Loss prediction system based on outsourcing employee behaviors

Abstract

The invention discloses a loss prediction system based on outsourcing employee behaviors, and relates to the field of employee loss management, and the system comprises a feature obtaining module which is used for obtaining the current static position feature and any dynamic behavior feature of an outsourcing employee; the loss probability prediction module is used for receiving the current static position feature and any dynamic behavior feature as input of a loss probability prediction model and outputting the loss probability of the outsourcing employee; the loss probability characterization is the demission possibility of the outsourcing employee in a future preset time period; the loss prediction module is used for receiving the loss probability as the input of a loss prediction model and outputting the loss of the outsourcing employee; the defect that in the prior art, the demission probability cannot be accurately predicted is effectively overcome, the problem that in the prior art, the demission loss cannot be quantified is solved, and by introducing the position weight, the influence of different position employees on an enterprise can be distinguished, and more accurate loss evaluation is achieved.

Description

Technical Field

[0001] The present invention relates to the field of employee turnover prediction, and specifically to a turnover prediction system based on the behavior of outsourced employees. Background Art

[0002] In the management of outsourced employees, employee turnover often has a non-negligible impact on enterprises. In particular, the departure of employees in some key positions may cause many problems such as a decline in production efficiency and an increase in recruitment costs.

[0003] The patent document with the patent publication number CN104809188B discloses a data mining and analysis method for enterprise talent turnover, which can analyze and predict enterprise talent turnover, obtain the law of talent turnover, and at the same time obtain the turnover rate of each employee and calculate the reference salary of the employee for employees with a high turnover rate.

[0004] However, in the prior art, there is still a lack of predicting the turnover probability of employees based on dynamic behavior characteristics and static position characteristics and evaluating the loss caused by the employee to the enterprise based on the turnover probability. In particular, the departure of employees in different positions has different impacts on enterprises. For example, the loss of high-position employees will cause greater losses to the company. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a turnover prediction system based on the behavior of outsourced employees, and solves the technical problems proposed in the background art by constructing a turnover probability prediction model and a turnover loss prediction model.

[0006] To achieve the above object, the present invention is realized through the following technical solutions:

[0007] A turnover prediction system based on the behavior of outsourced employees, comprising:

[0008] A feature acquisition module, configured to acquire the current static position feature and any dynamic behavior feature of the outsourced employee;

[0009] A turnover probability prediction module, configured to receive the current static position feature and any dynamic behavior feature as the input of the turnover probability prediction model, and output the turnover probability of the outsourced employee; the turnover probability is characterized as: the possibility of the outsourced employee leaving the job within a future predetermined time period;

[0010] A turnover loss prediction module, configured to receive the turnover probability as the input of the turnover loss prediction model, and output the turnover loss of the outsourced employee.

[0011] In some of the embodiments, the training process of the turnover probability prediction model includes:

[0012] S1. Collect the historical behavior vectors of a number of outsourced employees. The characteristics of the historical behavior vectors include: static position characteristics constructed from employee codes and positions, and dynamic behavior characteristics constructed from job performance, salary changes, and turnover status, all marked with timestamps.

[0013] S2. Using the employee code as an index, define job performance, salary changes, and position as input features, and turnover status as the target label to construct the first training sample.

[0014] S3. Aggregate a number of first training samples to construct a turnover probability prediction training set.

[0015] S4. Perform multiple rounds of iteration on the turnover probability prediction training set through N decision trees of the GBDT model to fit the residuals to obtain the turnover probability prediction model.

[0016] In some of these embodiments, performing multiple rounds of iteration on the turnover probability prediction training set through N decision trees of the GBDT model to fit the residuals to obtain the turnover probability prediction model includes:

[0017] S4-1. Sample a number of first training samples from the turnover probability prediction training set, and based on their input features, use the first decision tree to generate the first-round prediction values.

[0018] The expression for using the first decision tree to generate the first-round prediction values is:

[0019] represents the first-round prediction value, T 1 represents the first decision tree, x i represents the input features of the i-th first training sample;

[0020] S4-2. Calculate the first-round residuals between the first-round prediction values and the target labels in a number of first training samples.

[0021] The calculation expression for the first-round residuals is:

[0022] represents the first-round residuals, L i represents the turnover status in the first training sample;

[0023] S4-3. Use the second decision tree to fit the first-round residuals and generate the second-round prediction values based on the first-round residuals.

[0024] The expression for the second-round prediction values is:

[0025] represents the second-round prediction value of the i-th first training sample, γ1 Denotes the learning rate, representing the correction of the second decision tree to the first-round residuals, T 2 (x i) Denotes the fitting prediction value of the second decision tree for the i-th first training sample;

[0026] S4-4. Iteratively train S4-1 to S4-3 until the t-th decision tree to generate the t-th round of prediction values based on the (t - 1)-th round of residuals;

[0027] The expression of the N-th round of prediction values is:

[0028] Denotes the t-th round of prediction value of the i-th first training sample, Denotes the (t - 1)-th round of prediction value of the i-th first training sample, T t (x i ) Denotes the fitting prediction value of the t-th decision tree for the i-th first training sample;

[0029] S4-5. When the (N - 1)-th round of residuals converges to the threshold, define the cumulative value from the first-round prediction values to the N-th round of prediction values as the optimal churn probability prediction value to complete the training of the churn probability prediction model;

[0030] The expression of the churn probability prediction value is: Denotes the optimal churn probability prediction value.

[0031] In some embodiments, sample a number of first training samples from the churn probability prediction training set, and based on their input features, use the first decision tree to generate the first-round prediction values, including:

[0032] S4-1-1. Calculate the information gain of each input feature in the first training sample; wherein, the calculation of the information gain is based on the difference between the churn state entropy and the conditional entropy;

[0033] The calculation expression of the information gain is: IG(X j ) = H(L) - H(L|X j );

[0034] IG(X j ) Denotes the information gain of the input feature X j ;

[0035] H(L) Denotes the entropy of the churn state L;

[0036] H(L|X j ) Denotes the conditional entropy of the churn state L given the input feature X j ;

[0037] The expression of the churn state entropy is:

[0038] P(L k ) represents the probability of the churn state L k ;

[0039] The expression of the conditional entropy is as follows:

[0040] v ∈ V j represents the specific value of the input feature X j , and V j is the set of all specific values;

[0041] P(X j = v) represents the probability when the specific value of the input feature X j is v;

[0042] H(L|X j = v) represents the entropy of the churn state L under the condition that the input feature X j takes the value v;

[0043] S4-1-2. Select the input feature with the largest information gain, and divide the first training samples according to this input feature to generate two child nodes;

[0044] The expression for selecting the input feature with the largest information gain is as follows:

[0045] X best represents the input feature with the largest information gain, indicating the selection of the input feature with the largest information gain;

[0046] The expression for generating two child nodes is as follows:

[0047] D 1 = {x ∈ D | X best (x) = v 1}, D 2 = {x ∈ D | X best (x) = v 2};

[0048] D represents the dataset corresponding to the several first training samples after sampling, D 1 and D 2 represent the divided child nodes, respectively containing X best (x) = v 1 and X best (x) = v 2 of the input features, v 1 and v 2 represent the values of the input feature X bes t;

[0049] S4-1-3. Calculate the information gain of each input feature in the child node respectively;

[0050] The expression for calculating the information gain of each input feature in the child node is: IG(X j , D k ) = H(L, D k ) - H(L|X j , D k );

[0051] IG(X j , D k ) Information gain of the input feature X j in the child node D k , H(L, D k ) Entropy of the churn state L in the child node D k , H(L|X j , D k ) Conditional entropy of the churn state L given the input feature X k in the child node D j ;

[0052] S4-1-4. In each child node, select the feature with the largest information gain, continue to partition the first training sample to generate new child nodes until the child nodes become leaf nodes;

[0053] S4-1-5. In the leaf node, generate the first-round prediction value according to the majority churn state distribution;

[0054] The expression for generating the first-round prediction value is:

[0055] P(L k |D leaf ) represents the probability of the churn state L k in the subset of data in the leaf node;

[0056] In some embodiments, the training process of the churn loss prediction model includes:

[0057] M1. Obtain a loss prediction training set; wherein, the loss prediction training set contains a number of second training samples, the input feature of the second training sample is the optimal churn probability prediction value, and the target label is the historical churn loss;

[0058] M2. Receive the optimal churn probability prediction value as the feature variable of the MLP model and output the historical churn prediction loss;

[0059] M3. Define a loss function to calculate the error loss between the historical churn loss and the historical churn prediction loss;

[0060] The loss function is as follows:

[0061] Loss represents the error loss, n is the number of the second training samples, represents the historical churn loss, represents the historical churn prediction loss;

[0062] M4. Perform backpropagation on the MLP model and update the model parameters to minimize the error loss;

[0063] M5. When iterating to minimize the mean squared error loss, export the MLP model with the current model parameters as the churn loss prediction model.

[0064] In some of the embodiments, obtaining the loss prediction training set includes:

[0065] M1-1. Calculate the historical churn loss of the position corresponding to each employee code;

[0066] M1-2. Using the employee code as the index, define the optimal churn probability prediction value corresponding to the employee code as the input feature, and define the historical churn loss of the corresponding position as the target label to construct the second training sample;

[0067] M1-3. Aggregate a number of second training samples to construct the loss prediction training set.

[0068] In some of the embodiments, calculating the historical churn loss of the position corresponding to each employee code includes:

[0069] M1-1-1. Define the basic vacuum time and the position weight;

[0070] M1-1-2. Calculate the position vacuum time of the position according to the basic vacuum time and the position weight;

[0071] The expression of the position vacuum time is: T zkw = T jczk × W zw ;

[0072] T zkw represents the position vacuum time, T jczk represents the basic vacuum time of the position, W zw represents the position weight;

[0073] M1-1-3. Obtain the monthly average recruitment loss, the monthly average productivity loss, and the monthly average salary of a number of positions;

[0074] M1-1-4. Calculate the position churn loss of the position according to the position weight and the monthly average salary;

[0075] The expression for calculating the job loss loss of this position is: C ls = W zw × G yg ;

[0076] C ls represents the job loss loss, and G yg represents the average monthly salary;

[0077] M1-1-5. Determine the historical loss loss after several job losses;

[0078] The expression for determining the historical loss loss after several job losses is: C lsls = T zkw × (C zp + C sc + C ls - G yg )

[0079] C lsls represents the historical loss loss, C zp represents the average monthly recruitment loss, and C sc represents the average monthly productivity;

[0080] In some embodiments, receiving the optimal churn probability prediction value as a feature variable of the MLP model and outputting the historical churn prediction loss includes:

[0081] M2-1. Receive the optimal churn probability prediction value as the input of the input layer of the MLP model, and after weighted bias processing by the first fully connected layer, obtain the input layer output value;

[0082] M2-2. Take the input layer output value as the input of the hidden layer, and after weighted bias of the second fully connected layer and non-linear transformation of the activation layer, obtain the hidden layer output value;

[0083] M2-3. Sequentially take several hidden layer output values as the input of the next hidden layer, and after weighted bias of multiple fully connected layers and non-linear transformation of the activation layer, obtain the output value of the last hidden layer;

[0084] M2-4. Pass the output value of the last hidden layer to the output layer, and perform a linear transformation through the fully connected layer of the output layer to output the historical churn prediction loss;

[0085] In some embodiments, performing backpropagation on the MLP model and iteratively updating the model parameters to minimize the mean square error loss includes:

[0086] M4-1. Calculate the gradient of the loss function with respect to the output layer;

[0087] M4-2. Backpropagate and calculate the gradients of several hidden layers in sequence until the input layer;

[0088] M4-3. According to the gradients of the output layer, several hidden layers, and the input layer, use the gradient descent algorithm to iteratively update the model parameters;

[0089] M4-4. According to the iteratively updated model parameters, calculate the error loss for each iterative update until the error loss is minimized.

[0090] The present invention provides a turnover prediction system based on the behavior of outsourced employees, which has the following beneficial effects:

[0091] Through the feature acquisition module, the present invention comprehensively collects the static position features and dynamic behavior features of outsourced employees, and uses the GBDT model for multiple rounds of iterative training to output the turnover probability of employees, which can accurately capture the turnover tendency of employees and effectively solve the defect in the prior art that the turnover probability cannot be accurately predicted; and after obtaining the turnover probability of employees, the system calculates the loss that may be brought about by the employee's turnover through the turnover loss prediction model combined with the position weight; not only solves the problem in the prior art that the turnover loss cannot be quantified, but also by introducing the position weight, it can distinguish the impact of employees in different positions on the enterprise and achieve a more accurate loss assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 It is a structural block diagram of a turnover prediction system based on the behavior of outsourced employees according to the present invention;

[0093] Figure 2 It is an application flowchart of a turnover prediction system based on the behavior of outsourced employees according to the present invention;

[0094] Figure 3 It is a training schematic diagram of the turnover probability prediction model described in the present invention;

[0095] Figure 4 It is a construction schematic diagram of the historical turnover loss described in the present invention.

[0096] DETAILED DESCRIPTION OF THE EMBODIMENTS

[0097] Next, the prior art and related concepts involved in the embodiments of the present invention will be described:

[0098] First, the prior art and related concepts involved in the embodiments of the present invention are described:

[0099] GBDT model: Gradient boosted decision tree, an integrated learning method that gradually improves the accuracy of the model through multiple rounds of iterations. Each round of decision trees is fitted based on the residuals of the previous round, gradually correcting the prediction error of the churn status.

[0100] MLP model: Multilayer Perceptron, a classic neural network model, consists of multiple layers of fully connected layers and activation layers. The neurons in each layer are connected to all the neurons in the next layer, forming a complex feature transformation process.

[0101] Information gain: A classic feature selection criterion for classification tasks, especially in classification and regression tree (CART) models, it is a common choice. Information gain measures the contribution of a feature to reducing the uncertainty (or entropy) of a dataset; information gain is sensitive to data with imbalanced categories (i.e., large distribution differences between categories), and in employee turnover prediction, different features may have different effects on the turnover status, so information gain can better capture the relationship between such features and classification results.

[0102] Fully connected layer: Responsible for connecting all nodes between the input layer, hidden layer, and output layer. At each layer, the fully connected layer performs linear transformation on the input data through weighted calculation and bias processing. This process helps the model extract and aggregate features, gradually mapping complex input features to target outputs, i.e. employee turnover loss.

[0103] Activation layer: It is a nonlinear transformation layer after the fully connected layer. In the present invention, the activation layer is responsible for performing nonlinear transformation on the linear output values ​​after the fully connected layer to increase the expressive power of the model. Through activation functions (such as ReLU or Sigmoid, etc.), the model can better handle complex nonlinear relationships and help predict churn losses more accurately. The activation layer plays an important role between each hidden layer, allowing the model to evolve from simple linear combinations to deep learning of complex relationships.

[0104] Example 1: Please refer to Figures 1 to 2 The present invention provides a system for predicting employee turnover based on outsourced employee behavior, comprising:

[0105] The feature acquisition module is used to obtain the current static position features and any dynamic behavior features of the outsourced employee.

[0106] The churn probability prediction module is used to receive the current behavior characteristics as the input of the churn probability prediction model and output the churn probability of the outsourced employee. The churn probability is characterized by the possibility of the outsourced employee leaving within a predetermined time period in the future.

[0107] The attrition loss prediction module is used to receive the attrition probability as the input of the attrition loss prediction model and output the attrition loss of the outsourced employee.

[0108] In this embodiment, based on the current behavioral characteristics of outsourced employees, the probability of their turnover in a future period is predicted. Then, based on the turnover probability, the turnover loss brought to the company after the outsourced employees leave is further predicted. This method can help enterprises predict the turnover risk of outsourced employees and the losses brought by it.

[0109] Specifically, the feature acquisition module collects the current static position features and dynamic behavioral features of outsourced employees, providing a sufficient input data basis for subsequent turnover probability prediction. Then, the turnover probability prediction module receives the above-collected behavioral features as input and outputs the turnover probability of outsourced employees through the turnover probability prediction model. This turnover probability represents the possibility of an employee leaving within a certain predetermined future period, helping the enterprise to identify the turnover risk of employees in advance. Synchronously, the turnover loss prediction module, based on the predicted turnover probability, further calculates the possible losses caused by employee turnover through the turnover loss prediction model. Loss assessment can help enterprises quantify the impact of employee turnover on the company and provide important references for management decisions.

[0110] Example 2: Refer to Figures 1 to 3 , the technical solution of this Example 2 is different from that of Example 1 in that the training process of the turnover probability prediction model in Example 1 is disclosed, and this training process includes:

[0111] S1. Collect historical behavior vectors of a number of outsourced employees. The features of the historical behavior vectors include: static position features constructed by employee codes and positions, and dynamic behavioral features constructed by work performance, salary changes, and turnover status, all marked with timestamps.

[0112] Specifically, the historical behavior vector is a multi-dimensional vector, which contains multi-dimensional historical features. In terms of categories, this multi-dimensional historical feature includes static position features and dynamic behavioral features. Among them, the static position features are employee codes and positions, and the dynamic behavioral features are dynamic behavioral features constructed by work performance, salary changes, and turnover status, all marked with timestamps.

[0113] Among the collected outsourced employees, the ratio of departing employees to current employees is controlled at 1:1. This facilitates obtaining the same number of turnover statuses in different states.

[0114] Furthermore, work performance refers to: changes in the amount of tasks, changes in project participation, and changes in performance scores.

[0115] From the perspective of company managers, changes in the task volume indicate significant changes in the amount of tasks received by employees in the short term. For example, an increase in task volume leads to increased stress, while a decrease in task volume may indicate a weakening of an employee's responsibilities or a decline in the importance of their position. Changes in project participation indicate that there are changes in the frequency of an employee's participation in projects, or in other words, their interaction with the company. For example, has there been a recent reduction in participation in important projects, or has the employee been more assigned to minor tasks? Changes in performance ratings indicate that an employee's performance rating has changed significantly in the short term, such as a sudden drop. Changes in interaction frequency indicate whether there are changes in the number of interactions between an employee and the team, with a reduction in communication with colleagues and superiors.

[0116] Salary changes refer to: whether there is a salary change, the magnitude of the salary increase if there is a change, and the duration of no salary change.

[0117] Furthermore, the granularity of the timestamp when the above dynamic behavioral characteristics change should be accurate to the specific date when the dynamic behavioral characteristics change, so that the model can better capture the temporal pattern in the time dimension. The time window of the timestamp can be defined as the past year or three years, so that when obtaining salary changes, salary changes such as annual bonuses or annual appraisals that only occur within the annual time window can be collected.

[0118] In this embodiment, by collecting the historical behavior vectors of outsourced employees, a training set for the churn probability prediction model is constructed. During the training process, the model captures the temporal pattern among an employee's work performance, salary changes, position, and churn status, and then learns how to output the churn probability based on behavioral characteristics, thereby enhancing the utility of the model in accurately capturing the churn pattern of employees.

[0119] S2. Using the employee code as an index, define work performance, salary changes, and position as input features, and define the churn status as the target label to construct the first training sample.

[0120] S3. Aggregate a number of first training samples to construct a training set for churn probability prediction.

[0121] Specifically, in the first training sample, the input features are the work performance, salary changes, and positions of the outsourced employees corresponding to the employee codes, and the target label is the turnover status of the outsourced employees. Moreover, each dynamic behavior feature is marked with the timestamp when the dynamic behavior feature changes. Therefore, when the model is trained, it can not only capture the state laws between work performance, salary changes, and turnover status, but also capture the temporal laws that when work performance and salary changes occur, the turnover status also changes. As a static position feature, outsourced employees in different positions may have different turnover risks. For example, outsourced employees in high positions may have a lower turnover risk due to the relationship between responsibilities and rewards, while employees in low positions may have a higher turnover rate. Therefore, as an important static position feature, the position can also be included in the input features.

[0122] S4. Use N decision trees of the GBDT model to perform multiple rounds of iteration on the turnover probability prediction training set to fit the residuals and obtain the turnover probability prediction model.

[0123] Furthermore, S4 specifically includes:

[0124] S4-1. Sample a number of first training samples from the turnover probability prediction training set, and based on their input features, use the first decision tree to generate the first-round prediction values.

[0125] The expression for generating the first-round prediction values using the first decision tree is:

[0126] represents the first-round prediction value, T 1 represents the first decision tree, x i represents the input feature of the i-th first training sample.

[0127] Even further, S4-1 specifically includes:

[0128] S4-1-1. Calculate the information gain of each input feature in the first training sample. Among them, the calculation of the information gain is based on the difference between the turnover state entropy and the conditional entropy.

[0129] The calculation expression of the information gain is: IG(X j ) = H(L) - H(L|X j ).

[0130] IG(X j ) represents the information gain of the input feature X j . The information gain characterizes the degree to which the uncertainty of the turnover state L is reduced after dividing the first training sample according to the input feature X j .

[0131] H(L) represents the entropy of the churn state L. The entropy of the churn state L (the target label) characterizes the overall uncertainty of the churn state L.

[0132] H(L∣X j ) represents the conditional entropy of the churn state L given the input feature X j . This conditional entropy characterizes the uncertainty of the churn state L after knowing the value of the input feature X j .

[0133] The expression for the entropy of the churn state is:

[0134] P(L k ) represents the probability of the churn state L k . That is, the proportion of the k-th churn state K. Since a number of first training samples are sampled, there will be multiple different churn states, and P(L k ) is exactly used to characterize the proportion of different churn states, such as churn and non-churn.

[0135] The expression for the conditional entropy is:

[0136] v∈V j represents the specific value of the input feature X j , and V j is the set of all specific values. For example, if X j is a categorical feature (job level), then v is a specific job level.

[0137] P(X j =v) represents the probability when the input feature X j takes the specific value v. That is, in the current training sample, the proportion of the value of the input feature X j .

[0138] H(L∣X j =v) represents the entropy of the churn state L under the condition that the input feature X j takes the value v.

[0139] S4-1-2. Select the input feature with the largest information gain, and divide the first training samples according to this input feature to generate two child nodes.

[0140] The expression for selecting the input feature with the largest information gain is:

[0141] X best represents the input feature with the largest information gain, represents selecting the input feature with the largest information gain.

[0142] The expression for generating two child nodes is:

[0143] D 1 ={x ∈ D | X best (x) = v 1}, D 2 ={x ∈ D | X best (x) = v 2}}。

[0144] D represents the data set corresponding to a number of first training samples after sampling, D 1 and D 2 represent the divided child nodes, that is, the two sub-data sets corresponding to the child nodes, respectively containing X best (x) = v 1 and X best (x) = v 2 of the input features, v 1 and v 2 represent the values of the input feature X best with the largest information gain.

[0145] S4-1-3. Calculate the information gain of each input feature in the child nodes respectively.

[0146] The expression for calculating the information gain of each input feature in the child nodes is: IG(X j , D k ) = H(L, D k ) - H(L | X j , D k ).

[0147] IG(X j , D k ) is the information gain of the input feature X j in the child node D k , H(L, D k ) is the entropy of the churn state L in the child node D k , and H(L | X j , D k ) is the conditional entropy of the churn state L given the input feature X k in the child node D j .

[0148] S4-1-4. Within each child node, select the feature with the largest information gain, continue to divide the first training samples, and generate new child nodes until the child nodes become leaf nodes.

[0149] S4-1-5. Within the leaf nodes, generate the first-round prediction values according to the majority churn state distribution. The first-round prediction values are used to characterize the class distribution of the majority churn state within the leaf nodes.

[0150] The expression for generating the first-round predicted values is as follows:

[0151] P(L k ∣D leaf ) represents the probability of the churn status L k in the subset of data at the leaf node. It should be noted that the so-called subset of data at the leaf node cannot be regarded as the aforementioned first training sample. The first training sample is a complete sample composed of features of different categories (static job features corresponding to employee codes and positions, and dynamic behavior features of job performance marked with time stamps and salary changes marked with time stamps) and the churn status. During the training process of the GBDT model, with the division of each sub-node, a number of sampled first training samples are used to divide the dataset. Eventually, at the leaf node, there may only be specific categories of certain features left.

[0152] When the sample data reaches the leaf node, usually the samples at the leaf node have been completely divided according to specific features. At this time, the samples in the node may only belong to a certain category (such as a specific position or a certain type of salary change). At this time, these samples no longer have the nature of "first training sample" with multiple categories and multiple features, but are more used as the final classification basis of the model. Therefore, the samples at this stage are more suitable as the "subset of data at the leaf node" of the model, rather than the multi-feature samples for continued training.

[0153] In each iteration of training, by calculating the contribution (information gain) of each input feature to the churn status, the feature that can most reduce the uncertainty of the churn status is selected for division, so that the decision tree can be effectively split to improve the accuracy of predicting the employee churn status.

[0154] S4-2. Calculate the first-round residuals between the first-round predicted values and the target labels in a number of first training samples.

[0155] The calculation expression for the first-round residuals is as follows:

[0156] represents the first-round residuals, and L i represents the churn status in the first training sample.

[0157] S4-3. Use the second decision tree to fit the first-round residuals and generate the second-round predicted values based on the first-round residuals.

[0158] The expression for the second-round predicted values is as follows:

[0159] represents the second-round predicted value of the i-th first training sample, which is obtained by adding the first-round predicted value to the second decision tree T2 Generated by fitting the first-round residuals. γ 1 Denotes the learning rate, representing the correction of the first-round residuals by the second decision tree, T 2 (x i ) represents the fitting prediction value of the second decision tree for the i-th first training sample.

[0160] S4-4. Iteratively train S4-1 to S4-3 until the t-th decision tree to generate the t-th round of prediction values based on the (t - 1)-th round of residuals.

[0161] The expression for the N-th round of prediction values is:

[0162] represents the t-th round of prediction value of the i-th first training sample, represents the (t - 1)-th round of prediction value of the i-th first training sample, T t (x i ) represents the fitting prediction value of the t-th decision tree for the i-th first training sample.

[0163] S4-5. When the (N - 1)-th round of residuals converges to the threshold, define the cumulative value from the first-round prediction value to the N-th round of prediction values as the optimal churn probability prediction value to complete the training of the churn probability prediction model.

[0164] The expression for the churn probability prediction value is: represents the optimal churn probability prediction value.

[0165] In this embodiment, during the training process of the churn degree prediction model, the GBDT model (Gradient Boosting Decision Tree model) is used, and multiple rounds of decision trees are iteratively trained to gradually fit the residuals of the churn state of the outsourced employees, and finally the optimal churn degree prediction value is obtained. This training process continuously corrects the prediction error through decision trees and gradually improves the prediction accuracy of the model.

[0166] Embodiment 3: Refer to Figures 1 to 4 , the technical solution of this Embodiment 3 is different from that of Embodiment 2 in that the training process of the churn loss prediction model is disclosed, and this training process includes:

[0167] M1. Obtain the loss prediction training set. Among them, the loss prediction training set contains a number of second training samples, the input feature of the second training sample is the optimal churn probability prediction value, and the target label is the historical churn loss.

[0168] Further, M1 specifically includes:

[0169] M1-1. Calculate the historical churn loss corresponding to the position of each employee code.

[0170] Even further, M1-1 specifically includes:

[0171] M1-1-1. Define the basic vacancy time and position weight. The basic vacancy time refers to the vacancy time of a position without considering weight correction, and the position weight reflects the importance or irreplaceability of the position.

[0172] M1-1-2. Calculate the position vacancy time of this position according to the basic vacancy time and position weight.

[0173] The expression for the position vacancy time is: T zkw =T jczk ×W zw .

[0174] T zkw represents the position vacancy time, T jczk represents the basic vacancy time of the position, and W zw represents the position weight.

[0175] For example, if the defined basic vacancy time is 2 months and the position weight of an ordinary employee position is defined as 0.6, then the position vacancy time of the ordinary employee position may be 1.2 months; while the position weight of a team leader level may be 0.8, then its position vacancy time may be 1.6 months.

[0176] M1-1-3. Obtain the monthly average recruitment loss, monthly average productivity loss, and monthly average salary of several positions.

[0177] The monthly average recruitment loss can obtain the costs of the recruitment process such as recruitment advertisements, interview expenses, and HR working hours from the company's historical records or market benchmark data. The monthly average productivity loss can be calculated by analyzing the decline in the company's output during the position vacancy period or the reduction in efficiency caused by other employees due to additional work burdens. The monthly average salary can obtain the monthly average values of the basic salary, benefits, and bonuses of this position from the company's HR system or industry market benchmarks. The relevant data sources are the company's internal systems (HR, finance, performance, etc.), publicly available market data or reports in the industry, etc.

[0178] M1-1-4. Calculate the position turnover loss of this position according to the position weight and monthly average salary.

[0179] The expression for calculating the position turnover loss of this position is: C ls =W zw ×G yg .

[0180] C ls represents the position turnover loss, and G yg represents the monthly average salary.

[0181] M1-1-5. Determine the historical loss due to turnover after several positions are lost;

[0182] The expression for determining the historical loss due to turnover after several positions are lost is: C lsls = T zkw × (C zp + C sc + C ls - G yg );

[0183] C lsls represents the historical loss due to turnover, C zp represents the average monthly recruitment loss, and C sc represents the average monthly productivity.

[0184] Through the above steps, once an outsourced employee of a certain position leaves, the historical loss due to turnover formed by them for the company is used as the training data (target label) of the MLP model. The MLP model captures the association between the turnover probability including the static position characteristics of the position and the historical loss due to turnover, and thus can, based on the captured association, give the possible loss due to turnover formed by the outsourced employee of the position for the company once they leave, thereby helping the company's managers formulate different management methods based on the possible loss due to turnover.

[0185] In this embodiment, the historical loss due to turnover caused by position turnover is calculated through key indicators such as position vacancy time, position weight, average monthly recruitment loss, and productivity loss. By quantifying the impact of qualitative position turnover on the company, a historical loss due to turnover that better meets the current management needs can be obtained.

[0186] M1-2. Using the employee code as an index, define the optimal turnover probability prediction value corresponding to the employee code as the input feature, and define the historical loss due to turnover of the corresponding position as the target label to construct a second training sample.

[0187] M1-3. Aggregate several second training samples to construct a loss prediction training set.

[0188] In this embodiment, by establishing a record of the historical loss due to turnover of positions, the employee code is associated with the historical loss due to turnover of positions to construct a loss prediction training set. This step provides the basic data for the training of the loss prediction model and ensures that the model can capture the relationship between employee turnover and historical losses.

[0189] M2. Receive the optimal turnover probability prediction value as the feature variable of the MLP model and output the predicted historical loss due to turnover.

[0190] Among them, M2 specifically includes:

[0191] M2-1. Receive the optimal churn probability prediction value as the input of the input layer of the MLP model. After the weighted bias processing of the first fully connected layer, obtain the output value of the input layer.

[0192] Use the optimal churn probability prediction value output by the GBDT model as input features (i.e., the input layer of the MLP model), and pass these input features to the first fully connected layer in the MLP model. The optimal churn probability prediction value serves as a feature to help the MLP model extract information from it to predict historical churn losses.

[0193] M2-2. Use the output value of the input layer as the input of the hidden layer. After the weighted bias of the second fully connected layer and the non-linear transformation of the activation layer, obtain the output value of the hidden layer.

[0194] The output of the input layer extracts preliminary features through the first fully connected layer (weighted and bias processing), and then through the non-linear activation function, obtain the output value of this hidden layer.

[0195] M2-3. Sequentially use the output values of several hidden layers as the input of the next hidden layer. After the weighted bias of multiple fully connected layers and the non-linear transformation of the activation layer, obtain the output value of the last hidden layer.

[0196] The output of the hidden layer serves as the input of the next hidden layer. Through a series of non-linear transformations of fully connected layers and activation functions, finally obtain the output of the last hidden layer. Through multiple weighted, bias, and non-linear transformations of the hidden layer, features are gradually extracted, enabling the MLP model to model complex relationships.

[0197] M2-4. Pass the output value of the last hidden layer to the output layer. After the linear transformation of the fully connected layer of the output layer, output the historical churn prediction loss.

[0198] The output of the last hidden layer passes through the fully connected layer of the output layer and finally undergoes a linear transformation to map the extracted features to the historical churn prediction loss.

[0199] M3. Define a loss function to calculate the error loss between the historical churn loss and the historical churn prediction loss.

[0200] The loss function is:

[0201] Loss represents the error loss, n is the number of the second training samples, represents the historical churn loss, represents the historical churn prediction loss. The goal of this loss function is to minimize the difference between the true historical churn loss and the predicted historical churn prediction loss, enabling the MLP model to more accurately predict the loss of employee churn.

[0202] M4. Perform backpropagation on the MLP model and update the model parameters to minimize the error loss.

[0203] Among them, M4 specifically includes:

[0204] M4-1. Calculate the gradient of the loss function with respect to the output layer.

[0205] By calculating the gradient of the loss function with respect to the output layer, the direction and magnitude of updating the output layer parameters are calculated.

[0206] M4-2. Perform backpropagation calculations on the gradients of several hidden layers in turn until the input layer. The gradients are gradually passed from the output layer to each hidden layer, the gradients of each layer are calculated, and by backpropagating layer by layer, the model parameters of each layer are adjusted in turn to gradually reduce the error.

[0207] M4-3. According to the gradients of the output layer, several hidden layers, and the input layer, use the gradient descent algorithm to iteratively update the model parameters. Use the gradient descent algorithm to update the weight and bias parameters of each layer according to the calculated gradients to optimize the MLP model.

[0208] M4-4. According to the iteratively updated model parameters, calculate the error loss of each iterative update until the error loss is minimized. Repeat the above steps until the model converges (the error cannot be significantly reduced) or the set number of training epochs is reached, until the optimal solution is found, that is, the churn loss prediction model is found.

[0209] M5. When iterating to minimize the mean square error loss, export the MLP model with the current model parameters as the churn loss prediction model.

[0210] In this embodiment, according to the optimal churn prediction value of the outsourced employees, a churn loss prediction model is established. By inputting the optimal churn prediction value, the economic loss caused by employee churn to the company can be output, providing a decision-making basis for enterprise management.

[0211] In summary, the present invention provides a churn prediction system based on the behavior of outsourced employees and its related training method. By collecting the static job characteristics and dynamic behavior characteristics of outsourced employees, the future possible turnover probability (churn probability) is predicted, and based on this churn probability combined with the job weight, the loss caused by employee churn to the company is further predicted. The system iteratively trains multiple decision trees using the GBDT model to gradually improve the accuracy of churn probability prediction, and at the same time uses the MLP model for loss prediction to quantify the impact caused by job churn, so as to identify the churn risk of employees in advance, reasonably evaluate the loss caused by employee churn, and thus provide support for management and decision-making to help reduce operational risks.

[0212] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wired (such as infrared, wireless, microwave, etc.) means.

[0213] The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available media can be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0214] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division of an underwater terrain change analysis system and method for waterways. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0215] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. A churn prediction system based on outsourced employee behavior, characterized in that: include: A feature acquisition module is used to acquire the current static position features and any dynamic behavior features of the outsourced employee; A churn probability prediction module is used to receive the current static position characteristics and any dynamic behavior characteristics as inputs of a churn probability prediction model, and output the churn probability of the outsourced employee; the churn probability is characterized by the possibility of the outsourced employee leaving within a predetermined time period in the future; The attrition loss prediction module is used to receive the attrition probability as the input of the attrition loss prediction model and output the attrition loss of the outsourced employee.

2. A churn prediction system based on outsourced employee behavior according to claim 1, characterized in that: The training process of the churn probability prediction model includes: S1. Collecting historical behavior vectors of several outsourced employees, wherein the features of the historical behavior vectors include: static position features constructed by employee code and position, and dynamic behavior features constructed by work performance, salary changes, and turnover status, all of which are marked with timestamps; S2, using employee code as index, defining work performance, salary change and position as input features, and defining turnover status as target label to construct the first training sample; S3, collecting a number of first training samples to construct a churn probability prediction training set; S4. Perform multiple rounds of iterations on the churn probability prediction training set through N decision trees of the GBDT model to fit the residuals to obtain the churn probability prediction model.

3. The outsourcing employee behavior-based attrition prediction system according to claim 2, characterized in that: The N decision trees of the GBDT model are used to perform multiple rounds of iterations on the churn probability prediction training set, including: S4-1, sampling a number of first training samples from the churn probability prediction training set, and generating a first round of prediction values ​​using the first decision tree based on the input features thereof; The expression for generating the first round of prediction values ​​using the first decision tree is: represents the first round of prediction value, T1 represents the first decision tree, x i represents the input features of the i-th first training sample; S4-2, calculating the first round residual between the first round prediction value and the target labels in the first training samples; The calculation expression of the first round residual is: represents the first round residual, L i represents the churn status in the first training sample; S4-3, use the second decision tree to fit the first round residuals to generate the second round prediction values ​​based on the first round residuals; The expression of the second round of prediction value is: represents the second round prediction value of the first training sample of the ith class, γ1 represents the learning rate, which represents the correction of the first round residual by the second lesson decision tree, T2(x i ) represents the fitted prediction value of the second lesson decision tree for the i-th first training sample; S4-4, iteratively train S4-1 to S4-3 until the t-th decision tree is reached, and generate the t-th round prediction value based on the t-1-th round residual; The expression of the Nth round prediction value is: represents the t-th round prediction value of the i-th first training sample, represents the t-1th round prediction value of the i-th first training sample, T t (x i ) represents the fitted prediction value of the t-th decision tree for the i-th first training sample; S4-5. When the residual error of the N-1th round converges to the threshold, the accumulated value from the first round prediction value to the Nth round prediction value is defined as the optimal churn probability prediction value, so as to complete the training of the churn probability prediction model; The expression of the predicted value of the churn probability is: Represents the optimal churn probability prediction value.

4. The outsourcing employee behavior-based attrition prediction system according to claim 3, characterized in that: The method of using the first decision tree to generate a first round of prediction values ​​includes: S4-1-1, calculating the information gain of each input feature in the first training sample; wherein the information gain is calculated based on the difference between the churn state entropy and the conditional entropy; The calculation expression of the information gain is: IG(X j )=H(L)-H(L|X j ); IG(X j ) represents the input feature X j Information gain of H(L) represents the entropy of the churn state L; H(L|X j ) indicates that given input feature X j The conditional entropy of the post-churn state L; The expression of the loss state entropy is: P(L k ) indicates the loss state L k The probability of The expression of the conditional entropy is: v∈V j Represents the input feature X j The specific value of V j is the set of all specific values; P(X j =v) represents the input feature X j The probability when the specific value is v; H(L|X j =v) indicates that in the input feature X j The entropy of the loss state L under the condition of value v; S4-1-2, selecting an input feature with the largest information gain, dividing the first training sample according to the input feature, and generating two child nodes; The expression for selecting the input feature with the largest information gain is: X best represents the input feature with the largest information gain, Indicates selecting the input feature with the largest information gain; The expression for generating two child nodes is: D1={x∈D∣X best (x)=v1},D2={x∈D∣X best (x)=v2}; D represents the data set corresponding to the first training samples after sampling, D1 and D2 represent the child nodes after division, respectively containing X best (x) = v1 and X best (x) = input feature v2, v1 and v2 represent the input feature X with the largest information gain best The value of S4-1-3, calculate the information gain of each input feature in the child node respectively; The expression for calculating the information gain of each input feature in the subnode is: IG(X j ,D k )=H(L,D k )-H(L|X j ,D k ); IG(X j ,D k ) Input feature X j In child node D k The information gain in H(L,D k ) Child node D k The entropy of the churn state L, H(L|X j ,D k ) in child node D k In the example, given the input feature X j The conditional entropy of the post-churn state L; S4-1-4, in each child node, select the feature with the largest information gain, continue to divide the first training sample, and generate new child nodes until the child node becomes a leaf node; S4-1-5. In the leaf node, the first round of prediction values ​​are generated according to the majority loss state distribution; The expression for generating the first round of prediction values ​​is: P(L k ∣D leaf ) indicates that in the sub-dataset of the leaf node, the churn state L k probability.

5. The outsourcing employee behavior-based attrition prediction system according to claim 4, characterized in that: The training process of the churn loss prediction model includes: M1. Obtain a loss prediction training set; wherein the loss prediction training set includes a plurality of second training samples, the input feature of the second training sample is the optimal loss probability prediction value, and the target label is the historical loss; M2, receiving the optimal churn probability prediction value as the characteristic variable of the MLP model, and outputting the historical churn prediction loss; M3. Define the loss function to calculate the error loss of historical churn loss and historical churn prediction loss; The loss function is: Loss represents the error loss, n is the number of the second training samples, represents the historical loss, represents the historical churn predicted loss; M4, perform back propagation on the MLP model and update the model parameters to minimize the error loss; M5, when iterating to minimize the mean square error loss, deriving the MLP model with current model parameters as the churn loss prediction model.

6. The outsourcing employee behavior-based attrition prediction system according to claim 5, characterized in that: Get the loss prediction training set, including: M1-1. Calculate the historical loss of each employee code corresponding to the position; M1-2, using the employee code as an index, defining the optimal loss probability prediction value corresponding to the employee code as an input feature, defining the historical loss of the corresponding position as a target label, and constructing the second training sample; M1-3. Gather a number of second training samples to construct the loss prediction training set.

7. The outsourcing employee behavior-based attrition prediction system according to claim 6, characterized in that: Calculate the historical attrition loss of each employee code corresponding to the position, including: M1-1-1, define basic vacuum time and position weight; M1-1-2. Calculate the position vacuum time of the position based on the basic vacuum time and position weight; The expression of the position vacuum time is: T zkw =T jczk ×W zw ; T zkw Represents the position vacuum time, T jczk Indicates the basic vacuum time of the position, W zw Indicates the weight of the position; M1-1-3, obtain the average monthly recruitment loss, average monthly productivity loss and average monthly salary of several positions; M1-1-4. Calculate the job turnover loss of the job based on the job weight and the average monthly salary; The expression for calculating the job loss of the job is: C ls =W zw ×G yg ; C ls represents job loss, G yg represents the average monthly salary; M1-1-5. Determine the historical loss after the loss of a number of positions; The expression for determining the historical loss after the loss of a number of positions is: C lsls =T zkw ×(C zp +CG sc +C ls -G yg ) C lsls represents the historical loss, C zp represents the average monthly recruitment loss, C sc Represents the average monthly productivity.

8. The outsourcing employee behavior-based attrition prediction system according to claim 7, characterized in that: The output history loss prediction loss includes: M2-1, receiving the optimal churn probability prediction value as the input of the MLP model input layer, and obtaining the input layer output value after weighted bias processing of the first fully connected layer; M2-2, using the output value of the input layer as the input of the hidden layer, and obtaining the output value of the hidden layer after the weighted bias of the second fully connected layer and the nonlinear transformation of the activation layer; M2-3, sequentially use the output values ​​of several hidden layers as the input of the next hidden layer, and obtain the output value of the last hidden layer through the weighted bias of multiple fully connected layers and the nonlinear transformation of the activation layer; M2-4. The output value of the last hidden layer is passed to the output layer, and a linear transformation is performed through the fully connected layer of the output layer to output the historical churn prediction loss.

9. The outsourcing employee behavior-based attrition prediction system according to claim 8, characterized in that: Perform backpropagation on the MLP model and iteratively update the model parameters to minimize the mean squared error loss, including: M4-1. Calculate the gradient of the loss function relative to the output layer; M4-2, back-calculate the gradients of several hidden layers in sequence until the input layer; M4-3. Use the gradient descent algorithm to iteratively update the model parameters according to the gradients of the output layer, several hidden layers, and the input layer; M4-4. Calculate the error loss of each iterative update based on the iteratively updated model parameters until the error loss is minimized.

Images