Satisfaction prediction model construction method and device, electronic equipment and storage medium
By pre-screening univariate regression and classification model groups and combining them with the XGBoost algorithm, feature data is accurately selected, which solves the problems of redundancy and omission caused by feature selection in traditional methods and improves the accuracy of the satisfaction prediction model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA MOBILE GRP BEIJING
- Filing Date
- 2022-04-28
- Publication Date
- 2026-05-29
AI Technical Summary
In the process of machine learning modeling, existing technologies and traditional feature selection methods lead to redundant and missed feature information, resulting in low accuracy in predicting customer satisfaction.
By pre-screening univariate regression and univariate classification model groups, the original numerical and discrete feature data are filtered to construct a satisfaction prediction model. The XGBoost model construction algorithm is used to combine numerical and discrete feature modeling index data.
It reduces the interference of redundant feature information when building the satisfaction prediction model, improves the prediction accuracy, and can accurately predict customer satisfaction.
Smart Images
Figure CN117056870B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of machine learning modeling technology, and in particular to a method, apparatus, electronic device and storage medium for constructing a satisfaction prediction model. Background Technology
[0002] Traditional methods for feature selection in current big data analysis of user satisfaction mainly include manual experience-based screening, mean comparison, and linear graphing to check correlations. However, machine learning modeling requires continuous experimentation and selection to build an optimal model. Using traditional feature selection methods in machine learning modeling can lead to some features being of little value to the optimal model, or even producing side effects. This results in feature redundancy and missed features, leading to low model prediction accuracy and ultimately failing to accurately predict customer satisfaction. Summary of the Invention
[0003] This application provides a method, apparatus, electronic device, and storage medium for constructing a satisfaction prediction model, aiming to build a satisfaction prediction model with high prediction accuracy and accurately predict customer satisfaction.
[0004] Firstly, this application provides a method for constructing a satisfaction prediction model, including:
[0005] Determine the business domain indicator data, and classify the business domain indicator data according to the data characteristics to obtain the original numerical feature data and the original discrete feature data.
[0006] The original numerical feature data is filtered by a pre-screened univariate regression model group to obtain numerical feature modeling index data.
[0007] The original discrete feature data is filtered by a pre-screened univariate classification model group to obtain discrete feature modeling index data.
[0008] A satisfaction prediction model is constructed by combining the numerical feature modeling index data and the discrete feature modeling index data with the model building algorithm.
[0009] In one embodiment, the step of filtering the original numerical feature data through a pre-screened univariate regression model group to obtain numerical feature modeling index data includes:
[0010] Using the original numerical feature data as independent variables and the operational domain indicator data as dependent variables, the pre-screened univariate regression model group is constructed, wherein the pre-screened univariate regression model group includes multiple first pre-screening models.
[0011] Determine the regression sum of squares ratio value of each of the first pre-screened models in the pre-screened univariate regression model group;
[0012] The numerical feature modeling index data is obtained based on the regression sum of squares ratio values of each of the first pre-screened models.
[0013] The step of obtaining the numerical feature modeling index data based on the regression sum of squares ratio values of each of the first pre-screened models includes:
[0014] Calculate the average of the regression sum of squares ratio values of each of the first pre-screened models to obtain the target regression sum of squares ratio value of the pre-screened univariate regression model group;
[0015] Based on the target regression sum of squares ratio value and the first preset regression sum of squares ratio value, the characteristic index data of the first potential satisfaction prediction model are determined.
[0016] By combining the actual business information of the indicator data, the feature indicator data of the first potential satisfaction prediction model is filtered to obtain the numerical feature modeling indicator data.
[0017] After filtering the original numerical feature data using a pre-screened univariate regression model group to obtain numerical feature modeling index data, the process further includes:
[0018] Based on the target regression sum of squares ratio value of the pre-screened univariate regression model group and the second preset regression sum of squares ratio value, the target numerical feature data of the pre-screened univariate regression model group is determined.
[0019] The target numerical feature data is used to replace the original numerical feature data of the pre-screened univariate regression model group as the independent variable to construct the target pre-screened univariate regression model group.
[0020] The target numerical feature data is filtered by the target pre-screening univariate regression model group to obtain the replaced numerical feature modeling index data.
[0021] The process of filtering the original discrete feature data using a pre-screened univariate classification model group to obtain discrete feature modeling index data includes:
[0022] Using the original discrete feature data as independent variables and the operational domain indicator data as dependent variables, the pre-screening univariate classification model group is constructed, wherein the pre-screening univariate classification model group includes multiple second pre-screening models.
[0023] Determine the balanced score value of each of the second pre-screened models in the pre-screened univariate classification model group;
[0024] The discrete feature modeling index data is obtained based on the balance scores of each of the second pre-screening models.
[0025] The step of obtaining the discrete feature modeling index data based on the balanced scores of each of the second pre-screened models includes:
[0026] Calculate the average of the balance scores of each of the second pre-screening models to obtain the target balance score of the pre-screened univariate classification model group;
[0027] Based on the target balance score and the first preset balance score, determine the feature index data of the second potential satisfaction prediction model;
[0028] By combining clustering statistical analysis algorithms and actual business information from indicator data, the feature indicator data of the second potential satisfaction prediction model is filtered to obtain the discrete feature modeling indicator data.
[0029] After filtering the original discrete feature data using a pre-screened univariate classification model group to obtain discrete feature modeling index data, the process further includes:
[0030] Based on the target balanced score value and the second preset balanced score value of the pre-screened univariate classification model group, the target discrete feature data of the pre-screened univariate classification model group are determined.
[0031] The target discrete feature data is used to replace the original discrete feature data of the pre-screened univariate classification model group as the independent variable to construct the target pre-screened univariate classification model group.
[0032] The discrete feature data of the target is filtered by the target pre-screening univariate classification model group to obtain the replaced discrete feature modeling index data.
[0033] Secondly, this application provides a device for constructing a satisfaction prediction model, comprising:
[0034] The classification module is used to determine business domain indicator data and classify the business domain indicator data according to data characteristics to obtain original numerical feature data and original discrete feature data.
[0035] The first screening module is used to screen the original numerical feature data through a pre-screened univariate regression model group to obtain numerical feature modeling index data.
[0036] The second filtering module is used to filter the original discrete feature data through a pre-filtered univariate classification model group to obtain discrete feature modeling index data.
[0037] The model building module is used to build a satisfaction prediction model by combining the numerical feature modeling index data and the discrete feature modeling index data according to the model building algorithm.
[0038] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the satisfaction prediction model construction method described in the first aspect.
[0039] Fourthly, this application also provides a non-transitory computer-readable storage medium, the non-transitory computer-readable storage medium including a computer program, which, when executed by the processor, implements the satisfaction prediction model construction method described in the first aspect.
[0040] Fifthly, this application also provides a computer program product, which includes a computer program that, when executed by the processor, implements the satisfaction prediction model construction method described in the first aspect.
[0041] The satisfaction prediction model construction method, apparatus, electronic device, and storage medium provided in this application, in the process of constructing the satisfaction prediction model, accurately select feature information from the original numerical feature data and the original discrete feature data by pre-screening univariate regression model groups and pre-screening univariate classification model groups, thereby reducing the interference of redundant feature information during the construction of the satisfaction prediction model and constructing a satisfaction prediction model with high prediction accuracy. The satisfaction prediction model with high prediction accuracy can accurately predict customer satisfaction. Attached Figure Description
[0042] To more clearly illustrate the technical solutions of this application, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0043] Figure 1 This is one of the flowcharts for the satisfaction prediction model construction method provided in this application;
[0044] Figure 2 This is the second flowchart of the satisfaction prediction model construction method provided in this application;
[0045] Figure 3 This is the third flowchart of the satisfaction prediction model construction method provided in this application;
[0046] Figure 4 This is the fourth flowchart of the satisfaction prediction model construction method provided in this application;
[0047] Figure 5 This is the fifth flowchart of the satisfaction prediction model construction method provided in this application;
[0048] Figure 6 This is a schematic diagram of the structure of the satisfaction prediction model construction device provided in this application;
[0049] Figure 7 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation
[0050] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0051] Combination Figures 1 to 7 This application describes the method, apparatus, electronic device, and storage medium for constructing a satisfaction prediction model. Figure 1 This is one of the flowcharts for the satisfaction prediction model construction method provided in this application; Figure 2 This is the second flowchart of the satisfaction prediction model construction method provided in this application; Figure 3 This is the third flowchart of the satisfaction prediction model construction method provided in this application; Figure 4 This is the fourth flowchart of the satisfaction prediction model construction method provided in this application; Figure 5 This is the fifth flowchart of the satisfaction prediction model construction method provided in this application; Figure 6 This is a schematic diagram of the structure of the satisfaction prediction model construction device provided in this application; Figure 7 This is a schematic diagram of the structure of the electronic device provided in this application.
[0052] This application provides an embodiment of a method for constructing a satisfaction prediction model. It should be noted that although the logical order is shown in the flowchart, under certain data conditions, the steps shown or described may be completed in a different order than that shown here.
[0053] This application uses an electronic device as an example to illustrate the execution subject, and the model building system is used as one of the manifestations of the electronic device, without limitation.
[0054] Reference Figure 1 , Figure 1 This is one of the flowcharts for the satisfaction prediction model construction method provided in this application. The satisfaction prediction model construction method provided in this application includes:
[0055] Step S10: Determine the business domain indicator data, and classify the business domain indicator data according to the data characteristics to obtain the original numerical feature data and the original discrete feature data.
[0056] It should be noted that, in order to broaden the prediction scope and optimize user perceived satisfaction, this application embodiment uses B-domain (Business Support System) indicator data and O-domain (Operation Support System) indicator data, combined with multi-period user survey sample data, to model using machine learning algorithms and identify user groups with perceived dissatisfaction through quantitative evaluation methods. Therefore, this application embodiment can be understood as a method for constructing a satisfaction prediction model based on business domain indicator data and operation domain indicator data.
[0057] Furthermore, the business domain indicator data consists of basic user characteristic information, which includes, but is not limited to, subjective characteristic information, terminal characteristic information, brand characteristic information, tariff characteristic information, data package characteristic information, and data service characteristic information. In one embodiment, the business domain indicator data is shown in Table 1. Table 1 contains the business domain indicator data.
[0058] Table 1 Business Domain Indicator Data
[0059]
[0060] Furthermore, the operational domain indicator data mainly includes, but is not limited to, the overall network quality poor frequency statistics, MOS abnormal event indicators, WJTHD abnormal event indicators, network coverage table indicators, and internet access quality table indicators.
[0061] Furthermore, the overall network quality degradation frequency statistics include, but are not limited to, the number of wireless degradation frequencies, the number of large packet degradation frequencies, the number of small packet degradation frequencies, TCP connection establishment confirmations, the first HTTP packet response, and end-to-end issues. In one embodiment, the overall network quality degradation frequency statistics are shown in Table 2. Table 2 contains the overall network quality degradation frequency statistics.
[0062] Table 2. Statistical data on the number of times the network quality was poor.
[0063]
[0064] Furthermore, the MOS and WJTHD abnormal event indicator data include, but are not limited to, the number of MOS abnormal events, the number of WJT and GMHD abnormal events, the number of users dissatisfied with coverage, the number of users dissatisfied with internet access, the number of users dissatisfied with VoLTE, GMS dropped calls, GMS short-term second calls, VLR network disconnection, and GMS black holes. In one embodiment, the MOS and WJTHD abnormal event indicator data are shown in Table 3. Table 3 contains the MOS and WJTHD abnormal event indicator data.
[0065] Table 3 shows the abnormal event index data for MOS and WJTHD.
[0066]
[0067] Furthermore, the network coverage table indicator data includes, but is not limited to, outgoing cell time, outgoing cell time, incoming cell time, and incoming cell time. In one embodiment, the network may be a 5G network; therefore, the network coverage table indicator data is shown in Table 4. Table 4 contains the network coverage table indicator data.
[0068] Table 4 Network Coverage Table Indicator Data
[0069]
[0070] Furthermore, the Internet quality table index data includes unrestricted uplink coding average rate, downlink coding average rate, uplink jitter, downlink jitter, latency, and packet loss. In one embodiment, the Internet quality table index data is shown in Table 5. Table 5 contains the Internet quality table index data.
[0071] Table 5. Internet Quality Indicator Data
[0072]
[0073] Furthermore, the user complaint raw data and business domain indicator data in this embodiment contain a total of 1700+ feature information, which are further supplemented by 100+ operational domain indicator data.
[0074] Specifically, the model building system determines business domain indicator data and classifies it according to data characteristics. In this embodiment, the data characteristics include continuous data characteristics and discrete data characteristics. Continuous data characteristics include, but are not limited to, tariff characteristics, package traffic characteristics, and traffic service characteristics. Discrete data characteristics include, but are not limited to, subjective characteristics, terminal characteristics, user level characteristics, and brand characteristics. This can be understood as follows: the model building system classifies the business domain indicator data according to continuous and discrete data characteristics, obtaining feature data for continuous data characteristics and feature data for discrete data characteristics. The feature data for continuous data characteristics is numerical feature data, and the feature data for discrete data characteristics is discrete feature data. Therefore, the model building system obtains the original numerical feature data and the original discrete feature data of the business domain indicator data based on the data characteristics.
[0075] Step S20: The original numerical feature data is filtered by a pre-screened univariate regression model group to obtain numerical feature modeling index data.
[0076] Furthermore, the model building system uses the original numerical feature data as independent variables and the operational domain indicator data as dependent variables to construct a pre-screened univariate regression model group, where the operational domain indicator data can be understood as satisfaction values. Therefore, it can be further understood that the model building system uses the original numerical feature data as independent variables and the satisfaction values as dependent variables to construct a pre-screened univariate regression model group, where the pre-screened univariate regression model group includes multiple first pre-screening models. Next, the model building system determines the regression sum of squares ratio value of each first pre-screening model in the pre-screened univariate regression model group. Finally, the model building system obtains the numerical feature modeling indicator data based on the regression sum of squares ratio values of each first pre-screening model, as described in steps S201 to S203.
[0077] Step S30: The original discrete feature data is filtered by a pre-screened univariate classification model group to obtain discrete feature modeling index data.
[0078] Furthermore, the model building system uses the original discrete feature data as independent variables and the operational domain indicator data as dependent variables to construct a pre-screened univariate classification model group, where the operational domain indicator data can be understood as satisfaction values. Therefore, it can be further understood that the model building system uses the original discrete feature data as independent variables and the satisfaction values as dependent variables to construct a pre-screened univariate classification model group, where the pre-screened univariate classification model group includes multiple second pre-screening models. Next, the model building system determines the balance score value of each second pre-screening model in the pre-screened univariate classification model group. Finally, the model building system obtains the discrete feature modeling indicator data based on the balance score values of each second pre-screening model, as described in steps S301 to S303.
[0079] Step S40: Based on the model building algorithm, combine the numerical feature modeling index data and the discrete feature modeling index data to construct a satisfaction prediction model.
[0080] It should be further noted that, after obtaining the numerical feature modeling index data and the discrete feature modeling index data, the embodiments of this application need to construct a satisfaction prediction model through a model building algorithm. The model building algorithm includes, but is not limited to, support vector machine model building algorithm, Bayesian classification model building algorithm, XGBoost model building algorithm and convolutional neural network model building algorithm.
[0081] Furthermore, considering the precision and recall of various model building algorithms on the test set, the XGBoost model building algorithm has a higher precision and recall than the Support Vector Machine, Bayesian classification, and Convolutional Neural Network models. Therefore, this embodiment uses the XGBoost model building algorithm, which has better precision and recall, as the final model building algorithm. The model building system uses the XGBoost model building algorithm combined with numerical and discrete feature modeling index data to build a satisfaction prediction model. The objective loss function for training the k-th tree in the XGBoost model building algorithm is Obj. k :
[0082]
[0083] Among them, item It is the loss between the actual value and the cumulative prediction results up to the (k-1)th tree. Term f k (x i ) is the k-th tree pair of sample x i The prediction results. Term Ω(f) K L1 and L2 regularization are used to control model complexity and can prevent overfitting.
[0084] Furthermore, to achieve faster computation, this embodiment of the application performs a second-order Taylor expansion on the target loss function and optimizes it using first-order and second-order gradients, as follows:
[0085]
[0086]
[0087]
[0088] According to the Taylor expansion transformation formula:
[0089]
[0090] The XGBoost model building algorithm is an optimized version of the Boosting ensemble algorithm. Boosting ensembles are a commonly used statistical learning method that trains multiple classifiers by changing the weights of training samples during training, ultimately obtaining the optimal classifier. After each training round, the weights of correctly classified training samples are reduced, while the weights of misclassified samples are increased. After multiple training iterations, some misclassified training samples receive more attention, while the weights of correctly classified training samples approach zero, resulting in multiple simple classifiers. These classifiers are then combined to obtain a final model.
[0091] The XGBoost model building algorithm has been optimized and improved based on the Boosting ensemble algorithm as follows: (1) In each iteration, the decision tree is used as the base learner for serial training and optimization; (2) The loss function is expanded by second-order Taylor and optimized by first-order and second-order gradients of the model building algorithm, which makes the calculation faster; (3) A regularization term is added to the loss function to reduce the model variance and control the complexity of the model; (4) Feature sampling is supported, which not only prevents overfitting but also reduces the computation; (5) When there are missing values in the sample, the splitting direction can be learned automatically; (6) Parallel computation is supported.
[0092] In one embodiment, the XGBoost model building algorithm includes four decision trees: decision tree 1, decision tree 2, decision tree 3, and decision tree 4, as well as model training for decision trees 1, 2, 3, and 4. The implementation flow of the XGBoost model building algorithm is as follows: After sample data 1 is trained by decision tree 1 and its model, the model output is 4 points, and the calculated residual is input into decision tree 2. After training by decision tree 2 and its model, the model output is 2 points, and the calculated residual is input into decision tree 3. After training by decision tree 3 and its model, the model output is -2 points, and the calculated residual is input into decision tree 4. After training by decision tree 4 and its model, the model output is 1 point. Finally, the ensemble output of the XGBoost model building algorithm is 4 + 2 - 2 + 1 = 5 points.
[0093] In one embodiment, the actual effect of the XGBoost model building algorithm on the dual rate of voice call quality of surveyed users is shown in Table 6. Table 6 shows the actual effect of the dual rate of voice call quality of surveyed users.
[0094] In one embodiment, the actual effect of the XGBoost model building algorithm on the double rate of mobile internet quality of surveyed users is shown in Table 7. Table 7 shows the actual effect of the double rate of mobile internet quality of surveyed users.
[0095] Table 6 shows the actual effect of the survey on the dual rates of voice call quality among users.
[0096]
[0097] Table 7. Actual Results of the Survey on Mobile Internet Quality Rates
[0098]
[0099] The satisfaction prediction model construction method provided in this application, during the construction process, accurately selects feature information from the original numerical feature data and the original discrete feature data by pre-screening univariate regression model groups and pre-screening univariate classification model groups. This reduces the interference of redundant feature information during the construction of the satisfaction prediction model, thereby constructing a satisfaction prediction model with high prediction accuracy. The satisfaction prediction model with high prediction accuracy can accurately predict customer satisfaction.
[0100] Reference Figure 2 , Figure 2 This is the second flowchart of the satisfaction prediction model construction method provided in this application. Step S20 includes steps S201 to S203:
[0101] Step S201: Using the original numerical feature data as the independent variable and the operational domain indicator data as the dependent variable, construct the pre-screened univariate regression model group, wherein the pre-screened univariate regression model group includes multiple first pre-screened models.
[0102] Step S202: Determine the regression sum of squares ratio of each of the first pre-screened models in the pre-screened univariate regression model group;
[0103] Step S203: Obtain the numerical feature modeling index data based on the regression sum of squares ratio values of each of the first pre-screened models.
[0104] For raw numerical feature data (such as active days, monthly package fees, and number of manual calls), the model building system uses the raw numerical feature data as independent variables and operational domain indicator data as dependent variables to construct a pre-screening univariate regression model set. The operational domain indicator data can be understood as satisfaction values. Therefore, it can be understood that the model building system uses the raw numerical feature data as independent variables and satisfaction values as dependent variables to construct a pre-screening univariate regression model set. This pre-screening univariate regression model set includes multiple first-level pre-screening models, which include, but are not limited to, linear pre-screening models, second-order multinomial pre-screening models, exponential pre-screening models, logarithmic pre-screening models, and power-law pre-screening models.
[0105] Furthermore, the model building system determines the first regression sum of squares ratio of the linear pre-screening model, the second regression sum of squares ratio of the second-order multinomial pre-screening model, the third regression sum of squares ratio of the exponential pre-screening model, the fourth regression sum of squares ratio of the logarithmic pre-screening model, and the fifth regression sum of squares ratio of the power pre-screening model.
[0106] Furthermore, the model building system obtains numerical feature modeling index data based on the first regression sum of squares ratio of the linear pre-screening model, the second regression sum of squares ratio of the second-order multinomial pre-screening model, the third regression sum of squares ratio of the exponential pre-screening model, the fourth regression sum of squares ratio of the logarithmic pre-screening model, and the fifth regression sum of squares ratio of the power pre-screening model, as described in steps S2031 to S2033.
[0107] This application embodiment determines numerical feature modeling index data by pre-screening the regression sum of squares ratio values of each first pre-screened model in the univariate regression model group, thereby obtaining numerical feature modeling index data with low interference from redundant feature information, providing a data foundation for constructing a satisfaction prediction model with high prediction accuracy, and thus accurately predicting customer satisfaction through a satisfaction prediction model with high prediction accuracy.
[0108] Further, steps S2031 to S2033 are described as follows:
[0109] Step S2031: Calculate the average value of the regression sum of squares ratio of each first pre-screening model to obtain the target regression sum of squares ratio of the pre-screened univariate regression model group.
[0110] Step S2032: Determine the feature index data of the first potential satisfaction prediction model based on the target regression sum of squares ratio value and the first preset regression sum of squares ratio value.
[0111] Step S2033: Combine the actual business information of the indicator data to filter the feature indicator data of the first potential satisfaction prediction model and obtain numerical feature modeling indicator data.
[0112] Specifically, the model building system calculates the average of the first regression sum of squares ratio of the linear pre-screening model, the second regression sum of squares ratio of the second-order multinomial pre-screening model, the third regression sum of squares ratio of the exponential pre-screening model, the fourth regression sum of squares ratio of the logarithmic pre-screening model, and the fifth regression sum of squares ratio of the power pre-screening model, to obtain the target regression sum of squares ratio of the pre-screened univariate regression model group.
[0113] Furthermore, the model building system compares the target regression sum of squares ratio with the first preset regression sum of squares ratio to obtain a comparison result. The comparison result can be that the target regression sum of squares ratio is greater than or equal to the first preset regression sum of squares ratio, or the comparison result can be that the target regression sum of squares ratio is less than the first preset regression sum of squares ratio. The first preset regression sum of squares ratio is set according to the actual situation. In the embodiments of this application, the first preset regression sum of squares ratio is preferably 0.3.
[0114] Furthermore, the model building system identifies indicator data whose target regression sum of squares ratio is greater than or equal to a first preset regression sum of squares ratio as feature indicator data for the first potential satisfaction prediction model. Specifically, indicator data with a target regression sum of squares ratio greater than or equal to 0.3 are identified as feature indicator data for the first potential satisfaction prediction model. Next, the model building system combines the actual business information of the indicator data to filter the feature indicator data for the first potential satisfaction prediction model, resulting in numerical feature modeling indicator data. The final number of numerical feature modeling indicator data is 45.
[0115] In one embodiment, the first preset regression sum of squares ratio is 0.3. The original numerical feature data includes active days index data, monthly package fee index data, and number of manual calls index data. The target regression sum of squares ratio for the active days index data is 0.2, which is less than 0.3; the target regression sum of squares ratio for the monthly package fee index data is 0.35, which is greater than 0.3; and the target regression sum of squares ratio for the number of manual calls index data is 0.4, which is greater than 0.3. Therefore, the model building system determines the monthly package fee index data and the number of manual calls index data from the original numerical feature data as numerical feature modeling index data.
[0116] This application embodiment obtains numerical feature modeling indicator data with low redundancy interference by using the regression sum of squares ratio values and actual business information of the indicator data of each first pre-screened model in the pre-screened univariate regression model group. This provides a data foundation for constructing a satisfaction prediction model with high prediction accuracy, thereby enabling accurate prediction of customer satisfaction through the satisfaction prediction model with high prediction accuracy.
[0117] Reference Figure 3 , Figure 3 This is the third flowchart of the satisfaction prediction model construction method provided in this application, which includes steps S50 to S70 after step S20:
[0118] Step S50: Determine the target numerical feature data of the pre-screened univariate regression model group based on the target regression sum of squares ratio value of the pre-screened univariate regression model group and the second preset regression sum of squares ratio value.
[0119] Step S60: Replace the original numerical feature data of the pre-screened univariate regression model group with the target numerical feature data as the independent variable to construct the target pre-screened univariate regression model group.
[0120] Step S70: The target numerical feature data is filtered through the target pre-screening univariate regression model group to obtain the replaced numerical feature modeling index data.
[0121] Specifically, the model building system compares the target regression sum of squares ratio with a second preset regression sum of squares ratio to obtain a comparison result. The comparison result can be that the target regression sum of squares ratio is greater than or equal to the second preset regression sum of squares ratio, or it can be that the target regression sum of squares ratio is less than the second preset regression sum of squares ratio. The second preset regression sum of squares ratio is set according to actual conditions; in this embodiment, the second preset regression sum of squares ratio is preferably 0.5. Further, if the comparison result is determined to be that the target regression sum of squares ratio is greater than or equal to the second preset regression sum of squares ratio, the model building system determines the mean value of the output values of the pre-screened univariate regression model group as the target numerical feature data. That is, if the comparison result is determined to be that the target regression sum of squares ratio is greater than or equal to 0.5, the model building system determines the mean value of the output values of the pre-screened univariate regression model group as the target numerical feature data.
[0122] Furthermore, the model building system replaces the original numerical feature data of the pre-screened univariate regression model group with the target numerical feature data as the independent variable, and uses the satisfaction value as the dependent variable to construct the target pre-screened univariate regression model group.
[0123] Furthermore, the model building system follows the screening method in steps S201 to S203 above, and filters the target numerical feature data through the target pre-screening univariate regression model group to obtain the replaced numerical feature modeling index data.
[0124] Furthermore, the model building system combines the XGBoost model building algorithm from step S40, the replaced discrete feature modeling index data, and the replaced numerical feature modeling index data to build a satisfaction prediction model.
[0125] This application embodiment transforms the original numerical feature data by using the target regression sum of squares ratio value and the second preset regression sum of squares ratio value, thereby obtaining numerical feature modeling index data with low interference from redundant feature information. This provides a data foundation for constructing a satisfaction prediction model with high prediction accuracy, thus enabling accurate prediction of customer satisfaction through the high prediction accuracy satisfaction prediction model.
[0126] Reference Figure 4 , Figure 4 This is the fourth flowchart of the satisfaction prediction model construction method provided in this application. Step S30 includes steps S301 to S303:
[0127] Step S301: Using the original discrete feature data as the independent variable and the operational domain indicator data as the dependent variable, construct the pre-screened univariate classification model group, wherein the pre-screened univariate classification model group includes multiple second pre-screening models.
[0128] Step S302: Determine the balanced score value of each of the second pre-screened models in the pre-screened univariate classification model group;
[0129] Step S303: Obtain the discrete feature modeling index data based on the balanced scores of each of the second pre-screened models.
[0130] For raw discrete feature data (such as ID card with N identification indicators, three major brand affiliation indicators, dual-SIM phone identification indicators, dual-SIM type indicators, customer star rating indicators, customer type indicators, customer gender indicators, dual-SIM support indicators, terminal brand indicators, and broadband user status indicators), the screening method in this application embodiment is as follows: The model building system uses the raw discrete feature data as independent variables and the operational domain indicator data as dependent variables to construct a pre-screening univariate classification model group, where the operational domain indicator data can be understood as satisfaction values. Therefore, it can be further understood that the model building system uses the raw discrete feature data as independent variables and the satisfaction values as dependent variables to construct a pre-screening univariate classification model group, where the pre-screening univariate classification model group includes multiple second pre-screening models, including but not limited to C4.5 decision pre-screening models, CART decision pre-screening models, Bayesian pre-screening models, logistic regression pre-screening models, and support vector machine pre-screening models.
[0131] Furthermore, the model building system determines the first balance score of the C4.5 decision pre-screening model, the second balance score of the CART decision pre-screening model, the third balance score of the Bayesian pre-screening model, the fourth balance score of the logistic regression pre-screening model, and the fifth balance score of the support vector machine pre-screening model.
[0132] Furthermore, the model building system obtains discrete feature modeling index data based on the first balanced score of the C4.5 decision pre-screening model, the second balanced score of the CART decision pre-screening model, the third balanced score of the Bayesian pre-screening model, the fourth balanced score of the logistic regression pre-screening model, and the fifth balanced score of the support vector machine pre-screening model, as described in steps S3031 to S3033.
[0133] This application embodiment obtains discrete feature modeling index data with low redundancy interference by balancing the scores of each second pre-screened model in the pre-screened univariate classification model group. This provides a data foundation for constructing a satisfaction prediction model with high prediction accuracy, thereby enabling accurate prediction of customer satisfaction.
[0134] Further, steps S3031 to S3033 are described as follows:
[0135] Step S3031: Calculate the average of the balance scores of each of the second pre-screening models to obtain the target balance score of the pre-screening univariate classification model group.
[0136] Step S3032: Determine the feature index data of the second potential satisfaction prediction model based on the target balance score and the first preset balance score.
[0137] Step S3033: Combining the clustering statistical analysis algorithm and the actual business information of the indicator data, the feature indicator data of the second potential satisfaction prediction model is filtered to obtain the discrete feature modeling indicator data.
[0138] Specifically, the model building system calculates the average of the first balance score of the C4.5 decision pre-screening model, the second balance score of the CART decision pre-screening model, the third balance score of the Bayesian pre-screening model, the fourth balance score of the logistic regression pre-screening model, and the fifth balance score of the support vector machine pre-screening model to obtain the target balance score of the pre-screened univariate classification model group.
[0139] Furthermore, the model building system compares the target balance score with the first preset balance score to obtain a comparison result. The comparison result can be that the target balance score is greater than or equal to the first preset balance score, or the comparison result can be that the target balance score is less than the first preset balance score. The first preset balance score is set according to the actual situation. In this embodiment, the first preset balance score is preferably 0.3.
[0140] Furthermore, the model building system identifies indicator data whose target balance score is greater than or equal to the first preset balance score as feature indicator data for the second potential satisfaction prediction model. Specifically, indicator data with a target balance score greater than or equal to 0.3 are identified as feature indicator data for the second potential satisfaction prediction model. Next, the model building system combines clustering statistical analysis algorithms with actual business information from the indicator data to filter the feature indicator data for the second potential satisfaction prediction model, resulting in discrete feature modeling indicator data. The final number of discrete feature modeling indicator data is 14.
[0141] In one embodiment, the first preset balance score value is 0.3. The original discrete feature data includes three major brand affiliation index data, customer star rating index data, and customer gender index data. The target balance score value for the three major brand affiliation index data is 0.5, which is greater than 0.3; the target balance score value for the customer star rating index data is 0.2, which is less than 0.3; and the target balance score value for the customer gender index data is 0.37, which is greater than 0.3. Therefore, the model building system determines the three major brand affiliation index data and the customer gender index data from the original discrete feature data as discrete feature modeling index data.
[0142] This application embodiment obtains discrete feature modeling indicator data with low redundancy interference by using the balanced scores of each second pre-screening model in the pre-screened univariate classification model group, clustering statistical analysis algorithm, and actual business information of indicator data. This provides a data foundation for constructing a satisfaction prediction model with high prediction accuracy, thereby enabling accurate prediction of customer satisfaction through the high-precision satisfaction prediction model.
[0143] Reference Figure 5 , Figure 5 This is the fifth flowchart of the satisfaction prediction model construction method provided in this application, which includes steps S80 to S100 after step S30:
[0144] Step S80: Determine the target discrete feature data of the pre-screened univariate classification model group based on the target balanced score value and the second preset balanced score value.
[0145] Step S90: Replace the original discrete feature data of the pre-screened univariate classification model group with the target discrete feature data as the independent variable to construct the target pre-screened univariate classification model group.
[0146] Step S100: The discrete feature data of the target is filtered by the target pre-screening univariate classification model group to obtain the replaced discrete feature modeling index data.
[0147] Specifically, the model building system compares the target balance score with the second preset balance score to obtain a comparison result. The comparison result can be that the target balance score is greater than or equal to the second preset balance score, or the comparison result can be that the target balance score is less than the second preset balance score. The second preset balance score is set according to the actual situation. In this embodiment, the second preset balance score is preferably 0.5.
[0148] Furthermore, if the comparison result is determined to be an indicator data whose target balance score is greater than or equal to the second preset balance score, the model building system will determine the indicator data of the mode value of the output values of the pre-screened univariate classification model group as the target discrete feature data for screening the univariate classification model group. That is, if the comparison result is determined to be a target balance score greater than or equal to 0.5, the model building system will determine the indicator data of the mode value of the output values of the pre-screened univariate classification model group as the target discrete feature data for screening the univariate classification model group.
[0149] Furthermore, the model building system replaces the original discrete feature data of the pre-screened univariate classification model group with the target discrete feature data as the independent variable, and uses the satisfaction value as the dependent variable to construct the target pre-screened univariate classification model group.
[0150] Furthermore, the model building system follows the screening method in steps S301 to S303 above, and filters the discrete feature data of the target through the target pre-screening univariate classification model group to obtain the replaced discrete feature modeling index data.
[0151] Furthermore, the model building system combines the XGBoost model building algorithm from step S40, the replaced discrete feature modeling index data, and the replaced numerical feature modeling index data to build a satisfaction prediction model.
[0152] This application embodiment transforms the original discrete feature data by using the target balanced score and the second preset balanced score, thereby obtaining discrete feature modeling index data with low interference from redundant feature information. This provides a data foundation for constructing a satisfaction prediction model with high prediction accuracy, thus enabling accurate prediction of customer satisfaction.
[0153] It should be further noted that as the number of input metrics increases, the harmonic mean balance F-score of the model's two rates (precision and recall) shows an increasing trend. However, after the number of metrics increases to 46, the model's performance essentially stabilizes. When further metrics such as standard GPRS tariffs and VIP annual packages are added, the balance F-score shows virtually no increase, indicating that the 46 metrics essentially represent the upper limit of the model's performance; these are the most effective metrics.
[0154] Furthermore, the apparatus for constructing a satisfaction prediction model provided in this application is described, and the apparatus for constructing a satisfaction prediction model and the method for constructing a satisfaction prediction model can be referred to in correspondence with each other.
[0155] like Figure 6 As shown, Figure 6 This is a schematic diagram of the structure of the satisfaction prediction model building device provided in this application. The satisfaction prediction model building device includes:
[0156] The classification module 601 is used to determine the business domain indicator data and classify the business domain indicator data according to the data characteristics to obtain the original numerical feature data and the original discrete feature data.
[0157] The first screening module 602 is used to screen the original numerical feature data through a pre-screened univariate regression model group to obtain numerical feature modeling index data.
[0158] The second filtering module 603 is used to filter the original discrete feature data through a pre-filtered univariate classification model group to obtain discrete feature modeling index data.
[0159] The model building module 604 is used to build a satisfaction prediction model based on the model building algorithm, combining numerical feature modeling index data and discrete feature modeling index data.
[0160] Furthermore, the first screening module 602 is also used for:
[0161] Using the original numerical feature data as independent variables and the operational domain indicator data as dependent variables, the pre-screened univariate regression model group is constructed, wherein the pre-screened univariate regression model group includes multiple first pre-screening models.
[0162] Determine the regression sum of squares ratio value of each of the first pre-screened models in the pre-screened univariate regression model group;
[0163] The numerical feature modeling index data is obtained based on the regression sum of squares ratio values of each of the first pre-screened models.
[0164] Furthermore, the first screening module 602 is also used for:
[0165] Calculate the average of the regression sum of squares ratio values of each of the first pre-screened models to obtain the target regression sum of squares ratio value of the pre-screened univariate regression model group;
[0166] Based on the target regression sum of squares ratio value and the first preset regression sum of squares ratio value, the characteristic index data of the first potential satisfaction prediction model are determined.
[0167] By combining the actual business information of the indicator data, the feature indicator data of the first potential satisfaction prediction model is filtered to obtain the numerical feature modeling indicator data.
[0168] Furthermore, the first screening module 602 is also used for:
[0169] Based on the target regression sum of squares ratio value of the pre-screened univariate regression model group and the second preset regression sum of squares ratio value, the target numerical feature data of the pre-screened univariate regression model group is determined.
[0170] The target numerical feature data is used to replace the original numerical feature data of the pre-screened univariate regression model group as the independent variable to construct the target pre-screened univariate regression model group.
[0171] The target numerical feature data is filtered by the target pre-screening univariate regression model group to obtain the replaced numerical feature modeling index data.
[0172] Furthermore, the second screening module 602 is also used for:
[0173] Using the original discrete feature data as independent variables and the operational domain indicator data as dependent variables, the pre-screening univariate classification model group is constructed, wherein the pre-screening univariate classification model group includes multiple second pre-screening models.
[0174] Determine the balanced score value of each of the second pre-screened models in the pre-screened univariate classification model group;
[0175] The discrete feature modeling index data is obtained based on the balance scores of each of the second pre-screening models.
[0176] Furthermore, the second screening module 602 is also used for:
[0177] Calculate the average of the balance scores of each of the second pre-screening models to obtain the target balance score of the pre-screened univariate classification model group;
[0178] Based on the target balance score and the first preset balance score, determine the feature index data of the second potential satisfaction prediction model;
[0179] By combining clustering statistical analysis algorithms and actual business information from indicator data, the feature indicator data of the second potential satisfaction prediction model is filtered to obtain the discrete feature modeling indicator data.
[0180] Furthermore, the second screening module 602 is also used for:
[0181] Based on the target balanced score value and the second preset balanced score value of the pre-screened univariate classification model group, the target discrete feature data of the pre-screened univariate classification model group are determined.
[0182] The target discrete feature data is used to replace the original discrete feature data of the pre-screened univariate classification model group as the independent variable to construct the target pre-screened univariate classification model group.
[0183] The discrete feature data of the target is filtered by the target pre-screening univariate classification model group to obtain the replaced discrete feature modeling index data.
[0184] The specific embodiments of the satisfaction prediction model construction device provided in this application are basically the same as the embodiments of the satisfaction prediction model construction method described above, and will not be repeated here.
[0185] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a satisfaction prediction model construction method, which includes:
[0186] Determine the business domain indicator data, and classify the business domain indicator data according to the data characteristics to obtain the original numerical feature data and the original discrete feature data.
[0187] The original numerical feature data is filtered by a pre-screened univariate regression model group to obtain numerical feature modeling index data.
[0188] The original discrete feature data is filtered by a pre-screened univariate classification model group to obtain discrete feature modeling index data.
[0189] A satisfaction prediction model is constructed by combining the numerical feature modeling index data and the discrete feature modeling index data with the model building algorithm.
[0190] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0191] On the other hand, this application also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer is able to execute the satisfaction prediction model construction method provided by the above methods. This method includes:
[0192] Determine the business domain indicator data, and classify the business domain indicator data according to the data characteristics to obtain the original numerical feature data and the original discrete feature data.
[0193] The original numerical feature data is filtered by a pre-screened univariate regression model group to obtain numerical feature modeling index data.
[0194] The original discrete feature data is filtered by a pre-screened univariate classification model group to obtain discrete feature modeling index data.
[0195] A satisfaction prediction model is constructed by combining the numerical feature modeling index data and the discrete feature modeling index data with the model building algorithm.
[0196] Furthermore, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the satisfaction prediction model construction methods provided above, the method comprising:
[0197] Determine the business domain indicator data, and classify the business domain indicator data according to the data characteristics to obtain the original numerical feature data and the original discrete feature data.
[0198] The original numerical feature data is filtered by a pre-screened univariate regression model group to obtain numerical feature modeling index data.
[0199] The original discrete feature data is filtered by a pre-screened univariate classification model group to obtain discrete feature modeling index data.
[0200] A satisfaction prediction model is constructed by combining the numerical feature modeling index data and the discrete feature modeling index data with the model building algorithm.
[0201] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0202] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0203] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for constructing a satisfaction prediction model, characterized in that, include: Determine the business domain indicator data, and classify the business domain indicator data according to the data characteristics to obtain the original numerical feature data and the original discrete feature data. The original numerical feature data is filtered by a pre-screened univariate regression model group to obtain numerical feature modeling index data. The original discrete feature data is filtered by a pre-screened univariate classification model group to obtain discrete feature modeling index data. A satisfaction prediction model is constructed based on the model building algorithm, combined with the numerical feature modeling index data and the discrete feature modeling index data. The process of filtering the original discrete feature data using a pre-screened univariate classification model group to obtain discrete feature modeling index data includes: Using the original discrete feature data as independent variables and the operational domain indicator data as dependent variables, the pre-screening univariate classification model group is constructed, wherein the pre-screening univariate classification model group includes multiple second pre-screening models. Determine the balanced score value of each of the second pre-screened models in the pre-screened univariate classification model group; The discrete feature modeling index data is obtained based on the balance scores of each of the second pre-screening models.
2. The method for constructing a satisfaction prediction model according to claim 1, characterized in that, The process of filtering the original numerical feature data using a pre-screened univariate regression model group to obtain numerical feature modeling index data includes: Using the original numerical feature data as independent variables and the operational domain indicator data as dependent variables, the pre-screened univariate regression model group is constructed, wherein the pre-screened univariate regression model group includes multiple first pre-screening models. Determine the regression sum of squares ratio value of each of the first pre-screened models in the pre-screened univariate regression model group; The numerical feature modeling index data is obtained based on the regression sum of squares ratio values of each of the first pre-screened models.
3. The method for constructing a satisfaction prediction model according to claim 2, characterized in that, The step of obtaining the numerical feature modeling index data based on the regression sum of squares ratio values of each of the first pre-screened models includes: Calculate the average of the regression sum of squares ratio values of each of the first pre-screened models to obtain the target regression sum of squares ratio value of the pre-screened univariate regression model group; Based on the target regression sum of squares ratio value and the first preset regression sum of squares ratio value, the characteristic index data of the first potential satisfaction prediction model are determined; By combining the actual business information of the indicator data, the feature indicator data of the first potential satisfaction prediction model is filtered to obtain the numerical feature modeling indicator data.
4. The method for constructing a satisfaction prediction model according to any one of claims 1 to 3, characterized in that, After filtering the original numerical feature data using a pre-screened univariate regression model group to obtain numerical feature modeling index data, the process further includes: Based on the target regression sum of squares ratio value of the pre-screened univariate regression model group and the second preset regression sum of squares ratio value, the target numerical feature data of the pre-screened univariate regression model group is determined. The target numerical feature data is used to replace the original numerical feature data of the pre-screened univariate regression model group as the independent variable to construct the target pre-screened univariate regression model group. The target numerical feature data is filtered by the target pre-screening univariate regression model group to obtain the replaced numerical feature modeling index data.
5. The method for constructing a satisfaction prediction model according to claim 1, characterized in that, The step of obtaining the discrete feature modeling index data based on the balanced scores of each of the second pre-screened models includes: Calculate the average of the balance scores of each of the second pre-screening models to obtain the target balance score of the pre-screened univariate classification model group; Based on the target balance score and the first preset balance score, determine the feature index data of the second potential satisfaction prediction model; By combining clustering statistical analysis algorithms and actual business information from indicator data, the feature indicator data of the second potential satisfaction prediction model is filtered to obtain the discrete feature modeling indicator data.
6. The method for constructing a satisfaction prediction model according to claim 1 or 5, characterized in that, After filtering the original discrete feature data using a pre-screened univariate classification model group to obtain discrete feature modeling index data, the process further includes: Based on the target balanced score value and the second preset balanced score value of the pre-screened univariate classification model group, the target discrete feature data of the pre-screened univariate classification model group are determined. The target discrete feature data is used to replace the original discrete feature data of the pre-screened univariate classification model group as the independent variable to construct the target pre-screened univariate classification model group. The discrete feature data of the target is filtered by the target pre-screening univariate classification model group to obtain the replaced discrete feature modeling index data.
7. A device for constructing a satisfaction prediction model, characterized in that, include: The classification module is used to determine business domain indicator data and classify the business domain indicator data according to data characteristics to obtain original numerical feature data and original discrete feature data. The first screening module is used to screen the original numerical feature data through a pre-screened univariate regression model group to obtain numerical feature modeling index data. The second filtering module is used to filter the original discrete feature data through a pre-filtered univariate classification model group to obtain discrete feature modeling index data. The model building module is used to build a satisfaction prediction model based on the model building algorithm, combining the numerical feature modeling index data and the discrete feature modeling index data. The process of filtering the original discrete feature data using a pre-screened univariate classification model group to obtain discrete feature modeling index data includes: Using the original discrete feature data as independent variables and the operational domain indicator data as dependent variables, the pre-screening univariate classification model group is constructed, wherein the pre-screening univariate classification model group includes multiple second pre-screening models. Determine the balanced score value of each of the second pre-screened models in the pre-screened univariate classification model group; The discrete feature modeling index data is obtained based on the balance scores of each of the second pre-screening models.
8. An electronic device, the electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the satisfaction prediction model construction method according to any one of claims 1 to 6.
9. A non-transitory computer-readable storage medium, the non-transitory computer-readable storage medium comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the satisfaction prediction model construction method according to any one of claims 1 to 6.