User selection method, apparatus, device, storage medium, and program product

Abstract

The application provides a user selection method and device, electronic equipment, computer readable storage medium and computer program product, comprising: dividing users into a first user group and a second user group based on the intention degree of the users for a first event, the intention degree of a first user in the first user group being greater than or equal to an intention degree threshold, the intention degree of a second user in the second user group being less than the intention degree threshold; generating a first user instance based on a first feature of the first user in the first user group, and generating a second user instance based on a second feature of the second user in the second user group; determining a first influence degree of the first feature on the intention degree based on the first user instance, and determining a second influence degree of the second feature on the intention degree based on the second user instance; and selecting a third user with an influence degree greater than an influence degree threshold from the second user group based on the first influence degree and the second influence degree. In this way, potential users can be selected from users with low intention degrees.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a user selection method, apparatus, electronic device, computer-readable storage medium, and computer program product. Background Technology

[0002] In related technologies, machine learning models are widely used in various user segmentation and decision-making processes. For example, machine learning models can be used to predict users' purchasing behavior and obtain users' purchase intention propensity scores. Users whose purchase intention propensity scores exceed a threshold are designated as target users, while all other users are designated as non-target users. Summary of the Invention

[0003] This application provides a user selection method, apparatus, electronic device, and computer-readable storage medium that can select potential users from users with low intent.

[0004] The technical solution of this application embodiment is implemented as follows: This application provides a user selection method, the method including: Based on the user's intention for the first event, the user is divided into a first user group and a second user group. The intention of the first user in the first user group is greater than or equal to the intention threshold, and the intention of the second user in the second user group is less than the intention threshold. A first user instance is generated based on the first feature of the first user in the first user group, and a second user instance is generated based on the second feature of the second user in the second user group; Based on the first user instance, determine the first degree of influence of the first feature on the intention, and based on the second user instance, determine the second degree of influence of the second feature on the intention. Based on the first degree of influence and the second degree of influence, a third user whose degree of influence is greater than the influence threshold is selected from the second user group.

[0005] This application provides a user selection device, the device comprising: The grouping module is used to divide the users into a first user group and a second user group based on the users' intention to the first event. The intention of the first user in the first user group is greater than or equal to the intention threshold, and the intention of the second user in the second user group is less than the intention threshold. The generation module is used to generate a first user instance based on a first feature of a first user in the first user group, and to generate a second user instance based on a second feature of a second user in the second user group; The determining module is configured to determine, based on the first user instance, a first degree of influence of the first feature on the intention, and based on the second user instance, a second degree of influence of the second feature on the intention; The selection module is used to select a third user from the second user group whose influence level is greater than the influence level threshold, based on the first influence level and the second influence level.

[0006] This application provides an electronic device, including: Memory is used to store executable instructions or computer programs. When a processor executes computer-executable instructions or computer programs stored in the memory, it implements the user selection method provided in the embodiments of this application.

[0007] This application provides a computer-readable storage medium storing computer-executable instructions or computer programs. When the computer-executable instructions or computer programs are executed by a processor, the processor will execute the user selection method provided in this application.

[0008] This application provides a computer program product including a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions or computer program from the computer-readable storage medium and executes the computer-executable instructions or computer program, causing the electronic device to perform the user selection method provided in this application.

[0009] The embodiments of this application have the following beneficial effects: After grouping users based on their intention to a first event into a first user group with an intention greater than or equal to an intention threshold and a second user group with an intention less than the intention threshold, a first user instance is determined based on a first feature, and a second user instance is determined based on a second feature. Then, based on the first user instance, the first degree of influence of the first feature on the intention is determined, and based on the second user instance, the second degree of influence of the second feature on the intention is determined. Based on the first and second degree of influence, a third user with an influence greater than the influence threshold is selected from the second user group. In this way, compared with the scheme that selects users based solely on intention, generating multiple user instances based on the features of different users enriches the diversity of features, making it easier to determine the degree of influence of features on intention. For users with low intention, users are selected from among those with low intention based on the degree of influence. This allows for user selection based on the degree of influence of features on intention, making the user selection process more in-depth and comprehensive, avoiding the omission of users who should be selected, and achieving the effect of selecting potential users from among those with low intention. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the architecture of the user selection system 100 provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application; Figure 3 This is a flowchart illustrating the user selection method provided in an embodiment of this application; Figure 4 This is a schematic diagram of the model structure of the intention model provided in the embodiments of this application; Figure 5 This is a flowchart illustrating the process of determining the first feature and the second feature provided in an embodiment of this application; Figure 6 This is a schematic diagram of the model structure of the discrimination model provided in the embodiments of this application; Figure 7 This is a schematic diagram of the process for determining the first degree of influence provided in an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0013] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0014] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0015] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0016] 1) Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or computers-controlled machines to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. AI technology is a comprehensive discipline involving a wide range of fields, encompassing both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics.

[0017] 2) Client, also known as user terminal, refers to the program that provides local services to users in contrast to the server. Except for some applications that can only run locally, it is generally installed on ordinary client machines and needs to cooperate with the server to run. That is, there needs to be a corresponding server and service program in the network to provide the corresponding services. Thus, a specific communication connection needs to be established between the client and the server to ensure the normal operation of the application.

[0018] During the research process, the inventors discovered that when promoting to users, the relevant technology first predicts the user's intention level, then uses users with high intention level as target users and users with low intention level as non-target users, thus promoting only to target users. However, during the prediction process, some features have a high degree of influence on the prediction results. This means that the model only uses these important features to predict users. However, users have other features. Although other features have a lower impact on the prediction results, they may affect the actual behavior of users. Therefore, when predicting users as non-target users based on important features, other non-important features are not fully analyzed. As a result, potential users may also exist among non-target users. Thus, the target user selection scheme in related technologies may cause the omission of target users and make it impossible to select potential users from non-target users.

[0019] Based on this, embodiments of this application provide a user selection method. For users with low intent, i.e., non-target users, after determining the degree of influence of user characteristics on intent, users are selected from users with low intent based on the degree of influence. In this way, the selection of users with low intent can be combined with the degree of influence of characteristics on intent, making the user selection process more in-depth and comprehensive, avoiding the omission of users who should be selected, and thus enabling the selection of potential users from users with low intent.

[0020] See Figure 1 , Figure 1 This is a schematic diagram of the architecture of the user selection system 100 provided in this application embodiment. In order to realize the application scenario of user selection, the terminal (terminal 400 is shown as an example) is connected to the server 200 through the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two. The terminal 400 is used for the user to use the client 401 and display on the display interface (display interface 401-1 is shown as an example). The terminal 400 and the server 200 are connected to each other through wired or wireless networks.

[0021] The terminal 400 is used to send the user's intention regarding the first event to the server 200. Server 200 is configured to: receive user intent for a first event; divide users into a first user group and a second user group based on user intent for the first event, wherein the intent of the first user in the first user group is greater than or equal to an intent threshold, and the intent of the second user in the second user group is less than an intent threshold; generate a first user instance based on a first feature of the first user in the first user group, and generate a second user instance based on a second feature of the second user in the second user group; determine a first degree of influence of the first feature on intent based on the first user instance, and determine a second degree of influence of the second feature on intent based on the second user instance; and select a third user from the second user group whose influence is greater than an influence threshold based on the first and second degree of influence.

[0022] In some embodiments, server 200 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Terminal 400 can be a smartphone, tablet, laptop, desktop computer, set-top box, smart voice interaction device, smart home appliance, vehicle terminal, aircraft, and mobile devices (e.g., mobile phones, portable music players, personal digital assistants, dedicated messaging devices, portable gaming devices, smart speakers, and smartwatches), but is not limited to these. Terminal devices and servers can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment.

[0023] See Figure 2 , Figure 2 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. In practical applications, the electronic device can be... Figure 1 The server 200 or terminal 400 shown are described in the following document. Figure 2 , Figure 2 The illustrated electronic device includes at least one processor 410, a memory 450, at least one network interface 420, and a user interface 430. The various components in terminal 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to implement communication between these components. In addition to a data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 2 The general labeled all buses as Bus System 440.

[0024] Processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0025] User interface 430 includes one or more output devices 431 that enable the presentation of media content, including one or more speakers and / or one or more visual displays. User interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0026] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices physically located away from the processor 410.

[0027] The memory 450 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 450 described in this application embodiment is intended to include any suitable type of memory.

[0028] In some embodiments, memory 450 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0029] Operating system 451 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc. Presentation module 453 is configured to enable the presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 431 associated with user interface 430 (e.g., a display screen, a speaker, etc.). The input processing module 454 is used to detect and translate one or more user inputs or interactions from one or more input devices 432.

[0030] In some embodiments, the apparatus provided in this application can be implemented in software. Figure 2 A user selection device 455 stored in memory 450 is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: a grouping module 4551, a generation module 4552, a determination module 4553, and a selection module 4554. These modules are logically linked and can therefore be arbitrarily combined or further separated according to their implemented functions. The functions of each module will be described below.

[0031] In other embodiments, the apparatus provided in this application can be implemented in hardware. As an example, the user selection apparatus provided in this application can be a processor in the form of a hardware decoding processor, which is programmed to execute the user selection method provided in this application. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0032] In some embodiments, the terminal or server can implement the user selection method provided in this application embodiment by running a computer program. For example, the computer program can be a native program or software module in an operating system; it can be a native application (APP), that is, a program that needs to be installed in the operating system to run, such as an instant messaging APP or a web browser APP; it can also be a mini-program, that is, a program that only needs to be downloaded into the browser environment to run; or it can be a mini-program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module or plugin.

[0033] Based on the above description of the user selection system and electronic device provided in the embodiments of this application, the user selection method provided in the embodiments of this application is described below. In actual implementation, the user selection method provided in the embodiments of this application can be implemented by the terminal or the server alone, or by the terminal and the server working together, so that... Figure 1 The following description uses the example of server 200 executing the user selection method provided in this embodiment of the application independently. See also... Figure 3 , Figure 3 This is a flowchart illustrating the user selection method provided in the embodiments of this application. Below, it will be discussed in conjunction with... Figure 3 The steps shown are explained.

[0034] Step 101: Based on the user's intention for the first event, the server divides the user into a first user group and a second user group. The intention of the first user in the first user group is greater than or equal to the intention threshold, and the intention of the second user in the second user group is less than the intention threshold.

[0035] In actual implementation, the first event could be a user purchasing a product or downloading an app, etc. This application does not limit this; the user's intention for the first event refers to the probability that the user will perform the first event. If the user's intention for the first event is 1, that is, the probability that the user will perform the first event is 1, it means that the user will definitely perform the first event. If the user's intention for the first event is 0, that is, the probability that the user will perform the first event is 0, it means that the user will definitely not perform the first event. Among these, the intentions of multiple users regarding the first event can be predicted using a trained intention model; specifically, see [link to relevant documentation]. Figure 4 , Figure 4 This is a schematic diagram of the model structure of the intention model provided in the embodiments of this application, based on Figure 4 The intention model includes a feature extraction layer and an intention prediction layer. The process of obtaining a user's intention for the first event can be as follows: For each user, the user's attribute information is input into the intention model. The feature extraction layer of the intention model extracts user features based on the user's attribute information, obtaining user features. The intention prediction layer of the intention model predicts the user's intention for the first event based on the user features, obtaining the predicted user intention for the first event. The user's attribute information can include the user's age, gender, occupation, number of likes in the past 30 days, number of searches in the past 30 days, etc.

[0036] In actual implementation, after determining the user's intention for the first event, multiple users are grouped based on each user's intention and intention threshold to obtain a first user group and a second user group. Specifically, for each user, the user's intention is compared with the intention threshold. When the comparison result indicates that the user's intention is greater than or equal to the intention threshold, the user is designated as the first user; when the comparison result indicates that the user's intention is less than the intention threshold, the user is designated as the second user. The intention threshold is preset, and this embodiment does not limit it.

[0037] Step 102: Generate a first user instance based on the first feature of the first user in the first user group, and generate a second user instance based on the second feature of the second user in the second user group.

[0038] In practice, a first user instance is generated based on the first characteristic of the first user in the first user group. That is, multiple user instances are generated based on the first characteristics of different first users. Before generating the first user instance based on the first characteristic of the first user in the first user group, please refer to [link to previous section]. Figure 5 , Figure 5 This is a flowchart illustrating the determination of the first feature and the second feature provided in an embodiment of this application, based on... Figure 5 Before generating a first user instance based on the first feature of the first user in the first user group, the first feature and the second feature can be determined through the following steps.

[0039] Step 201: The server performs an importance analysis on the user's features to determine the importance of each feature.

[0040] It should be noted that the importance analysis of user features here refers to the importance analysis of the features of each user. The importance of each feature refers to the impact of the corresponding feature on the prediction result when predicting whether the user's intention exceeds the intention threshold. For example, if the confidence of the prediction result decreases significantly when predicting whether the user's intention exceeds the intention threshold based on other features that do not include a certain feature, it means that the feature is very important to the prediction process of whether the user's intention exceeds the intention threshold, that is, it is highly important, such as exceeding the pre-set importance threshold.

[0041] In practice, the process of performing importance analysis on user features to obtain the importance of each feature can be as follows: For each of the user's multiple features, perform the following processing to obtain the importance of that feature: Remove the corresponding feature from the user's multiple features to obtain a third feature; Based on the third feature, predict the probability that the user's intention exceeds the intention threshold; Based on the probability, determine the importance of the feature.

[0042] It should be noted that for each feature, the user's multiple features are removed to obtain a third feature. Here, the third feature refers to all features other than the first one. Then, based on the third feature, the probability that the user's intention exceeds the intention threshold is predicted. The process of determining the importance of a feature based on the probability can be as follows: determine the predicted user's intention for the first event, and determine the user's label based on whether the user's intention exceeds the intention threshold. The absolute value of the difference between the predicted probability and the user's label is taken as the importance of the corresponding feature.

[0043] It should be noted that the user's label is used to indicate whether the user's intent exceeds the intent threshold, that is, to indicate the probability that the user's intent exceeds the intent threshold. Since the user's intent is already determined, whether the user's intent exceeds the intent threshold is also determined, that is, the probability that the user's intent exceeds the intent threshold is also determined. Therefore, the user's label is 1 or 0. When the user's intent exceeds the intent threshold, that is, the probability that the user's intent exceeds the intent threshold is 1, the user's label is 1; when the user's intent does not exceed the intent threshold, that is, the probability that the user's intent exceeds the intent threshold is 0, the user's label is 0.

[0044] Based on this, if the probability of predicting a user's intention exceeds the intention threshold is greater than the absolute value of the difference between the predicted probability and the user's label, the prediction result is less accurate. In other words, if the prediction result is based on features other than a certain feature, the prediction result is less accurate (equivalent to the prediction result is less accurate if the prediction is not based on that feature). Therefore, the importance of the corresponding feature is higher.

[0045] Thus, after removing features and using the remaining features to make predictions, the importance of features is determined based on the difference between the predicted and actual results. This quantifies the impact of feature removal on the model's prediction accuracy, which in turn quantifies the importance of features. This makes it easier to rank the importance of features and thus identify important features.

[0046] In practice, the process of analyzing the importance of user features and determining the importance of each feature is achieved through a discriminant model. This discriminant model determines whether a user's intent exceeds an intent threshold, i.e., it determines the probability that a user's intent exceeds the intent threshold. Furthermore, the discriminant model shares the same variables as the intent model described earlier, meaning it processes the same type of features. For example, see [link to relevant documentation]. Figure 6 , Figure 6 This is a schematic diagram of the model structure of the discrimination model provided in the embodiments of this application, based on Figure 6 The discriminative model includes a feature extraction layer and a prediction layer. The process of determining the probability that a user's intent exceeds an intent threshold can be as follows: For each user, the user's attribute information is input into the discriminative model. The feature extraction layer of the discriminative model extracts user features based on the user's attribute information, obtaining user features. The prediction layer of the discriminative model then predicts the probability that the user's intent exceeds the intent threshold based on the user features, thus obtaining the probability that the user's intent exceeds the intent threshold. The user's attribute information can include the user's age, gender, occupation, number of likes in the past 30 days, number of searches in the past 30 days, etc.

[0047] It should be noted that after removing a feature, the same discriminant model is retrained using all the remaining features. Then, the performance changes of the two discriminant models before and after removing the feature are compared. If the model performance drops significantly after removing the corresponding feature, it indicates that the feature is very important for the model's prediction.

[0048] Step 202: Based on the importance of each feature, select features from the user's features whose importance is greater than or equal to the importance threshold as important features.

[0049] It should be noted that the importance threshold here can be set, and this application embodiment does not limit it; for example, after determining the importance of each feature, multiple features are sorted based on their importance, and based on the sorting results, the target number of features, such as the top 20%, is selected. Here, the importance threshold can be the minimum importance among the features corresponding to the target number of features.

[0050] It should be noted that the important features here refer to important feature types. Therefore, for different users, the important features are the same, that is, the important feature types are the same, but the values ​​of the important features are different, that is, the feature values ​​under the important feature types are different. For example, the user's feature types include the user's age, gender, occupation, number of likes in the past 30 days, and number of searches in the past 30 days. Here, the important features can be the number of likes and the number of searches in the past 30 days, and the corresponding feature values ​​are the specific number of likes and the specific number of searches in the past 30 days.

[0051] Step 203: From the important features, determine the first feature of the first user and the second feature of the second user.

[0052] It should be noted that since important features are selected based on user characteristics, determining the first feature of the first user and the second feature of the second user from the important features is equivalent to taking the important features of the first user as the first feature and the important features of the second user as the second feature. At the same time, the number of first features can be one or more, and the number of second features can also be one or more. Following the example above, the user's feature types include the user's age, gender, occupation, number of likes in the past 30 days, and number of searches in the past 30 days. Here, the important features can be the number of likes and searches in the past 30 days. Therefore, the first feature of the first user is the number of likes and searches in the past 30 days. The second feature of the second user is the number of likes and searches in the past 30 days.

[0053] In this way, by selecting important features from multiple user characteristics, and then using these important features to conduct subsequent user selection processes, not only can the efficiency of the subsequent user selection process be improved, but the accuracy of the selected users can also be improved.

[0054] In actual implementation, there are several ways to generate a first user instance based on the first feature of the first user in the first user group. Here, we will take a few of them as examples to explain the process of generating a first user instance based on the first feature of the first user in the first user group.

[0055] In some embodiments, there are multiple first features, and the feature types of different first features are different; thus, the process of generating a first user instance based on the first features of a first user in a first user group may be as follows: for the first features of a first user, perform multiple selection processes to obtain multiple combined features; wherein, for each selection process, perform the following operations to obtain a combined feature: for each feature type, determine the first feature of each first user under the feature type, and select one first feature from the determined multiple first features; combine the first features selected under each feature type to obtain a combined feature; wherein the feature type corresponding to the combined feature is the same as the feature type of the first feature; and determine the first user instance based on the multiple combined features.

[0056] It should be noted that there can be one or more first features. When there are multiple first features, the feature types corresponding to different first features are different. For each feature type, the first feature of each first user under the feature type is determined, thereby determining the feature set under each feature type. One feature type corresponds to one feature set, and one feature set includes the feature values ​​of all first users under this feature type. Then, in each selection process, a first feature is randomly selected from each feature set, thereby using the selected multiple first features as combined features. At the same time, since multiple selection processes are performed, multiple combined features are obtained. The feature type corresponding to the combined feature is the same as the feature type of the first feature, which means that the number of feature types corresponding to the combined feature is the same as the number of feature types of the first feature, that is, the number of features included in the combined feature is the same as the number of first features.

[0057] For example, the feature types of the first feature include the number of likes and searches a user has within 30 days. There are three first users: first user A has 20 likes and 18 searches within 30 days; first user B has 1 like and 1 search within 30 days; and first user C has 5 likes and 0 searches within 30 days. For each feature type, the first feature for each first user under that feature type is determined, meaning the number of likes within 30 days includes 20, 1, and 5, and the number of searches within 30 days is 18, 1, and 0. Then, from the multiple first features determined under different feature types, one first feature is selected, and the selected feature under each feature type is... Combining the first feature yields nine combined features: {20 likes in 30 days, 18 searches in 30 days}, {20 likes in 30 days, 1 search in 30 days}, {20 likes in 30 days, 0 searches in 30 days}, {1 like in 30 days, 1 search in 30 days}, {1 like in 30 days, 18 searches in 30 days}, {1 like in 30 days, 0 searches in 30 days}, {5 likes in 30 days, 18 searches in 30 days}, {5 likes in 30 days, 1 search in 30 days}, and {5 likes in 30 days, 0 searches in 30 days}.

[0058] It should be noted that the number of selections can be preset, as long as it does not exceed P. n This can be done in one step, where P is the number of first users and n is the number of first features.

[0059] In practice, for the process of determining the first user instance based on multiple combined features, since the combined features are generated based on the first feature, and the first user includes other features in addition to the first feature, for each first user, starting from the first combined feature among the multiple combined features, the first feature of the first user is replaced with the corresponding combined feature in turn to complete one replacement. Based on the feature formed after the replacement (which is equivalent to the concatenation feature of the corresponding combined feature and the other features mentioned above), a first user instance is determined. In this way, the number of replacement processes is performed as many times as there are combined features, thus obtaining multiple first user instances. In this context, for each first user, the number of combined features is the same as the number of corresponding first user instances. As for the process of determining the first user instance based on multiple combined features, the first user instance determined here includes the first user instances corresponding to each of the multiple first users. For example, if k first user instances are generated based on k combined features for each first user, then the process of determining the first user instance based on multiple combined features is to take the k first user instances corresponding to each of the multiple first users as the first user instance determined above, where k is an integer greater than 1.

[0060] It should be noted that a user instance is not an actual user, but a virtual user, which is equivalent to indicating a new set of features. For example, as mentioned above, for each first user, starting from the first combined feature among multiple combined features, the first feature of the first user is replaced with the corresponding combined feature in turn, thereby obtaining multiple first user instances. Since there are multiple combined features, after replacing the first feature respectively, one combined feature and the remaining features will form a new set of features, and this new set of features will be used as a user instance.

[0061] Following the example above, the user's feature types include age, gender, occupation, number of likes in the past 30 days, and number of searches in the past 30 days. The first feature is the number of likes in the past 30 days and the number of searches in the past 30 days. There are 9 combined features. There are 3 users. Therefore, for user A, user A's age (18), gender (female), and occupation (student) are combined with these 9 combined features to form 9 new feature groups: {age=18, gender=female, occupation=student, number of likes in the past 30 days=20, number of searches in the past 30 days=18}, {age=18, gender=female, occupation=student, number of likes in the past 30 days=20, number of searches in the past 30 days=1}, {age=18, gender=female, occupation=student, number of likes in the past 30 days=20, number of searches in the past 30 days=0 ... 8. {Gender=Female, Occupation=Student, 30-day likes=1, 30-day search count=1}, {Age=18, Gender=Female, Occupation=Student, 30-day likes=1, 30-day search count=18}, {Age=18, Gender=Female, Occupation=Student, 30-day likes=1, 30-day search count=0}, {Age=18, Gender=Female, Occupation=Student, 30-day likes=5, 30-day search count=18}, {Age=18, Gender=Female, Occupation=Student, 30-day likes=5, 30-day search count=18}, {Age=18, Gender=Female, Occupation=Student, 30-day likes=5, 30-day search count=1}, {Age=18, Gender=Female, Occupation=Student, 30-day likes=5, 30-day search count=0}, thus these 9 new features are used as 9 first user instances corresponding to the first user A; Each first user can correspond to 9 first user instances. Since there are 3 first users here, there are a total of 27 first user instances. In other words, the number of first user instances generated based on the first characteristics of different first users is 27.

[0062] Similarly, the process of generating a second user instance based on the second feature of the second user in the second user group is similar to the process of generating a first user instance based on the first feature of the first user in the first user group described above. Here, the embodiments of this application will not be repeated.

[0063] In this way, a feature is randomly selected from different feature types, and the selected features are combined to obtain combined features. Based on the combined features, multiple user instances are generated. By combining different types of features, the diversity of user features is increased, which makes it easier to determine the degree of influence of features on intent based on user instances in subsequent processes.

[0064] In other embodiments, the number of first features is n, where n is a positive integer. The process of generating a first user instance based on the first features of the first user in the first user group can be as follows: for each first user, perform k update processes on the first features of the first user to obtain k first user instances; wherein, for each update process, perform the following operations to obtain a first user instance: update m first features of the first user to m new first features to obtain multiple first user instances, where m is a positive integer less than or equal to n, and k is an integer greater than 1.

[0065] Here, updating the first feature refers to updating the feature value of the first feature. The m corresponding to different update processes can be different or the same. When the m corresponding to different update processes is the same, the feature value of the first feature of the first user instance obtained by different update processes is different, which is equivalent to the new first feature obtained by different update processes being different. At the same time, when the m corresponding to different update processes is the same, m and n are equal. When the m corresponding to different update processes is not the same, in multiple update processes, the m corresponding to some update processes can be the same, while the m corresponding to other update processes is different. This application embodiment does not limit this.

[0066] For example, consider a first user with the characteristics {age=18, gender=female, occupation=student, number of likes in 30 days=20, number of searches in 30 days=18}. The first characteristics are age, number of likes in 30 days, and number of searches in 30 days, i.e., n=3. Here, if m is the same, the three first characteristics of the first user need to be updated, that is, the age, number of likes in 30 days, and number of searches in 30 days need to be updated respectively. For example, the first user's characteristics {age=18, gender=female, occupation=student, number of likes in 30 days=20, number of searches in 30 days=18} are updated to {age=20, gender=female, occupation=student, number of likes in 30 days=18, number of searches in 30 days=1}. This update process is performed k times to obtain k new characteristics corresponding to the first user, i.e., k first user instances. If m is not the same, then in the first update process, one of the first features of the first user can be updated, that is, the number of likes of the first user in the past 30 days can be updated separately. For example, if the first user's features are {age=18, gender=female, occupation=student, number of likes in the past 30 days=20, number of searches in the past 30 days=18}, it can be updated to {age=18, gender=female, occupation=student, number of likes in the past 30 days=18, number of searches in the past 30 days=18}. In the second update process, two of the first user's first features can be updated, that is, the number of likes and the number of searches in the past 30 days can be updated separately. For example, if the first user's features are {age=18, gender=female, occupation=student, number of likes in the past 30 days=20, number of searches in the past 30 days=18}, it can be updated to {age=18, gender=female, occupation=student, number of likes in the past 30 days=18, number of searches in the past 30 days=18}. The first user's initial characteristics are updated from {age=18, gender=female, occupation=student, likes=18 in 30 days, search count=1 in 30 days} to {age=18, gender=female, occupation=student, likes=20 in 30 days, search count=18 in 30 days}. In the third update, the three primary characteristics of the first user are updated: age, likes in 30 days, and search count in 30 days. For example, if the first user's characteristics are {age=18, gender=female, occupation=student, likes=20 in 30 days, search count=18}, they are updated to {age=20, gender=female, occupation=student, likes=18 in 30 days, search count=1 in 30 days}. This update process is performed k times to obtain k new sets of characteristics corresponding to the first user, resulting in k instances of the first user.

[0067] At the same time, each first user can obtain k first user instances. For each first user, after obtaining k first user instances, the k first user instances corresponding to each of the multiple first users are used as the first user instances determined above.

[0068] Similarly, the process of generating a second user instance based on the second feature of the second user in the second user group is similar to the process of generating a first user instance based on the first feature of the first user in the first user group described above. Here, the embodiments of this application will not be repeated.

[0069] In this way, for each first user, the first feature of the first user is updated multiple times, thereby generating multiple user instances. By updating the features, the diversity of user features is increased, which makes it easier to determine the degree of influence of features on intent based on user instances in subsequent processes.

[0070] Step 103: Based on the first user instance, determine the first degree of influence of the first feature on the intention, and based on the second user instance, determine the second degree of influence of the second feature on the intention.

[0071] It should be noted that the degree of influence of a feature on intention refers to the relevance of that feature to the user's decision (i.e., whether the user performs the first action) when determining the user's intention, which is the weight of the corresponding feature when predicting intention. For example, the higher the degree of influence of a feature on intention, the greater the impact of feature changes on the prediction result, which means that the prediction result of intention depends more on the corresponding feature.

[0072] In practice, multiple first user instances can be generated for each first user. These multiple first user instances are the multiple first user instances corresponding to each first user, that is, the set of user instances generated based on the multiple first user instances corresponding to each first user. Based on the first user instances, the degree of influence of the first feature on the intention is determined. In other words, based on the multiple first user instances corresponding to each first user, the degree of influence of the first feature of the corresponding first user on the intention of that first user is determined. Here, each first user corresponds to a first degree of influence, that is, there is a one-to-one correspondence between the first user and the first degree of influence. Similarly, multiple second user instances can be generated for each second user. These multiple second user instances are the multiple second user instances corresponding to each second user, that is, the set of user instances generated based on the multiple second user instances corresponding to each second user. Based on these multiple second user instances, the second influence degree of the second feature on the intention is determined. In other words, based on the multiple second user instances corresponding to each second user, the influence degree of the second feature of the corresponding second user on the intention of that second user is determined. Here, each second user corresponds to a second influence degree, that is, there is a one-to-one correspondence between the second user and the second influence degree.

[0073] In actual implementation, see Figure 7 , Figure 7This is a flowchart illustrating the determination of the first degree of influence provided in an embodiment of this application, based on... Figure 7 The process of determining the degree of influence of the first feature on the intention based on the first user instance, where there are multiple first user instances, can be achieved through the following steps.

[0074] Step 1031: For each first user instance, determine the intention degree of the first user instance for the first event.

[0075] In actual implementation, as mentioned above, the intention of the first user instance for the first event is also the probability of the first user instance executing the first event. The intention of the first user instance for the first event can be predicted by the trained intention model. This will not be elaborated on in the embodiments of this application.

[0076] Step 1032: Average the intention of multiple first user instances to obtain the average intention of multiple first user instances.

[0077] In practice, the process of averaging the intentions of multiple first user instances to obtain the average intention of multiple first user instances can be as follows: sum the intentions of multiple first user instances to obtain the sum of intentions; and use the ratio of the sum of intentions to the number of first user instances as the average intention.

[0078] Step 1033: Determine the degree of dispersion of the intention of multiple first user instances relative to the average intention, and use the degree of dispersion as the first degree of influence.

[0079] In practice, determining the dispersion of intent among multiple first user instances relative to the average intent can be achieved by: subtracting the intent of each first user instance from the average intent to obtain an intent difference value; summing the squares of the intent differences for each first user instance to obtain a sum of intent values; determining the ratio of the sum of intent values ​​to the number of first user instances, and then taking the square root of the ratio to obtain the dispersion, i.e.: ...Formula (1); in, To represent the degree of dispersion, n is the number of first user instances. Let be the intent level of the i-th first user instance. This represents the average level of interest.

[0080] It should be noted that since the first features of different first user instances are different, that is, the feature values ​​of the first features of different first user instances are different, if the dispersion of the intention degree determined based on different first features relative to the average intention degree is very high, it means that the difference in the intention degree determined based on different first features will be large. In other words, the difference in the first feature has a great impact on the intention degree, and thus the degree of influence of the first feature on the intention degree will also be great. It is equivalent to reflecting the magnitude of the influence of the first feature on the intention degree by the fluctuation of the intention degree based on the difference in the first feature. Therefore, the dispersion of the intention degree of multiple first user instances relative to the average intention degree can be regarded as the first degree of influence.

[0081] Similarly, the process of determining the second degree of influence of the second feature on the intention based on the second user instance may involve: for each second user instance, determining the intention of the second user instance for the first event; averaging the intentions of multiple second user instances to obtain the average intention of the multiple second user instances; determining the dispersion of the intentions of the multiple second user instances relative to the corresponding average intention, and using this dispersion as the second degree of influence; wherein, the process of determining the dispersion of the intentions of the multiple second user instances relative to the average intention is similar to the process of determining the dispersion of the intentions of the multiple first user instances relative to the average intention described above, and will not be elaborated further in this embodiment.

[0082] In this way, the influence of features on intention is determined by the fluctuations in intention caused by different features, and the influence of features on intention is quantified, thereby improving the accuracy of determining the influence of corresponding features on intention.

[0083] Step 104: Based on the first degree of influence and the second degree of influence, select a third user from the second user group whose degree of influence is greater than the influence threshold.

[0084] In practice, as mentioned above, there is a one-to-one correspondence between the first degree of influence and the first user, and a one-to-one correspondence between the second degree of influence and the second user. Therefore, the process of selecting a third user whose degree of influence is greater than the influence threshold from the second user group based on the first degree of influence and the second degree of influence can be as follows: sum the first degree of influence of each first user to obtain the sum of influence; determine the ratio of the sum of influence to the number of first users, and determine the influence threshold based on the ratio; select the second user whose second degree of influence is greater than the influence threshold from the second users as the third user.

[0085] It should be noted that the process of determining the impact threshold based on the ratio can be achieved by multiplying the ratio by a pre-set first coefficient to obtain the impact threshold, i.e.: ...Formula (2); in, It is the threshold of influence, where Q is the number of first users. It is the degree of first influence of the i-th first user. It is a pre-set first coefficient, which is generally between 0 and 1, and can be 0.7 in this case.

[0086] In practice, after determining the impact threshold, the second impact level of each second user is compared with the impact threshold to obtain the comparison result. From multiple comparison results, the comparison result with the second impact level greater than the impact threshold is selected, and the second user corresponding to the selected comparison result is taken as the third user.

[0087] It should be noted that if the second user's second influence is greater, it means that the second user's intention is more significantly affected by the fluctuation of the important feature, i.e., the second feature. In other words, the other remaining features (i.e., features other than the second feature) have a relatively weak determining effect on the second user's intention. At the same time, the influence threshold is determined based on the first user's first influence. Therefore, the influence threshold can be considered as the average influence of the important feature, i.e., the first feature, on the first user. If the third user's second influence is greater than the influence threshold, it means that the influence of the important feature on the third user's intention is greater than the influence of the important feature on the intention of most first users. In other words, the influence of the other remaining features on the third user's intention is less than the influence of the other remaining features on the intention of most first users. Since the first user has been identified as the target user based on the important features and other remaining features, it can be assumed that the important features and other remaining features have been fully analyzed and utilized relative to the first user. However, the influence of the other remaining features on the intention of the third user is less than the influence of the other remaining features on the intention of most of the first users. Therefore, it can be assumed that the other remaining features have not been fully analyzed and utilized for the third user. In short, the influence of the important features on the third user is too great, that is, it exceeds the influence threshold. It is equivalent to obtaining the intention of the third user based solely on the important features. Therefore, the other remaining features have not affected the intention of the third user, so it can be assumed that the other remaining features have not been fully analyzed and utilized. Based on this, the third user can be regarded as a potential user, so as to further accurately identify users in the subsequent process and determine whether the user will execute the first event.

[0088] Thus, after determining the influence threshold based on the first user's level of influence, the second user whose second level of influence is greater than the influence threshold is selected from the second users as the third user, i.e., the potential user. This not only allows for comprehensive selection of users and avoids the omission of target users, but also makes it simpler and more effective to locate potential users, reduce the waste of promotional resources, and improve the conversion rate and return on investment of promotional activities.

[0089] In some embodiments, as described above, the user's intention for the first event is predicted by a trained intention model; therefore, after selecting a third user from the second user group whose influence is greater than the influence threshold based on the first and second influence levels, the execution result of the third user for the first event can be determined, and the execution result is used to indicate whether the third user executes the first event; based on the execution result, the label of the third user is determined, and the label is used to indicate the probability that the third user executes the first event; based on the label and the third user's intention, the intention model is trained again.

[0090] It should be noted that the execution result of the third user in relation to the first event refers to whether the third user has executed the first event. For example, after selecting the third user, promotional information such as advertisements for the first event can be sent to the third user. When feedback information indicating that the third user has executed the first event is received, the execution result indicating that the third user has executed the first event is determined. If no feedback information indicating that the third user has executed the first event is received within the target period, such as within one month, the execution result indicating that the third user has not executed the first event is determined. Alternatively, staff can conduct real-world promotions to the third user, such as by phone, and then obtain the promotional results uploaded by the staff for the third user, which is also the execution result. The promotional result is used to indicate whether the third user has executed the first event or not.

[0091] In actual implementation, after determining the execution result of the third user in response to the first event, in the process of determining the label of the third user, if the third user executes the first event, the probability of the third user executing the first event is determined to be 1, that is, the label of the third user is determined to be 1; if the third user does not execute the first event, the probability of the third user executing the first event is determined to be 0, that is, the label of the third user is determined to be 0. Then, based on the intention of the label and the intention of the third user, the intention model is trained again. That is, the difference between the intention of the label and the intention of the third user is obtained, and the intention model is trained again based on the difference. The intention of the third user is predicted based on the intention model. As mentioned above, the process of predicting the intention of the third user based on the intention model is not repeated here.

[0092] It should be noted that when the intention model first predicted the intention of the third user, the predicted intention was less than the intention threshold, meaning the third user was predicted not to perform the first event. However, since the third user is a potential user, a label for the third user is determined based on whether or not they perform the first event. The intention model is then retrained based on this label and the third user's intention. In this way, by using the third user's actual behavior (whether or not they perform the first event) as the label to retrain the model, not only is the model's prediction accuracy further improved, but the use of more precise and specific behavioral data also enhances the model's generalization ability.

[0093] Applying the above embodiments of this application, after grouping users based on their intention for the first event to obtain a first user group with an intention greater than or equal to an intention threshold and a second user group with an intention less than the intention threshold, a first user instance is determined based on a first feature, and a second user instance is determined based on a second feature. Then, based on the first user instance, the first degree of influence of the first feature on the intention is determined, and based on the second user instance, the second degree of influence of the second feature on the intention is determined. Thus, based on the first degree of influence and the second degree of influence, a third user with an influence degree greater than the influence threshold is selected from the second user group. In this way, compared with the scheme of selecting users based solely on intention, generating multiple user instances based on the features of different users enriches the diversity of features, thereby facilitating the determination of the degree of influence of features on intention. Furthermore, for users with low intention, users are selected from among those with low intention based on the degree of influence. In this way, for users with low intention, user selection can be combined with the degree of influence of features on intention, making the user selection process more in-depth and comprehensive, avoiding the omission of users who should be selected, and achieving the effect of selecting potential users from among those with low intention.

[0094] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.

[0095] The research revealed that machine learning models are widely used in various user segmentation and decision-making processes. For example, machine learning models can predict user purchasing behavior, obtaining a purchase intention propensity score. Users with purchase intention propensity scores exceeding a threshold are designated as target users, while all other users are considered non-target users. Promotional activities are then only applied to target users. However, only target users have real performance data regarding the promotional activities. Since non-target users haven't received any promotional activity, there's no data on their performance. Therefore, the model cannot continuously adjust and correct prediction biases by collecting real performance data. Consequently, because the model hasn't fully explored a segment of users, it assigns scores below the threshold to these users, leading to their rejection. The model then cannot continue learning from the real performance data of these rejected users until it has fully explored them.

[0096] In practical applications, common methods for achieving a good balance between exploration and exploitation include greedy strategies and methods based on upper confidence bounds. However, these methods do not describe how to distinguish between truly underexplored samples and samples that the model has explored sufficiently but was rejected due to low decision scores. Generally, it is often assumed that all rejected samples are underexplored and in an "exploratory state." This assumption is often invalid in real-world problems. Samples already in an "explosive state" but rejected solely due to low decision scores may constitute a large proportion of all rejected samples and cannot be ignored. Thus, a continuous learning scheme based on the assumption that all rejected samples represent "exploratory" samples may not only be inefficient, but also, due to the high proportion of "explosive" samples among the rejected samples, despite significant investment in re-exploration, may never fully explore the samples that truly need exploration.

[0097] Based on this, this application provides a user selection method that selects users whose model scores are too low due to insufficient model exploration in the past from those who have been rejected. This allows for continued exploration of these users whose model exploration was insufficient, that is, it enables the selection of potential users from those who have been rejected.

[0098] In practical implementation, assuming a promotional scenario, the current strategy being implemented is to use a machine learning model M. 0 意向度 (Intention Model) For each user in the full target audience Q (multiple users), a score is first assigned regarding their intention to purchase a certain product (the first event). For users whose intention score is higher than a pre-set threshold (intention threshold), further promotional actions will be taken.

[0099] Here, we assume that the set of users in Q whose intention level exceeds the threshold is Q. + (First user group), while the set of users below the threshold is Q. - (Second User Group) The goal is to be in Q - Identify a sub-customer group Q that the current model M has not fully explored. - 探索 (Third user), then in the future, Q can be used independently. - 探索 Design an exploration strategy for users in the Q group, for example, for Q... - 探索 A random sample is taken, and then for each person in the sample, Q is used. + The same promotional actions are applied to users in the same context, and then the model is retrained based on feedback to complete the process for Q. - 探索 A thorough exploration of the potential customer base.

[0100] The following explains how to obtain Q - Find Q in - 探索 Please provide an explanation.

[0101] The first step is to train a linear machine learning discriminative model M. 1 判别 The model (discriminative model) uses variables from model M. 0 意向度 For the same variables in the model, M 1 判别 The model is used to identify the probability that a user's intent exceeds an intent threshold, that is, to indicate whether a user's intent exceeds the intent threshold.

[0102] The second step is to apply the model M. 1 判别 A univariate importance analysis was performed on each variable characteristic in the variables, specifically, from M... 1 判别 Remove the univariate feature X from the variable set; then, retrain the same linear discriminant model M using all the remaining variable features except X. -X 判别 Then compare model M before and after removing variable X. 1 判别 and model M -X 判别 If removing variable X significantly reduces model performance, it indicates that the variable is crucial for model prediction. The importance of each variable can be quantified by the degree of performance degradation. Assuming model M... 1判别 There are n variables X1, X2, ..., X n Their corresponding importance values ​​(level of importance) are Z and Z, respectively. x1 Z x2 ,…, Z xn .

[0103] The third step is to place the 20% of variables (important features) with the highest importance values ​​calculated in the previous step into set R. Then R contains [n...] There are 20% variables, let them be {X1, X2, ..., X...} n 20% Meanwhile, let M... 1 判别 The set of all remaining variables in the set of variables excluding those in set R is ~R. From Q... + The dataset (first user group) randomly samples 100 users, and the different combinations of values ​​of variables in the included set R are denoted as {D}. {1} D {2} D {100}}; For example, R = {30-day likes, 30-day search count}, ~R = {user gender, user age, user occupation}, then {D {1} ={30-day likes=20, 30-day search count=18}D {2} ={30-day likes=1, 30-day search count=1}, ...D {100} ={30-day likes=5, 30-day search count=0}.

[0104] Step 4, via Q + The data in the dataset is used to determine the threshold for the perturbation score (the threshold for the degree of influence). Specifically, this is done by analyzing the Q... + For each user U (the first user) in the dataset, for example, first user U = {user gender = male, user age = 43, user occupation = employee, 30-day likes = 10, 30-day search count = 12}, according to {D {1} D {2} ,…,D {100} This is expanded to generate 100 new user instances (the first user instance), for example, generating instance U. D（1） ={User Gender=Male, User Age=43, User Occupation=Employee, 30-day Likes=20, 30-day Searches=18}, Generate Instance U D（2） ={User gender=male, User age=43, User occupation=employee, 30-day likes=1, 30-day search count=1}; ...; Generate instance U D（100）={User gender=male, User age=43, User occupation=employee, 30-day likes=5, 30-day search count=0}; That is, it is equivalent to replacing the user's 30-day likes and 30-day search count with the instance D extracted above, to obtain 100 new user instances. Then, 100 instances U are generated for each user U. D(1) U D(2) ,…, U D(100) Call model M 0 意向度 For each generated instance, predict a corresponding intention level, i.e., s. D(1) s D(2) ,…, s D(100) Next, calculate the dispersion of this set of scores for user U, that is: ...Formula (3); Next, based on the above formula (2), the perturbation threshold is determined.

[0105] Step 5, from Q - Q is determined in (second user group) - 探索 Sub-customer group. Specifically, for Q - For each user U (the second user) in the dataset, according to {D {1} D {2} ,…,D {100}}, expand it to generate 100 new user instances (second user instances); then for each generated instance U D(1) U D(2) ,…, U D(100) Call model M 0 意向度 For each generated instance, predict a corresponding intention level, i.e., s. D(1) s D(2) ,…, s D(100) Then, combining the above formula (3), we can calculate the dispersion of this set of scores for user U.

[0106] It should be noted that, The larger the value, the greater the influence of fluctuations in the features in set R on the user's intention score. In other words, the remaining features (features belonging to set ~R) have a relatively low determining effect on the user's intention score. It is equivalent to these features in set ~R not being fully "utilized". Therefore, the user should be marked as a user who needs to be "explored".

[0107] Based on this, if If the value exceeds the disturbance threshold, then user U is placed into Q.- 探索 In the subgroup, because in this case, it means that among all the features of the current user U, the features in set ~R are compared with the features in set R in M. 0 意向度 The model's degree of determination on user intention compared to Q + The average level of all known users in the "exploitation" state is lower, so they should be classified into the "exploration" state user group.

[0108] Applying the above embodiments of this application, after grouping users based on their intention for the first event to obtain a first user group with an intention greater than or equal to an intention threshold and a second user group with an intention less than the intention threshold, a first user instance is determined based on a first feature, and a second user instance is determined based on a second feature. Then, based on the first user instance, the first degree of influence of the first feature on the intention is determined, and based on the second user instance, the second degree of influence of the second feature on the intention is determined. Thus, based on the first degree of influence and the second degree of influence, a third user with an influence degree greater than the influence threshold is selected from the second user group. In this way, compared with the scheme of selecting users based solely on intention, generating multiple user instances based on the features of different users enriches the diversity of features, thereby facilitating the determination of the degree of influence of features on intention. Furthermore, for users with low intention, users are selected from among those with low intention based on the degree of influence. In this way, for users with low intention, user selection can be combined with the degree of influence of features on intention, making the user selection process more in-depth and comprehensive, avoiding the omission of users who should be selected, and achieving the effect of selecting potential users from among those with low intention.

[0109] The following description continues to illustrate the exemplary structure of the user selection device 455 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 2 As shown, the software modules stored in the user selection device 455 of the memory 450 may include: Grouping module 4551 is used to divide the user into a first user group and a second user group based on the user's intention to the first event, wherein the intention of the first user in the first user group is greater than or equal to the intention threshold, and the intention of the second user in the second user group is less than the intention threshold. The generation module 4552 is used to generate a first user instance based on a first feature of a first user in the first user group, and to generate a second user instance based on a second feature of a second user in the second user group; The determining module 4553 is used to determine, based on the first user instance, the first degree of influence of the first feature on the intention, and based on the second user instance, the second degree of influence of the second feature on the intention. Selection module 4554 is used to select a third user from the second user group whose influence level is greater than the influence level threshold based on the first influence level and the second influence level.

[0110] In some embodiments, the apparatus further includes an analysis module, which is configured to perform importance analysis on the user's features to obtain the importance level of each feature; based on the importance level of each feature, select features from the user's features whose importance level is greater than or equal to an importance level threshold as important features; and determine a first feature of the first user and a second feature of the second user from the important features.

[0111] In some embodiments, the analysis module is further configured to perform the following processing for each of the multiple features of the user to obtain the importance of the feature: removing the corresponding feature from the multiple features of the user to obtain a third feature; predicting the probability that the user's intention exceeds the intention threshold based on the third feature; and determining the importance of the feature based on the probability.

[0112] In some embodiments, the number of first features is multiple, and the feature types of different first features are different; the generation module 4552 is further configured to perform multiple selection processes for the first features of the first user to obtain multiple combined features; wherein, for each selection process, the following operations are performed to obtain a combined feature: for each feature type, determine the first feature of each first user under the feature type, and select one first feature from the determined multiple first features; combine the first features selected under each feature type to obtain a combined feature; wherein, the feature type corresponding to the combined feature is the same as the feature type of the first feature; and determine the first user instance based on the multiple combined features.

[0113] In some embodiments, the number of the first features is n, where n is a positive integer. The generation module 4552 is further configured to perform k update processes on the first features of each first user to obtain k first user instances. For each update process, the following operations are performed to obtain a first user instance: update m first features of the first user to m new first features to obtain a first user instance, where m is a positive integer less than or equal to n, and k is an integer greater than 1.

[0114] In some embodiments, the number of first user instances is multiple, and the determining module 4553 is further configured to, for each first user instance, determine the intention degree of the first user instance for the first event; average the intention degrees of the multiple first user instances to obtain the average intention degree of the multiple first user instances; determine the dispersion degree of the intention degree of the multiple first user instances relative to the average intention degree, and use the dispersion degree as the first influence degree.

[0115] In some embodiments, the determining module 4553 is further configured to: subtract the intention degree of each first user instance from the average intention degree to obtain an intention degree difference; sum the squares of the intention degree differences of each first user instance to obtain a sum of intention degrees; determine the ratio of the sum of intention degrees to the number of first user instances, and perform square root processing on the ratio to obtain the dispersion degree.

[0116] In some embodiments, there is a one-to-one correspondence between the first degree of influence and the first user, and a one-to-one correspondence between the second degree of influence and the second user; the selection module 4554 is further configured to sum the first degree of influence of each first user to obtain a sum of degree of influence; determine the ratio of the sum of degree of influence to the number of first users, and determine the degree of influence threshold based on the ratio; and select a second user from the second users whose second degree of influence is greater than the degree of influence threshold as the third user.

[0117] In some embodiments, the user's intention for the first event is predicted by a trained intention model; the apparatus further includes a training module, which is configured to determine the execution result of the third user for the first event, the execution result indicating whether the third user executes the first event; based on the execution result, determine a label for the third user, the label indicating the probability that the third user executes the first event; and retrain the intention model based on the label and the third user's intention.

[0118] This application provides a computer program product comprising computer-executable instructions or a computer program stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions or computer program from the computer-readable storage medium and executes the computer-executable instructions or computer program, causing the electronic device to perform the user selection method described in this application embodiment, for example, as... Figure 3 The user selection method is shown.

[0119] This application provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the user selection method provided in this application embodiment, for example, such as... Figure 3 The user selection method is shown.

[0120] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0121] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0122] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0123] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0124] It should be noted that in this application embodiment, data such as user attribute information is involved. When this application embodiment is applied to a specific product or technology, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0125] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A user selection method, characterized in that, The method includes: Based on the user's intention for the first event, the user is divided into a first user group and a second user group. The intention of the first user in the first user group is greater than or equal to the intention threshold, and the intention of the second user in the second user group is less than the intention threshold. The first event includes the user purchasing goods or the user downloading an application. A first user instance is generated based on the first feature of the first user in the first user group, and a second user instance is generated based on the second feature of the second user in the second user group. There are multiple first features, and the feature types of different first features are different. The feature types of the first features include the number of likes and searches of the user within a preset time period. Based on the first user instance, a first degree of influence of the first feature on the intention is determined, and based on the second user instance, a second degree of influence of the second feature on the intention is determined, wherein the first degree of influence corresponds one-to-one with the first user and the second degree of influence corresponds one-to-one with the second user; An influence threshold is determined based on the average value of the first influence levels of each of the first users, and users whose second influence level is greater than the influence threshold are selected from the second user group as the third users; The step of generating a first user instance based on a first feature of a first user in the first user group includes: cross-combining first features of the same feature type of different first users in the first user group to obtain multiple combined features, and generating a first user instance based on the combined features.

2. The method according to claim 1, characterized in that, Before generating a first user instance based on a first feature of a first user in the first user group, the method further includes: An importance analysis is performed on the user's features to determine the importance level of each feature; Based on the importance of each feature, features with an importance greater than or equal to an importance threshold are selected from the user's features as important features; From the aforementioned key features, determine the first feature of the first user and the second feature of the second user.

3. The method according to claim 2, characterized in that, The step of performing importance analysis on the user's features to obtain the importance level of each feature includes: For each of the user's multiple characteristics, the following processing is performed to determine the importance of that characteristic: From the user's multiple features, remove the corresponding feature to obtain the third feature; Based on the third feature, the probability that the user's intention exceeds the intention threshold is predicted; The importance of the feature is determined based on the probability.

4. The method according to claim 1, characterized in that, The step of cross-combining the first features of the same feature type among different first users in the first user group to obtain multiple combined features includes: For the first feature of the first user, perform multiple selection processes to obtain multiple combined features; For each selection process, the following operations are performed to obtain a combined feature: For each of the feature types, a first feature for each first user under the feature type is determined, and one first feature is selected from the determined plurality of first features; the first features selected under each feature type are combined to obtain a combined feature; wherein the feature type corresponding to the combined feature is the same as the feature type of the first feature.

5. The method according to claim 1, characterized in that, The number of the first user instances is multiple, and the step of determining the degree of first influence of the first feature on the intention based on the first user instances includes: For each of the first user instances, determine the degree of intent of the first user instance toward the first event; The average intention of the multiple first user instances is obtained by averaging the intention of the multiple first user instances. Determine the degree of dispersion of the intention of the plurality of first user instances relative to the average intention, and use the degree of dispersion as the first degree of influence.

6. The method according to claim 5, characterized in that, Determining the dispersion of the intention levels of the plurality of first user instances relative to the average intention level includes: The intention degree of each first user instance is subtracted from the average intention degree to obtain the intention degree difference value; The sum of the squares of the intention differences for each of the first user instances is obtained by summing the intentions. Determine the ratio of the sum of intent to the number of the first user instances, and take the square root of the ratio to obtain the dispersion degree.

7. The method according to claim 1, characterized in that, The step of determining the influence threshold based on the average of the first influence levels of each of the first users includes: The sum of the first degree of influence for each of the first users is obtained. Determine the ratio of the sum of the impact levels to the number of the first users, and determine the impact level threshold based on the ratio.

8. The method according to claim 1, characterized in that, The user's intention for the first event is predicted by a trained intention model; After selecting a third user from the second user group whose influence level is greater than the influence level threshold based on the first influence level and the second influence level, the method further includes: Determine the execution result of the third user in response to the first event, the execution result being used to indicate whether the third user executed the first event; Based on the execution result, a tag for the third user is determined, the tag being used to indicate the probability that the third user will execute the first event; Based on the tags and the third user's intention, the intention model is trained again.

9. An electronic device, characterized in that, include: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the user selection method according to any one of claims 1 to 8.

10. A computer program product comprising computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, the user selection method according to any one of claims 1 to 8 is implemented.

Images