Round selection model training method, round selection method and electronic equipment

Abstract

The invention provides a circling model training method, a circling method and electronic equipment, and relates to the field of artificial intelligence. The method comprises the steps that the electronic equipment can classify and process acquired training data to obtain intervention group data corresponding to a first sales mode and control group data corresponding to a second sales mode; the electronic device can input the intervention group data into the first basic model, and a first prediction model corresponding to the first sales mode is obtained through training. The electronic device can input the control group data into the first basic model, and a second prediction model corresponding to the second sales mode is obtained through training. And the electronic equipment can input all training data into the classification basic model for training to obtain a third prediction model. The electronic equipment can construct a first weight coefficient and a second weight coefficient by using a third prediction model, and performs weighted fusion on the first prediction model and the second prediction model to obtain a circling model. According to the method provided by the invention, the efficiency of customer group selection accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a circle selection model training method, a circle selection method and an electronic device. Background Art

[0002] In the field of telemarketing, recommending the right products to the right customer groups at the right time and with the right means can maximize customer conversion rates. Accurately selecting the right customers and thus forming the right customer groups can effectively improve the utilization of customer resources.

[0003] In traditional telemarketing scenarios, different customer groups can be screened by setting different business rules, and different telemarketing methods can be used for different customer groups. The screening method based on business rules does not fully utilize the advantages of big data. For this reason, the existing technology introduces the use of machine learning response models to predict conversion rates.

[0004] The machine learning model uses whether the customer is successfully converted as the conversion result, and is trained using the customer features obtained from the customer data. The machine learning model can output the predicted probability of the customer's successful conversion. It is generally believed that the higher the predicted probability, the greater the possibility of the customer's successful conversion. Furthermore, based on the predicted probability, customers can be grouped into different customer groups, and different telemarketing methods can be used for different customer groups.

[0005] However, although this machine learning model uses the advantages of big data, it does not introduce the causal relationship between the telemarketing method and the customer's conversion results into the prediction of the machine learning model, so there is a problem of low selection accuracy. Summary of the invention

[0006] The present application provides a circle selection model training method, a circle selection method and an electronic device to solve the problem of low circle selection accuracy in the prior art.

[0007] In a first aspect, the present application provides a circle selection model training method, comprising:

[0008] Acquire training data, and process the training data to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method;

[0009] Input the intervention group data into the first basic model, and train to obtain a first prediction model corresponding to the first sales method; input the control group data into the first basic model, and train to obtain a second prediction model corresponding to the second sales method;

[0010] Inputting all the training data into the classification basic model to obtain a third prediction model for predicting sales methods through training;

[0011] A circle selection model is constructed based on the first prediction model, the second prediction model, and the weight coefficients generated by the third prediction model.

[0012] In a second aspect, the present application provides a circle selection method, comprising:

[0013] Acquire a customer data set, wherein the customer data set includes customer characteristic data of at least one customer to be selected;

[0014] Inputting the customer feature data in the customer data set into the selection model in the first aspect and any possible design of the first aspect, and calculating the comprehensive impact value of each of the to-be-selected customers;

[0015] According to the comprehensive influence value of each of the customers to be selected, the customers to be selected are selected into different customer groups.

[0016] In a third aspect, the present application provides a circle selection model training device, comprising:

[0017] An acquisition module, used for acquiring training data and processing the training data to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method;

[0018] A training module is used to input the intervention group data into the first basic model to train a first prediction model corresponding to the first sales method; input the control group data into the first basic model to train a second prediction model corresponding to the second sales method; input all the training data into the classification basic model to train a third prediction model for predicting the sales method; and construct a circled model based on the first prediction model, the second prediction model, and the weight coefficient generated by the third prediction model.

[0019] In a fourth aspect, the present application provides a circle selection device, comprising:

[0020] An acquisition module, used for acquiring a customer data set, wherein the customer data set includes customer characteristic data of at least one customer to be selected;

[0021] A selection module is used to input the customer feature data in the customer data set into the selection model in the first aspect of the claim and any possible design of the first aspect, and calculate the comprehensive influence value of each of the customers to be selected; according to the comprehensive influence value of each of the customers to be selected, the customers to be selected are selected into different customer groups.

[0022] In a fifth aspect, the present application provides an electronic device, including: a memory and a processor;

[0023] The memory is used to store computer programs; the processor is used to execute the first aspect and the selection model training method in any possible design of the first aspect, or the second aspect and the selection method in any possible design of the second aspect according to the computer program stored in the memory.

[0024] In a sixth aspect, the present application provides a computer-readable storage medium, which stores a computer program. When at least one processor of an electronic device executes the computer program, the electronic device executes the first aspect and the selection model training method in any possible design of the first aspect, or the second aspect and the selection method in any possible design of the second aspect.

[0025] In a seventh aspect, the present application provides a computer program product, comprising a computer program. When at least one processor of an electronic device executes the computer program, the electronic device executes the circle selection model training method in the first aspect and any possible design of the first aspect, or the circle selection method in the second aspect and any possible design of the second aspect.

[0026] The selection model training method, selection method and electronic device provided in the present application realize the identification of the causal relationship between the sales method and the conversion success rate by inputting the intervention group data into the first basic model to train the first prediction model corresponding to the first sales method, and by inputting the control group data into the first basic model to train the second prediction model corresponding to the second sales method; the construction of the selection model is realized by weighted summing the first prediction model and the second prediction model using the weight coefficient generated by the third prediction model; based on the selection model, the present application realizes the calculation of the comprehensive influence value of the customer to be selected after different sales methods intervene in the customer to be selected, and based on the comprehensive influence value, the present application realizes the judgment of whether the influence of different sales methods on the conversion success rate is positive, and the calculation of the gain of different sales methods on the conversion success rate; based on the gain, the present application selects the customer to be selected into the customer groups corresponding to different sales methods and sales frequencies, improves the accuracy of customer group selection, realizes precision marketing, and improves the conversion success rate of the customer group. The customer group selection method of the present application achieves the effect of investing a small amount of sales resources in customers with high conversion intention, and investing more sales resources in customers with low conversion intention but high gain, thereby optimizing the sales resource delivery strategy, achieving accurate delivery of sales resources, reducing waste of sales resources, and improving the effectiveness and utilization of sales resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] 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.

[0028] Figure 1 A machine learning model training and selection flow chart for existing technologies;

[0029] Figure 2 A flowchart of a circle selection model training method provided in one embodiment of the present application;

[0030] Figure 3 A flowchart of a method for generating training data provided in one embodiment of the present application;

[0031] Figure 4 A flowchart of a circle selection model construction provided in one embodiment of the present application;

[0032] Figure 5 A flowchart of a circle selection method provided in one embodiment of the present application;

[0033] Figure 6 A flowchart of a circled example provided in one embodiment of the present application;

[0034] Figure 7 A schematic diagram of the structure of a circle selection model training device provided in one embodiment of the present application;

[0035] Figure 8 A schematic diagram of the structure of a circle selection device provided in one embodiment of the present application;

[0036] Fig. 9 A schematic diagram of the hardware structure of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0037] 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 the embodiments. 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.

[0038] It should be noted that the customer information (including but not limited to customer device information, customer personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the customer or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for customers to choose to authorize or refuse.

[0039] In the field of telemarketing, the goal of telemarketing is to select customers who have not been successfully converted before telemarketing, but will be successfully converted after telemarketing, and select them into a customer group. For those customers who have a high conversion willingness, a small amount of sales resources can promote conversion. For those customers who have a low conversion willingness, it is necessary to allocate sales resources to them according to their actual situation to increase the conversion success rate of these customers. Therefore, customer group selection is a key step in growing telemarketing. Only by accurately selecting the appropriate customer group and recommending the right products to the right customer group at the right time and with the right means can we maximize the customer conversion rate, avoid wasting customer resources, improve the utilization efficiency of customer resources, and thus achieve the optimization of overall interests. Among them, customer group is the abbreviation of customer group.

[0040] In traditional telemarketing scenarios, some business rules are usually used to screen customers. However, this method does not take advantage of big data, cannot fully tap the value of data in customer selection, lacks scientificity, and often cannot make accurate analysis of customer groups and telemarketing methods.

[0041] Therefore, based on the use of big data, a method of using a machine learning model to predict conversion rates is used, so that customer group selection can be achieved based on the prediction results. The machine learning model can use whether a customer is successfully converted as a label. The machine learning model can extract feature data from data such as the customer's identity information and business information. The machine learning model can use the label and the feature data to implement model training. The predicted probability output by the machine learning model is used to indicate the conversion success rate of the customer after receiving telemarketing. It is generally believed that the higher the predicted probability, the greater the possibility of successful conversion. Therefore, in the process of customer group segmentation, the electronic device can divide the customer group into multiple groups based on the value of the predicted probability. Since the conversion success rate of the group with a high predicted probability is high, telemarketers usually focus on telemarketing for the group with a high predicted probability, thereby further improving the conversion success rate of this group of customers. Optionally, the above-mentioned process of implementing customer group selection through a machine learning model can be as follows Figure 1 shown.

[0042] Although this method has achieved the utilization of big data by using a machine learning model, it has not fully analyzed the inherent causal relationship between telemarketing methods and conversion results. Even for a single machine learning model such as a linear regression model, some causal relationships can be identified, but there is still the problem of being unable to quantify the strength of the causal relationship. Therefore, the existing technology for the allocation of sales resources still has the problem of inaccurate customer group selection, which is prone to mismatch of sales resources and waste of sales resources and customer resources. To this end, further optimizing the customer group selection method, achieving accurate delivery of sales resources, and improving the conversion success rate are still current problems that need to be solved.

[0043] In response to the shortcomings of traditional methods, this application innovatively proposes a customer selection solution based on a weighted counterfactual causal inference model. The core of this method is the introduction of an advanced Uplift causal inference model. The Uplift causal inference model can predict the additional improvement in the return on investment (ROI) when intervening with customers using different sales methods. This application uses the Uplift causal inference model to accurately calculate the sensitivity of the population to marketing, thereby simulating and optimizing revenue budgets and delivery strategies to maximize the effectiveness of marketing promotion.

[0044] This application is not limited to measuring and predicting "incremental lift", but avoids blindly investing marketing budget in "natural conversion" groups. Our method deeply explores the potential causal relationship between telemarketing methods and conversion goals, accurately determines the impact of telemarketing factors on conversion success rate, and quantifies their causal strength. Based on these in-depth analyses, we can accurately match telemarketing lists with the most suitable telemarketing methods to achieve optimal resource allocation.

[0045] This application can accurately calculate the impact of telemarketing on the conversion success rate by using a selection model built based on a weighted counterfactual causal inference model, and then segment the customer base based on these values ​​and allocate the most appropriate sales resources to them. For customers with high conversion intention, we can adopt a more accurate and resource-saving strategy; for customers with low conversion intention but high potential gain, we increase investment to tap their greater value. It can be seen that the customer group selection method of this application not only improves scientificity and logic, but also achieves remarkable results in practical applications. By adopting the method of this application, the conversion rate in the telemarketing process has been significantly improved, which fully proves the accuracy and efficiency of this application in customer group selection. Compared with traditional methods, this application has undoubtedly brought new breakthroughs and progress in the field of customer group selection.

[0046] This application uses an electronic device as an execution subject to execute the circle selection model training method and circle selection method of the following embodiments. Specifically, the execution subject can be a hardware device of the electronic device, or a software application that implements the following embodiments in the electronic device, or a computer-readable storage medium installed with the software application that implements the following embodiments, or a code that implements the software application of the following embodiments.

[0047] The technical solution of the present application is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0048] Figure 2 FIG. 1 is a flow chart of a method for training a circle selection model provided by an embodiment of the present application. Figure 2 As shown, with the electronic device as the execution subject, the method of this embodiment may include the following steps:

[0049] S101. Acquire training data, and process the training data to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method.

[0050] In this embodiment, the electronic device can obtain a large amount of training data. Optionally, each training data may include three parts: sales method, actual conversion label, and sample data. Optionally, the sales method is the sales method actually used by the training data. Optionally, the actual conversion label may be the actual conversion result of the training data after being interfered with by the sales method. The actual conversion result may include conversion success and conversion failure. The actual conversion label may be the result after the actual conversion result is digitized. For example, the actual conversion label corresponding to the successful conversion may be 1, and the actual conversion label corresponding to the failed conversion may be 0. Optionally, the sample data may be customer features.

[0051] The electronic device can classify and process the training data to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method. For example, the first sales method can be an outbound call made by an automatic voice response service (Interactive Voice Response, IVR). The second sales method can be manual dialing. Optionally, the intervention group data and the control group data can also include an impact label. The impact label is used to determine whether the impact of different sales methods on the customer is consistent. Optionally, when two sales methods are used to interfere with the customer, the more similar the conversion results are, the closer the impact value is to 0, and the greater the difference in conversion results is, the closer the impact value is to 1.

[0052] In one example, the specific process of the electronic device obtaining the intervention group data corresponding to the first sales method may include:

[0053] Step 11: Group all training data according to the sales methods in the training data to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method.

[0054] In this step, after acquiring a large amount of training data, the electronic device can group the training data according to the sales methods included in each training data. Since the sales methods in this application include a first sales method and a second sales method, this application can divide the training data into intervention group data corresponding to the first sales method and control group data corresponding to the second sales method.

[0055] Optionally, the intervention group data in this step may include three parts: a first sales method, a first actual conversion label, and sample data. Also, the control group data in this step may include three parts: a second sales method, a second actual conversion label, and sample data.

[0056] Step 12: Based on the intervention group data, the control group data and the second basic model, generate a first impact label corresponding to each intervention group data.

[0057] In this step, the electronic device may be preset with a second basic model. Optionally, the second basic model may be a lightweight and fast iterative two-causal inference model. The setting of the lightweight and fast iterative two-causal inference model can realize the rapid training of the first basic model and the second basic model, thereby improving the model training efficiency and facilitating the periodic optimization of the model. Optionally, the two causal inference models may be models such as decision trees, random forests, GBDT, LightGBM and XGBoost. Optionally, the two causal inference models are preferably lightGBM. The lightGBM model has the advantages of fast training speed and the ability to automatically handle missing values ​​and discrete values. The lightGBM adopts a negative gradient optimization algorithm, and each round of iteration will fit the residual of the previous round of iterations, so that the prediction result of the current t step is optimized in the negative gradient direction of the target loss function to the prediction value of the previous step t-1, so that the target loss can be continuously reduced through each iteration. Therefore, the lightGBM algorithm is used in the uplift causal inference model of the present application, which can quickly realize the training of the causal inference model, thereby quickly obtaining the final circle selection model.

[0058] The electronic device can use the second basic model to process the intervention group data and the control group data to obtain the first impact label corresponding to each intervention group data. The first impact label can be used to indicate the difference between the conversion result of each training sample in the intervention group data using the first sales method and the conversion result of the second sales method. When the conversion result of the intervention using the first sales method is consistent with the conversion result of the intervention using the second sales method, the first impact label can be close to 0. When the conversion result of the intervention using the first sales method is inconsistent with the conversion result of the intervention using the second sales method, the first impact label can be close to 1 or -1. Optionally, when the first impact label can be close to 1, it can be used to indicate that the first sales method has produced a positive result. When the first impact label can be close to -1, it can be used to indicate that the second sales method has produced a positive result.

[0059] Optionally, the process of the electronic device calculating the first impact label may include the following steps:

[0060] Step 12.1: Input the control group data into the second basic model, and train to obtain a control group model for predicting the conversion success rate after intervention using the second sales method.

[0061] In this step, the control group data is actually interfered with by the second sales method. The control group data used in this step includes sample data and the second actual conversion label. The second actual conversion label is determined according to the conversion result after the control group data is interfered with by the second sales method. Optionally, the second actual conversion label may include 0 and 1. 1 is used to indicate a successful conversion, and 0 is used to indicate a failed conversion.

[0062] The electronic device can input the sample data in the control group data and the second actual conversion label into the second basic model for training to obtain a control group model. The control group model can be used to predict the conversion success rate after intervention using the second sales method.

[0063] In one implementation, if the conversion success rate is close to 1, it means that the conversion success rate of the sample data after being interfered with by the second sales method is high. If the conversion success rate is close to 0, it means that the conversion success rate of the sample data after being interfered with by the second sales method is low.

[0064] In another implementation, the conversion success rate can also be used to indicate whether the second sales method has a positive effect on the sample data. If the conversion success rate is close to 1, it means that the sample data has a positive effect when the second sales method is used for interference, which can promote sales. If the conversion success rate is close to 0, it means that the sample data has a negative effect when the second sales method is used for interference, which cannot promote sales.

[0065] Step 12.2: Input the intervention group data into the trained control group model to predict the first counterfactual result corresponding to the conversion success rate of each intervention group data after the second sales method is used for intervention.

[0066] In this step, the intervention group data are actually interfered with using the first sales method. The intervention group data used in this step includes sample data and a first actual conversion label. The first actual conversion label is determined according to the conversion result after the intervention group data is intervened using the first sales method. Optionally, the actual conversion label may include 0 and 1. 1 is used to indicate a successful conversion, and 0 is used to indicate a failed conversion.

[0067] The electronic device can input the sample data in the intervention group data into the control group model trained in step 2.1. The control group model can predict the first counterfactual result corresponding to each intervention group data. Optionally, the first counterfactual result is used to indicate the conversion success rate of the sample data in the intervention group data after the sample data in the intervention group data is intervened using the second sales method.

[0068] Since the intervention group data actually used the first sales method for intervention, using the second sales method for intervention is inconsistent with the facts. Therefore, the predicted conversion success rate generated when the intervention group data is intervened using the second sales method can be named the first counterfactual result. The first counterfactual result is a theoretical reference result under hypothetical conditions.

[0069] The closer the first counterfactual result is to 1, the higher the conversion success rate of the sample data of the intervention group after being interfered with by the second sales method. The closer the conversion success rate is to 0, the smaller the impact of the second sales method on the sample data of the intervention group.

[0070] Step 12.3: Determine the first impact label corresponding to each intervention group data according to the first actual conversion label and the first counterfactual result of each intervention group data.

[0071] In this step, the electronic device can obtain the first actual conversion label corresponding to each intervention group data from the intervention group data. The electronic device can obtain the first counterfactual result corresponding to each intervention group data from step 2.2. The electronic device can determine the first impact label corresponding to each intervention group data by comparing the first actual conversion label and the first counterfactual result of each intervention group data.

[0072] In one implementation, the electronic device may determine the first impact label corresponding to the intervention group data according to the difference between the first actual conversion label and the first counterfactual result of each intervention group data.

[0073] For example, when the first actual conversion label is 1, it means that the intervention group data was successfully converted after the intervention using the first sales method. When the first counterfactual result is 0.9, it means that the intervention group data has a very high conversion success rate after the intervention using the second sales method. At this time, it means that the intervention group data may have a strong willingness to consume. Therefore, the use of different sales methods has little impact on the customer. The first impact label calculated based on the difference between the first actual conversion label and the first counterfactual result can be 0.1. The first impact label is close to 0, indicating that different sales methods have little impact on the intervention group data.

[0074] For another example, when the first actual conversion label is 1, it means that the intervention group data was successfully converted after the intervention using the first sales method. When the first counterfactual result is 0.1, it means that the conversion success rate of the intervention group data after the intervention using the second sales method is low. At this time, it means that the intervention group data can be successfully converted with a high probability because the correct sales method was used for intervention. Therefore, the use of different sales methods has a greater impact on the customer. The first impact label calculated based on the difference between the first actual conversion label and the first counterfactual result can be 0.9. The first impact label is close to 1, indicating that different sales methods have a great impact on the intervention group data. Optionally, the first sales method is the correct sales method. In order to illustrate the positive impact of the first sales method, the first impact label can be +0.9.

[0075] For another example, when the first actual conversion label is 0, it means that the intervention group data failed to convert after the first sales method was used for intervention. When the first counterfactual result is 0.9, it means that the conversion success rate of the intervention group data after the second sales method was used for intervention is extremely high. At this time, it means that the failure of the conversion of the intervention group data is very likely due to the use of the wrong sales method for intervention. Therefore, the use of different sales methods has a greater impact on the customer. The first impact label calculated based on the difference between the first actual conversion label and the first counterfactual result can be -0.9. The first impact label is close to -1, indicating that different sales methods have a great impact on the intervention group data. Optionally, the first sales method is an incorrect sales method, and the second sales method is the correct sales method. In order to illustrate the positive impact of the second sales method, the first impact label can be -0.9.

[0076] For another example, when the first actual conversion label is 0, it means that the intervention group data failed to convert after the first sales method was used for intervention. When the first counterfactual result is 0.1, it means that the conversion success rate of the intervention group data after the second sales method was used for intervention is low. At this time, it means that the intervention group data may have a weak willingness to consume. Therefore, the use of different sales methods has little impact on the customer. The first impact label calculated based on the difference between the first actual conversion label and the first counterfactual result can be -0.1. The first impact label is close to 0, indicating that different sales methods have little impact on the intervention group data.

[0077] Step 13: Add the first impact label to the intervention group data to obtain the intervention group data corresponding to the first sales method after the first impact label is added.

[0078] In this step, after the first impact label is added to the intervention group data, the intervention group data may include four parts: the first sales method, the first actual conversion label, the sample data, and the first impact label.

[0079] In another example, the specific process of the electronic device obtaining the control group data corresponding to the second sales method may include:

[0080] Step 21: group all training data according to the sales methods in the training data to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method.

[0081] In this step, after acquiring a large amount of training data, the electronic device can group the training data according to the sales methods included in each training data. Since the sales methods in this application include a first sales method and a second sales method, this application can divide the training data into intervention group data corresponding to the first sales method and control group data corresponding to the second sales method.

[0082] Optionally, the intervention group data in this step may include three parts: a first sales method, a first actual conversion label, and sample data. Also, the control group data in this step may include three parts: a second sales method, a second actual conversion label, and sample data.

[0083] Step 22: Based on the intervention group data, the control group data and the second basic model, generate a second impact label corresponding to each control group data.

[0084] In this step, the electronic device may be preset with a second basic model. Optionally, the second basic model is the same as the second basic model in step 12 above.

[0085] The electronic device can use the second basic model to process the intervention group data and the control group data to obtain a second impact label corresponding to each control group data. The second impact label can be used to indicate the difference between the conversion result of each training sample in the control group data when the first sales method is used for intervention and the conversion result of the second sales method is used for intervention. When the conversion result of the intervention using the first sales method is consistent with the conversion result of the intervention using the second sales method, the second impact label can be close to 0. When the conversion result of the intervention using the first sales method is inconsistent with the conversion result of the intervention using the second sales method, the second impact label can be close to 1 or -1. Optionally, when the second impact label can be close to 1, it can be used to indicate that the first sales method has produced a positive result. When the second impact label can be close to -1, it can be used to indicate that the second sales method has produced a positive result.

[0086] Optionally, the process of the electronic device calculating the second impact label may include the following steps:

[0087] Step 22.1: Input the intervention group data into the second basic model, and train the intervention group model for predicting the conversion success rate after the intervention using the first sales method.

[0088] In this step, the intervention group data are actually interfered with using the first sales method. The intervention group data used in this step includes sample data and a first actual conversion label. The first actual conversion label is determined according to the conversion result after the intervention group data is intervened using the first sales method. Optionally, the first actual conversion label may include 0 and 1. 1 is used to indicate a successful conversion, and 0 is used to indicate a failed conversion.

[0089] The electronic device can input the sample data in the intervention group data and the first actual conversion label into the first basic model for training to obtain the intervention group model. The intervention group model can be used to predict the conversion success rate after the intervention using the first sales method.

[0090] In one implementation, if the conversion success rate is close to 1, it means that the conversion success rate of the sample data after being interfered with by the first sales method is high. If the conversion success rate is close to 0, it means that the conversion success rate of the sample data after being interfered with by the first sales method is low.

[0091] In another implementation, the conversion success rate can also be used to indicate whether the interference of the first sales method on the sample data is positive. If the conversion success rate is close to 1, it means that the sample data has a positive impact when the first sales method is used for interference, which can promote sales. If the conversion success rate is close to 0, it means that the sample data has a negative impact when the first sales method is used for interference, which cannot promote sales.

[0092] Step 22.2: Input the control group data into the trained intervention group model to predict the second counterfactual result corresponding to the conversion success rate of each control group data after the intervention using the first sales method.

[0093] In this step, the control group data is actually interfered with by the second sales method. The control group data used in this step includes sample data and a second actual conversion label. The second actual conversion label is determined according to the conversion result after the control group data is interfered with by the second sales method. Optionally, the actual conversion label may include 0 and 1. 1 is used to indicate a successful conversion, and 0 is used to indicate a failed conversion.

[0094] The electronic device may input the sample data in the control group data into the intervention group model trained in step 2.1. The intervention group model may predict a second counterfactual result corresponding to each control group data. Optionally, the second counterfactual result is used to indicate the conversion success rate of the sample data in the control group data after the sample data in the control group data is intervened using the first sales method.

[0095] Since the control group data actually used the second sales method for intervention, the use of the first sales method for intervention is inconsistent with the facts. Therefore, the predicted conversion success rate generated when the control group data is intervened using the first sales method can be named the second counterfactual result. The second counterfactual result is a theoretical reference result under hypothetical conditions.

[0096] The closer the second counterfactual result is to 1, the higher the conversion success rate of the sample data of the control group data after being interfered with by the first sales method. The closer the conversion success rate is to 0, the smaller the impact of the first sales method on the sample data of the control group data.

[0097] Step 22.3: Determine the second impact label corresponding to each control group data according to the second actual conversion label and the second counterfactual result of each control group data, wherein the second actual conversion label is determined according to the conversion result after intervening the control group data using the second sales method.

[0098] In this step, the electronic device can obtain the second actual conversion label corresponding to each control group data from the control group data. The electronic device can obtain the second counterfactual result corresponding to each control group data from step 2.2. The electronic device can determine the second impact label corresponding to each control group data by comparing the second actual conversion label and the second counterfactual result of each control group data.

[0099] In one implementation, the electronic device may determine the second impact label corresponding to the control group data according to the difference between the second counterfactual result of each control group data and the second actual conversion label.

[0100] For example, when the second actual conversion label is 1, it means that the control group data was successfully converted after the second sales method was used for intervention. When the second counterfactual result is 0.9, it means that the conversion success rate of the control group data was extremely high after the first sales method was used for intervention. At this time, it means that the control group data may have a strong willingness to consume. Therefore, the impact of using different sales methods on the customer is small. The second impact label calculated based on the difference between the second counterfactual result and the second actual conversion label can be -0.1. The second impact label is close to 0, indicating that different sales methods have little impact on the control group data.

[0101] For another example, when the second actual conversion label is 1, it means that the control group data was successfully converted after the second sales method was used for intervention. When the second counterfactual result is 0.1, it means that the conversion success rate of the control group data was low after the first sales method was used for intervention. At this time, it means that the control group data was able to be successfully converted with a high probability because the correct sales method was used for intervention. Therefore, the use of different sales methods has a greater impact on the customer. The second impact label calculated based on the difference between the second counterfactual result and the second actual conversion label can be -0.9. The second impact label is close to -1, indicating that different sales methods have a great impact on the control group data. Optionally, the second sales method is the correct sales method. In order to illustrate the positive impact of the second sales method, the second impact label can be -0.9.

[0102] For another example, when the second actual conversion label is 0, it means that the control group data failed to convert after the second sales method was used for intervention. When the second counterfactual result is 0.9, it means that the conversion success rate of the control group data is extremely high after the first sales method is used for intervention. At this time, it means that the failure of the conversion of the control group data is very likely due to the use of the wrong sales method for intervention. Therefore, the use of different sales methods has a greater impact on the customer. The second impact label calculated based on the difference between the second counterfactual result and the second actual conversion label can be 0.9. The second impact label is close to 1, indicating that different sales methods have a great impact on the control group data. Optionally, the second sales method is an incorrect sales method, and the first sales method is the correct sales method. In order to illustrate the positive impact of the first sales method, the second impact label can be 0.9.

[0103] For another example, when the second actual conversion label is 0, it means that the control group data failed to convert after the second sales method was used for intervention. When the second counterfactual result is 0.1, it means that the conversion success rate of the control group data after the first sales method was used for intervention is low. At this time, it means that the control group data may have a weak willingness to consume. Therefore, the use of different sales methods has little impact on the customer. The second impact label calculated based on the difference between the second counterfactual result and the second actual conversion label can be 0.1. The second impact label is close to 0, indicating that different sales methods have little impact on the control group data.

[0104] Step 23: Add the second impact label to the control group data to obtain the control group data corresponding to the second sales method after the second impact label is added.

[0105] In this step, after the second impact label is added to the intervention group data, the intervention group data may include four parts: the second sales method, the second actual conversion label, the sample data, and the second impact label.

[0106] S102, input the intervention group data into the first basic model, and train to obtain a first prediction model corresponding to the first sales method; input the control group data into the first basic model, and train to obtain a second prediction model corresponding to the second sales method.

[0107] In this embodiment, the electronic device may be preset with a first basic model. Optionally, the first basic model may be a lightweight and fast iterative two-causal inference model. The setting of the lightweight and fast iterative two-causal inference model can realize the rapid training of the first basic model and the second basic model, thereby improving the model training efficiency and facilitating the periodic optimization of the model. Optionally, the two causal inference models may be models such as decision trees, random forests, GBDT, LightGBM and XGBoost. Optionally, the two causal inference models are preferably lightGBM. The lightGBM model has the advantages of fast training speed and the ability to automatically handle missing values ​​and discrete values. The lightGBM adopts a negative gradient optimization algorithm, and each round of iteration will fit the residual of the previous round of iterations, so that the prediction result of the current t step is optimized in the negative gradient direction of the target loss function to the prediction value of the previous step t-1, so that the target loss can be continuously reduced through each iteration. Therefore, the lightGBM algorithm is used in the uplift causal inference model of the present application, which can quickly realize the training of the causal inference model, thereby quickly obtaining the final circle selection model.

[0108] The intervention group data may include sample data and a first impact label. The electronic device may input the sample data and the first impact label in the intervention group data into the first basic model and train to obtain a first prediction model. The first prediction model is used to predict the impact of different sales methods on the customer. When the prediction result of the first prediction model is close to 1, it means that the use of the first sales method has a positive impact. When the prediction result of the first prediction model is close to -1, it means that the use of the second sales method has a positive impact. When the prediction result of the first prediction model is close to 0, it means that the use of different sales methods has little impact on the customer.

[0109] The control group data may include sample data and a second impact label. The electronic device may input the sample data and the second impact label in the control group data into the first basic model and train to obtain a second prediction model. The second prediction model is used to predict the impact of different sales methods on the customer. When the prediction result of the second prediction model is close to 1, it indicates that the use of the first sales method has a positive impact. When the prediction result of the second prediction model is close to -1, it indicates that the use of the second sales method has a positive impact. When the prediction result of the second prediction model is close to 0, it indicates that the use of different sales methods has little impact on the customer.

[0110] In one example, since the first impact label in the intervention group data and the second impact label in the control group data are substantially the same. Therefore, the electronic device can write the first impact label and the second impact label into the training data as impact labels. The electronic device can input the sample data and impact labels in all the training data into the first basic model and train to obtain a prediction model. The function implemented by the prediction model is the same as the effect implemented by the selection model in step S104. Optionally, the weighted selection model after distinguishing the sales method in the present application adds more influences of sales methods on the selected customers compared to the prediction model in this example. Therefore, the selection model of the present application can better implement analysis based on sales methods and further improve the prediction accuracy.

[0111] S103: Input all training data into the classification basic model, and train to obtain a third prediction model for predicting sales methods.

[0112] In this embodiment, each training data may include a sales method and sample data. The electronic device may input all the training data into a classification basic model for training to obtain a third prediction model. Optionally, the classification basic model may be a binary classification model. Since the present application only includes the first sales method and the second sales method, the setting of the binary classification model can realize the prediction of the sales method of the training data. That is, the electronic device can use the third prediction model to predict the sales method corresponding to the sample data. Optionally, the classification basic model may include logistic regression, support vector machine, decision tree, random forest, gradient boosting tree, etc.

[0113] In one example, the electronic device may obtain training data with an actual conversion label of 1. The electronic device may input the sales method and sample data of the training data into a classification basic model for training to obtain a third prediction model. Optionally, by using the training data with an actual conversion label of 1, the sample data may be classified more accurately. This method may effectively avoid the situation where sample data 1 is classified into the first sales method, but the first sales method has a reverse effect on sample data 1.

[0114] In another example, the electronic device may obtain training data whose actual conversion label is 1 and whose impact label is close to 1 or -1. Optionally, the impact label close to 1 or -1 may be a value of the impact label greater than a first threshold or less than a second threshold. For example, the first threshold may be 0.9, and the second threshold may be -0.9. The electronic device may input the sales method and sample data of these training data into a classification basic model for training to obtain a third prediction model. Optionally, the training data whose actual conversion label is 1 and whose impact label is close to 1 or -1 is training data that has been successfully converted after using the correct sales method. The third prediction model trained using the training data can more accurately classify the sample data and select a sales method with a more positive impact for the sample data.

[0115] S104: construct a circle selection model according to the first prediction model, the second prediction model, and the weight coefficients generated by the third prediction model.

[0116] In this embodiment, the electronic device can use the third prediction model to construct the first weight coefficient and the second weight coefficient. The electronic device can use the first weight coefficient and the second weight coefficient to perform weighted fusion on the first prediction model and the second prediction model to obtain the circle selection model.

[0117] In one example, the specific implementation process of the above steps may include:

[0118] Step 1: Generate a first weight coefficient of the first prediction model and a second weight coefficient of the second prediction model according to the prediction result of the third prediction model for the selected customers.

[0119] In this step, the electronic device may assume a set of customers to be selected. The electronic device may input each customer to be selected in the set into the third prediction model to obtain a prediction result. The prediction result is used to indicate the sales method that needs to intervene in the customer to be selected. The electronic device may count the proportion of customers to be selected who use the first sales method and the proportion of customers to be selected who use the second sales method in the set of customers to be selected. The electronic device may use the proportion of customers to be selected using the first sales method as the first weight coefficient and the proportion of customers to be selected using the second sales method as the second weight coefficient.

[0120] Step 2: According to the first weight coefficient and the second weight coefficient, the first prediction model and the second prediction model are weightedly fused to construct a circle selection model.

[0121] In this step, the electronic device can use the first weight coefficient as the weight of the first prediction model, and use the second weight coefficient as the weight of the second prediction model. The electronic device can perform weighted fusion on the first prediction model and the second prediction model according to the first weight coefficient and the second weight coefficient to obtain a circled selection model. Since both the first prediction model and the second prediction model determine that the first sales method has a positive impact when the prediction result is close to 1, and determine that the second sales method has a positive impact when the prediction result is close to -1, and determine that different sales methods have little impact on customers when the prediction result is close to 0. Therefore, the prediction result of the weighted circled selection model has the same meaning as the prediction results of the first prediction model and the second prediction model.

[0122] The electronic device can obtain a large amount of training data in the circle selection model training method provided in the present application. The electronic device can classify and process the training data to obtain the intervention group data corresponding to the first sales method and the control group data corresponding to the second sales method. The electronic device can input the intervention group data into the first basic model and train to obtain the first prediction model corresponding to the first sales method. The electronic device can input the control group data into the first basic model and train to obtain the second prediction model corresponding to the second sales method. The electronic device can input all the training data into the classification basic model for training to obtain a third prediction model. The electronic device can use the third prediction model to construct a first weight coefficient and a second weight coefficient. The electronic device can use the first weight coefficient and the second weight coefficient to perform weighted fusion on the first prediction model and the second prediction model to obtain a circle selection model. In the present application, the fusion calculation obtains the comprehensive impact value when different sales methods interfere with the circled customers. The customer group circle selection achieved based on the comprehensive impact value improves the circle selection accuracy, improves the precise intervention of the circled customers, improves the utilization rate of sales resources, and improves the conversion success rate of the customers to be circled.

[0123] Figure 3 A flow chart of a method for generating training data provided by an embodiment of the present application is shown. Figure 2 Based on the example, Figure 3 As shown, with the electronic device as the execution subject, the method of this embodiment is used to process the customer data to obtain training data, and the specific process may include the following steps:

[0124] S201. Obtain customer data of each customer, and determine the sales method corresponding to the first moment in the customer data.

[0125] In this step, the electronic device can obtain customer data of multiple customers from the business system. The customer data can be all the data of a customer from the time of registration in the business system to the current moment. Alternatively, the electronic device can also obtain customer data from the outside. The customer data can include all the data of the customer from the time of registration to the current moment. The electronic device can identify the first moment when the sales behavior occurs from the customer data. For example, when the sales method is IVR outbound call, the first moment can be the moment when the IVR outbound call behavior occurs. That is, the first moment can be the moment when the customer is contacted through the IVR outbound call.

[0126] Optionally, the customer data of each customer may include multiple sales behaviors. Optionally, multiple sales behaviors of a customer may correspond to one or more sales methods. Optionally, the moment when each sales row occurs may be the first moment. For example, when three sales behaviors appear in the customer data of a customer on January 5, April 7, and October 9, respectively, the three moments are all the first moments. For another example, the sales method corresponding to January 5 may be IVR outbound calling, and the sales methods corresponding to April 7 and October 9 may be manual calling.

[0127] The electronic device may determine the first moment according to the sales behavior in the customer data. Also, the electronic device may determine the corresponding sales method according to the sales behavior corresponding to each first moment. Optionally, the first moment may be recorded as T.

[0128] Optionally, a mark S may be used to record the sales method, wherein the first sales method may be recorded as S=1, and the second sales method may be recorded as S=0.

[0129] Optionally, since this application is mainly used for the selection of sales methods, S is used to mark the sales methods in this application. In other schemes, S may also correspond to other policies, measures or activities. For example, S=1 may correspond to participation in activity A, and S=0 may correspond to participation in activity B. Then, the corresponding training data will no longer include the sales methods, but needs to include participation in activities.

[0130] Optionally, since this application mainly includes two sales methods, manual dialing and outbound dialing, this application subsequently divides the training data into two groups, intervention group data and control group data, and the third prediction model also realizes binary classification. In actual use, when three or more sales methods, activities, policies, and measures are included, the above technical solution can also be expanded to divide the training data into multiple groups, and the third prediction model realizes multiple classification.

[0131] S202: Obtain conversion results after the first moment from the customer data, and generate actual conversion labels according to the conversion results.

[0132] In this step, after determining at least one first moment included in the customer data, the electronic device may obtain a conversion result after the first moment from the customer data. The conversion result is used to indicate whether the customer makes a purchase after the customer is intervened with using the sales method. The conversion result may include two types: conversion success and conversion failure. If the customer makes a purchase, the conversion is successful. Otherwise, if the customer does not make a purchase, the conversion fails.

[0133] Optionally, the conversion result may be obtained at a third time after the intervention is performed. Optionally, the third time may be 1 day. Optionally, when the first moment is T, the moment corresponding to the third time after the intervention is performed may be recorded as T+1. That is, the electronic device may obtain the conversion result at T+1.

[0134] The electronic device may also generate an actual conversion tag based on the conversion result. The actual conversion tag may be recorded as Y. When the conversion is successful, the actual conversion tag may be recorded as Y=1. When the conversion fails, the actual conversion tag may be recorded as Y=0.

[0135] S203: Acquire customer information before the first moment from the customer data, and perform feature extraction on the customer information to obtain sample data.

[0136] In this step, based on the first moment, the electronic device can obtain customer information before the first moment. The customer information may include identity information, cash withdrawal, loan, adjustment, login, MOT event, etc. The electronic device can perform feature extraction on the customer information to obtain sample data.

[0137] Optionally, the process of acquiring the sample data may specifically include the following steps:

[0138] Step 31: Obtain basic customer information before a first moment from the customer data, and perform feature extraction on the basic customer information to obtain first feature data.

[0139] In this step, the electronic device may obtain the basic customer information before the first moment from the customer data. Optionally, the moment before the first moment may be the second moment. Optionally, the time interval between the second moment and the first moment may be the first time interval. Optionally, when the first time interval is 2 days, the second moment is T-2.

[0140] Optionally, the basic customer information may include customer information extracted from data sources such as registration, customer profile, price reduction, credit, quota adjustment, cash withdrawal, real name, and repayment. The electronic device may perform feature extraction on the basic customer information to obtain first feature data. Specifically, after obtaining the basic customer information, the electronic device may input the basic customer information into a preset feature extraction model to extract the first feature data.

[0141] Optionally, the electronic device can extract the field information to be used from the customer basic information according to the preset fields. The electronic device can also convert the field information into a code according to a preset conversion method. The electronic device can combine the codes of these fields to obtain the final code information. The electronic device can also input the code information into a preset feature extraction model to process and obtain the first feature data. Optionally, the first feature data can be a feature vector.

[0142] Step 32: Obtain customer reception information before the first moment from the customer data, and perform feature extraction on the customer reception information to obtain second feature data.

[0143] In this step, the electronic device may obtain the customer reception information before the first moment from the customer data. Optionally, the moment of obtaining the customer reception information and the moment of obtaining the customer basic information in step 31 may be different moments. Optionally, the moment of obtaining the customer reception information may be a moment before the first moment or may be a third moment. Optionally, the time interval between the third moment and the first moment may be a second time interval. Optionally, when the second time interval is 1 day, the second moment is T-1.

[0144] Optionally, the customer receiving information may include key moment (MOT) event information such as red envelopes, coupons, telemarketing, push, and text messages. Optionally, due to the time limit of data warehouse technology (Extract-Transform-Load, ETL), the MOT event information usually needs to be extracted at the third moment T-1.

[0145] After extracting the customer acceptance information, the electronic device may perform feature extraction on the customer acceptance information to obtain second feature data.

[0146] Optionally, the electronic device can extract the field information to be used from the customer acceptance information according to a preset field. The electronic device can also convert the field information into a code according to a preset conversion method. The electronic device can combine the codes of these fields to obtain the final code information. The electronic device can also input the code information into a preset feature extraction model to process and obtain the second feature data. Optionally, the second feature data can be a feature vector.

[0147] Step 33: Fusion extract the first feature data and the second feature data to obtain sample data.

[0148] In this step, the fusion process may be to splice the first feature data and the second feature data to obtain the final customer feature data. Alternatively, feature extraction may be performed on the spliced ​​first feature data and the second feature data to obtain the final customer feature data. Alternatively, the first feature data and the second feature data may be fused by weighted fusion or the like to obtain the final customer feature data.

[0149] Optionally, in the model training phase, the customer feature data is sample data. Optionally, in the model use phase, the customer feature data can be input into the selection model to achieve the prediction of the impact value.

[0150] S204: Use the sales method, actual conversion labels and sample data to form training data.

[0151] In this step, the electronic device can combine the sales method, actual conversion label and sample data corresponding to a sales behavior obtained in the above step into a training data. Optionally, multiple sales behaviors of a customer can correspond to multiple training data.

[0152] The present application provides a method for training a selection model, in which an electronic device can obtain customer data from a business system. The electronic device can identify the first moment when a sales behavior occurs from the customer data. The electronic device can obtain the sales method corresponding to the first moment. The electronic device can obtain the conversion result after the first moment from the customer data. The electronic device can also generate an actual conversion label based on the conversion result. Based on the first moment, the electronic device can obtain customer information before the first moment. The electronic device can perform feature extraction on the customer information to obtain sample data. The electronic device can combine the sales method, actual conversion label and sample data corresponding to a sales behavior obtained in the above steps into a training data. The training data of the present application realizes the mining of customer information of customers to be selected, and constructs a wide range of rich and diverse features, so that the selection model can improve the selection accuracy and the conversion success rate of customers to be selected on the basis of absorbing these features.

[0153] Figure 4 A flowchart of constructing a circle selection model provided by an embodiment of the present application is shown. Figure 2 and Figure 3 Based on the embodiment shown. Figure 4 As shown, with the electronic device as the execution subject, the method of this embodiment may include the following steps:

[0154] S301, data splitting.

[0155] In this embodiment, after acquiring the training data set, the electronic device can group the training data according to the sales method included in each training data in the training data set. The sales method can include a first sales method and a second sales method. The recording of the sales method can be implemented using a mark S. The first sales method can be recorded as S=1. The second sales method can be recorded as S=0. The recording of the training data can be implemented using a mark X.

[0156] The electronic device can extract the training data of S=1 to form the intervention group data, and extract the training data of S=0 to form the control group data. The intervention group data can be recorded as X1, and the control group data can be recorded as X0.

[0157] S302. Training of control group model and intervention group model.

[0158] In this embodiment, the electronic device can obtain a second basic model. The second basic model can be lightgbm. The lightgbm algorithm has a fast training speed and can automatically handle missing values ​​and discrete values. Lightgbm uses a negative gradient optimization algorithm. Each round of iteration will fit the residual of the previous round of iteration, so that the prediction result of the current step is optimized in the negative gradient direction of the target loss function to the prediction value of the previous step, thereby continuously reducing the target loss through each iteration. Optionally, the second basic model of the present application can implement the use of the lightgbm by directly calling the interface of the open source lightgbm. For example, the lightgbm version can be 2.2.3.

[0159] The electronic device can use the control group data X0 to train the lghSgbm model to obtain the control group model μ0. Specifically, the electronic device can input the sample data in the control group data together with the second actual conversion label into the second basic model for training to obtain the control group model. The control group model can be used to predict the conversion success rate of the sample data after the second sales method is used to intervene in the sample data. The control group model μ0 can be:

[0160] μ0=E[Y|X,S=0]

[0161] Where X is used to indicate the training data, and S=0 is used to indicate the second sales method. X, S=0 is used to indicate the training data X using the second sales method, that is, the control group data X0. Y is used to indicate the actual conversion label of the training data X. The control group model μ0 is used to predict the conversion success rate after the intervention using the second sales method.

[0162] The electronic device can use the intervention group data X1 to train the lghSgbm model to obtain the intervention group model μ0. Specifically, the electronic device can input the sample data in the intervention group data and the first actual conversion label into the second basic model for training to obtain the intervention group model. The intervention group model can be used to predict the conversion success rate of the sample data after the sample data is intervened by the first sales method. The intervention group model μ1 can be:

[0163] μ1=E[Y|X,S=1]

[0164] Among them, X is used to indicate the training data, and S=1 is used to indicate the first sales method. X, S=1 is used to indicate the training data X using the first sales method, that is, the intervention group data X1. Y is used to indicate the actual conversion label of the training data X. The intervention group model μ0 is used to predict the conversion success rate after the intervention using the first sales method.

[0165] S303: Influence label calculation.

[0166] In this embodiment, the electronic device can input the intervention group data X1 into the control group model μ0 to obtain the first counterfactual result The first counterfactual result is used to indicate the conversion success rate after the intervention group data is intervened using the second sales method. The electronic device may be used to indicate the i-th intervention group data according to the first actual conversion label of the i-th intervention group data. and the first counterfactual result The difference between and determines the first impact label of the i-th intervention group data. The calculation formula of the first impact label can be:

[0167]

[0168] in, The label used to indicate the first actual transformation of the i-th intervention group data. Used to indicate the first counterfactual outcome for the i-th intervention group data. The first effect label used to indicate the data for the ith intervention group.

[0169] The electronic device can input the control group data X0 into the control group model μ1 to obtain the second counterfactual result The second counterfactual result is used to indicate the conversion success rate after the control group data is intervened by the first sales method. The electronic device can be used to indicate the i-th control group data according to the first actual conversion label of the i-th control group data. and the first counterfactual result The difference between and determines the second impact label of the i-th control group data. The calculation formula of the second impact label can be:

[0170]

[0171] in, The second actual conversion label used to indicate the i-th control group data. Used to indicate the second counterfactual outcome for the i-th control group data. The second impact label used to indicate the i-th control group data.

[0172] The electronic device may add the first impact label to the intervention group data, and add the second impact label to the control group data.

[0173] S304: Generate a first prediction model and a second prediction model.

[0174] In this embodiment, the electronic device can use the intervention group data to train the first basic model to obtain the first prediction model. The electronic device can use the control group data to train the first basic model to obtain the second prediction model. Optionally, the first basic model can be a lighSgbm model.

[0175] Specifically, the electronic device can input the sample data and the first impact label in the intervention group data into the lighSgbm model to train and obtain the first prediction model. The first prediction model τ1(x) can be:

[0176] τ1(x)=E[D 1 |X=x]

[0177] Among them, D 1 The first impact label used to indicate the intervention group data. X is the training data. x is the training sample in the training data. τ1(x) is the first prediction model, which is used to predict the impact value of the customer who uses the first sales method to interfere on the first sales method or the second sales method.

[0178] Furthermore, the electronic device can input the sample data in the control group data and the second impact label into the lighSgbm model to train and obtain a second prediction model. The second prediction model τ0(x) can be:

[0179] τ0(x)=E[D 0 |X=x]

[0180] Among them, D 0The second impact label used to indicate the control group data. X is the training data. x is the training sample in the training data. τ0(x) is the second prediction model, which is used to predict the impact value of customers who use the second sales method to interfere on the first sales method or the second sales method.

[0181] S305: training the third prediction function.

[0182] In this embodiment, the electronic device may input the training data X into the classification basic model to train and obtain the third prediction function. Optionally, the classification basic model may be a logistic regression model (LR) or any binary classification model. The electronic device may generate a calculation formula for the first weight coefficient based on the prediction result of the selected customer according to the third prediction model:

[0183] g(x)=P(S=1|X ′ =x ′ )

[0184] Wherein, S=1 is used to indicate that the sales method indicated by the prediction result is the first sales method. ′ is the set of customers to be selected. ′ is the customer characteristic data of the customers to be selected. The customer characteristic data is the sample data. g(x) is used to calculate the proportion of customers to be selected for the first sales method whose prediction result is the third prediction function among all the customers to be selected. The value of g(x) is in the interval [0,1].

[0185] According to the first weight coefficient, the electronic device can determine the second weight coefficient as (1-g(x)).

[0186] S306, constructing a circle selection model.

[0187] In this embodiment, the electronic device can use the first weight coefficient and the second weight coefficient to weight the first prediction model and the second prediction model to obtain a final selection model. The formula of the selection model can be:

[0188] CASE=g(x)τ0(x)+(1-g(x))τ1(x)

[0189] Where g(x) is the first weight coefficient. τ0(x) is the first prediction model. (1-g(x)) is the second weight coefficient. τ1(x) is the second prediction model. CASE is the circle selection model.

[0190] S307, customer group selection.

[0191] In this embodiment, based on the selection model constructed in step S306, the electronic device can predict the comprehensive impact value of each customer to be selected. The comprehensive impact value is the gain of different sales methods on the customer to be selected.

[0192] If the comprehensive impact value is greater than or equal to 0, it means that the first sales method has a positive impact on the selected customer and the effect is better than the second sales method. Therefore, for the selected customers with a comprehensive impact value greater than 0, the first sales method should be used for marketing.

[0193] If the comprehensive impact value is less than 0, it means that the second sales method has a positive impact on the selected customer and the effect is better than the first sales method. Therefore, for the selected customers with a comprehensive impact value less than 0, the second sales method should be used for marketing.

[0194] In this way, the sales method for the selected customers can be determined based on the comprehensive impact value.

[0195] In addition, the electronic device can also sort the comprehensive influence values ​​of all the customers to be selected from high to low. The electronic device can divide the customers to be selected into 10 groups with equal frequency. For the first five groups, the electronic device can use the first sales method. For the last five groups, the electronic device can use the second sales method. The electronic device can also set different sales frequencies for different groups.

[0196] Figure 5 FIG. 1 is a flow chart of a circle selection method provided by an embodiment of the present application. Figures 2 to 4 Based on the embodiment shown. Figure 5 As shown, with the electronic device as the execution subject, the method of this embodiment may include the following steps:

[0197] S401 , obtaining a customer data set, wherein the customer data set includes customer characteristic data of at least one customer to be selected.

[0198] In this embodiment, the electronic device can obtain customer information of the customers to be selected, and perform feature extraction on the customer information to obtain customer feature data. The customer feature data of all the customers to be selected can constitute the customer data set.

[0199] In one example, the process of obtaining customer characteristic data may specifically include the following steps:

[0200] Step 41: Obtain basic customer information before the current moment from the customer data of each to-be-selected customer in the customer data set, and perform feature extraction on the basic customer information to obtain first feature data.

[0201] In this step, the electronic device may obtain the customer data of each to-be-circled customer from the customer data set. The electronic device may obtain the customer basic information before the first moment from each customer data. Optionally, the moment of obtaining the customer basic information may be the fourth moment before the current moment. Optionally, the time interval between the fourth moment and the time before the current moment may be the first time. Optionally, the first time may be 2 days. Optionally, when the current moment is recorded as T, the fourth moment may be recorded as T-2.

[0202] Optionally, the customer basic information may include registration, customer profile, price reduction, credit, quota adjustment, withdrawal, real name, repayment and other data. The process of the electronic device extracting the features of the customer basic information to obtain the first feature data is similar to Figure 3 The embodiments shown are identical.

[0203] Step 42: Acquire customer reception information before the current moment from the customer data of the customer to be selected, and perform feature extraction on the customer reception information to obtain second feature data.

[0204] In this step, the electronic device may continue to obtain the customer reception information before the first moment from the customer data in step 41. Optionally, the moment of obtaining the customer reception information may be a moment different from the moment of obtaining the customer basic information. Optionally, the moment of obtaining the customer reception information may be the fifth moment before the current moment. Optionally, the time interval between the fifth moment and the time before the current moment may be a second time interval. Optionally, the second time interval may be 1 day. Optionally, when the current moment is recorded as T, the fifth moment may be recorded as T-1.

[0205] Optionally, the customer received information may include data such as red envelopes, coupons, telemarketing, push notifications, and SMS push notifications. The process of the electronic device extracting features from the customer received information to obtain the second feature data is similar to the process of the electronic device extracting features from the customer received information to obtain the second feature data. Figure 3 The embodiments shown are identical.

[0206] Step 43: Fusion extract the first feature data and the second feature data to obtain customer feature data.

[0207] In this step, the fusion process may be to splice the first feature data and the second feature data. For example, when the first feature data is [2,1,2] and the second feature data is [2,3,2], the spliced ​​customer feature data may be [2,1,2,2,3,2]. Alternatively, feature extraction may be performed on the spliced ​​first feature data and the second feature data to obtain the final customer feature data. Optionally, the electronic device may use an automatic learning feature extraction method such as a deep neural network to extract the customer feature data. Alternatively, the first feature data and the second feature data may be fused by weighted fusion or the like to obtain the final customer feature data. For example, when the first feature data is [2,1,2] and the second feature data is [2,3,2], assuming the weight ratio is 1:1, the customer feature data may be [4,4,4]

[0208] S402: Input the customer data set into the selection model to calculate the comprehensive impact value of each customer.

[0209] In this embodiment, the electronic device can input the customer feature data in the customer data set into the selection model. The selection model can predict the comprehensive impact value of the customer when different sales methods are used to intervene in the selected customer based on each customer feature data. When the comprehensive impact value is close to 1, it is determined that the first sales method has a positive impact, when the comprehensive impact value is close to -1, it is determined that the second sales method has a positive impact, and when the comprehensive impact value is close to 0, it is determined that the different sales methods have little impact on the customer.

[0210] In one example, the calculation process of the impact value may specifically include the following steps:

[0211] Step 51: The electronic device inputs the customer characteristic data of each to-be-selected customer in the customer data set into a third prediction function to obtain a sales method corresponding to each to-be-selected customer.

[0212] Step 52: The electronic device counts the proportion of the customers to be selected corresponding to the first sales method among all the customers to be selected, and uses the proportion as the first weight coefficient. The electronic device counts the proportion of the customers to be selected corresponding to the second sales method among all the customers to be selected, and uses the proportion as the second weight coefficient.

[0213] Step 53: The electronic device inputs the customer feature data of each to-be-selected customer in the customer data set into the first prediction model to obtain a first impact value. The electronic device inputs the customer feature data of each to-be-selected customer in the customer data set into the second prediction model to obtain a second impact value.

[0214] Step 54: The electronic device calculates a first product of the first weight coefficient and the first influence value. And the electronic device calculates a second product of the second weight coefficient and the second influence value. The electronic device calculates the sum of the first product and the second product to obtain a comprehensive influence value. Optionally,

[0215] S403. According to the comprehensive influence value of each customer, the customers to be selected are grouped into different customer groups.

[0216] In this embodiment, the electronic device can select the customer to be selected according to the comprehensive influence value to obtain two customer groups. The two customer groups correspond to two sales methods. The electronic device can also divide the customers to be selected in the two customer groups in more detail according to the threshold value to obtain multiple customer group groups. Each customer group group can correspond to a different sales frequency.

[0217] In one example, after determining a first threshold, the electronic device may divide the customers to be selected into two customer groups with different sales methods according to the first threshold.

[0218] Step 61: When the comprehensive influence value of the customer is greater than or equal to the first threshold, the customer is circled into the first customer group.

[0219] Step 62: When the influence value of the customer is less than the first threshold, the customer is circled into the second customer group.

[0220] The sales methods of the first customer group and the second customer group are different. For example, the sales method of the first customer group is the first sales method, and the sales method of the second customer group is the second sales method.

[0221] For example, the first threshold may be 0. The electronic device may select the customers to be selected whose comprehensive influence value is greater than or equal to 0 into the first customer group, and select the customers to be selected whose comprehensive influence value is less than 0 into the second customer group.

[0222] In another example, the electronic device can also divide the customer groups more finely according to the comprehensive influence value. The specific steps may include:

[0223] Step 71: Calculate the gain difference between the influence value of customers in the first customer group or the second customer group and the first threshold.

[0224] Step 72: grouping the customers to be selected in the first customer group or the second customer group according to at least one difference threshold and gain difference, and grouping the customers to be selected into a plurality of customer group groups according to the grouping; wherein different customer group groups correspond to different sales frequencies.

[0225] In this step, the electronic device can also obtain at least one preset difference threshold. The number of the difference thresholds determines the number of customer groupings. For example, when 1 difference threshold is included, the two customer groups can be further divided into 4 customer groupings. For example, when the difference threshold is h, the impact value can be divided into the following four intervals: [-1, -h], (-h, 0], (0, h], (h, 1]. For another example, when 2 differences are included, the two customer groups can be further divided into 6 customer groupings. For another example, when two difference thresholds h1 and h2 are included, assuming that h1 is less than h2, the six intervals can be: [-1, -h2], (-h2, h1], (-h1, 0], (0, h1], (-h1, h2], (h2, 1]. Each interval can correspond to a group.

[0226] Optionally, groups that are usually close to 1 and -1 will arrange a higher sales frequency. For example, the sales frequency of the [-1, -h] group can be one manual call per week, the sales frequency of the (-h, 0] group can be one manual call per month, the sales frequency of the (0, h] group can be one outbound call per month, and the sales frequency of the (h, 1] group can be one outbound call per week.

[0227] The selection method provided in the present application can obtain customer information of the customer to be selected by the electronic device, and extract features from the customer information to obtain customer feature data. The customer feature data of all the customers to be selected can constitute the customer data set. The electronic device can input the customer feature data in the customer data set into the selection model to calculate the comprehensive impact value of each customer. The electronic device can select the customer to be selected based on the comprehensive impact value to obtain multiple customer groups. In the present application, by using the selection model, the selection accuracy is improved, and the conversion success rate when using the customer group is improved.

[0228] Figure 6 A flowchart of a circle selection example provided by an embodiment of the present application is shown. Figures 2 to 5 Based on the embodiment shown. Figure 6 As shown, with the electronic device as the execution subject, the method of this embodiment may include the following steps:

[0229] S501, ESL data acquisition.

[0230] In this embodiment, ESL generally stands for "Extract, Transform, Load". The electronic device can pull a customer list and customer data of each customer in the customer list from a data warehouse. The customer data may include customer list data, customer historical behavior data, customer identity data, customer rating event data (Mean Opinion Score, MOS), etc.

[0231] S502: variable processing.

[0232] In this embodiment, the electronic device can determine the actual conversion label Y and sales method S of the customer based on the customer data. The electronic device can use the sparksql component to process the customer data, obtain more than 500 features of each customer, and form a wide feature table for the customer. In the processing of the wide feature table, it is necessary to pay attention to the time node, and obtain the features in the wide feature table at time T. Each row in the wide feature table needs to correspond to the time T. The setting of the time T can prevent feature crossing. For example, some features can be shown in Table 1.

[0233] Table 1

[0234] Is it a basic variable? Chinese name event type yes Coupon Type Coupons real time yes Coupon Status Coupons real time yes The number of days since the last release Coupons real time yes Number of push notifications from customers in the past 3 days Push push

[0235] S503, Uplift modeling.

[0236] In this embodiment, the electronic device can use a row of the data wide table as training data. The electronic device can build a circle selection model through Uplift modeling based on the training data.

[0237] The data table organized by the electronic device may include the sales method S. The electronic device may use the sales method S as a dependent variable and other variables X in the training data as independent variables to train a third prediction model. The third prediction model may be a classification model. Based on the third prediction model, the electronic device may calculate the first weight coefficient and the second weight coefficient.

[0238] The electronic device can split the training data into control group data and intervention group data according to the sales method S. The electronic device can use the control group data and the intervention group data to train the lightgbm model respectively to obtain a control group model and an intervention group model.

[0239] The electronic device can use the intervention group data and the control group model to perform counterfactual prediction. The electronic device can use the control group data and the intervention group model to perform counterfactual prediction. The electronic device can train a first prediction model τ1(x) and a second prediction model τ0(x) based on the counterfactual prediction results.

[0240] The electronic device can perform weighted summation of the first prediction model τ1(x) and the second prediction model τ0(x) according to the first weight coefficient and the second weight coefficient to obtain a selection model g(x)τ0(x)+(1-g(x))τ1(x). The selection model can be used to predict the comprehensive impact value of each customer to be selected. The comprehensive impact value can be the gain of different sales methods to the customer to be selected.

[0241] S504: Calculate gain.

[0242] S505, customer group selection.

[0243] In this embodiment, the electronic device can sort the customers to be selected according to the comprehensive influence value. The electronic device can divide the customers to be selected into 10 groups with equal frequency. The electronic device can allocate the first 5 groups to the first sales method and the last 5 groups to the second sales method.

[0244] S506, KAFKA push number.

[0245] In this embodiment, Kafka is a popular distributed stream processing platform. The electronic device can push the selected customer group to the business system through Kafka. Since the sales method of this embodiment is actually an outbound call, the business system of this embodiment can be an outbound call system for telemarketing. The outbound call system can make outbound calls based on the list of customers to be selected in the customer group.

[0246] After selecting customer group A according to the above selection method, the proportion of successfully converted customers increased by 7%, the average unit output increased by 7 yuan, and the average value of each item increased by 100 yuan.

[0247] Figure 7 A schematic diagram of the structure of a circle selection model training device provided by an embodiment of the present application is shown. Figure 7 As shown, the circle selection model training device 10 of this embodiment is used to implement the operation corresponding to the electronic device in any of the above method embodiments. The circle selection model training device 10 of this embodiment is used to:

[0248] An acquisition module 11 is used to acquire training data and process the training data to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method;

[0249] The training module 12 is used to input the intervention group data into the first basic model to train and obtain the first prediction model corresponding to the first sales method; input the control group data into the first basic model to train and obtain the second prediction model corresponding to the second sales method; input all the training data into the classification basic model to train and obtain the third prediction model for predicting the sales method; and construct a circle selection model based on the first prediction model and the second prediction model, as well as the weight coefficient generated by the third prediction model.

[0250] Optionally, the processing module 12 is used to:

[0251] Generate a first weight coefficient of the first prediction model and a second weight coefficient of the second prediction model according to the prediction result of the third prediction model on the selected customers;

[0252] The first prediction model and the second prediction model are weightedly fused according to the first weight coefficient and the second weight coefficient to construct a circled model.

[0253] Optionally, the acquisition module 11 is used to:

[0254] According to the sales methods in the training data, all the training data are grouped to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method;

[0255] Based on the intervention group data, the control group data and the second basic model, a first impact label corresponding to each intervention group data is generated;

[0256] The first impact label is added to the intervention group data to obtain the intervention group data corresponding to the first sales method after the first impact label is added.

[0257] Optionally, the acquisition module 11 is used to:

[0258] The control group data is input into the second basic model to train a control group model for predicting the conversion success rate after intervention using the second sales method;

[0259] The intervention group data is input into the trained control group model to predict the first counterfactual result corresponding to the conversion success rate of each intervention group data after the second sales method is used for intervention;

[0260] Determine a first impact label corresponding to each intervention group data according to a first actual conversion label and a first counterfactual result of each intervention group data;

[0261] The first actual conversion label is determined according to the conversion result after the intervention group data is intervened by using the first sales method.

[0262] Optionally, the acquisition module 11 is used to:

[0263] According to the sales methods in the training data, all the training data are grouped to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method;

[0264] Based on the intervention group data, the control group data and the second basic model, a second impact label corresponding to each control group data is generated;

[0265] The second impact label is added to the control group data to obtain the control group data corresponding to the second sales method after the second impact label is added.

[0266] Optionally, the acquisition module 11 is used to:

[0267] The intervention group data is input into the second basic model, and the intervention group model is trained to predict the conversion success rate after the intervention using the first sales method;

[0268] The control group data is input into the trained intervention group model to predict the second counterfactual result corresponding to the conversion success rate of each control group data after the intervention using the first sales method;

[0269] Determine a second impact label corresponding to each control group data according to the second actual conversion label of each control group data and the second counterfactual result;

[0270] The second actual conversion label is determined according to the conversion result after intervening the control group data using the second sales method.

[0271] Optionally, the acquisition module 11 is used to:

[0272] Obtain customer data for each customer and determine the sales method corresponding to the first moment in the customer data;

[0273] Obtain the conversion results after the first moment from the customer data, and generate actual conversion labels based on the conversion results;

[0274] Obtain customer information before the first moment from the customer data, and perform feature extraction on the customer information to obtain sample data;

[0275] Use sales methods, actual conversion labels, and sample data to form training data.

[0276] Optionally, the acquisition module 11 is used to:

[0277] Obtaining basic customer information before a first moment from the customer data, and performing feature extraction on the basic customer information to obtain first feature data;

[0278] Acquire customer reception information before the first moment from the customer data, and perform feature extraction on the customer reception information to obtain second feature data;

[0279] The first feature data and the second feature data are fused and extracted to obtain sample data.

[0280] The circle selection model training device 10 provided in the embodiment of the present application can execute the above method embodiment. Its specific implementation principle and technical effects can be found in the above method embodiment, and this embodiment will not be repeated here.

[0281] Figure 8 FIG. 1 shows a schematic diagram of a structure of a circle selection device provided by an embodiment of the present application. Figure 8As shown, the selection device 20 of this embodiment is used to implement the operation corresponding to the electronic device in any of the above method embodiments. The selection device 20 of this embodiment is used to:

[0282] An acquisition module 21 is used to acquire a customer data set, wherein the customer data set includes customer characteristic data of at least one customer to be selected;

[0283] The selection module 22 is used to input the customer feature data in the customer data set into Figure 2 The selection model is used to calculate the comprehensive influence value of each customer to be selected; according to the comprehensive influence value of each customer to be selected, the customers to be selected are selected into different customer groups.

[0284] Optionally, the acquisition module 21 is used to:

[0285] Acquire basic customer information before the current moment from the customer data of each to-be-selected customer in the customer data set, and perform feature extraction on the basic customer information to obtain first feature data;

[0286] Acquire customer reception information before the current moment from the customer data of the customer to be selected, and perform feature extraction on the customer reception information to obtain second feature data;

[0287] The first characteristic data and the second characteristic data are fused and extracted to obtain customer characteristic data of the customer to be selected.

[0288] Optionally, the circle selection module 22 is used to:

[0289] When the comprehensive influence value of the customer to be selected is greater than or equal to the first threshold, the customer to be selected is selected into the first customer group;

[0290] When the comprehensive influence value of the customer to be selected is less than the first threshold, the customer to be selected is selected into the second customer group;

[0291] Among them, the sales methods corresponding to the first customer group and the second customer group are different.

[0292] Optionally, the circle selection module 22 is used to:

[0293] Calculate the gain difference between the influence value of the customers to be selected in the first customer group or the second customer group and the first threshold;

[0294] According to at least one difference threshold and a gain difference, the customers to be selected in the first customer group or the second customer group are grouped, and the customers to be selected are grouped into a plurality of customer group groups according to the grouping;

[0295] Among them, different customer groups correspond to different sales frequencies.

[0296] The circle selection device 20 provided in the embodiment of the present application can execute the above method embodiment. Its specific implementation principle and technical effects can be found in the above method embodiment, and this embodiment will not be repeated here.

[0297] Fig. 9 FIG. 1 shows a hardware structure diagram of an electronic device provided in an embodiment of the present application. Fig. 9 As shown, the electronic device 30 is used to implement the operations corresponding to the electronic device in any of the above method embodiments. The electronic device 30 of this embodiment may include: a memory 31 and a processor 32.

[0298] The memory 31 is used to store computer programs. The memory 31 may include a high-speed random access memory (RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a disk or an optical disk.

[0299] The processor 32 is used to execute the computer program stored in the memory to implement the circle selection model training method or the circle selection method in the above embodiment. For details, please refer to the relevant description in the above method embodiment. The processor 32 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the invention can be directly embodied as being executed by a hardware processor, or executed by a combination of hardware and software modules in the processor.

[0300] Optionally, the memory 31 may be independent or integrated with the processor 32 .

[0301] When the memory 31 is a device independent of the processor 32, the electronic device 30 may further include a bus 33. The bus 33 is used to connect the memory 31 and the processor 32. The bus 33 may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus.

[0302] The electronic device provided in this embodiment can be used to execute the above-mentioned circle selection model training method or circle selection method, and its implementation method and technical effect are similar, which will not be repeated in this embodiment.

[0303] The present application also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, it is used to implement the methods provided in the various embodiments described above.

[0304] Among them, the computer-readable storage medium can be a computer storage medium or a communication medium. The communication medium includes any medium that facilitates the transmission of a computer program from one place to another. The computer storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer. For example, a computer-readable storage medium is coupled to a processor so that the processor can read information from the computer-readable storage medium and write information to the computer-readable storage medium. Of course, the computer-readable storage medium can also be a component of the processor. The processor and the computer-readable storage medium can be located in an application-specific integrated circuit (ASIC). In addition, the ASIC can be located in a client device. Of course, the processor and the computer-readable storage medium can also exist in a communication device as discrete components.

[0305] Specifically, the computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0306] The present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor of a device can read the computer program from the computer-readable storage medium, and at least one processor executes the computer program so that the device implements the methods provided in the various embodiments described above.

[0307] An embodiment of the present application also provides a chip, which includes a memory and a processor, the memory is used to store computer programs, and the processor is used to call and run the computer programs from the memory, so that a device equipped with the chip executes the methods in various possible implementation modes as described above.

[0308] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0309] Among them, each module can be physically separated, for example, installed in different locations of a device, or installed on different devices, or distributed on multiple network units, or distributed on multiple processors. Each module can also be integrated together, for example, installed in the same device, or integrated in a set of codes. Each module can exist in the form of hardware, or can also exist in the form of software, or can also be implemented in the form of software plus hardware. The present application can select some or all of the modules according to actual needs to achieve the purpose of the present embodiment.

[0310] When each module is implemented as an integrated module in the form of a software function module, it can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, electronic device, or network device, etc.) or a processor to perform some steps of the methods of each embodiment of the present application.

[0311] It should be understood that, although the various steps in the flowcharts in the above-described embodiments are sequentially displayed according to the indications of the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps is not strictly limited in order, and they can be executed in other orders. Moreover, at least a portion of the steps in the figure may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily to be carried out sequentially, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0312] 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 them. Although the present application has been described in detail with reference to the above embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the above embodiments can still be modified, or some or all of the technical features can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A circle selection model training method, characterized in that: The method comprises: Acquire training data, and process the training data to obtain intervention group data corresponding to the first sales method and control group data corresponding to the second sales method; Input the intervention group data into the first basic model, and train to obtain a first prediction model corresponding to the first sales method; input the control group data into the first basic model, and train to obtain a second prediction model corresponding to the second sales method; Inputting all the training data into the classification basic model to obtain a third prediction model for predicting sales methods through training; A circle selection model is constructed based on the first prediction model, the second prediction model, and the weight coefficients generated by the third prediction model.

2. The method according to claim 1, characterized in that The step of constructing a selection model based on the first prediction model, the second prediction model, and the weight coefficients generated by the third prediction model includes: Generate a first weight coefficient of the first prediction model and a second weight coefficient of the second prediction model according to the prediction result of the third prediction model for the selected customers; The first prediction model and the second prediction model are weightedly fused according to the first weight coefficient and the second weight coefficient to construct the circled model.

3. The method according to claim 1, characterized in that The processing of the training data to obtain intervention group data corresponding to the first sales method includes: According to the sales methods in the training data, the training data are grouped to obtain intervention group data corresponding to a first sales method and control group data corresponding to a second sales method; Based on the intervention group data, the control group data and the second basic model, generating a first impact label corresponding to each intervention group data; The first impact label is added to the intervention group data to obtain the intervention group data corresponding to the first sales method after the first impact label is added.

4. The method according to claim 3, characterized in that Based on the intervention group data, the control group data and the second basic model, generating a first impact label corresponding to each intervention group data includes: Inputting the control group data into a second basic model, and training to obtain a control group model for predicting the conversion success rate after intervention using a second sales method; Inputting the intervention group data into the trained control group model, predicting a first counterfactual result corresponding to the conversion success rate of each intervention group data after the intervention using the second sales method; Determine a first impact label corresponding to each of the intervention group data according to the first actual conversion label of each of the intervention group data and the first counterfactual result; The first actual conversion tag is determined according to the conversion result after intervening the intervention group data using the first sales method.

5. The method according to claim 1, characterized in that The processing of the training data to obtain control group data corresponding to the second sales method includes: According to the sales methods in the training data, the training data are grouped to obtain intervention group data corresponding to a first sales method and control group data corresponding to a second sales method; Based on the intervention group data, the control group data and the second basic model, generate a second impact label corresponding to each control group data; The second influence tag is added to the control group data to obtain the control group data corresponding to the second sales method after the second influence tag is added.

6. The method according to claim 5, characterized in that Based on the intervention group data, the control group data and the second basic model, generating a second impact label corresponding to each control group data, including: Inputting the intervention group data into a second basic model, and training to obtain an intervention group model for predicting the conversion success rate after intervention using the first sales method; Input the control group data into the trained intervention group model to predict a second counterfactual result corresponding to the conversion success rate of each control group data after the intervention using the first sales method; Determining a second impact label corresponding to each control group data according to the second actual conversion label of each control group data and the second counterfactual result; The second actual conversion tag is determined according to the conversion result after intervening the control group data using the second sales method.

7. The method according to any one of claims 1 to 6, characterized in that The obtaining of training data comprises: Obtain customer data for each customer, and determine a sales method corresponding to a first moment in the customer data; Acquire conversion results after the first moment from the customer data, and generate actual conversion labels according to the conversion results; Acquire customer information before the first moment from the customer data, and perform feature extraction on the customer information to obtain sample data; The training data is composed of the sales method, the actual conversion label and the sample data.

8. The method according to claim 7, characterized in that The acquiring the customer information before the first moment from the customer data and performing feature extraction on the customer information to obtain sample data specifically includes: Acquire basic customer information before the first moment from the customer data, and perform feature extraction on the basic customer information to obtain first feature data; Acquire customer reception information before the first moment from the customer data, and perform feature extraction on the customer reception information to obtain second feature data; The first feature data and the second feature data are fused and extracted to obtain the sample data.

9. A circle selection method, characterized in that: The method comprises: Acquire a customer data set, wherein the customer data set includes customer characteristic data of at least one customer to be selected; Inputting the customer feature data in the customer data set into the selection model described in any one of claims 1 to 8 to calculate the comprehensive impact value of each customer to be selected; According to the comprehensive influence value of each of the customers to be selected, the customers to be selected are selected into different customer groups.

10. The method according to claim 9, characterized in that The obtaining of the customer data set specifically includes: Acquire basic customer information before the current moment from the customer data of each of the to-be-selected customers in the customer data set, and perform feature extraction on the basic customer information to obtain first feature data; Acquire customer reception information before the current moment from the customer data of the customer to be selected, and perform feature extraction on the customer reception information to obtain second feature data; The first feature data and the second feature data are fused and extracted to obtain the customer feature data of the customer to be selected.

11. The method according to claim 9 or 10, characterized in that: The step of selecting the customers to be selected into different customer groups according to the comprehensive influence value of each customer to be selected specifically includes: When the comprehensive influence value of the customer to be selected is greater than or equal to a first threshold, the customer to be selected is selected into a first customer group; When the comprehensive influence value of the customer to be selected is less than the first threshold, the customer to be selected is selected into the second customer group; The sales methods corresponding to the first customer group and the second customer group are different.

12. The method according to claim 11, characterized in that The method further comprises: Calculating a gain difference between the influence value of the to-be-selected customers in the first customer group or the second customer group and the first threshold; According to at least one difference threshold and the gain difference, the customers to be selected in the first customer group or the second customer group are grouped, and the customers to be selected are grouped into a plurality of customer group groups according to the grouping; Among them, the sales frequencies corresponding to different customer groupings are different.

13. An electronic device, characterized in that: The electronic device comprises: a memory and a processor; The memory is used to store a computer program; the processor is used to implement the selection model training method as described in any one of claims 1 to 8, or the selection method as described in any one of claims 9 to 12 according to the computer program stored in the memory.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, is used to implement the selection model training method described in any one of claims 1 to 8, or the selection method described in any one of claims 9 to 12.

15. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, it implements the selection model training method described in any one of claims 1 to 8, or the selection method described in any one of claims 9 to 12.

Images