Outbound survey user screening method, device, equipment, medium and program product

By using the TabNet model to identify outbound call survey users, the problems of low user cooperation completion and low survey efficiency in the existing technology are solved, and the accuracy and efficiency of surveys are improved.

CN119052374APending Publication Date: 2024-11-29CHINA MOBILE GRP BEIJING +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411189646.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-28
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing user screening methods for out-of-call surveys have resulted in low user cooperation completion and low overall survey efficiency.

Method used

By obtaining the feature data set of users to be screened, the user identification model based on the TabNet model is used to identify the target users used for out-of-call research from the users to be screened. The model is iterative training and cluster analysis based on the historical feature data of all users. As a result, the dependence on feature engineering is reduced and the sample imbalance problem is solved.

Benefits of technology

The user's cooperation in the survey and the overall efficiency of outbound call surveys have been improved. By identifying users with a high probability of cooperation in completing the survey, the accuracy of the survey results has been improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119052374A_ABST
    Figure CN119052374A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of communication, and provides an outbound survey user screening method and device, equipment, a storage medium and a program product, and the method comprises the steps: obtaining a feature data set, such as a classified feature, a call statistical feature and a voice outbound feature, of a to-be-screened user, inputting the information into a user identification model to identify a target user for outbound investigation; wherein the user identification model is obtained by performing iterative training on a preset TabNet model based on a target sample; the target sample is obtained by performing clustering analysis and random sampling on a sample data set constructed based on the historical feature data of the total user. According to the method, the target user for outbound investigation is identified through the TabNet model, and the sample data set is subjected to clustering analysis and sampling at any time in the model training stage, so that the problem of sample imbalance is solved, the identification accuracy of the outbound investigation user is improved, and the cooperation completion degree of the user for investigation and the outbound investigation efficiency are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a method, device, equipment, storage medium and program product for screening outbound call survey users. Background Art

[0002] For service providers, proactively contacting users to evaluate service quality and understand users' expectations and evaluations of business services is an important way to improve business standards and service quality. At present, screening rules are generally constructed from dimensions such as user levels, and users participating in the survey are determined through simple rule screening, or users participating in the survey are randomly selected from all users. After determining the users participating in the survey, they are contacted through voice outbound calls and other methods, and user evaluations are obtained in the form of questions and answers, and user evaluations are recorded. Judging from the survey results, the proportion of users cooperating to complete the survey evaluation is very low, and the overall survey efficiency is low. Summary of the invention

[0003] The embodiments of the present application provide a method, device, equipment, storage medium and program product for screening outbound survey users, which are used to solve the defects of the existing survey user screening method, which leads to low cooperation completion of the screened users for the survey and low overall survey efficiency.

[0004] The present application embodiment provides a method for screening outbound call survey users, including: Acquire a feature data set of the user to be screened; the feature data set includes classification features, call statistics features and voice outbound call features; Inputting the feature data set into a user identification model, and using the user identification model to identify target users for outbound call survey from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

[0005] In one embodiment, obtaining a feature data set of a user to be screened includes: Obtaining a basic data set of users to be screened; the basic data set includes economic data, call statistics data, and voice outbound call data; Based on the correlation between each feature data and user accessibility, feature extraction is performed on the basic data set to obtain the feature data set of the user to be screened; the correlation is obtained by performing a correlation analysis on each feature data and user accessibility.

[0006] In one embodiment, before inputting the feature data set into a user identification model and using the user identification model to identify target users for outbound call survey from the users to be screened, the method further includes: Obtaining a historical basic data set of all users; the historical basic data set includes historical economic data, historical call statistics data, and historical voice outbound call data; Constructing a basic feature vector for each user in the total number of users according to the historical basic data set; Based on the basic feature vector, the Spearman coefficient method is used to perform correlation analysis on each basic feature data corresponding to the historical basic data set to obtain the correlation between the first feature data and each second feature data; the first feature data is a user easy-to-visit feature in each basic feature data, and the second feature data is any item in each basic feature data except the first feature data; the user easy-to-visit feature is determined based on known easy-to-visit users; Selecting historical feature data from each of the basic feature data according to the correlation to construct a sample data set; Performing cluster analysis and random sampling on the sample data set to obtain a target sample; The preset TabNet model is iteratively trained using the target samples.

[0007] In one embodiment, the iterative training of the preset TabNet model using the target sample includes: Dividing the target sample into a training set, a test set and a validation set; Based on a preset first number of iterations, the preset TabNet model is iteratively trained using the training set and the test set to obtain model parameters and evaluation parameters in each round of iteration; the evaluation parameters include precision and recall; The model parameters with the best evaluation parameters are verified using the verification set, the model parameters are updated according to the verification results, and the step of iteratively training the preset TabNet model based on the preset first number of iterations and using the training set and the test set is returned and executed until a preset iteration termination condition is met; the iteration termination condition includes that the number of iterations reaches a second number of iterations or the model parameters meet a preset accuracy; the second number of iterations is greater than the first number of iterations.

[0008] In one embodiment, performing cluster analysis and random sampling on the sample data set to obtain a target sample includes: Based on the user label of each user in the total number of users, the sample data set is divided into a positive sample set and a negative sample set; wherein the user label is used to characterize the completion degree of the user's outbound call survey; Performing cluster analysis on the negative sample set using a preset cluster analysis algorithm, and clustering the negative sample set into negative sample subsets of multiple categories; Based on a preset sampling ratio, each of the negative sample subsets is randomly sampled to downsample the negative sample set to obtain a target sample.

[0009] In one embodiment, constructing a basic feature vector of each user in the total number of users according to the historical basic data set includes: Dividing the historical basic data set according to a preset time interval to obtain multiple historical basic data subsets; Counting characteristic values ​​of each of the basic characteristic data based on the historical basic data subset; A basic feature vector of each user in the total number of users is constructed according to the feature value.

[0010] The present application also provides an outbound call survey user screening device, including: A feature extraction module is used to obtain a feature data set of the user to be screened; the feature data set includes classification features, call statistics features and voice outbound call features; A user screening module, used to input the feature data set into a user identification model, and use the user identification model to identify target users for outbound call surveys from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

[0011] An embodiment of the present 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 when the processor executes the computer program, the steps of the outbound call survey user screening method as described in any one of the above are implemented.

[0012] The embodiment of the present application also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the outbound call survey user screening method as described in any one of the above are implemented.

[0013] The embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of any of the above-mentioned methods for screening users for outbound surveys.

[0014] The outbound survey user screening method, device, equipment, storage medium and program product provided in the embodiments of the present application obtains a feature data set of users to be screened and inputs it into a user identification model, and uses the user identification model to identify target users for outbound surveys from users to be screened, and the user identification model is obtained by iteratively training a preset TabNet model based on a target sample, and the target sample is obtained by clustering analysis and random sampling of a sample data set constructed from historical feature data of all users. By identifying target users for outbound surveys through the TabNet model, the model's reliance on feature engineering is reduced, and the impact of a small number of irrelevant features on model performance is avoided. In addition, in the model training stage, the sample data set is clustered and sampled at any time to solve the problem of sample imbalance, improve the accuracy of identifying outbound survey users, and help improve the user's cooperation with the survey and the efficiency of outbound surveys. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the present application or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 It is a flowchart of the method for screening outbound survey users provided in an embodiment of the present application; Figure 2 It is a structural schematic diagram of an outbound call survey user screening device provided in an embodiment of the present application; Figure 3 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0018] The embodiment of the present application provides a method for screening users for an outbound survey. A user identification model is constructed based on a TabNet model. The constructed user identification model is used to identify target users for an outbound survey. Users with a high probability of cooperating in completing the outbound survey are screened out, thereby increasing the proportion of users who cooperate in completing the survey, thereby improving the overall efficiency of the outbound survey.

[0019] Specifically, Figure 1 is a flow chart of the outbound call survey user screening method provided in the embodiment of the present application, such as Figure 1 As shown, the outbound call survey user screening method provided in the embodiment of the present application includes the following steps: Step 100, obtaining a feature data set of the user to be screened; the feature data set includes classification features, call statistics features and voice outbound call features; Step 200, inputting the feature data set into a user identification model, and using the user identification model to identify target users for outbound call survey from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

[0020] First, a feature data set of the users to be screened is obtained, which includes the classified features, call statistics features and voice outbound call features of each user to be screened. The obtained feature data set is input into a user identification model, and the user identification model is used to identify target users for outbound call surveys from the users to be screened.

[0021] Optionally, the users to be screened include multiple users, and the users to be screened may be all users or a part of all users. For the feature data set, each user to be screened may correspond to one feature data set, or all users to be screened may correspond to one feature data set, and the feature data set includes feature data of each user to be screened, and this is not specifically limited.

[0022] Optionally, the user identification model is constructed based on the TabNet model and pre-trained. Specifically, the preset basic TabNet model is iteratively trained based on the target sample to obtain the pre-trained user identification model. The target sample is obtained by clustering analysis and random sampling of the constructed sample data set, and the sample data set is constructed based on the historical feature data of all users.

[0023] That is, in this embodiment, the screening of outbound survey users includes model training and model application stages. In the model training stage, a user identification model is constructed based on the TabNet model, a sample data set is constructed based on the historical feature data of all users, and cluster analysis and random sampling are performed on the sample data set to obtain a target sample. Finally, the preset basic TabNet model is iteratively trained using the target sample to obtain a pre-trained user identification model. On the one hand, in the historical data of outbound survey users, since the proportion of users who cooperate to complete the survey is relatively low, the problem of sample imbalance can be solved by clustering analysis and random sampling of the sample data set. On the other hand, by constructing a user identification model based on the TabNet model, the tabular data can be encoded, thereby obtaining a method that can characterize the tabular data. This representation learning can effectively reduce the reliance on feature engineering. In the process of identifying outbound survey users, a small number of features with weak correlation with user accessibility do not affect the prediction accuracy of the model.

[0024] In the model application stage, for the users to be screened, the feature data of each user is obtained to form a feature data set of the users to be screened, and the feature data set is input into a pre-trained user recognition model. The feature data of each user is predicted through the user recognition model, so as to identify whether each user is a target user for outbound call survey.

[0025] In this embodiment, by obtaining a feature data set of users to be screened and inputting it into a user identification model, the user identification model is used to identify target users for outbound surveys from the users to be screened, and the user identification model is obtained by iteratively training a preset TabNet model based on a target sample, and the target sample is obtained by clustering analysis and random sampling of a sample data set constructed from historical feature data of all users. By identifying target users for outbound surveys through the TabNet model, the model's reliance on feature engineering is reduced, and the impact of a small number of irrelevant features on model performance is avoided. In addition, during the model training stage, the sample data set is clustered and sampled at any time to solve the problem of sample imbalance, improve the accuracy of identifying outbound survey users, and help improve the user's cooperation with the survey and the efficiency of outbound surveys.

[0026] In one embodiment, the feature data set of the to-be-screened user is obtained by extracting features from the basic data of the to-be-screened user. In step 100, obtaining the feature data set of the to-be-screened user includes: Step 110, obtaining a basic data set of the users to be screened; the basic data set includes economic data, call statistics data and voice outbound call data; Step 120, based on the correlation between each feature data and user accessibility, feature extraction is performed on the basic data set to obtain the feature data set of the to-be-screened user; the correlation is obtained by performing correlation analysis on each feature data and user accessibility.

[0027] When obtaining the feature data set of the user to be screened, first obtain the basic data set of the user to be screened, the basic data set includes the economic data, call statistics data and voice outbound call data of the user to be screened, and based on the correlation between each feature data and the user accessibility, perform feature extraction on the basic data set to obtain the feature data set of the user to be screened, the features in the feature data set are features with a high correlation with the user accessibility, wherein the correlation between the feature data and the user accessibility is obtained by performing a correlation analysis between each feature data and the user accessibility.

[0028] Furthermore, feature extraction is performed on the basic data set, including feature extraction on the economic segmentation data to obtain the economic segmentation features of the users to be screened, feature extraction on the call statistics data to obtain the call statistics features of the users to be screened, and feature extraction on the outbound voice call data to obtain the outbound voice call features of the users to be screened.

[0029] Optionally, the feature data extracted from the basic data set is the features with a greater correlation with the user's accessibility that are screened out based on the correlation analysis of the sample data during the model training phase, including classification features, call statistics features, and voice outbound call features. In one embodiment, the feature data set of the user to be screened includes feature items extracted from the basic data set, and the variation of each feature item, which includes but is not limited to the mean and variance. Specifically, feature extraction is performed on the basic data set, and feature items related to the user's accessibility are extracted therefrom, and each feature item is divided and counted according to a preset time granularity to obtain the statistical value of each feature item at the time granularity, and the variation of any feature item is calculated based on the statistical value of the feature item, thereby obtaining a feature data set containing each feature item and its variation.

[0030] In one embodiment, the data in the basic data set includes the total number of calls (the user is the caller or the called party), the total duration of calls, the number of calling calls, the duration of calling calls, the number of called calls, the duration of called calls, the number of called calls when the caller is a fixed-line phone, the duration of called calls when the caller is a fixed-line phone, the number of called calls when the caller is the same fixed-line phone, the duration of called calls when the caller is the same fixed-line phone, the number of called calls when the caller is a fixed-line phone + the number of calls is a preset number, the number of called calls when the caller is a fixed-line phone + the number of calls is a preset number, the duration of called calls when the caller is a fixed-line phone + the number of calls is a preset number, the caller is a fixed-line phone + the number of calls is a preset number + the duration of the call is greater than the preset duration The number of called calls, the number of called calls with fixed-line + preset number of calls + 0 user callback number, the number of called calls with fixed-line + preset number of calls + 0 user callback number + call duration greater than preset duration, the ratio of called calls with fixed-line as the caller, the ratio of called calls with unfamiliar fixed-line as the caller, the average call duration with fixed-line, the average call duration with unfamiliar fixed-line, the ratio of calling calls, the ratio of called calls, the ratio of unfamiliar calls, the number of days of calls, the duration of called calls at different time granularities, the duration of calls at different time granularities, the average ARPU (Average Revenue PerUser, average revenue per user), voice call duration, number of voice calls, local call duration, number of local calls, non-roaming duration, non-roaming number, non-roaming outgoing call duration, number of non-roaming outgoing calls, non-roaming incoming call duration, number of non-roaming incoming calls, non-roaming incoming call local duration, number of non-roaming incoming calls to local, number of non-roaming operator incoming calls, non-roaming operator incoming call duration, billing duration and calling billing duration, etc.

[0031] In some embodiments, the basic data set obtained includes the economic data and call statistics of n-1, n-2...nm months of the users to be screened and the voice outbound call data of the nth month. The feature data set obtained by feature extraction based on the basic data set includes 33-dimensional call features, 48-dimensional economic data features and 48-dimensional economic data change features, totaling 129-dimensional data.

[0032] Before inputting the feature data set of the to-be-screened users into the user identification model and using the user identification model to identify the target users for outbound call survey from the to-be-screened users, model training needs to be performed first. Therefore, before step 200, the following may also be included: Step 101, obtaining a historical basic data set of all users; the historical basic data set includes historical economic data, historical call statistics data and historical voice outbound call data; Step 102, constructing a basic feature vector of each user in the total number of users according to the historical basic data set; Step 103: Based on the basic feature vector, the Spearman coefficient method is used to perform correlation analysis on each basic feature data corresponding to the historical basic data set to obtain the correlation between the first feature data and each second feature data; the first feature data is a user easy-to-visit feature in each basic feature data, and the second feature data is any item in each basic feature data except the first feature data; the user easy-to-visit feature is determined based on known easy-to-visit users; Step 104, selecting historical feature data from each of the basic feature data according to the correlation to construct a sample data set; Step 105, performing cluster analysis and random sampling on the sample data set to obtain a target sample; Step 106: Iteratively train the preset TabNet model using the target sample.

[0033] In the model training stage, first obtain the historical basic data set of all users, which includes the historical economic data, historical call statistics data and historical voice outbound call data of all users. According to the obtained historical basic data set, the basic feature vector of each user in the total number of users is constructed. Based on the basic feature vector, the Spearman coefficient method is used to perform correlation analysis on the basic feature data corresponding to the historical basic data set to obtain the correlation between the first feature data and the second feature data. Among them, the first feature data is the user's easy-to-visit feature in each basic feature data corresponding to the historical basic data set, and the second feature data is any item in each basic feature data except the first feature data. Further, the user's easy-to-visit feature is determined based on known easy-to-visit users. In one embodiment, the user's easy-to-visit feature can be determined based on the user label of the easy-to-visit user. The user label is used to characterize the user's outbound call survey completion, and the user's easy-to-visit feature includes the outbound call survey completion. By performing correlation analysis on other feature data and the user's easy-to-visit feature, features with a greater correlation with whether the user is easy to access can be screened out, and the screened features are used to construct a sample data set for model training.

[0034] Furthermore, historical feature data are selected from each basic feature data according to the correlation to construct a sample data set, and cluster analysis and random sampling are performed on the sample data set to obtain a target sample. Finally, the preset TabNet model is iteratively trained using the target sample.

[0035] Optionally, the feature dimensions of the segmentation data and call data are high, and most of the features have low correlation. If the full amount of basic feature data is used for model training, the model will have the following problems: Low prediction performance: It is difficult for the model to extract useful information from a large number of irrelevant features, which will reduce the model's prediction accuracy and reliability. Overfitting: If the feature vector contains many irrelevant features, the model will learn noise from the irrelevant features instead of effective information that is truly relevant to the prediction target, which will cause the model to overfit the training data. Difficult to explain the model: Because there are a large number of features that are irrelevant to the prediction target, it will make it more complicated to understand the internal decisions of the model, resulting in a waste of computing resources and computing time.

[0036] Based on this, during the model training phase, the Spearman coefficient method is used to select features that are highly correlated with the prediction target. The Spearman coefficient correlation analysis is a non-parametric statistical method used to evaluate the degree of association between two variables. It is particularly suitable for processing nonlinear relationships or data that do not satisfy the normal distribution. In the Spearman correlation analysis, the feature data is first used as a variable and sorted according to the value of the variable. Then, the Spearman rank correlation coefficient between each feature data and the user-accessible feature is calculated in sequence according to the sorting order. It is usually expressed as .

[0037] ; (1)

[0038] In formula 1, For the i The eigenvalues ​​of the features x of samples, is the total sample size, that is, the number of all users, is the average value of feature x of n samples; It is i The characteristic value of the user's accessible feature of samples, is the average value of the user-accessible features of n samples, is the Spearman rank correlation coefficient, ranging from [-1,1]. Indicates perfect positive correlation, that is, when the value of one variable increases, the value of the other variable always increases; Indicates perfect negative correlation, that is, when the value of one variable increases, the value of the other variable always decreases; It means there is no correlation, that is, there is no linear relationship between the values ​​of the two variables.

[0039] The Spearman correlation has relatively loose requirements on data types. The main advantage of this method is that it does not need to assume that the data follows a specific distribution. Therefore, it is more robust and applicable to various types of data sets. It is usually used to explore the monotonic relationship between variables, and is not limited to the detection of linear relationships.

[0040] Furthermore, when constructing the basic feature vector of each user in the total number of users based on the historical basic data set, it is based on extracting the feature value of the historical basic data. Based on this, step 102 specifically includes: Step 112, dividing the historical basic data set according to a preset time interval to obtain a plurality of historical basic data subsets; Step 122, counting the characteristic values ​​of each of the basic characteristic data based on the historical basic data subset; Step 132: construct a basic feature vector for each user in the total number of users according to the feature value.

[0041] The historical basic data set is divided into multiple historical basic data subsets according to the preset time intervals. The characteristic values ​​of each basic characteristic data are counted based on the historical basic data subsets, and the basic characteristic vectors of each user in the total number of users are constructed based on the characteristic values.

[0042] Among them, the characteristic value of each basic characteristic data of any user is used as the element value of the vector to construct the basic characteristic vector of the user. In this way, the basic characteristic vector of each user in the total number of users is constructed. According to the constructed basic characteristic vectors of each user, the Spearman coefficient is used to perform the correlation analysis between the characteristic items and the user accessibility of each basic characteristic data, and the correlation between any basic characteristic data and the user accessibility is calculated. In this way, the correlation coefficient between each basic characteristic data and the user accessibility is calculated in turn, and then according to the correlation coefficient, the characteristic data with greater correlation with the prediction target is screened to construct a sample data set.

[0043] Optionally, correlation analysis is performed based on the constructed basic feature vectors, and feature items related to user accessibility are screened out from the basic feature data to construct a sample data set, which also includes changes such as the mean and variance of each feature item screened out from the basic feature data.

[0044] Taking the menstrual data at the monthly granularity as an example, when constructing the basic feature vector, based on the historical basic data set containing n months of menstrual data, the variance of the menstrual data of n-1 months, n-2 months, ...nm months is calculated as the change in the menstrual data, which reflects the stability of the user. The variance is calculated as shown in the following formula 2: ; (2)

[0045] in, is the overall variance, X is the variable value of the menstrual data, based on the above-obtained 48-dimensional features, X is the characteristic value of the menstrual characteristics from n-1 to nm months, is the overall mean of the classified data, that is, the mean of each characteristic item of the classified data from n-1 to nm months, N is the total number of cases, which is m months in this embodiment.

[0046] In this way, the changes in the feature values ​​of each feature item selected from the basic feature data are calculated in turn, and the calculated changes are used as element values ​​to construct a sample feature vector of the user, and a sample data set is constructed based on the sample feature vector.

[0047] Furthermore, for the constructed sample data set, the sample imbalance problem is solved by cluster analysis and random sampling. Step 105 includes: Step 115, based on the user label of each user in the total number of users, the sample data set is divided into a positive sample set and a negative sample set; wherein the user label is used to characterize the completion degree of the user's outbound call survey; Step 125, performing cluster analysis on the negative sample set using a preset cluster analysis algorithm, and clustering the negative sample set into negative sample subsets of multiple categories; Step 135 , based on a preset sampling ratio, randomly sample each of the negative sample subsets to downsample the negative sample set to obtain a target sample.

[0048] Based on the user labels of all users, the sample data set is divided into a positive sample set and a negative sample set, where the user label is used to characterize the user's outbound survey completion. A preset clustering analysis algorithm is used to perform clustering analysis on the negative sample set, and the negative sample set is clustered into multiple categories of negative sample subsets. Based on the preset sampling ratio, each negative sample subset is randomly sampled, so that the negative samples are adopted to obtain the target samples.

[0049] In one embodiment, the degree of completion of the outbound survey by users can be summarized into four categories, namely, not connected, refused to answer after connected, refused midway after the survey started, and completed all the survey content. In conventional analysis, users who have completed all the survey content are generally regarded as positive samples, and users with other survey results are uniformly regarded as negative samples. However, based on the analysis of actual survey results, it can be seen that the proportion of customers who refused midway after the survey started and those who completed all the survey content when they participated in the survey again later is similar. Therefore, in this embodiment, these two types of users are defined as positive samples, and users with other survey results are defined as negative samples, which increases the number of effective positive samples while making the distinction between positive and negative samples higher.

[0050] Due to the imbalanced distribution ratio of target users and non-target users in user outbound surveys, if the sample data set is directly used for model training, it will lead to problems such as data skew and overfitting. For data skew, when the distribution of positive and negative samples is uneven, both machine learning and deep learning models may tend to favor categories with more samples, because this will allow the model to obtain lower losses on the training set, resulting in poor performance of the model on categories with fewer samples due to lack of sufficient data to learn the characteristics of this category. For the overfitting problem, the imbalanced data distribution may cause the model to over-focus on categories with a larger number and overfit on categories with a smaller number, resulting in the model being unable to generalize to the target users of the test set, and thus performing poorly on the test set.

[0051] Based on this, the embodiment of the present application uses a clustering analysis algorithm such as K-means to undersample negative samples so that the ratio of positive and negative samples is balanced. Specifically, based on a clustering analysis algorithm such as K-means, the distribution law of samples is obtained through clustering analysis, and then a certain number of target samples that can reflect the common characteristics of negative samples are randomly selected from each type of samples according to the distribution law of samples.

[0052] The K-means clustering analysis algorithm is used to analyze the Euclidean distance between data objects and cluster centers in space. The specific calculation method is: ; (3)

[0053] Where x is the characteristic of the surveyed user, For the i Cluster centers, m To investigate the dimensions of user characteristics, , They are x and No. j The attribute value of a feature item.

[0054] The formula for calculating the sum of squared error (SSE) of the entire sample data set is: ; (4)

[0055] Among them, SSE is a key indicator for evaluating the prediction accuracy and clustering effect of the model. By minimizing SSE, the model can be optimized to improve the prediction accuracy and clustering quality. The size of SSE indicates the quality of the clustering result, and k is the number of clusters.

[0056] Calculate the average value of the data objects in each cluster as the new cluster center, and perform clustering iterations in this way until the cluster center no longer changes or the maximum number of iterations is reached.

[0057] According to the clustering results obtained by the K-means clustering analysis algorithm, the similarity of intra-class features is high, and the similarity of inter-class features is low. For example, all users in the sample data set are clustered into 50 categories, and a certain number of samples are randomly selected from each category in proportion as the training data set, i.e., the target samples. Compared with random sampling, clustering based on the clustering analysis algorithm and random extraction of negative samples help ensure that the selected negative samples are diverse, thereby better representing the overall distribution and common laws of non-target users.

[0058] Optionally, in one embodiment, during the iterative model training process, model evaluation parameters such as loss value, precision rate and recall rate are recorded at the end of each round of training, and finally the model parameters of the round with the highest precision rate are verified. Step 106, iteratively training the preset TabNet model using the target sample, specifically includes: Step 116, dividing the target sample into a training set, a test set and a validation set; Step 126, based on a preset first number of iterations, iteratively train the preset TabNet model using the training set and the test set to obtain model parameters and evaluation parameters in each round of iteration; the evaluation parameters include precision and recall; Step 136, using the verification set to verify the model parameters with the best evaluation parameters, updating the model parameters according to the verification results, returning to and executing the step of iteratively training the preset TabNet model based on the preset first number of iterations using the training set and the test set until the preset iteration termination condition is met; the iteration termination condition includes that the number of iterations reaches the second number of iterations or the model parameters meet the preset accuracy; the second number of iterations is greater than the first number of iterations.

[0059] In the model training stage, the target samples are first divided into a training set, a test set, and a validation set. For example, for the training set and the test set, the division method in this embodiment is to split them by month. Specifically, the historical feature data of the target users and non-target users in a certain month (month n) are used as the training set and the test set, and the data ratio is 8:2. The historical feature data of the subsequent months (such as n+1, n+2 months...) are used as the validation set for model training.

[0060] Based on the preset first iteration number, the preset TabNet model is iteratively trained using the training set and the test set, and the model parameters and evaluation parameters in each iteration process are recorded, where the evaluation parameters include precision and recall.

[0061] After completing the iterative training for the first number of iterations, the model parameters with the best rating parameters are verified using the validation set, and the model parameters are updated according to the verification results. Then, the process returns to and executes the step of training the TabNet model based on the preset first number of iterations using the training set and the test set. The model parameters updated based on the validation set are used as the initial model parameters, and the iterative training for the first number of iterations is performed again until the preset iteration termination conditions are met.

[0062] Optionally, the iteration termination condition includes but is not limited to the number of iterations reaching the second number of iterations, or the accuracy of the model parameters meets the preset accuracy requirements. In one embodiment, the second number of iterations is greater than the first number of iterations, and optionally, the second number of iterations is an integer multiple of the first number of iterations.

[0063] In some embodiments, the loss function used for iterative training of the TabNet model is a cross entropy loss function, as shown in the following formula 5: ; (5)

[0064] In formula 5, is the predicted value of the model (target users are 1, non-target users are 0), is the user label of the user (target user is 1, non-target user is 0), is the function value of the cross entropy loss function.

[0065] In this embodiment, negative samples are downsampled through cluster analysis and random sampling, and a certain amount of samples are randomly selected from each category of negative samples, so that the sample data used for model training can reflect the common characteristics of negative samples, which not only ensures that negative samples have diversity, but also better represents the overall distribution and common laws of users who failed outbound call surveys. At the same time, the balance of positive and negative samples can be achieved, which can prevent the overfitting phenomenon of the trained user recognition model and ensure the prediction performance of the model.

[0066] Furthermore, when constructing a sample data set for model training, the Spearman coefficient is used to select the feature data with the greatest correlation with the prediction target from all feature data to construct the sample data set, which can effectively reduce the negative impact of data noise on model performance.

[0067] The following is a description of an outbound call survey user screening device provided in an embodiment of the present application. The outbound call survey user screening device described below and the outbound call survey user screening method described above can be referenced to each other.

[0068] Reference Figure 2 The outbound call survey user screening device provided in the embodiment of the present application includes: The feature extraction module 10 is used to obtain a feature data set of the user to be screened; the feature data set includes a classification feature, a call statistics feature and a voice outbound call feature; A user screening module 20, for inputting the feature data set into a user identification model, and using the user identification model to identify target users for outbound call surveys from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

[0069] In one embodiment, the feature extraction module 10 is further used for: Obtaining a basic data set of users to be screened; the basic data set includes economic data, call statistics data, and voice outbound call data; Based on the correlation between each feature data, feature extraction is performed on the basic data set to obtain the feature data set of the to-be-screened user; the correlation is obtained by performing correlation analysis on each feature data.

[0070] In one embodiment, the outbound survey user screening device further includes a model pre-training module for: Obtaining a historical basic data set of all users; the historical basic data set includes historical economic data, historical call statistics data, and historical voice outbound call data; Constructing a basic feature vector for each user in the total number of users according to the historical basic data set; Based on the basic feature vector, the Spearman coefficient method is used to perform correlation analysis between the basic feature data corresponding to the historical basic data set to obtain the correlation between the first feature data and the second feature data; the first feature data is the user easy-to-visit feature in each of the basic feature data, and the second feature data is any item in each of the basic feature data except the first feature data; the user easy-to-visit feature is determined based on known easy-to-visit users; Selecting historical feature data from each of the basic feature data according to the correlation to construct a sample data set; Performing cluster analysis and random sampling on the sample data set to obtain a target sample; The preset TabNet model is iteratively trained using the target samples.

[0071] In one embodiment, the model pre-training module is further used to: Dividing the target sample into a training set, a test set and a validation set; Based on a preset first number of iterations, the preset TabNet model is iteratively trained using the training set and the test set to obtain model parameters and evaluation parameters in each round of iteration; the evaluation parameters include precision and recall; The model parameters with the best evaluation parameters are verified using the verification set, the model parameters are updated according to the verification results, and the step of iteratively training the preset TabNet model based on the preset first number of iterations and using the training set and the test set is returned and executed until a preset iteration termination condition is met; the iteration termination condition includes that the number of iterations reaches a second number of iterations or the model parameters meet a preset accuracy; the second number of iterations is greater than the first number of iterations.

[0072] In one embodiment, the model pre-training module is further used to: Based on the user label of each user in the total number of users, the sample data set is divided into a positive sample set and a negative sample set; wherein the user label is used to characterize the completion degree of the user's outbound call survey; Performing cluster analysis on the negative sample set using a preset cluster analysis algorithm, and clustering the negative sample set into negative sample subsets of multiple categories; Based on a preset sampling ratio, each of the negative sample subsets is randomly sampled to downsample the negative sample set to obtain a target sample.

[0073] In one embodiment, the model pre-training module is further used to: Dividing the historical basic data set according to a preset time interval to obtain multiple historical basic data subsets; Counting characteristic values ​​of each of the basic characteristic data based on the historical basic data subset; A basic feature vector of each user in the total number of users is constructed according to the feature value.

[0074] Figure 3 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 3 As shown, the electronic device may include: a processor 310, a communications interface 320, a memory 330 and a communication bus 340, wherein the processor 310, the communications interface 320 and the memory 330 communicate with each other via the communication bus 340. The processor 310 may call the logic instructions in the memory 330 to execute the steps of the outbound call survey user screening method, which includes: Acquire a feature data set of the user to be screened; the feature data set includes classification features, call statistics features and voice outbound call features; Inputting the feature data set into a user identification model, and using the user identification model to identify target users for outbound call survey from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

[0075] In addition, the logic instructions in the above-mentioned memory 330 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0076] On the other hand, the present application further provides a computer program product, the computer program product including a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can execute the steps of the outbound call survey user screening method provided in the above embodiments, the method including: Acquire a feature data set of the user to be screened; the feature data set includes classification features, call statistics features and voice outbound call features; Inputting the feature data set into a user identification model, and using the user identification model to identify target users for outbound call survey from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

[0077] In another aspect, the present application further provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the outbound call survey user screening method provided in the above embodiments are implemented. The method includes: Acquire a feature data set of the user to be screened; the feature data set includes classification features, call statistics features and voice outbound call features; Inputting the feature data set into a user identification model, and using the user identification model to identify target users for outbound call survey from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

[0078] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0079] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for screening outbound survey users, characterized in that: include: Obtain the feature data set of the user to be screened; The feature data set includes classification features, call statistics features and voice outbound call features; Inputting the feature data set into a user identification model, and using the user identification model to identify target users for outbound call survey from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

2. The outbound survey user screening method according to claim 1 is characterized in that: The step of obtaining a feature data set of the user to be screened includes: Obtaining a basic data set of users to be screened; the basic data set includes economic data, call statistics data, and voice outbound call data; Based on the correlation between each feature data and user accessibility, feature extraction is performed on the basic data set to obtain the feature data set of the user to be screened; the correlation is obtained by performing a correlation analysis on each feature data and user accessibility.

3. The outbound survey user screening method according to claim 1 is characterized in that: Before inputting the feature data set into the user identification model and using the user identification model to identify target users for outbound call research from the users to be screened, the method further includes: Obtaining a historical basic data set of all users; the historical basic data set includes historical economic data, historical call statistics data, and historical voice outbound call data; Constructing a basic feature vector for each user in the total number of users according to the historical basic data set; Based on the basic feature vector, the Spearman coefficient method is used to perform correlation analysis on each basic feature data corresponding to the historical basic data set to obtain the correlation between the first feature data and each second feature data; the first feature data is a user easy-to-visit feature in each basic feature data, and the second feature data is any item in each basic feature data except the first feature data; the user easy-to-visit feature is determined based on known easy-to-visit users; Selecting historical feature data from each of the basic feature data according to the correlation to construct a sample data set; Performing cluster analysis and random sampling on the sample data set to obtain a target sample; The preset TabNet model is iteratively trained using the target samples.

4. The outbound survey user screening method according to claim 3 is characterized in that: The iterative training of the preset TabNet model using the target sample includes: Dividing the target sample into a training set, a test set and a validation set; Based on a preset first number of iterations, the preset TabNet model is iteratively trained using the training set and the test set to obtain model parameters and evaluation parameters in each round of iteration; the evaluation parameters include precision and recall; The model parameters with the best evaluation parameters are verified using the verification set, the model parameters are updated according to the verification results, and the step of iteratively training the preset TabNet model based on the preset first number of iterations and using the training set and the test set is returned and executed until a preset iteration termination condition is met; the iteration termination condition includes that the number of iterations reaches a second number of iterations or the model parameters meet a preset accuracy; the second number of iterations is greater than the first number of iterations.

5. The outbound survey user screening method according to claim 3 is characterized in that: The performing cluster analysis and random sampling on the sample data set to obtain a target sample includes: Based on the user label of each user in the total number of users, the sample data set is divided into a positive sample set and a negative sample set; wherein the user label is used to characterize the completion degree of the user's outbound call survey; Performing cluster analysis on the negative sample set using a preset cluster analysis algorithm, and clustering the negative sample set into negative sample subsets of multiple categories; Based on a preset sampling ratio, each of the negative sample subsets is randomly sampled to downsample the negative sample set to obtain a target sample.

6. The outbound survey user screening method according to claim 3 is characterized in that: The constructing a basic feature vector of each user in the total number of users according to the historical basic data set includes: Dividing the historical basic data set according to a preset time interval to obtain multiple historical basic data subsets; Counting characteristic values ​​of each of the basic characteristic data based on the historical basic data subset; A basic feature vector of each user in the total number of users is constructed according to the feature value.

7. An outbound survey user screening device, characterized in that: include: A feature extraction module is used to obtain a feature data set of the user to be screened; The feature data set includes classification features, call statistics features and voice outbound call features; A user screening module, used to input the feature data set into a user identification model, and use the user identification model to identify target users for outbound call surveys from the users to be screened; Among them, the user recognition model is obtained by iteratively training the preset TabNet model based on the target sample; the target sample is obtained by clustering analysis and random sampling of the sample data set; the sample data set is constructed based on the historical feature data of all users.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the outbound survey user screening method according to any one of claims 1 to 6 are implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the outbound call survey user screening method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the outbound call survey user screening method according to any one of claims 1 to 6 are implemented.