Target user determination method, program product, electronic equipment and storage medium

Abstract

The invention provides a target user determination method, a program product, electronic equipment and a storage medium. The method comprises the following steps: acquiring a feature vector of a to-be-predicted user; inputting the feature vector of the to-be-predicted user into a pre-trained user value causal model to obtain a user value increment value and a user cost increment value of the to-be-predicted user; determining the return rate of the to-be-predicted user according to the user value increment value and the user cost increment value; and determining a target user needing to be pushed according to the return rate of the to-be-predicted user. Through accurate prediction and selection of users with high return rates, marketing resources can be distributed more effectively, the resources are pushed to the users with positive influence as far as possible, and the problems of serious consumption of pushed resources and inaccurate pushed crowds are improved. And the return rate of different users can be calculated, so that the refinement degree of user management is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular, to a target user determination method, a program product, an electronic device, and a storage medium. Background Art

[0002] Push service refers to the process of actively sending service content or advertisements to users through various forms such as email, web page pop-ups and social media notifications, thereby achieving the effect of promotion. The current push solution increases the push amount to improve the push effect, such as increasing the number of pushes and the frequency of pushes, which makes users more likely to click on the push information. However, this method may cause serious consumption of push resources, and the push targets may not be the accurate ones. Summary of the invention

[0003] The purpose of the embodiments of the present application is to provide a target user determination method, program product, electronic device and storage medium to improve the problem of serious consumption of push resources and inaccurate push targets.

[0004] In the first aspect, an embodiment of the present application provides a method for determining a target user, including: obtaining a feature vector of a user to be predicted; the feature vector is used to represent basic information of the user to be predicted; the feature vector of the user to be predicted is input into a pre-trained user value causal model to obtain an incremental user value value and an incremental user cost value of the user to be predicted; the incremental user value value represents the change in revenue brought about by a case where the user to be predicted receives a push compared to a case where the user does not receive a push; the incremental cost value of the user to be predicted is used to represent the change in user cost when the user to be predicted receives a push compared to a case where the user does not receive a push; the rate of return of the user to be predicted is determined based on the incremental user value value and the incremental user cost value; and the target user to be pushed is determined based on the rate of return of the user to be predicted.

[0005] In the above implementation process, by inputting the feature vector of the user to be predicted into the pre-trained user value causal model, the incremental value of the user value and the incremental value of the user cost of the user to be predicted are obtained, and the rate of return of the user to be predicted is determined. By accurately predicting and selecting users with high rates of return, marketing resources can be allocated more effectively, and resources can be pushed to users with positive impact as much as possible, improving the problem of serious consumption of push resources and inaccurate push population. In addition, the rate of return can be calculated for different users, improving the refinement of user management.

[0006] Optionally, in an embodiment of the present application, the user value causal model includes a user value model; before inputting the feature vector of the user to be predicted into a pre-trained user value causal model to be predicted, the method also includes: obtaining intervention variables, sample feature vectors and result variables of sample users; wherein the intervention variables of the sample users include sending push messages and not sending push messages; the sample feature vectors are used to characterize the characteristic information of the sample users; the result variables include sample user values; the sample user values ​​are obtained by statistically analyzing the benefits brought by the sample users under the intervention variables; the intervention variables, sample feature vectors and sample user values ​​of the sample users are input into a pre-defined user value initial model, the user value initial model is trained to obtain the user value model; the user value initial model is a causal forest dual machine learning model.

[0007] In the above implementation process, the user value model can more accurately predict the impact of intervention variables (whether to send push messages to sample users) on user value, identify which user groups are most likely to gain value from push messages, and thus optimize resource allocation. And by calculating the value of sample users, the accuracy of push messages can be improved, the interference to users can be reduced, and the user experience can be improved.

[0008] Optionally, in an embodiment of the present application, the user value causal model also includes a user cost model; the outcome variable of the sample user also includes a sample user cost value; before inputting the feature vector of the user to be predicted into a pre-trained user value causal model to be predicted, the method also includes: inputting the intervention variable, sample feature vector and sample user cost value of the sample user into a pre-defined user cost initial model, training the user cost initial model, and obtaining the user cost model.

[0009] In the above implementation process, by training the user cost model, the costs of different user groups can be predicted more accurately, providing data support for cost control and budget planning. The user cost model can identify user groups with higher cost-effectiveness, optimize resource allocation, and improve the accuracy of target users.

[0010] Optionally, in an embodiment of the present application, after obtaining the intervention variables, sample feature vectors and result variables of the sample users, the method further includes: accumulating the result variables of the sample users according to a preset period to obtain accumulated result variables; performing data preprocessing on the sample feature vectors to obtain preprocessed feature vectors; the data preprocessing includes performing at least one of missing value processing, outlier processing and / or standardization processing on the sample feature vectors; the inputting the intervention variables, sample feature vectors and sample user values ​​of the sample users into a predefined user value initial model includes: inputting the intervention variables of the sample users, the preprocessed feature vectors and the accumulated result variables into a predefined user value initial model.

[0011] In the above implementation process, by accumulating the result variables, the impact of random fluctuations can be reduced and the stability and reliability of the prediction can be improved. Preprocessing feature vectors helps to improve the generalization ability of the model, so that the model performs better on new data, thereby improving the accuracy of model predictions, making the prediction of user value and user cost (user value increment and user cost increment) more accurate, and further optimizing the allocation of push resources.

[0012] Optionally, in an embodiment of the present application, training the initial user value model to obtain the user value model includes: using the user value model to predict a pre-acquired test set to obtain the incremental value of the test user in the test set; the test set includes a first test user and a second test user; the first test user intervention variable is a user who sends a push message; the second test user intervention variable is a user who does not send a push message; generating a first gain curve based on the incremental value of the first test user; generating a second gain curve based on the incremental value of the second test user; obtaining a gain coefficient based on the first gain curve and the second gain curve; during the training process, adjusting the initial user value model according to the gain coefficient to obtain the user value model.

[0013] In the above implementation process, the gain coefficient and gain curve provide a quantitative and visual tool for evaluating and optimizing the user value model and the user cost model. The gain coefficient can be used to better understand and improve the performance and parameters of the model.

[0014] Optionally, in an embodiment of the present application, determining the rate of return of the user to be predicted based on the user value increment value and the user cost increment value includes: determining the rate of return of the user to be predicted by the ratio of the user value increment value to the user cost increment value; or, obtaining the difference between the user value increment value and the user cost increment value, and determining the rate of return of the user to be predicted by the ratio of the difference to the user cost increment value; or, adjusting the user value increment value by using a preset risk factor to obtain an adjusted user value increment value; and determining the rate of return of the user to be predicted by the ratio of the adjusted user value increment value to the user cost increment value.

[0015] In the above implementation process, multiple methods of calculating the return rate of the predicted users are provided to evaluate the effect of push from different angles, which increases the flexibility and comprehensiveness of the evaluation. The calculation method of the return rate can be determined according to business needs, the purpose of calculation, and the aspects emphasized. In the process of determining the target users, both the incremental revenue and the cost are considered. The target users are not determined only by the incremental revenue, which improves the problem of low resource utilization.

[0016] Optionally, in an embodiment of the present application, determining the target users to be pushed based on the return rates of the users to be predicted includes: determining the users to be predicted whose return rates exceed a preset threshold as the target users to be pushed; or determining the target users to be pushed based on the ranking of the return rates of the users to be predicted.

[0017] In the above implementation process, by selecting users whose return rate exceeds the preset threshold for push, resource utilization is improved. And through the sorting method, target users are flexibly selected according to available resources and business goals, resource allocation is optimized to ensure that resources are used on users who are most likely to generate high returns.

[0018] In the second aspect, an embodiment of the present application also provides a target user determination device, including: a feature vector acquisition module, used to acquire a feature vector of a user to be predicted; the feature vector is used to represent basic information of the user to be predicted; a value and cost calculation module, used to input the feature vector of the user to be predicted into a pre-trained user value causal model, and obtain a user value increment and a user cost increment of the user to be predicted; the user value increment represents the change in the benefit brought by the user to be predicted when receiving a push compared with the case where the user does not receive a push; the user cost increment is used to represent the change in user cost when the user to be predicted receives a push compared with the case where the user does not receive a push; a rate of return calculation module, used to determine the rate of return of the user to be predicted based on the user value increment and the user cost increment; a target user determination module, used to determine the target user to be pushed based on the rate of return of the user to be predicted.

[0019] In a third aspect, an embodiment of the present application further provides a computer program product, including computer program instructions, which, when executed by a processor, execute the method provided by the first aspect or any one of the implementations of the first aspect.

[0020] In a fourth aspect, an embodiment of the present application further provides an electronic device, comprising: a processor and a memory, the memory storing computer program instructions, and the computer program instructions, when executed by the processor, execute the method provided by the first aspect or any one of the implementations of the first aspect.

[0021] In a fifth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method provided by the first aspect or any one of the implementations of the first aspect is executed.

[0022] By adopting a target user determination method, program product, electronic device and storage medium provided by the present application, the feature vector of the user to be predicted is input into a pre-trained user value causal model to obtain the incremental user value and incremental user cost of the user to be predicted, and the rate of return of the user to be predicted is determined. By accurately predicting and selecting users with high rates of return, marketing resources can be allocated more effectively, and resources can be pushed to users with positive impact as much as possible, improving the problem of serious consumption of push resources and inaccurate push population. In addition, the rate of return can be calculated for different users, improving the refinement of user management. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 A flowchart of a method for determining a target user provided in an embodiment of the present application;

[0025] Figure 2 A schematic diagram of the structure of a target user determination device provided in an embodiment of the present application;

[0026] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following embodiments of the technical solution of the present application are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present application, and are therefore only used as examples, and cannot be used to limit the scope of protection of the present application.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0029] In the description of the embodiments of the present application, the technical terms "first", "second", etc. are only used to distinguish different objects, and cannot be understood as indicating or implying relative importance or implicitly indicating the number, specific order or primary and secondary relationship of the indicated technical features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0030] Smart push service can determine the user's preference for the product or page based on the user's clicks, searches or purchases. If the number of clicks, searches or purchases reaches a certain threshold, it is determined that the user likes the product or page, and push service is provided to the user, such as pushing more products of the same category to the user, thereby achieving the purpose of promotion. In order to further improve the effect of promotion, the push intensity will be increased, such as increasing the number of pushes, increasing the push frequency, etc., to increase the probability of users clicking on push messages, thereby increasing the user's purchasing power.

[0031] However, the above method does not take the push cost into consideration. For example, the more push times are performed, the more computer resources and material resources are required, and the push operations need to be performed more frequently. Even if the purpose of the push is finally achieved, the push cost required in the process has exceeded the value of the product itself, resulting in a waste of resources.

[0032] The embodiment of the present application obtains the incremental value of the user value and the incremental value of the user cost of the user to be predicted by inputting the feature vector of the user to be predicted into the pre-trained user value causal model, determines the rate of return of the user to be predicted, and can more effectively allocate marketing resources by accurately predicting and selecting users with high rates of return, so that resources are pushed to users with positive impact as much as possible, improving the problem of serious consumption of push resources and inaccurate push population. In addition, the rate of return can be calculated for different users, improving the refinement of user management.

[0033] See also Figure 1 A flow chart of a target user determination method provided by an embodiment of the present application is shown. The target user determination method provided by an embodiment of the present application can be applied to an electronic device, which may include a physical device such as a server, a PC, a tablet computer, or a smart phone, or a virtual device such as a virtual machine or a container. The electronic device may be a single device, or a combination of multiple devices or a cluster of a large number of devices. The target user determination method may include:

[0034] Step S110: Acquire a feature vector of the user to be predicted; the feature vector is used to represent basic information of the user to be predicted.

[0035] Step S120: Input the feature vector of the user to be predicted into the pre-trained user value causal model to obtain the incremental user value value and incremental user cost value of the user to be predicted; the incremental user value value represents the change in the benefit brought by the case where the user to be predicted receives push notifications compared to the case where the user does not receive push notifications; the incremental user cost value is used to represent the change in user cost when the user to be predicted receives push notifications compared to the case where the user does not receive push notifications.

[0036] Step S130: Determine the rate of return of the user to be predicted according to the incremental value of the user value and the incremental value of the user cost.

[0037] Step S140: Determine the target user to be pushed according to the return rate of the user to be predicted.

[0038] In step S110, the users to be predicted, such as users registered in a certain APP (Application), or users who click on a certain webpage, or users who use a certain application, can all be objects to be pushed. However, in order to improve the accuracy of the push and save push resources, the users to be predicted can be screened or calculated before pushing, and the target users can be determined from the users to be predicted. The target users are the users who are determined to be pushed.

[0039] The feature vector of the user to be predicted is used to represent the basic information of the user to be predicted. For example, the feature vector can be a mathematical representation of the basic information of the user to be predicted, which is used to input into the machine learning model. The mathematical representation can be the original data itself, or it can be the data after preprocessing or conversion. The feature vector of the user to be predicted can be generated, for example, after the user's consent and in compliance with laws and regulations, by collecting the basic information of the user to be predicted. The basic information may include at least one of age, gender, income, consumption history, browsing behavior, etc. The collected basic information is organized into a structured format, such as an array or a list. This structure is called a feature vector. Exemplarily, the user feature vector includes three features: age, gender, and income. Then this feature vector can be expressed as: feature vector = [age, gender, income]. This step can be performed for each user to be predicted to obtain its corresponding feature vector, so as to perform subsequent predictions of user value increment and user cost increment.

[0040] In step S120, the user value causal model is used to obtain the incremental user value and incremental user cost of the user to be predicted based on the user's feature vector. The incremental user value can be understood as the user UE (where UE refers to UnitEconomics, a single economic model, representing the net profit obtained by the platform after a single user consumes on the platform), and the incremental user UE refers to the change in the revenue brought by the user to be predicted when the user receives the push compared to the case where the push is not received. The revenue here can be economic revenue, such as the net profit brought by the user to be predicted, or other revenue, such as clicks, forwardings, and playbacks. The incremental user cost refers to the change in user cost when the user to be predicted receives the push compared to the case where the push is not received, and the cost required for performance push. The cost can be understood as the economic cost used for the push, or it can be cost data converted from resource cost, time cost, or computing cost.

[0041] The user value causal model can be a general model, which includes two branches, one branch is used to calculate the incremental value of the user value of the user to be predicted, and the other branch is used to calculate the incremental value of the user cost of the user to be predicted. The user value causal model can also be multiple independent models, one model is used to calculate the incremental value of the user value of the user to be predicted, and the other model is used to calculate the incremental value of the user cost of the user to be predicted.

[0042] For a single model, dual-branch solution, the model will learn how to predict two outputs based on the input feature vector during training. This method only requires one model inference process to obtain two prediction results (user value increment and user cost increment), which improves computational efficiency. In addition, the two branches share the same input features (feature vectors of the user to be predicted), which can improve the consistency between the predicted user value increment and user cost increment.

[0043] In the case where the user value causal model is two independent models, the feature vectors of the user to be predicted are input into the two models respectively, and each model only focuses on its specific prediction task (one model predicts the incremental value of the user value of the user to be predicted, and the other model is used to calculate the incremental value of the user cost of the user to be predicted). And each model is trained independently to learn how to predict its corresponding output based on the input feature vector. The advantage of this approach is that each model can be optimized for its specific prediction task, which may improve the accuracy of the prediction, and each model can focus on capturing the features and patterns most relevant to its prediction task.

[0044] The user value causal model uses historical data during training. The historical data includes the user's sample feature vectors, the benefits brought by the user when accepting the push, and the benefits brought by the user when not accepting the push. The historical data also includes the user's information, as well as the user cost when the user accepts the push, and the user cost when the user does not accept the push. The model learns the relationship between the sample feature vector and the incremental value of the user value, and the relationship between the sample feature vector and the incremental value of the user, and the model also learns how the push affects the user. The training process of the two models will be explained later.

[0045] In step S130, the rate of return of the user to be predicted is determined according to the incremental value of user value and the incremental value of user cost. The rate of return of the user to be predicted may be the UE-ROI (user return on investment) of the user to be predicted. The rate of return of the user to be predicted is used to characterize the ratio or difference between the revenue generated after the push for the user to be predicted and the push cost paid. The greater the rate of return of the user to be predicted, the greater the revenue or value brought by the user to be predicted.

[0046] Methods for calculating the rate of return of the user to be predicted include, for example, calculating the ratio or difference between the incremental value of user value and the incremental value of user cost, or adjusting the incremental value of user value using risk factors, which will be introduced later.

[0047] In step S140, after calculating the return rate of the user to be predicted, the target user is determined from the users to be predicted based on the return rate of the user to be predicted. For example, the user to be predicted whose return rate exceeds a preset threshold can be determined as the target user to be pushed. The top several users to be predicted can also be determined as the target users to be pushed based on the ranking of the return rates of the users to be predicted.

[0048] In the above implementation process, by inputting the feature vector of the user to be predicted into the pre-trained user value causal model, the incremental value of the user value and the incremental value of the user cost of the user to be predicted are obtained, and the rate of return of the user to be predicted is determined. By accurately predicting and selecting users with high rates of return, marketing resources can be allocated more effectively, and resources can be pushed to users with positive impact as much as possible, improving the problem of serious consumption of push resources and inaccurate push population. In addition, the rate of return can be calculated for different users, improving the refinement of user management.

[0049] Optionally, in an embodiment of the present application, the user value causal model includes a user value model; the user value model is used to obtain the user cost increment value of the user to be predicted. Before inputting the feature vector of the user to be predicted into the pre-trained user value causal model to be predicted, the user value model needs to be trained first, and the steps of training the user value model include:

[0050] Obtain the sample user's intervention variables, sample feature vectors and result variables; wherein, the sample user's intervention variables include sending push messages and not sending push messages; the sample feature vector is used to characterize the characteristic information of the sample user; the result variable includes the sample user value; the sample user value is obtained by statistically analyzing the benefits brought by the sample user under the intervention variables.

[0051] As an implementation method, sample users may be first obtained and divided into an experimental group and a control group, wherein the users in the experimental group are users who sent push messages, and the intervention variable of the users in the experimental group is set to sending push messages; the users in the control group are users who did not send push messages, and the intervention variable of the users in the control group is set to not sending push messages.

[0052] Then, a sample feature vector of the sample user is obtained. The sample feature vector includes basic information such as age, gender, income, occupation, consumption history, browsing behavior, etc. of the sample user. The collected basic information can be organized into a structured format, such as an array or list, as a sample feature vector in the manner of step S110.

[0053] It is also necessary to collect the user's result variables, including sample user value, which refers to the benefits brought by the sample users; the sample user value is obtained by counting the benefits brought by the sample users under the intervention variables (sending push messages and not sending push messages). For example, if the intervention variable is not sending push messages, the benefits brought by the sample users without sending push messages are counted as the sample user value; if some pre-variables are sending push messages, the benefits brought by the sample users when sending push messages are counted as the sample user value.

[0054] Afterwards, the intervention variables, sample feature vectors, and outcome variables of the sample users can be cleaned, and missing values ​​and outliers can be processed to improve the quality of the training data. Features can be preprocessed, such as encoding categorical variables and standardizing or normalizing numerical variables.

[0055] Define the model structure: Select the causal forest dual machine learning model as the initial user value model. The causal forest dual machine learning model is a method that combines causal inference and machine learning techniques and is suitable for estimating the causal effects of interventions. It is also necessary to build a model framework, including defining the model input (sample feature vectors of sample users and intervention variables), output (sample user value), and model parameters and structure.

[0056] The sample user's intervention variables, sample feature vectors and sample user values ​​are input into a pre-defined user value initial model, and the user value initial model is trained to obtain a user value model; the user value initial model is a causal forest dual machine learning model.

[0057] The above data is used to train the initial user value model. During the training process, the model parameters are adjusted to optimize the performance, such as the number of forest trees, tree depth, learning rate, etc., based on the performance of the model on the training set. During the training process, the model will learn how to predict the sample user value based on the sample feature vector and intervention variables of the input sample user.

[0058] In the implementation process of the above embodiment: the user value model can more accurately predict the impact of the intervention variable (whether to send push messages to sample users) on user value, identify which user groups are most likely to gain value from push messages, and thus optimize resource allocation. And by calculating the value of sample users, the accuracy of push messages can be improved, the interference to users can be reduced, and the user experience can be improved.

[0059] Optionally, in an embodiment of the present application, the user value causal model also includes a user cost model; the user cost model is used to obtain the incremental user cost value of the user to be predicted.

[0060] Before inputting the feature vector of the user to be predicted into the pre-trained causal model of the user value to be predicted, the user cost model needs to be trained first. The process of training the user cost model is as follows:

[0061] The result variable of the sample user also includes the sample user cost value, which refers to the push cost used to send a push message to the sample user. It can be understood that if no push message is sent to the sample user (that is, the intervention variable of the sample user is no push message), then the user cost of the sample user can be ignored or assigned a fixed value. The sample user cost can be obtained through statistical calculation or calling an interface.

[0062] When training the user cost model, it is still necessary to obtain the sample feature vectors and intervention variables of the sample users. The method of obtaining them is as mentioned above.

[0063] For the model structure, the causal forest dual machine learning model can be selected as the initial model of user value. And build a model framework, including defining the model input (sample feature vector and intervention variables of sample users), output (sample user cost), and parameters and structure of the model. It should be noted that the model architecture of the user cost model and the user value model can be the same or different, and the input of the user cost model and the user value model can also be the same or different, but the prediction targets of the two are different, so the outputs are different. And the two models can be trained independently, each model can have its own objective function and optimization target. For example, the optimization target of the user value model is to minimize the prediction error of user value, while the optimization target of the user cost model is to minimize the prediction error of user cost. In actual operation, because the two models can be independent of each other's results, the two models can be trained in parallel to improve training efficiency.

[0064] The sample user's intervention variables, sample feature vectors, and sample user cost values ​​are input into the pre-defined user cost initial model, and the user cost initial model is trained to obtain the user cost model. During the training process, the user cost initial model learns how to predict the sample user cost based on the input sample feature vectors and intervention variables. After that, the trained user cost model can be evaluated and verified, and the model can be iteratively optimized based on the evaluation results to obtain the final user cost model.

[0065] In the implementation process of the above embodiment: by training the user cost model, the costs of different user groups can be more accurately predicted, providing data support for cost control and budget planning. The user cost model can identify user groups with higher cost-effectiveness, optimize resource allocation, and improve the accuracy of target users.

[0066] Optionally, in the embodiment of the present application, after obtaining the intervention variables, sample feature vectors, and result variables of the sample users, the method further includes:

[0067] The result variables of the sample users are accumulated according to a preset period to obtain an accumulated result variable. The preset period can be daily, weekly or monthly, etc., and this application does not limit this. Through the above accumulation operation, short-term fluctuations can be smoothed, making the result variable more stable and reliable. The result variable can be the sample user value or the sample user cost value.

[0068] Data preprocessing is performed on the sample feature vector to obtain a preprocessed feature vector; data preprocessing includes performing at least one of missing value processing, outlier processing and / or standardization processing on the sample feature vector, wherein missing value processing refers to identifying and processing missing values ​​in the sample feature vector, and the method of missing value processing may include deleting missing values, filling missing values ​​(using mean, median or mode, etc.). Outlier value processing refers to identifying and processing outliers in the sample feature vector, and the method of outlier value processing may include deleting outliers, limiting outliers or using statistical methods (such as IQR) for processing. Standardization processing refers to standardizing the sample feature vector, such as Z-score standardization. The data after standardization will have the characteristics of a standard normal distribution, such as a mean of 0 and a standard deviation of 1. Standardization processing helps the model learn and generalize better.

[0069] Exemplarily, the formula Z=(X-μ) / σ can be used for Z-score standardization. Where X represents the sample feature vector (such as age, income or consumption history, etc.), μ represents the mean of the feature (for example, if the sample characteristic of age is Z-score standardized, μ is the mean of the age of the sample users), and σ represents the standard deviation of the feature (following the above embodiment, it can be the standard deviation of the age of the sample users). For the calculation of the mean and standard deviation, it can be the mean or standard deviation of the sample users as a whole, or it can be the mean or standard deviation of the sample users after the intervention variable grouping, for example, the mean or standard deviation is calculated for the experimental group and the control group respectively.

[0070] The sample user's intervention variables, sample feature vectors, and sample user values ​​are input into the pre-defined user value initial model, including:

[0071] The sample user's intervention variables, preprocessed feature vectors, and accumulated result variables are input into a predefined user value initial model. As an implementation method, the preprocessed feature vectors, intervention variables, and accumulated result variables can be integrated together to form an input data set for the model. The data set is then input into a predefined user value initial model for training. The training process is described above.

[0072] It is understandable that in the process of training the user cost model, the result variables can also be accumulated and the sample feature vectors can be preprocessed to improve the generalization ability of the user cost model and make the model perform better on new data.

[0073] In the implementation process of the above embodiment: by accumulating the result variables, the influence of random fluctuations can be reduced, and the stability and reliability of the prediction can be improved. Preprocessing the feature vector helps to improve the generalization ability of the model, so that the model performs better on new data, thereby improving the accuracy of the model prediction, making the prediction of user value and user cost (user value increment and user cost increment) more accurate, and further optimizing the allocation of push resources.

[0074] Since the model has taken into account the intervention variable (i.e., whether the user receives the push message) during training, the user value model can learn the relationship between the user feature vector and the user value increment under different intervention conditions, and the user cost model can learn the relationship between the user feature vector and the user cost increment under different intervention conditions. In the process of predicting the user to be predicted, the user feature vector (such as age, gender, income, etc.) can be input to predict the user value increment and user cost increment.

[0075] Optionally, in the embodiment of the present application, training the initial user value model to obtain the user value model includes:

[0076] The user value model is used to predict the pre-acquired test set to obtain the incremental value of the test users in the test set; the test set includes a first test user and a second test user; the first test user intervention variable is the user who sends a push message; the second test user intervention variable is the user who does not send a push message. The test set obtained here can be a part pre-divided from the sample data, or it can be new data.

[0077] The test set is input into the user value model to obtain the incremental value of the test users in the test set. According to the incremental value, it is determined which test users are considered as "responders", for example, users whose incremental value of the test users exceeds a preset threshold are considered as responders.

[0078] A first gain curve is generated according to the incremental value of the first test user; the first gain curve may be a Qini curve, which shows the positive impact of push messages on user value. For example, on the Qini curve, the horizontal axis represents the proportion of users ranked from high to low according to the incremental value, and the vertical axis represents the proportion of cumulative responders.

[0079] According to the incremental value of the second test user, a second gain curve is generated; the second gain curve can also be a Qini curve, which shows the natural change of user value without push messages. Similarly, on the Qini curve, the horizontal axis represents the proportion of users ranked from high to low according to the incremental value, and the vertical axis represents the proportion of cumulative responders.

[0080] A gain coefficient is obtained according to the first gain curve and the second gain curve; the gain coefficient may be the ratio of the areas under the first gain curve and the second gain curve, the difference between specific points in the first gain curve and the second gain curve, or other statistical measures; the larger the value of the gain coefficient, the better the effect of the model in predicting the increment of user value.

[0081] During the training process, the initial user value model is adjusted according to the gain coefficient to obtain the user value model. The gain coefficient is used as feedback to adjust the initial user value model. This may involve adjusting the parameters, structure, or training strategy of the model to better capture the impact of push messages on user value. The model can also be optimized according to the gain coefficient to improve its prediction accuracy and causal inference ability. This may include feature engineering, model selection, or hyperparameter tuning.

[0082] In an optional embodiment, during the process of training the user cost model, the performance of the user cost model may be evaluated and the model may be optimized by drawing a first gain curve, a second gain curve, and calculating a gain coefficient.

[0083] In the implementation process of the above embodiment: the gain coefficient and the gain curve provide a quantitative and visual tool for evaluating and optimizing the user value model and the user cost model, and the performance and parameters of the model can be better understood and improved through the gain coefficient.

[0084] Optionally, in the embodiment of the present application, the rate of return of the user to be predicted is determined according to the incremental value of user value and the incremental value of user cost, including the following methods:

[0085] The first method is to determine the rate of return of the predicted user as the ratio of the incremental value of user value to the incremental value of user cost. This method directly compares the value generated per unit cost, such as the revenue that can be generated by each push, and intuitively reflects the benefits.

[0086] The second method is to obtain the difference between the incremental value of user value and the incremental value of user cost, and determine the ratio of the difference to the incremental value of user cost as the rate of return of the user to be predicted. If the result is positive, it means that the benefits brought by the investment exceed the cost; if it is negative, it means that the cost exceeds the benefits. This method takes into account the net benefit of the incremental value of value exceeding the incremental value of cost, and provides a net rate of return indicator.

[0087] The third method: Use the preset risk factor to adjust the user value increment to obtain the adjusted user value increment; determine the ratio of the adjusted user value increment to the user cost increment as the rate of return for the user to be predicted. For example, the adjusted user value increment = user value increment × risk factor, where the risk factor can be defined and calculated according to the specific situation. This method takes into account the impact of risk factors on expected value and is suitable for risk-sensitive decision-making environments.

[0088] In the implementation process of the above embodiment: multiple methods of calculating the rate of return of the predicted users are provided, and the effect of the push is evaluated from different angles, which increases the flexibility and comprehensiveness of the evaluation. The method of calculating the rate of return can be determined according to business needs, the purpose of the calculation, and the aspects emphasized. In the process of determining the target users, both the incremental revenue and the cost are considered, and the target users are not determined only by the incremental revenue, so as to improve the problem of low resource utilization.

[0089] Optionally, in the embodiment of the present application, determining the target user to be pushed according to the return rate of the user to be predicted includes:

[0090] The predicted users whose return rate exceeds the preset threshold are identified as the target users to be pushed. A preset threshold for the return rate can be set based on business goals and historical data. The return rate of the predicted user is compared with the preset threshold, and the predicted users whose return rate exceeds the preset threshold are identified as the target users to be pushed. This method is simple and direct, easy to implement, and can quickly identify the user group with the highest expected return.

[0091] Alternatively, the target users to be pushed are determined based on the ranking of the return rates of the users to be predicted. The return rates of all the users to be predicted are ranked. Based on business requirements and resource constraints, the top N% of users with the highest return rates are selected as target users. This method can more flexibly select target users based on the distribution of return rates and business goals.

[0092] After determining the target users, you can implement push strategies for the determined target user groups, monitor the push effects, and collect feedback data for further optimization of models and strategies.

[0093] In the implementation process of the above embodiment: by selecting users whose return rate exceeds a preset threshold for push, resource utilization is improved. And by using a sorting method, target users are flexibly selected according to available resources and business goals, resource allocation is optimized, and resources are ensured to be used on users who are most likely to generate high returns.

[0094] In an optional embodiment, during the process of training the causal model of the user value to be predicted, a causal effect estimate can also be calculated. The causal effect estimate is used to characterize the causal effect of the intervention variable on the outcome variable. The larger the causal effect estimate, the greater the impact of the intervention measure (such as paid push) on the outcome variable (such as user value UE). The intervention measure has a greater impact on user behavior or outcome variables, and this impact is significant and unlikely to be caused by random changes. It should be noted that the causal effect estimate can be positive or negative: Positive causal effect estimate: indicates that the intervention measure has a positive impact on the outcome variable, such as increasing the value or activity of the user. Negative causal effect estimate: indicates that the intervention measure has a negative impact on the outcome variable, which may be due to the intervention measure causing user dissatisfaction or increased costs.

[0095] A large causal effect estimate may indicate that without such intervention, the user's value or behavior may be significantly reduced, which emphasizes the necessity of intervention. Therefore, the target users are determined based on the causal effect estimate, and users whose causal effect estimate exceeds a preset value can be identified as target users.

[0096] The method of calculating the causal effect estimate includes: obtaining the difference between the outcome variable and the incremental value of user value, recorded as the first difference; obtaining the difference between the outcome variable and the incremental value of user cost, recorded as the second difference, and taking the ratio of the first difference to the second difference as the causal effect estimate.

[0097] See also Figure 2 The structure diagram of the target user determination device provided in the embodiment of the present application is shown; the embodiment of the present application provides a target user determination device 200, including:

[0098] A feature vector acquisition module 210 is used to acquire a feature vector of the user to be predicted; the feature vector is used to represent basic information of the user to be predicted;

[0099] The value and cost calculation module 220 is used to input the feature vector of the user to be predicted into a pre-trained user value causal model to obtain the user value increment value and user cost increment value of the user to be predicted; the user value increment value represents the change in the benefit brought by the user to be predicted when the user receives the push notification compared with the case where the user does not receive the push notification; the user cost increment value is used to represent the change in the user cost when the user to be predicted receives the push notification compared with the case where the user does not receive the push notification;

[0100] A rate of return calculation module 230, configured to determine the rate of return of the user to be predicted according to the user value increment value and the user cost increment value;

[0101] The target user determination module 240 is used to determine the target user to be pushed according to the return rate of the user to be predicted.

[0102] Optionally, in an embodiment of the present application, the user value causal model includes a user value model; the target user determination device 200 also includes a value model training module, which is used to obtain intervention variables, sample feature vectors and result variables of sample users; wherein the intervention variables of the sample users include sending push messages and not sending push messages; the sample feature vectors are used to characterize the characteristic information of the sample users; the result variables include sample user values; the sample user values ​​are obtained by counting the benefits brought by the sample users under the intervention variables; the intervention variables, sample feature vectors and sample user values ​​of the sample users are input into a pre-defined user value initial model, the user value initial model is trained to obtain the user value model; the user value initial model is a causal forest dual machine learning model.

[0103] Optionally, in an embodiment of the present application, the user value causal model also includes a user cost model; the result variable of the sample user also includes a sample user cost value; the target user determination device 200 also includes a cost model training module, which is used to input the intervention variables, sample feature vectors and sample user cost values ​​of the sample users into a predefined user cost initial model, train the user cost initial model, and obtain the user cost model.

[0104] Optionally, in an embodiment of the present application, the target user determination device 200 further includes a data preprocessing module, which is used to accumulate the result variables of the sample users according to a preset period to obtain accumulated result variables; perform data preprocessing on the sample feature vector to obtain a preprocessed feature vector; the data preprocessing includes performing at least one of missing value processing, outlier processing and / or standardization processing on the sample feature vector; the inputting the intervention variables, sample feature vectors and sample user values ​​of the sample users into a predefined user value initial model includes: inputting the intervention variables of the sample users, the preprocessed feature vectors and the accumulated result variables into a predefined user value initial model.

[0105] Optionally, in an embodiment of the present application, the target user determination device 200, the value model training module or the cost model training module, is also used to use the user value model to predict the pre-acquired test set to obtain the incremental value of the test user in the test set; the test set includes a first test user and a second test user; the first test user intervention variable is the user who sends a push message; the second test user intervention variable is the user who does not send a push message; based on the incremental value of the first test user, a first gain curve is generated; based on the incremental value of the second test user, a second gain curve is generated; based on the first gain curve and the second gain curve, a gain coefficient is obtained; during the training process, the initial user value model is adjusted according to the gain coefficient to obtain the user value model.

[0106] Optionally, in an embodiment of the present application, the target user determination device 200 and the rate of return calculation module 230 are further used to determine the ratio of the user value increment value to the user cost increment value as the rate of return of the user to be predicted; or, obtain the difference between the user value increment value and the user cost increment value, and determine the ratio of the difference to the user cost increment value as the rate of return of the user to be predicted; or, adjust the user value increment value using a preset risk factor to obtain an adjusted user value increment value; and determine the ratio of the adjusted user value increment value to the user cost increment value as the rate of return of the user to be predicted.

[0107] Optionally, in an embodiment of the present application, the target user determination device 200 and the target user determination module 240 are further used to determine the to-be-predicted user whose return rate exceeds a preset threshold as the target user that needs to be pushed; or, determine the target user that needs to be pushed based on the ranking of the return rates of the to-be-predicted users.

[0108] It should be understood that the device corresponds to the above-mentioned target user determination method embodiment and can execute the various steps involved in the above-mentioned method embodiment. The specific functions of the device can be found in the above description. To avoid repetition, the detailed description is appropriately omitted here. The device includes at least one software function module that can be stored in a memory in the form of software or firmware or fixed in the operating system (OS) of the device.

[0109] See also Figure 3 The electronic device 300 provided in the embodiment of the present application includes: a processor 310 and a memory 320, wherein the memory 320 stores machine-readable instructions executable by the processor 310, and when the machine-readable instructions are executed by the processor 310, the above method is executed.

[0110] Figure 3 Each component shown in can be implemented by hardware, software or a combination thereof. The electronic device 300 may be a physical device, such as a server, a PC, etc., or a virtual device, such as a virtual machine, a virtualized container, etc. Moreover, the electronic device 300 is not limited to a single device, but may also be a combination of multiple devices or a cluster consisting of a large number of devices.

[0111] An embodiment of the present application further provides a storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above method is executed.

[0112] Among them, the 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 red-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, disk or optical disk.

[0113] In several embodiments provided by the embodiments of the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and a part of the module, program segment or code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0114] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part.

[0115] The above description is only an optional implementation manner of the embodiments of the present application, but the protection scope of the embodiments of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in the embodiments of the present application, which should be covered within the protection scope of the embodiments of the present application.

Claims

1. A method for determining a target user, characterized in that: include: Obtain the feature vector of the user to be predicted; The feature vector is used to represent the basic information of the user to be predicted; Input the feature vector of the user to be predicted into a pre-trained user value causal model to obtain the incremental user value value and incremental user cost value of the user to be predicted; the incremental user value value represents the change in the benefit brought by the user to be predicted when the user receives the push notification compared with the case where the user does not receive the push notification; the incremental user cost value is used to represent the change in the user cost when the user to be predicted receives the push notification compared with the case where the user does not receive the push notification; Determining the rate of return of the user to be predicted according to the user value increment value and the user cost increment value; The target users to be pushed are determined according to the return rate of the users to be predicted.

2. The method according to claim 1, characterized in that: The user value causal model includes a user value model; before inputting the feature vector of the user to be predicted into the pre-trained user value causal model to be predicted, the method further includes: Obtaining intervention variables, sample feature vectors, and result variables of sample users; wherein the intervention variables of the sample users include sending push messages and not sending push messages; the sample feature vectors are used to characterize the feature information of the sample users; the result variables include sample user values; the sample user values ​​are obtained by counting the benefits brought by the sample users under the intervention variables; The intervention variables, sample feature vectors and sample user values ​​of the sample users are input into a predefined user value initial model, and the user value initial model is trained to obtain the user value model; the user value initial model is a causal forest dual machine learning model.

3. The method according to claim 2, characterized in that The user value causal model also includes a user cost model; the result variable of the sample user also includes a sample user cost value; before inputting the feature vector of the user to be predicted into the pre-trained user value causal model to be predicted, the method also includes: The intervention variables, sample feature vectors and sample user cost values ​​of the sample users are input into a predefined user cost initial model, and the user cost initial model is trained to obtain the user cost model.

4. The method according to claim 2, characterized in that: After obtaining the intervention variables, sample feature vectors, and result variables of the sample users, the method further includes: Accumulate the result variables of the sample users according to a preset period to obtain an accumulated result variable; Performing data preprocessing on the sample feature vector to obtain a preprocessed feature vector; the data preprocessing includes performing at least one of missing value processing, outlier processing and / or standardization processing on the sample feature vector; The step of inputting the intervention variables, sample feature vectors and sample user values ​​of the sample users into a predefined user value initial model includes: The intervention variables of the sample users, the preprocessed feature vectors and the accumulated result variables are input into a predefined user value initial model.

5. The method according to claim 2 or 3, characterized in that: The training of the initial user value model to obtain the user value model includes: The user value model is used to predict the pre-acquired test set to obtain the incremental value of the test user in the test set; the test set includes a first test user and a second test user; the first test user intervention variable is the user who sends the push message; the second test user intervention variable is the user who does not send the push message; Generate a first gain curve according to the incremental value of the first test user; generate a second gain curve according to the incremental value of the second test user; Obtaining a gain coefficient according to the first gain curve and the second gain curve; During the training process, the initial user value model is adjusted according to the gain coefficient to obtain the user value model.

6. The method according to claim 1, characterized in that The determining the rate of return of the user to be predicted according to the user value increment value and the user cost increment value includes: The ratio of the user value increment to the user cost increment is determined as the rate of return of the user to be predicted; Alternatively, the difference between the user value increment value and the user cost increment value is obtained, and the ratio of the difference to the user cost increment value is determined as the rate of return of the user to be predicted; Alternatively, the user value increment value is adjusted using a preset risk factor to obtain an adjusted user value increment value; and the ratio of the adjusted user value increment value to the user cost increment value is determined as the rate of return of the user to be predicted.

7. The method according to any one of claims 1 to 4, characterized in that: The step of determining the target user to be pushed according to the return rate of the user to be predicted includes: Determine the predicted user whose return rate exceeds a preset threshold as the target user to be pushed; Alternatively, the target users to be pushed are determined according to the ranking of the return rates of the users to be predicted.

8. A computer program product, characterized in that The method comprises computer program instructions, and when the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is executed.

9. An electronic device, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is executed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is executed.

Images