the step size S also serves as the length of the non-overlapping region between two consecutive sequence segments. Then, the N sequence segments of size P are embedded into D-dimensional vectors, and the embedded vectors are used as the training dataset;
[0053] The training dataset is sent to a linear differential equation solver to initialize the base coefficients θ0. The initialized base coefficients have corresponding frequency information and are embedded into base coefficient tokens. At the same time, the training dataset is embedded into time value tokens;
[0054] (2) Construct the model: Construct a multi-dimensional scoring prediction model for personnel (see Figure 2 ), which includes an encoder and a decoder. The specific processing process is as follows:
[0055] S21. The time value token is used as the input and sent to the self-attention module for preliminary feature extraction. The extracted features and the original time value token are sent to the token fusion module for feature fusion to reduce redundant features. After fusion, the fused time value token and the fused attention features are obtained. After adding the two and performing layer normalization, the result is input to the feed-forward neural network (FNN). The FNN consists of a linear layer Linear, an activation function GeLU, and a linear layer Linear. The features output by the feed-forward neural network (FNN) are added to the fused attention features, and after performing layer normalization, the time value token is updated. Subsequently, it is used as the time value token of the next encoder layer for input and repeated operations;
[0056] See Figure 3, the self-attention module consists of a pure convolutional structure. After the input is reshaped, it undergoes a Conv_step1 convolution operation followed by two Conv_step2 convolution operations, that is, convolutional feature extraction is performed in the feature dimension. Then, the output vector is reshaped and permuted, and two more Conv_step2 convolution operations are carried out to obtain new feature representations on the target values. Finally, after reshaping and permuting again, residual processing is performed with the input to obtain the output of the self-attention module. For the Conv_step1 convolution operation on a vector with an input dimension of 1×(M×D)×N, there is one filter in the Conv_step1 convolution operation, and the filter contains M convolutional kernels, each with a size of 3×3. For the Conv_step2 convolution operation on a vector with an input dimension of 1×(M×D)×N, there are M filters in the Conv_step2 convolution operation, each filter contains M convolutional kernels, and each convolutional kernel has a size of 1×1.
[0057] S22. The decoder layer receives the output of the corresponding encoder layer as the input to its cross-attention module K and V. The input to the cross-attention module Q comes from the base coefficient tokens. The decoder layer adds the output of the cross-attention module to the original base coefficient tokens and performs layer normalization before feeding them into the feed-forward neural network (FNN). The output features of the feed-forward neural network are further processed by one layer normalization and the feed-forward neural network (FNN), and finally, after a layer normalization process, the base coefficient tokens are updated and output to the next decoder layer.
[0058] See Figure 4 , the cross-attention module receives three inputs Q, K, and V. Q, K, and V represent Query, Key, and Value respectively. Specifically, for the input sequence X, there are three learnable weight matrices W Q , W K , W V , and Q, K, and V are obtained through matrix multiplication, that is, Q = XW Q , K = XW K , V = XW V ; then, the similarity matrix between Q and K is obtained through dot product operation, and then after normalization, it is multiplied by V to infer the common information P between Q and V QV , as shown in the following formula (3):
[0059]
[0060] In formula (3), d k is the scaling factor, and softmax is the softmax activation function.
[0061] Subsequently, by removing the common information, the difference information between Q and V is obtained. The specific processing process is shown in the following formula (4):
[0062] P = MLP(LN(Linear(V - P QV ) + Q)) + (Linear(V - P QV ) + Q) (4);
[0063] In formula (4), P represents the output of the cross - attention module, Linear represents the linear layer, LN represents layer normalization, and MLP represents the multi - layer perceptron;
[0064] S23. Both the encoder layer and the decoder layer have four layers, that is, the operations in steps S21 and S22 are repeated four times to output the updated base coefficient tokens;
[0065] (3) Model training: Input the training data set into the multi - dimensional scoring prediction model for personnel to train and obtain a trained multi - dimensional scoring prediction model for personnel;
[0066] The multi - dimensional scoring prediction model for personnel is trained using the Huber loss function. The specific Huber loss function is shown in the following formula (5):
[0067]
[0068] In formula (5), y is the true value, is the predicted value, δ is the hyper - parameter of the Huber loss function. Use the training data set to train the constructed multi - dimensional scoring prediction model for personnel, update the model parameters to minimize the Huber loss function until the Huber loss function converges or reaches the upper limit of the number of training times, and obtain a trained multi - dimensional scoring prediction model for personnel;
[0069] (4) Model fitting and training: Fit the basis function of the trained multi - dimensional scoring prediction model for personnel, and then input the time point to be predicted into the basis function to obtain the prediction result;
[0070] The basis function of the trained multi - dimensional scoring prediction model for personnel is f(t, θ, v), which is specifically shown in the following formula (6):
[0071]
[0072] In formula (6), t represents the time, θ is the output vector containing the base coefficients, v ∈ V is the frequency, and They are the basis coefficients of the sine and cosine functions with frequency v respectively; IQR represents the interquartile range, and MED is the abbreviation of Median, which is calculated based on each data point in the multi-dimensional scoring data of personnel obtained from non-uniform sampling; both a and b are affine adjustment coefficients.
[0073] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A multi-dimensional scoring and prediction method for personnel, characterized in that: Specifically, it includes the following steps: (1) Collect the training data set: Collect the historical data set of multi-dimensional ratings of personnel and perform data cleaning operations to obtain the initial data. Then, according to the non-uniform sampling mechanism, obtain the multi-dimensional time value data from the initial data to form the training data set; (2) Build the model: Build a multi-dimensional rating prediction model for personnel, including an encoder and a decoder; (3) Model training: Input the training data set into the multi-dimensional rating prediction model for personnel for training to obtain a trained multi-dimensional rating prediction model for personnel; (4) Model fitting and training: Fit the basis function of the trained multi-dimensional rating prediction model for personnel, and then input the time point to be predicted into the basis function to obtain the prediction result.
2. The multi-dimensional scoring prediction method for personnel according to claim 1, wherein: The distribution formula of the multi-dimensional time value data sampled under the non-uniform sampling mechanism is as follows (1): In formula (1), T represents the actual time value, and R represents the sample resolution. Divide the actual time by the sample resolution to convert the data with different time resolutions into a unified time step x that is convenient for processing in non-uniform sampling; The processing process of the non-uniform sampling mechanism is specifically shown in the following formula (2): In Equation (2), p non-usm (x) represents the value of the probability density function sampled at time step x, and the normalization constant σ is the width of non-uniform sampling; The non-uniform sampling mechanism specifically samples x and constructs (t, w) tuples from it, where w is the value corresponding to time t, for the input That is, for an input sequence of M variables of length L, after padding, the input sequence is divided into N sequence segments of size P, with a step size S during the division process. At the same time, the step size S also serves as the length of the non-overlapping region between two consecutive sequence segments. Then, the N sequence segments of size P are embedded into D-dimensional vectors, and the embedded vectors are used as the training dataset; Send the training data set to a linear differential equation solver to initialize the basis coefficient θ0. The initialized basis coefficient has corresponding frequency information and is embedded in the basis coefficient token. At the same time, the training data set is embedded in the time value token.
3. A multi-dimensional scoring prediction method for personnel according to claim 2, characterized in that: The processing process of the multi-dimensional rating prediction model for personnel is specifically as follows: S21. The time value token is sent as input to the self-attention module for preliminary feature extraction. The extracted features and the original time value token are sent to the token fusion module for feature fusion to reduce redundant features. After fusion, the fused time value token and the fused attention feature are obtained. After adding the two and performing layer normalization, it is input to the feed-forward neural network. The features output by the feed-forward neural network are added to the fused attention feature, and after performing layer normalization, the time value token is updated, and then used as the time value token of the next encoder layer for input and repeated operations; S22. The decoder layer receives the output of the corresponding encoder layer as the input of its cross-attention module K and V. The input of the cross-attention module Q comes from the basis coefficient token. The decoder layer adds the output of the cross-attention module to the original basis coefficient token and performs layer normalization, and then sends it to the feed-forward neural network. The output features of the feed-forward neural network are processed by another layer normalization and feed-forward neural network, and finally, after a layer normalization process, the basis coefficient token is updated and output to the next decoder layer; S23. Both the encoder layer and the decoder layer have four layers, that is, repeat the operations of steps S21 and S22 four times to output the updated basis coefficient token.
4. A multi-dimensional scoring and prediction method for personnel according to claim 3, characterized in that: The self-attention module consists of a pure convolutional structure. After the input is reshaped, it undergoes a Conv_step1 convolutional operation followed by two Conv_step2 convolutional operations, i.e., convolutional feature extraction is performed in the feature dimension. Then, the output vector is reshaped and flipped, and two more Conv_step2 convolutional operations are carried out, i.e., new feature representations are obtained on the target values. Finally, after reshaping and flipping again, residual processing is performed with the input to obtain the output of the self-attention module. For the Conv_step1 convolutional operation on a vector with an input dimension of 1×(M×D)×N, there is one filter in the Conv_step1 convolutional operation, and the filter contains M convolutional kernels, each with a size of 3×3. For the Conv_step2 convolutional operation on a vector with an input dimension of 1×(M×D)×N, there are M filters in the Conv_step2 convolutional operation, each filter contains M convolutional kernels, and each convolutional kernel has a size of 1×1.
5. A multi-dimensional scoring and prediction method for personnel according to claim 3, characterized in that: The described cross-attention module receives three inputs Q, K, and V, where Q, K, and V represent Query, Key, and Value respectively. Specifically, for the input sequence X, there are three learnable weight matrices W Q , W K , W V . Q, K, and V are obtained through matrix multiplication, i.e., Q = ZW Q , K = ZW K , V = XW V . Then, the similarity matrix between Q and K is obtained through dot product operation, and then multiplied by V after normalization to infer the common information P QV between Q and V, as shown in the following formula (3): In formula (3), d k is a scaling factor, and softmax is the softmax activation function; Subsequently, by removing the common information, the difference information between Q and V is obtained. The specific processing process is shown in the following formula (4): P = MLP(LN9Linear9V - P QV ) + Q)) + (Linear(V - P Qv ) + Q)(4); In formula (4), P represents the output of the cross-attention module, Linear represents the linear layer, LN represents layer normalization, and MLP represents the multi-layer perceptron.
6. A multi-dimensional scoring prediction method for personnel according to claim 3, characterized in that: The multi-dimensional score prediction model for personnel uses the Huber loss function for training. The specific Huber loss function is shown in the following formula (5): In formula (5), y is the true value, is the predicted value, δ is the hyperparameter of the Huber loss function. The constructed multi-dimensional score prediction model for personnel is trained using the training data set, and the model parameters are updated to minimize the Huber loss function until the Huber loss function converges or reaches the upper limit of the number of training times, and a trained multi-dimensional score prediction model for personnel is obtained.
7. A multi-dimensional scoring prediction method for personnel according to claim 6, characterized in that: The basis function of the trained multi-dimensional score prediction model for personnel is f(t,θ,v), as shown in the following formula (6): In Equation (6), t represents the time, θ is the output vector containing the basis coefficients, v ∈ V is the frequency, and are the basis coefficients of the sine and cosine functions with frequency v, respectively; IQR represents the interquartile range, and MED is the abbreviation of Median, which is calculated for each data point in the multi-dimensional scoring data of personnel obtained from non-uniform sampling; both a and b are affine adjustment coefficients.