Object classification method and apparatus, computer device and storage medium

Abstract

The application relates to an object classification method and device, computer equipment, a storage medium and a computer program product. An initial decision tree is generated based on training operation data and a training label corresponding to a training object, and the initial decision tree is used as a current decision tree; the training operation data is input into the current decision tree to obtain a current round prediction result, a prediction error is generated based on the training label, the current round prediction result and a historical round prediction result; a backward decision tree of the current decision tree is generated based on the training operation data and the prediction error, the backward decision tree is used as the current decision tree, the step of inputting the training operation data into the current decision tree to obtain the current round prediction result is returned to be executed until a first convergence condition is met, and a plurality of backward decision trees are obtained; and an object classification model generated based on the initial decision tree and each backward decision tree is used to obtain a prediction label corresponding to a target object by fusing prediction results of target operation data of the target object of each decision tree, so that the object classification accuracy can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to an object classification method and device, computer equipment, storage medium and computer program product. BACKGROUND

[0002] With the development of computer technology, more and more application programs emerge in the network. In order to maintain the network environment and protect the physical and mental health of the object, the operation permission of different types of objects in the same application program can be distinguished.

[0003] In the traditional technology, different types of objects are usually identified by experience rules, for example, for the same application program, the operation time of different types of objects is different, and the object operating the application program at a specific time is identified as a specific type of object corresponding to the specific time. However, the object classification method based on experience rules has great subjectivity and low classification accuracy. SUMMARY

[0004] Therefore, it is necessary to provide an object classification method, device, computer equipment, computer readable storage medium and computer program product capable of improving classification accuracy for the above technical problems.

[0005] The present application provides an object classification method. The method comprises:

[0006] Obtaining training operation data and training labels corresponding to a training object; the training operation data is obtained based on operation data of the training object in a target application, and the training labels are used to determine the operation permission of the training object in the target application;

[0007] Generating an initial decision tree based on the training operation data and the training labels, and taking the initial decision tree as a current decision tree;

[0008] Inputting the training operation data into the current decision tree to obtain a current prediction result, and generating a prediction error based on the training labels, the current prediction result and a historical prediction result;

[0009] Generating a backward decision tree of the current decision tree based on the training operation data and the prediction error, taking the backward decision tree as the current decision tree, returning to the step of inputting the training operation data into the current decision tree to obtain the current prediction result, until a first convergence condition is met, and obtaining a plurality of backward decision trees;

[0010] An object classification model is generated based on the initial decision tree and each backward decision tree; the object classification model is used to input the target operation data corresponding to the target object into the initial decision tree and each backward decision tree, and to fuse the prediction results of the initial decision tree and each backward decision tree to obtain the prediction label corresponding to the target object.

[0011] This application also provides an object classification apparatus. The apparatus includes:

[0012] The training data acquisition module is used to acquire training operation data and training labels corresponding to the training object; the training operation data is obtained based on the operation data of the training object in the target application, and the training labels are used to determine the operation permissions of the training object in the target application.

[0013] An initial decision tree generation module is used to generate an initial decision tree based on the training operation data and the training labels, and to use the initial decision tree as the current decision tree.

[0014] The prediction error calculation module is used to input the training operation data into the current decision tree to obtain the prediction result of this round, and generate the prediction error based on the training label, the prediction result of this round and the prediction results of the previous rounds.

[0015] The backward decision tree generation module is used to generate a backward decision tree of the current decision tree based on the training operation data and the prediction error, use the backward decision tree as the current decision tree, return to the step of inputting the training operation data into the current decision tree to obtain the prediction result of this round, until the first convergence condition is met, and obtain multiple backward decision trees;

[0016] An object classification model generation module is used to generate an object classification model based on the initial decision tree and each backward decision tree. The object classification model is used to input the target operation data corresponding to the target object into the initial decision tree and each backward decision tree, and fuse the prediction results of the initial decision tree and each backward decision tree to obtain the prediction label corresponding to the target object.

[0017] This application provides an object classification method. The method includes:

[0018] Obtain the target operation data corresponding to the target object; the target operation data is obtained based on the operation data of the target object in the target application.

[0019] The target operation data is input into an object classification model to obtain a predicted label corresponding to the target object; the predicted label is used to determine the operation permissions of the target object in the target application, and the predicted label is obtained by fusing the prediction results of each decision tree in the object classification model;

[0020] The training process of the object classification model includes the following steps:

[0021] Obtain the training operation data and training labels corresponding to the training object;

[0022] An initial decision tree is generated based on the training operation data and the training labels, and the initial decision tree is used as the current decision tree.

[0023] The training operation data is input into the current decision tree to obtain the prediction result of this round, and the prediction error is generated based on the training label, the prediction result of this round and the prediction results of the previous rounds.

[0024] Based on the training operation data and the prediction error, a backward decision tree is generated for the current decision tree. The backward decision tree is used as the current decision tree. The step of inputting the training operation data into the current decision tree to obtain the prediction result of this round is returned to the execution until the first convergence condition is met, and multiple backward decision trees are obtained.

[0025] The object classification model is generated based on the initial decision tree and each backward decision tree.

[0026] This application also provides an object classification apparatus. The apparatus includes:

[0027] The data acquisition module is used to acquire target operation data corresponding to the target object; the target operation data is obtained based on the operation data of the target object in the target application;

[0028] The label prediction module is used to input the target operation data into the object classification model to obtain the predicted label corresponding to the target object; the predicted label is used to determine the operation permission of the target object in the target application, and the predicted label is obtained by fusing the prediction results of each decision tree in the object classification model;

[0029] The training process of the object classification model includes the following steps:

[0030] Obtain training operation data and training labels corresponding to the training objects; generate an initial decision tree based on the training operation data and training labels, and use the initial decision tree as the current decision tree; input the training operation data into the current decision tree to obtain the prediction result for this round, and generate a prediction error based on the training labels, the prediction result for this round, and the prediction results for previous rounds; generate a backward decision tree based on the training operation data and the prediction error, and use the backward decision tree as the current decision tree; return to the step of inputting the training operation data into the current decision tree to obtain the prediction result for this round, until the first convergence condition is met, resulting in multiple backward decision trees; generate the object classification model based on the initial decision tree and each backward decision tree.

[0031] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps described in the object classification method above.

[0032] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps described in the object classification method above.

[0033] A computer program product includes a computer program that, when executed by a processor, implements the steps described in the object classification method above.

[0034] The aforementioned object classification method, apparatus, computer device, storage medium, and computer program product acquire training operation data and training labels corresponding to the training objects. The training operation data is obtained based on the training objects' operation data in the target application, and the training labels are used to determine the operation permissions of the training objects in the target application. An initial decision tree is generated based on the training operation data and training labels. This initial decision tree is used as the current decision tree. The training operation data is input into the current decision tree to obtain the prediction result for this round. A prediction error is generated based on the training labels, the prediction result for this round, and the prediction results from previous rounds. A backward decision tree is generated based on the training operation data and prediction error. This backward decision tree is used as the current decision tree. The process returns to the step of inputting the training operation data into the current decision tree to obtain the prediction result for this round, continuing until the first convergence condition is met, resulting in multiple backward decision trees. An object classification model is generated based on the initial decision tree and each backward decision tree. Subsequently, the target operation data corresponding to the target object can be input into the object classification model. The object classification model outputs the predicted label corresponding to the target object by fusing the prediction results of the initial decision tree and each backward decision tree. In this way, an object classification model can be trained based on training data and training labels. Classifying objects using this model improves classification accuracy and efficiency. Furthermore, the object classification model includes multiple decision trees. Each backward decision tree learns the error between the predictions of the forward decision trees and the training labels. The backward decision trees can improve the model's accuracy based on the forward decision trees. Therefore, by fusing the predictions from all decision trees, the object classification model can obtain more accurate predicted labels, further improving object classification accuracy. Attached Figure Description

[0035] Figure 1 This is a diagram illustrating the application environment of an object classification method in one embodiment.

[0036] Figure 2 This is a flowchart illustrating an object classification method in one embodiment;

[0037] Figure 3This is a schematic diagram illustrating the statistical results of the age of an object before and after real-name registration in one embodiment;

[0038] Figure 4 This is a flowchart illustrating the object classification method in another embodiment;

[0039] Figure 5 This is a schematic diagram of the process for identifying game players in one embodiment;

[0040] Figure 6 This is a schematic diagram illustrating the training and application of an ensemble model in one embodiment;

[0041] Figure 7 This is a structural block diagram of an object classification device in one embodiment;

[0042] Figure 8 This is a structural block diagram of an object classification device in another embodiment;

[0043] Figure 9 This is a structural block diagram of the object classification device in another embodiment;

[0044] Figure 10 This is an internal structural diagram of a computer device in one embodiment;

[0045] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including 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. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0048] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0049] The solutions provided in this application involve technologies such as machine learning in artificial intelligence, and are specifically illustrated through the following embodiments:

[0050] The object classification method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another server.

[0051] Terminal 102 and server 104 can cooperate to execute the object classification method provided in this embodiment. For example, the server obtains training operation data and training labels corresponding to the training objects from the terminal. The training operation data is obtained based on the operation data of the training objects in the target application, and the training labels are used to determine the operation permissions of the training objects in the target application. The server generates an initial decision tree based on the training operation data and training labels, uses the initial decision tree as the current decision tree, inputs the training operation data into the current decision tree to obtain the prediction result for this round, generates a prediction error based on the training labels, the prediction result for this round, and the prediction results for previous rounds, generates a backward decision tree based on the training operation data and prediction error, uses the backward decision tree as the current decision tree, and returns to the step of inputting the training operation data into the current decision tree to obtain the prediction result for this round until the first convergence condition is met, resulting in multiple backward decision trees. The server generates an object classification model based on the initial decision tree and each backward decision tree, and sends the object classification model to the terminal. The terminal acquires the target operation data corresponding to the target object, inputs the target operation data into the initial decision tree and each backward decision tree in the object classification model, and the object classification model fuses the prediction results of the initial decision tree and each backward decision tree to obtain the predicted label corresponding to the target object. The object classification model outputs the predicted label. The terminal can determine the target object's operation permissions in the target application based on the predicted label, thereby controlling the target object's operations in the target application.

[0052] Terminal 102 and server 104 can also be used independently to execute the object classification method provided in the embodiments of this application. For example, the terminal trains an object classification model based on training data, and the terminal obtains the predicted label corresponding to the target object based on the object classification model.

[0053] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle systems. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The target application is installed on the terminal. The server 104 can be implemented using a standalone server, a server cluster consisting of multiple servers, or a cloud server.

[0054] In one embodiment, such as Figure 2 As shown, an object classification method is provided. Taking the application of this method to a computer device as an example, the computer device can be one of the above-mentioned... Figure 1 Terminal 102 or server 104 in the reference. Figure 2 The object classification method includes the following steps:

[0055] Step S202: Obtain the training operation data and training labels corresponding to the training object; the training operation data is obtained based on the operation data of the training object in the target application, and the training labels are used to determine the operation permissions of the training object in the target application.

[0056] In this context, "target application" refers to an application that requires differentiated access permissions for different types of users. Target applications can be games, video applications, e-commerce applications, etc. For example, in a game application, adult and minor players have different access permissions, and minor players' viewing time is limited. Similarly, in a video application, adults and minors also have different access permissions, such as limiting the viewing time and types of videos minors can watch.

[0057] Training operation data is generated based on the operation data of the training object in the target application. This data reflects the training object's usage of the target application. Training labels are used to determine the training object's operational permissions within the target application. For example, a first label indicates unrestricted operational permissions, while a second label indicates restricted permissions.

[0058] Specifically, computer devices can obtain training operation data and training labels corresponding to training objects locally or from other terminals or servers. Based on this data, they can then train a model to obtain an object classification model composed of multiple decision trees. Subsequently, the computer device can classify any object based on the object classification model, thereby controlling the object's operation in the target application and achieving the purpose of maintaining the network environment and protecting objects.

[0059] In one embodiment, a computer device can extract the operation data corresponding to the training object from the application logs reported by the target application to obtain training operation data.

[0060] Step S204: Generate an initial decision tree based on the training operation data and training labels, and use the initial decision tree as the current decision tree.

[0061] A decision tree is a tree-like structure where each internal node represents a test on an attribute or feature, each branch represents a test output, and each leaf node represents a category. The initial decision tree refers to the first decision tree obtained through training.

[0062] Specifically, computer equipment can generate an initial decision tree based on training operation data and training labels. The training operation data can be used as input data for the initial decision tree, and the training labels as the expected output. Through multiple iterations, the initial decision tree is finally obtained. It can be understood that the training operation data includes multiple training operation features. Through multiple iterations, several features with significant influence on the labels can be selected as decision tree nodes for the initial decision tree. Through further iterations, the connections and parameters between the decision tree nodes can be determined, ultimately resulting in an initial decision tree. After inputting data into the initial decision tree, based on the decision tree nodes and node parameters traversed by the input data, the initial decision tree can output prediction results.

[0063] Step S206: Input the training operation data into the current decision tree to obtain the prediction result of this round, and generate the prediction error based on the training label, the prediction result of this round and the prediction results of the previous round.

[0064] Step S208: Generate a backward decision tree for the current decision tree based on the training operation data and prediction error. Use the backward decision tree as the current decision tree and return to the step of inputting the training operation data into the current decision tree to obtain the prediction result of this round, until the first convergence condition is met, and obtain multiple backward decision trees.

[0065] Here, the historical round prediction results refer to the prediction results obtained before the current round prediction results. Different rounds of training are used to train different decision trees. For example, the first round of training is used to train the backward decision tree B of the initial decision tree A, the second round of training is used to train the backward decision tree C of decision tree B, the third round of training is used to train the backward decision tree D of decision tree C, and so on, ultimately resulting in multiple ordered decision trees.

[0066] Prediction error characterizes the gap between the training labels and existing predictions. The newly generated decision tree is used to reduce the gap between the training labels and existing predictions, thus bringing the predictions closer to the true results. The current prediction result, previous prediction results, and existing prediction results all refer to the predictions output by the trained decision tree.

[0067] Specifically, in the first round of backward decision tree training, the computer device can use the initial decision tree as the current decision tree, the training data as the input data, and the current decision tree outputs the prediction result of this round (i.e., the prediction result of the first round). Based on the training labels and the prediction result of this round, a prediction error is generated. The training data is then used as the input data for the next decision tree, and the prediction error is used as the expected output of the next decision tree, thus fitting the next decision tree (i.e., the backward decision tree corresponding to the initial decision tree). In the second round of backward decision tree training, the computer device can use the decision tree fitted in the first round as the new current decision tree, the training data as the input data, and the current decision tree outputs the prediction result of this round (i.e., the prediction result of the second round). Based on the training labels, the prediction result of this round, and the prediction results of previous rounds (i.e., the prediction result of the first round), a new prediction error is generated. The training data is then used as the input data for the next decision tree, and the new prediction error is used as the expected output of the next decision tree, thus fitting the next decision tree (i.e., the new backward decision tree). In the third round of backward decision tree training, the computer can use the decision tree fitted in the second round as the new current decision tree, and the training data as the input data for the current decision tree. The current decision tree outputs the prediction result of this round (i.e., the prediction result of the third round). Based on the training labels, the prediction result of this round, and the prediction results of previous rounds (i.e., the prediction results of the first and second rounds), a new prediction error is generated. The training data is then used as the input data for the next decision tree, and the new prediction error is used as the expected output of the next decision tree. The next decision tree (i.e., the new backward decision tree) is then fitted. This process continues, and after multiple rounds of training, multiple backward decision trees can be generated.

[0068] The first convergence condition can be at least one of the following: the number of backward decision trees is greater than a preset number, the prediction difference is less than a preset difference, etc. For example, if the first convergence condition is to generate 100 backward decision trees, then training will stop after the 100th backward decision tree is obtained. The prediction difference refers to the difference between the training label and the predicted label, obtained based on the training label and the existing prediction results. If the first convergence condition is that the prediction difference is less than 0.001, then training will stop when the difference between the training label and the predicted label is less than 0.001, and no new backward decision trees will be generated. The preset number and preset difference can be set according to actual needs.

[0069] It is understandable that, in the initial decision tree and each subsequent decision tree, due to different fitting objectives, any two decision trees are not exactly the same. For example, they may have different tree shapes or different node parameters.

[0070] Step S210: Generate an object classification model based on the initial decision tree and each backward decision tree; the object classification model is used to input the target operation data corresponding to the target object into the initial decision tree and each backward decision tree, and fuse the prediction results of the initial decision tree and each backward decision tree to obtain the prediction label corresponding to the target object.

[0071] Here, the target object refers to the object to be categorized and whose operation permissions need to be determined. The target operation data is obtained based on the operation data of the target object in the target application.

[0072] Specifically, after training to obtain the initial decision tree and each backward decision tree, the computer device can combine the initial decision tree and each backward decision tree to form an object classification model. Subsequently, when applying the model, the computer device can acquire the target operation data corresponding to the target object, input the target operation data into the object classification model, and feed the target operation data into the initial decision tree and each backward decision tree. After data processing by the decision trees, the initial decision tree and each backward decision tree output prediction results respectively. The object classification model then merges the prediction results from each decision tree to finally output the predicted label corresponding to the target object.

[0073] In one embodiment, the computer device can directly control the operation of a target object in a target application based on the predicted label. For example, if the predicted label is the target label, the operation of the target object in the target application is restricted. Alternatively, the computer device can determine the operation permissions of the target object in the target application based on the predicted label and other data, thereby controlling the operation of the target object in the target application.

[0074] The object classification method described above obtains training operation data and training labels corresponding to the training objects. The training operation data is obtained based on the operation data of the training objects in the target application, and the training labels are used to determine the operation permissions of the training objects in the target application. An initial decision tree is generated based on the training operation data and training labels. This initial decision tree is used as the current decision tree. The training operation data is input into the current decision tree to obtain the prediction result for this round. A prediction error is generated based on the training labels, the prediction result for this round, and the prediction results for previous rounds. A backward decision tree is generated based on the training operation data and prediction error. This backward decision tree is used as the current decision tree, and the process of inputting the training operation data into the current decision tree to obtain the prediction result for this round is repeated until the first convergence condition is met, resulting in multiple backward decision trees. An object classification model is generated based on the initial decision tree and each backward decision tree. Subsequently, the target operation data corresponding to the target object can be input into the object classification model. The object classification model outputs the prediction label corresponding to the target object by fusing the prediction results of the initial decision tree and each backward decision tree. In this way, an object classification model can be trained based on the training operation data and training labels. Classifying objects based on the object classification model can improve the accuracy and efficiency of object classification. Furthermore, the object classification model includes multiple decision trees. Each backward decision tree is obtained by learning the error between the prediction results of the forward decision tree and the training labels. The backward decision tree can improve the accuracy of the model based on the forward decision tree. Thus, the object classification model can obtain more accurate predicted labels by fusing the prediction results of each decision tree, thereby further improving the classification accuracy of objects.

[0075] In one embodiment, the training operation data includes at least one of the following: object attribute features of the training object, operation interaction features, operation duration features, device login features, registration time features, operating platform features, and target association features between the training object and the target associated object. The target associated object is the associated object among the various associated objects of the training object that has the target operation permission.

[0076] Among them, object attribute features are used to characterize the attribute information of an object. Object attribute features may include at least one of the following: object age, object gender, terminal identifier of the target application's running terminal, terminal model, terminal network, terminal operating region, and terminal operating system. Specifically, object attribute features may include various attribute information of the object before real-name authentication. The terminal operating region refers to the geographical area where the terminal is located, such as first-tier, second-tier, and third-tier cities.

[0077] Operational interaction features are derived from the interaction between a specific object and other objects within a target application. They characterize the frequency and proficiency of an object's operations within the target application. For example, if the target application is a game, the interaction could specifically be adversarial. Operational interaction features could include data such as win rate, draw rate, loss rate, win rate for each mode, total number of matches, total number of days with matches, average daily number of matches (total number of matches / number of days with matches), average daily number of matches on weekdays (total number of matches on weekdays / number of weekdays with matches), average daily number of matches on weekends and holidays (total number of matches on weekends and holidays / number of weekends and holidays with matches), percentage of average daily number of matches on weekends and holidays, average number of matches per hour on weekends and holidays, and percentage of average daily number of matches per hour on weekends and holidays.

[0078] Operation duration features are used to characterize the operation time of an object in a target application. Operation duration features can include at least one attribute information such as operation duration, average duration, and duration percentage. For example, operation duration features can include data such as average daily duration on weekdays, average daily duration on weekends and holidays, duration of each time period on weekdays, duration of each time period on weekends and holidays, and the percentage of time spent in the morning / noon / afternoon / evening / late night each day.

[0079] Device login characteristics are derived from the terminals and accounts used by the target user to log in to the target application. They characterize the frequency and specificity of logins. Device login characteristics can include the number of devices that have logged in historically, the number of individuals historically linked to the account, the number of accounts historically logged into by devices within the target time period, the number of accounts historically linked to the individual linked to the account within the target time period, and whether logins occurred on suspicious devices within the target time period. The target time period can be set according to actual needs, for example, during the summer vacation. Suspicious devices refer to devices that have historically allowed the target type of user to log in to the target application, such as devices historically used by minors to log into games.

[0080] Registration time features are used to characterize the registration time and duration of an object within the target application. Registration time features can include data such as registration duration and registration year.

[0081] Platform characteristics are used to characterize the activity level of an object operating the target application on different platforms. The platform can specifically include client-side (e.g., desktop) and mobile-side (e.g., smartphone). Platform characteristics can include data such as the number of active days on the client-side within a preset time period, the duration of active days on the client-side within a preset time period, the number of active days on the mobile-side within a preset time period, and the duration of active days on the mobile-side within a preset time period. The preset time period can be set according to actual needs; for example, the preset time period could be the past month.

[0082] A target associated object refers to an object's friends within the target application. A target associated object refers to an associated object among the training object's associated objects that possesses the target's operational permissions; it can also be considered a target-type friend of the training object within the target application. For example, a target associated object could be a minor friend of the training object within the target application. The target association features between the training object and target associated objects characterize the importance of the target associated objects to the training object. Target association features may include data such as the number of target associated objects, the proportion of target associated objects, and the total intimacy level of target associated objects. The intimacy level of target associated objects is generated based on the communication information between the training object and target associated objects, including at least one of the following: conversation information, gift-giving information, and invitation information.

[0083] In one embodiment, the operation interaction characteristics, operation duration characteristics, and device login characteristics can be operation characteristics corresponding to the target time period. For example, operation interaction characteristics include win rate, draw rate, and loss rate in July and August.

[0084] Specifically, the training operation data may include at least one dimension of training operation features, which may include at least one of the following: object attribute features, operation interaction features, operation duration features, device login features, registration time features, operating platform features, and target association features. It can be understood that each dimension of training operation features may also include at least one category of operation features; for example, target association features may include two categories of operation features: the number of target-associated objects and the proportion of target-associated objects.

[0085] In one embodiment, a computer device can obtain application logs reported by a target application, extract various training operation features of the same training object from the application logs, and obtain training operation data.

[0086] In one embodiment, the operational features used for model training are obtained through big data analysis of historical operational data of a large number of known object categories. For example, for game applications, refer to... Figure 3 As can be seen, the age difference between those before and after real-name registration is significant; therefore, the age before real-name registration can be used as an operational feature for model training. Referring to Tables 1 and 2, it can be seen that underage gamers changed their account bindings multiple times during holidays; therefore, device login features can be used as operational features for model training. Referring to Table 3, it can be seen that adults prefer playing PC games, while minors primarily play mobile games; therefore, the operating platform features can be used as operational features for model training.

[0087] Table 1

[0088]

[0089] Table 2

[0090] Month Average number of natural persons bound per capita in game A Average number of natural persons bound per capita in game B 202105 0.97 0.98 202106 0.97 0.98 202107 0.97 0.98 202108 0.96 0.98 202109 1.61 1.64

[0091] Table 3

[0092]

[0093] Positive samples represent underage players, while negative samples represent adult players.

[0094] In the above embodiments, the training operation data includes a variety of data, which helps to ensure the training effect of the model.

[0095] In one embodiment, obtaining the training operation data and training labels corresponding to the training object includes:

[0096] Statistical analysis is performed on the operation data of the training object in the target application to obtain initial operation data; the initial operation data includes multiple initial operation features; multiple target operation features are determined from each initial operation feature; the target operation features are cross-referenced to obtain cross operation data; training operation data is obtained based on the initial operation data and cross operation data.

[0097] Feature intersection refers to a composite feature formed by combining individual features. Feature intersection helps to represent nonlinear relationships.

[0098] Specifically, when acquiring training operation data, the computer device can perform statistical analysis on the operation data of the training object in the target application to obtain initial operation data, which is then used as training operation data. To further improve the training effect of the model, the computer device can also select some operation features from multiple initial operation features contained in the initial operation data as target operation features, perform feature crossing on the target operation features to obtain at least one cross operation feature, combine the various cross operation features to form cross operation data, and finally use the initial operation data and cross operation data as training operation data.

[0099] The selection of target operational features can be determined based on actual needs. For example, a computer device can randomly select multiple operational features from the initial operational features as target operational features. Alternatively, the computer device can obtain operational features belonging to different dimensions from the initial operational features as target operational features; feature crossing of operational features from different dimensions can increase the amount of information expressed by the crossed operational features. Furthermore, the computer device can also obtain multiple operational features with the greatest differences between different types of objects from the initial operational features as target operational features; the resulting crossed operational features help the model distinguish between different types of objects.

[0100] In one embodiment, the computer device can multiply the feature values ​​of different target operation features to obtain cross-operation features. Alternatively, the computer device can divide the feature value range of the target operation features to obtain multiple feature value intervals, combine the feature value intervals of different categories of target operation features to obtain multiple combined intervals, obtain initial features based on each combined interval, and update the initial features based on the feature value intervals to which the feature values ​​of different categories of target operation features belong to, thereby obtaining cross-operation features.

[0101] For example, the operation duration can be divided into three feature value intervals: A, B, and C. The target associated object ratio can be divided into two feature value intervals: 1 and 2. Combining these feature value intervals yields six merged intervals: A and 1, B and 1, C and 1, A and 2, B and 2, and C and 2. An initial feature (0, 0, 0, 0, 0, 0) is generated based on these six merged intervals. If the operation duration corresponding to a certain object falls into feature value interval A, and the target associated object ratio falls into feature value interval 2, then the cross-operation feature obtained by updating the initial feature based on this operation duration and target associated object ratio is (0, 0, 0, 1, 0, 0).

[0102] In the above embodiments, the training operation data includes the original operation data and the operation data obtained through feature cross-validation. Training the model based on such data helps to improve the accuracy of the model.

[0103] The model performance was tested experimentally with and without feature crossing. The results are shown in Table 4. As can be seen from Table 4, feature crossing helps ensure effective model training and improves prediction accuracy.

[0104] Table 4

[0105] Training data AUC value Accuracy Recall rate F1 value False positive rate of negative samples Without feature intersection 0.948 0.741074 0.6 0.663 0.047 With feature intersection 0.962 0.76159 0.62 0.669 0.039 Promotion range of model effect 1.48％ 2.77％ 3.33％ 0.90％ -17.02％

[0106] In this context, AUC refers to the area enclosed by the ROC curve and the X-axis; a value closer to 1 indicates better model optimization. Precision is the ratio of the number of correctly predicted classes to the total number of samples. Recall is the proportion of correctly predicted positive samples. F1 score is the harmonic mean of precision and recall. False positive rate is the proportion of samples predicted as positive out of the actual number of negative samples. For example, in gaming applications, the false positive rate could be the probability that an adult player is incorrectly identified as a minor.

[0107] In one embodiment, generating an initial decision tree based on training operation data and training labels includes:

[0108] A first basic decision tree is generated based on the training operation data and training labels; the training operation data is input into the first basic decision tree to obtain the initial prediction result; based on the initial prediction result and training labels, the decision tree parameters of the first basic decision tree are adjusted until the second convergence condition is met to obtain the initial decision tree.

[0109] Specifically, the base decision tree refers to the initial decision tree, the decision tree to be trained. When generating the initial decision tree, the computer device can first generate an initial decision tree as the first base decision tree based on the training operation data and training labels. Then, the first base decision tree is iteratively trained, and the decision tree parameters of the first base decision tree are gradually adjusted until the second convergence condition is met, indicating that the training is complete and the initial decision tree is obtained.

[0110] After obtaining the first basic decision tree, the computer device can input the training operation data into the first basic decision tree. The first basic decision tree outputs the initial prediction result. Then, the computer device performs backpropagation update based on the difference between the initial prediction result and the training label, adjusts the decision tree parameters of the first basic decision tree, and obtains a new first basic decision tree. The training operation data is then input into the new first basic decision tree to obtain a new initial prediction result. Based on the difference between the new initial prediction result and the training label, backpropagation update is performed again, and the decision tree parameters of the first basic decision tree are adjusted again. This process is repeated iteratively, with the goal of each adjustment being to make the difference between the initial prediction result and the training label smaller and smaller, until the second convergence condition is met, and the initial decision tree is obtained.

[0111] Adjusting the decision tree parameters can involve adjusting the tree's shape, such as adjusting the split points, or adjusting the node parameters. The second convergence condition can be at least one of the following: the difference between the initial prediction and the training labels is less than a preset difference, or the number of iterations is less than a preset number.

[0112] In one embodiment, a computer device can calculate the feature importance of various training operation features based on training operation data and training labels of multiple training objects. Based on the feature importance, it determines the optimal splitting feature and its corresponding optimal splitting point from among the various training operation features. Based on the optimal splitting point, the feature space is divided into two regions, and each training operation data point is assigned to one of the two regions according to the value of the optimal splitting feature. This division process is repeated for each region until a division stopping condition is met, resulting in multiple regions. The optimal splitting point used in each space division is used as a decision tree node, and the decision tree nodes are connected according to the division order to obtain the first basic decision tree. Feature importance is used to characterize the degree of influence of a feature on the prediction result; it can also be considered as the classification ability of the feature.

[0113] In the above embodiments, the first basic decision tree generated based on the training operation data and training labels has a basic and relatively complete decision tree form. An initial decision tree can be quickly trained on the first basic decision tree through iterative training.

[0114] In one embodiment, the training objects include multiple objects, and the training operation data includes multiple training operation features. A first basic decision tree is generated based on the training operation data and training labels, including:

[0115] The training operation features corresponding to the same category are binned to obtain feature bin sets corresponding to each category; each feature bin set includes multiple feature bins; the first split coefficient corresponding to each current feature bin is calculated based on the training labels corresponding to each current feature bin; the second split coefficient is determined based on the first split coefficient of each current feature bin corresponding to the same category, and the second split coefficient corresponding to each category is obtained; the target category is determined from each category based on the second split coefficient; the target feature bin is determined from each current feature bin corresponding to the target category based on the first split coefficient of each current feature bin corresponding to the target category; decision tree nodes are generated based on the target feature bins; the feature bins corresponding to the remaining categories are updated based on the target feature bins, and the updated feature bins are used as the current feature bins. The process of calculating the first split coefficient corresponding to each current feature bin based on the number of features and training labels corresponding to each current feature bin is repeated until a first preset condition is met, resulting in each decision tree node; and a first basic decision tree is generated based on each decision tree node.

[0116] Binning refers to grouping disordered features into ordered bins. Computer equipment can use methods such as equal-frequency binning, equal-length binning, chi-square binning, and custom binning. Equal-frequency binning results in bins with the same number of features. Equal-length binning results in bins with the same feature range length. Chi-square binning is a binning method based on the chi-square test. Custom binning refers to a binning method that defines the binning split points.

[0117] Specifically, computer equipment can bin the features of each training operation corresponding to the same category, obtaining feature bin sets for each category. For example, training operation data includes operation duration and the number of target associated objects. Binning the operation duration for each training object yields a feature bin set corresponding to that operation duration, and binning the number of target associated objects for each training object yields a feature bin set corresponding to that number. A feature bin set comprises multiple feature bins.

[0118] Using feature bins as units, the computer device can calculate the first split coefficient based on the number of features and training labels corresponding to the current feature bin, obtaining the first split coefficient for each current feature bin. The first split coefficient is used to determine the optimal split point for a certain category of features. The computer device can calculate the first split coefficient based on a custom formula or algorithm. Furthermore, the computer device can determine the second split coefficient based on the first split coefficients of each current feature bin corresponding to the same category, obtaining the second split coefficient for each category. The second split coefficient is used to determine the optimal splitting feature. For example, if the smaller the first split coefficient, the higher the discriminative power of the corresponding splitting method for different objects, then the first split coefficients corresponding to the same category can be sorted from smallest to largest. The computer device can obtain the minimum value as the second split coefficient, or it can obtain the average of the top-ranked first split coefficients as the second split coefficient. The computer device can determine the target category from each category based on the second split coefficient. For example, if the smaller the second split coefficient, the higher the discriminative power of the corresponding splitting method for different objects, and the higher the feature importance of the corresponding category's operational features, then the category corresponding to the smallest second split coefficient can be taken as the target category. The target category can be considered the currently found optimal splitting feature. Furthermore, based on the first split coefficients corresponding to the target category, the computer device determines the target feature bin from the current feature bins corresponding to the target category. For example, if the smaller the first split coefficient, the higher the distinguishability of the corresponding splitting method for different objects, then the feature bin corresponding to the smallest first split coefficient can be used as the target feature bin. The target feature bin can be considered as the optimal split point corresponding to the optimal splitting feature.

[0119] After obtaining the optimal split point corresponding to the optimal splitting feature, the computer device can generate decision tree nodes based on the target feature binning. This binning divides the feature space into two regions, with different regions corresponding to different branches of the decision tree nodes. For example, if the target feature binning is for operation durations of 9-11 hours, the decision tree node can be configured to determine if the operation duration is greater than 9 hours. Operation durations greater than 9 hours form one decision tree branch, and operation durations less than or equal to 9 hours form another. The decision tree node can be any feature value within the range corresponding to the target feature bin.

[0120] Computer equipment can update the feature bins corresponding to the remaining categories based on the target feature bins. That is, it divides the training operation data for each training object into different regions of the feature space and onto different decision tree branches, and updates the feature bins corresponding to the remaining categories based on the division results. For example, if a decision tree node indicates whether the operation duration is greater than 9 hours, before the update, feature bin A corresponding to the registration duration includes data for 10 training objects. After the update, for the first region or the first branch, feature bin A corresponding to the registration duration includes data for 6 training objects whose operation duration is greater than 9 hours. For the second region or the second branch, feature bin A corresponding to the registration duration includes data for 4 training objects whose operation duration is less than or equal to 9 hours.

[0121] The computer device can select any branch or region corresponding to the updated feature bin as the updated feature bin, obtaining the updated feature bins for each category. It can be understood that the categories corresponding to already generated decision tree nodes no longer need to participate in subsequent data processing. The computer device uses the updated feature bin as the current feature bin, repeats the above steps, determines the new optimal splitting feature and the optimal splitting point corresponding to the optimal splitting feature from the remaining categories, generates new decision tree nodes, and so on, until the first preset condition is met, resulting in multiple decision tree nodes.

[0122] Finally, the decision tree nodes are connected sequentially according to their generation order to obtain the first basic decision tree. It can be understood that if a decision tree branch does not generate a new decision tree node based on the split coefficient, then a leaf node can be directly connected to that branch. The node parameters for each decision tree node can be randomly initialized or generated based on the corresponding second split coefficient. The first basic decision tree is just an initialized decision tree; its shape and node parameters will be adjusted and changed through multiple iterations, such as merging several decision tree nodes or splitting to create new decision tree nodes.

[0123] The first preset condition may be at least one of the following: the decision tree depth is greater than a preset depth, or the binning range of the feature bins corresponding to the decision tree nodes is greater than a preset binning range.

[0124] In the above embodiments, based on the first split coefficient and the second split coefficient, the optimal split feature and the optimal split point can be quickly selected from the training operation features, thereby quickly generating decision tree nodes and obtaining the basic decision tree.

[0125] In one embodiment, the training operation features corresponding to the same category are binned to obtain feature bin sets corresponding to each category, including:

[0126] Initial binning is performed on each training operation feature corresponding to the current category to obtain multiple candidate bins, and the candidate bins are used as the current bins; based on the training labels corresponding to each current bin, the binning merging coefficients corresponding to adjacent current bins are calculated; based on the binning merging coefficients, adjacent current bins are merged to obtain multiple merged bins; the merged bins are used as the current bins, and the step of calculating the binning merging coefficients corresponding to adjacent current bins based on the training labels corresponding to each current bin is returned to the execution until the second preset condition is met, resulting in multiple feature bins corresponding to the current category; based on each feature bin corresponding to the current category, a feature bin set corresponding to the current category is obtained.

[0127] Specifically, during binning, initial binning can be performed first, followed by bin merging to progressively optimize the binning results and ultimately obtain a set of feature bins. For any given category, the computer device can perform initial binning on each training operation feature corresponding to that category, resulting in multiple candidate bins. For example, the training operation features corresponding to the same category can be sorted according to their feature values ​​from smallest to largest, and several binning split points can be randomly determined. Based on these split points, the value range of the training operation features can be divided into multiple feature value intervals, with each interval corresponding to a candidate bin. The training operation features of each training object are then placed into the corresponding candidate bin based on the feature value interval into which their feature values ​​fall.

[0128] For any given category, the computer device uses candidate bins as the current bins and calculates bin-merging coefficients for adjacent current bins based on the training labels corresponding to each current bin. The bin-merging coefficients represent the similarity of data distribution within adjacent feature bins. For example, the proportion of each training label can be calculated based on the number of labels in a feature bin, and the bin-merging coefficients are obtained based on the differences in the proportions of the same training label in adjacent feature bins. The computer device merges adjacent current bins based on these bin-merging coefficients, merging at least one group of adjacent feature bins with similar data distributions, while leaving other adjacent feature bins unchanged, thus obtaining multiple merged bins. For example, if the smaller the bin-merging coefficient, the more similar the data distribution of adjacent feature bins, then in one round of data processing, the adjacent feature bins with the smallest bin-merging coefficient can be merged to obtain one merged bin, while the other feature bins are used as merged bins, thus obtaining multiple merged bins. The computer device can use the merged bin as the current bin, repeat the above steps, calculate the new bin merging coefficient, and perform a new round of bin merging, and so on, until the second preset condition is met, resulting in multiple feature bins corresponding to the current category. Finally, the feature bins corresponding to the current category form the feature bin set corresponding to the current category.

[0129] The second preset condition can be at least one of the following: the number of feature bins is less than or equal to the preset number of bins, or the bin merging coefficient is greater than the preset coefficient. For example, if the preset number of bins is 10, and after a certain round of data processing, the number of merged bins decreases from 11 to 10, then bin merging can be stopped. Similarly, after a certain round of data processing, if the bin merging coefficients calculated based on the latest number of tags corresponding to each merged bin are all greater than the preset coefficient, then bin merging can be stopped.

[0130] In one embodiment, the bin consolidation coefficient can be calculated using the following formula:

[0131]

[0132]

[0133] Among them, A ij E represents the number of labels corresponding to the j-th training label in the i-th bin for a certain category of features. ij A represents ij The expected probability. N represents the total number of training labels for a certain category of features. i C represents the number of labels corresponding to the i-th bin for a certain category of features. j This represents the proportion of the j-th training label among all training labels for a specific category of features.

[0134] In the above embodiments, binning can combine similarly distributed data sets into one feature bin, so that the different feature bins in the final feature bin set have certain data differences. Such a feature bin set helps to improve the convergence speed of the model during model training.

[0135] In one embodiment, the feature binning set obtained by binning in the above manner is applied to model training, and the trained model is validated with data. The obtained underage authentication rate is 7.3%, and the false positive rate is 5.9%. Here, the underage authentication rate refers to the proportion of underage objects among the real-name authentication objects.

[0136] In one embodiment, the binning results for the duration feature are shown in Table 5. The binning results in Table 5 are obtained by clustering the operation durations of a large number of objects. Applying such binning results to model training and validating the trained model with data, the underage authentication rate was 6.4%, and the false positive rate was 5.1%.

[0137] Table 5

[0138]

[0139] In one embodiment, the binning results for the duration feature can be shown in Table 6. Table 6 shows equal-length binning, the purpose of which is to allow the model to learn the optimal split points for generating decision tree nodes. Applying these binning results to model training and validating the trained model with data yielded a minor authentication rate of 8.5% and a false positive rate of 4.2%.

[0140] Table 6

[0141]

[0142] In one embodiment, the first splitting coefficient corresponding to each current feature bin is calculated based on the training labels corresponding to each current feature bin, including:

[0143] A reference feature bin is determined from each current feature bin corresponding to the same category. Based on the reference feature bin, each current feature bin corresponding to the same category is divided into a first-class bin and a second-class bin. Based on the total number of labels corresponding to the first-class bin and the second-class bin, the label ratios corresponding to the first-class bin and the second-class bin, respectively, are obtained. Based on the number of labels corresponding to various training labels in the current-class bin and the total number of labels corresponding to the current-class bin, the label distribution coefficient corresponding to the current-class bin is obtained. Based on the label ratios and label distribution coefficients corresponding to the first-class bin and the second-class bin, the first split coefficient corresponding to the reference feature bin is obtained.

[0144] Specifically, when calculating the first split coefficient, a certain feature bin corresponding to a certain category can be used as a candidate split point. Based on the candidate split point, the training operation features corresponding to that category are divided into two classes of data. The label ratio and label distribution coefficient corresponding to any one class of data are calculated. The label ratio and label distribution coefficient corresponding to each class of data are then fused to obtain the first split coefficient corresponding to the candidate split point.

[0145] The computer device can select any current feature bin from the current feature bins corresponding to the same category as a reference feature bin. Based on the reference feature bin, it divides the current feature bins corresponding to the same category into a first category bin and a second category bin. For example, the current feature bins corresponding to the same category can be sorted in ascending order of feature values. The current feature bins that precede the reference feature bin are designated as the first category bins, and the current feature bins that follow the reference feature bin, along with the reference feature bin, are designated as the second category bins. Next, for each category bin, the computer device counts the number of training labels corresponding to each feature falling into the current category bin, obtaining the total number of labels for the first and second categories. Based on the total number of labels for the first and second categories, the device calculates the label ratios, obtaining the label ratios for the first and second categories, respectively. For example, if the total number of labels for the first category bin is 5 and the total number of labels for the second category bin is 10, then the label ratio for the first category bin is 5 / 15, and the label ratio for the second category bin is 10 / 15. Furthermore, for either the first or second type of binning, the computer device counts the number of labels corresponding to various training labels in the current binning class. Based on the number of labels corresponding to various training labels in the current binning class and the total number of labels corresponding to the current binning class, a label distribution coefficient is calculated to obtain the label distribution coefficients for the first and second type of binning, respectively. Finally, based on the label ratio and label distribution coefficients corresponding to the first and second type of binning, the first split coefficient corresponding to the reference feature binning is calculated. For example, the label ratio and label distribution coefficients corresponding to the same type of binning are fused to obtain the initial fusion result for the first and second type of binning, and then the initial fusion result for the first and second type of binning is fused again to obtain the first split coefficient.

[0146] By analogy, the first splitting coefficient corresponding to each current feature bin can eventually be calculated.

[0147] In one embodiment, the first splitting coefficient can be calculated using the following formula:

[0148]

[0149]

[0150] Where Gini(D, A) represents the first splitting coefficient corresponding to feature bin A. D1 represents the first type of bin, D2 represents the second type of bin, and D represents the set of feature bins corresponding to a certain category. Based on feature bin A, D can be divided into D1 and D2. This indicates the label ratio corresponding to the first category of boxes. This indicates the label proportion corresponding to the second category of boxes. Gini(D1) represents the label distribution coefficient corresponding to the first category of boxes, and Gini(D2) represents the label distribution coefficient corresponding to the second category of boxes. k Indicated in D i In this context, the percentage of labels corresponding to the k-th training label is as follows: For example, the first bin includes 100 features, of which 40 features correspond to the first label. That is, the total number of labels in the first bin is 100, the number of labels corresponding to the first label is 40, and the percentage of labels corresponding to the first label is 40 / 100 = 0.4.

[0151] In the above embodiments, the first split coefficient can be quickly calculated based on the label ratio and label distribution coefficient corresponding to the first type of bin and the second type of bin. The first split coefficient can characterize the ability of the corresponding split point to distinguish between different objects and different labels.

[0152] In one embodiment, a second splitting coefficient is determined based on the first splitting coefficient of each current feature bin corresponding to the same category, resulting in a second splitting coefficient for each category. A target category is then determined from each category based on the second splitting coefficient. Finally, a target feature bin is determined from each current feature bin corresponding to the target category based on the first splitting coefficient of each current feature bin corresponding to the target category. This includes:

[0153] From the first split coefficients corresponding to the same category, the smallest first split coefficient is selected as the second split coefficient, and the second split coefficients corresponding to each category are obtained. The category corresponding to the smallest second split coefficient is selected as the target category. From the first split coefficients corresponding to the target category, the current feature bin corresponding to the smallest first split coefficient is selected as the target feature bin.

[0154] Specifically, the smaller the first splitting coefficient, the higher the discriminative power of the corresponding splitting method for different objects. Therefore, the computer device can select the smallest first splitting coefficient from all the first splitting coefficients corresponding to the same category as the second splitting coefficient, thus obtaining the second splitting coefficients for each category. The smaller the second splitting coefficient, the higher the feature importance. Therefore, the computer device can select the category corresponding to the smallest second splitting coefficient as the target category, thus obtaining the optimal splitting feature. Furthermore, for each current feature bin corresponding to the target category, the computer device can select the current feature bin corresponding to the smallest first splitting coefficient as the target feature bin, thus obtaining the optimal splitting point corresponding to the optimal splitting feature.

[0155] In one embodiment, a prediction error is generated based on the training labels, the current prediction result, and the prediction results from previous rounds, including:

[0156] The current prediction result and the historical prediction results are merged to obtain the fused prediction result. The fused prediction result is then mapped to obtain the target prediction label. A prediction loss is generated based on the difference between the training label and the target prediction label. The loss gradient of the prediction loss is calculated, and the prediction error is obtained based on the loss gradient.

[0157] The mapping process is used to map the fused prediction results to a preset range in order to distinguish the final prediction results for different types of objects.

[0158] Specifically, when calculating the prediction error, the computer device first fuses the prediction results of the current round and the prediction results of previous rounds to obtain a fused prediction result. This fused prediction result is then mapped to a preset range to obtain the target prediction label. Next, the training label and the target prediction label are substituted into the loss function, and the gradient of the loss function is calculated to obtain the loss gradient of the prediction loss. Finally, the prediction error is obtained based on the loss gradient. The loss function is used to calculate the difference between the training label and the target prediction label.

[0159] In one embodiment, obtaining the prediction error based on the loss gradient includes: adjusting the loss gradient based on the loss adjustment step size to obtain the prediction error. Specifically, when generating the prediction error based on the loss gradient, the loss gradient can be adjusted based on the loss adjustment step size to obtain the prediction error. The loss adjustment step size is used to control the learning progress of the model and ensure the learning effect of the model.

[0160] In one embodiment, the prediction error can be calculated using the following formula:

[0161]

[0162]

[0163] Where ΔL represents the prediction error, also known as the prediction residual, and is the target for fitting the next decision tree. η represents the loss adjustment step size, also known as the learning rate. represents the loss function used to calculate the difference between the training labels and the target predicted labels; for example, a logarithmic function, an exponential function, etc., can be used. y represents the training labels. Indicates the target prediction label. f t (x) represents the prediction result of the t-th decision tree. The prediction result of the decision tree can be considered as a mapping from the input data to the output of the decision tree. When training the (t+1)-th decision tree (i.e., training the backward decision tree), the (t+1)-th decision tree should fit ΔL, that is, the prediction result of the (t+1)-th decision tree should be consistent with the input data. The closer the better. sigmoid represents the sigmoid activation function, used to map variables to the range of 0-1.

[0164] In the above embodiments, existing prediction results are first fused, and then the fused results are mapped to constrain the prediction results within a preset range, resulting in standardized prediction labels. The prediction error is obtained based on the loss gradient between the training labels and the predicted labels. Using this prediction error as the fitting target of the decision tree helps improve the overall model's prediction accuracy, making the predicted labels output by the model increasingly closer to the training labels.

[0165] In one embodiment, generating a backward decision tree based on training operation data and prediction error for the current decision tree includes:

[0166] Obtain the second basic decision tree; input the training operation data into the second basic decision tree to obtain the basic prediction results; generate loss information based on the difference between the basic prediction results and the prediction error; adjust the model parameters of the second basic decision tree based on the loss information until the third convergence condition is met, and obtain the backward decision tree of the current decision tree.

[0167] Specifically, similar to the process of generating the initial decision, when generating the backward decision tree, the computer device can acquire the second basic decision tree, use the training operation data as the input data of the second basic decision tree, use the prediction error as the expected output data of the second basic decision tree, and iteratively train the second basic decision tree until the third convergence condition is met, indicating that the training is complete and the backward decision tree is obtained.

[0168] Therefore, the computer device can input training data into the second basic decision tree to obtain basic prediction results. Based on the difference between the basic prediction results and the prediction error, loss information is generated. The loss information is then backpropagated to update the model parameters of the second basic decision tree, resulting in a new second basic decision tree. The training data is then input into the new second basic decision tree to obtain new basic prediction results. Based on the difference between the new basic prediction results and the prediction error, new loss information is generated and backpropagated to update the decision tree again. The decision tree parameters of the second basic decision tree are then adjusted again. This process is repeated iteratively, with the goal of each adjustment being to make the difference between the basic prediction results and the prediction error smaller and smaller until the third convergence condition is met, resulting in the backward decision tree of the current decision tree.

[0169] The second basic decision tree can be the same as or different from the first basic decision tree. The second basic decision tree can also be pre-set manually. It's understandable that even if the second and first basic decision trees are the same, the final trained decision trees will differ due to different training objectives. The third convergence condition can be at least one of the following: the difference between the basic prediction result and the prediction error is less than a preset difference, or the number of iterations is less than a preset number. It's understandable that the preset difference and number of iterations corresponding to different convergence conditions can be the same or different.

[0170] In the above embodiments, the training operation data is used as the input data of the backward decision tree, and the prediction error is used as the expected output of the backward decision tree. The prediction result of the backward decision tree obtained by training can be used as supplementary data for the prediction result of the forward decision tree. Thus, by fusing the prediction results of each decision tree, a more accurate prediction label can be obtained.

[0171] In one embodiment, such as Figure 4 As shown, an object classification method is provided. Taking the application of this method to a computer device as an example, the computer device can be one of the above-mentioned... Figure 1 Terminal 102 or server 104 in the reference. Figure 4 The object classification method includes the following steps:

[0172] Step S402: Obtain the target operation data corresponding to the target object; the target operation data is obtained based on the operation data of the target object in the target application.

[0173] Step S404: Input the target operation data into the object classification model to obtain the predicted label corresponding to the target object; the predicted label is used to determine the operation authority of the target object in the target application, and the predicted label is obtained by fusing the prediction results of each decision tree in the object classification model.

[0174] The training process of the object classification model includes the following steps: obtaining training operation data and training labels corresponding to the training objects; generating an initial decision tree based on the training operation data and training labels, and using the initial decision tree as the current decision tree; inputting the training operation data into the current decision tree to obtain the prediction result for this round, and generating a prediction error based on the training labels, the prediction result for this round, and the prediction results for previous rounds; generating a backward decision tree based on the training operation data and prediction error, using the backward decision tree as the current decision tree, and returning to the step of inputting the training operation data into the current decision tree to obtain the prediction result for this round, until the first convergence condition is met, resulting in multiple backward decision trees; and generating an object classification model based on the initial decision tree and each backward decision tree.

[0175] It is understood that the specific training process of the object classification model can be referred to the content of the aforementioned relevant embodiments, and will not be repeated here.

[0176] Specifically, computer devices can classify target objects based on a trained object classification model to determine the target object's operational permissions within the target application. The computer device can acquire target operation data corresponding to the target object, input this data into the object classification model, and then feed it into an initial decision tree and various backward decision trees. After data processing by the decision trees, the initial decision tree and each backward decision tree output prediction results. The object classification model then merges the prediction results from all decision trees to finally output the predicted label corresponding to the target object.

[0177] In one embodiment, after obtaining the prediction results output by the initial decision tree and each backward decision tree, the prediction results can be fused to obtain a fused prediction result, and then the fused prediction result can be mapped to obtain the prediction label.

[0178] The object classification method described above inputs the target operation data corresponding to the target object into an object classification model generated by multiple decision trees. The object classification model outputs the predicted label corresponding to the target object by fusing the prediction results of the initial decision tree and each backward decision tree. An object classification model can be trained based on training operation data and training labels. Classifying objects based on this model can improve the accuracy and efficiency of object classification. Furthermore, the object classification model includes multiple decision trees. Each backward decision tree is obtained by learning the error between the prediction results of the forward decision tree and the training label. The backward decision trees can improve the accuracy of the model based on the forward decision trees. Therefore, by fusing the prediction results of each decision tree, the object classification model can obtain a more accurate predicted label, further improving the accuracy of object classification.

[0179] In one embodiment, the method further includes:

[0180] When the predicted label is the target label, an object authentication request is generated and sent to the target terminal corresponding to the target object; the authentication information returned by the target terminal based on the object authentication request is obtained, and the authentication result of the target object is determined based on the authentication information; when the authentication result is that the authentication fails, the operation time of the target object in the target application is restricted.

[0181] The target label indicates that the model predicts a certain object belongs to the target type. The target type object has restricted access permissions within the target application. The object authentication request is used to determine the identity of the target object, further confirming whether it is indeed the target type object.

[0182] Specifically, to improve the accuracy of target object classification, computer devices can determine the target object's operational permissions in the target application based on a combination of predicted labels and authentication information. If the predicted label corresponding to the target object is not the target label, based on the object classification model, the target object is considered to have a low probability of belonging to the target type object, and its operations in the target application can be left unrestricted. If the predicted label corresponding to the target object is the target label, based on the object classification model, the target object is considered to have a high probability of belonging to the target type object. To further determine whether the target object is the target type object, the computer device can generate an object authentication request and send it to the target terminal corresponding to the target object. The target terminal can collect the target object's personal information as authentication information according to the object authentication request. For example, it can collect at least one of the target object's facial information, ID information, and communication contact information as authentication information, and feed the authentication information back to the computer device. The computer device can determine the target object's identity information based on the authentication information and obtain the authentication result of the target object. For example, it can match the authentication information with the registration information to obtain the authentication result. If the authentication result shows that the target object is not the target type object, the authentication result is determined to be successful; if the authentication result shows that the target object is the target type object, the authentication result is determined to be unsuccessful. Therefore, if the authentication result is that the authentication fails, the target object is determined to be a target type object, and it is necessary to restrict the operation of the target object in the target application. For example, restrict the operation time of the target object in the target application, and restrict the virtual resource transfer share of the target object in the target application (e.g., game currency, recharge, etc.).

[0183] In one embodiment, the authentication information can be facial information. The target object is authenticated based on the facial information. If the facial information matches the real-name identity information in the account registration information, it indicates that the target object is an adult and the authentication result is determined to be successful. If the facial information does not match the real-name identity information in the account registration information, the authentication result is determined to be unsuccessful, and it can be considered that the target object is a minor and the minor is using an adult's account in the target application.

[0184] In one embodiment, the predicted label can be represented by a probability. If the predicted label is greater than a preset probability, then the predicted label is determined to be the target label. The preset probability can be set according to actual needs, for example, to 0.5.

[0185] In the above embodiments, the object classification model is first used to preliminarily determine whether the target object is a target type object. If the preliminary determination result indicates that the target object is not a target type object, no further data processing is required, saving resources. If the preliminary determination result indicates that the target object is a target type object, the authentication information of the target object is further used to determine whether the target object is a target type object. If the authentication result again indicates that the target object is a target type object, then the target object is finally determined to be a target type object, thereby restricting the operation of the target object in the target application. Through the dual operation of model prediction and information authentication, the accuracy of object classification can be further improved.

[0186] In one specific embodiment, the object classification method of this application can be applied to game applications. According to national policies protecting minors, minors' actions in game applications are restricted. For example, minors can only play games for one hour each day from 8 PM to 9 PM on Fridays to Sundays or public holidays; they cannot play games at other times. (Reference) Figure 5 In the object classification method of this application, an ensemble model (i.e., an object classification model) is used to identify whether gamers playing during holiday periods are suspected minors. If the identification result shows that the gamer is not a suspected minor, then the gamer is not restricted from playing games. If the identification result shows that the gamer is a suspected minor, then a facial recognition pop-up is applied to the suspected minor during non-game hours, and real-name authentication is performed based on the information collected from the facial recognition pop-up. If the real-name authentication is successful, the gamer can be considered an adult, and the gamer is not restricted from playing games. If the real-name authentication fails, the gamer can be considered a minor, and the gamer is restricted from playing games.

[0187] refer to Figure 6 An ensemble model for identifying suspicious underage players is trained based on positive and negative samples and the player features corresponding to each sample. The final trained ensemble model can be called a high-risk model. The player features of the player to be identified are input into the high-risk model, and the high-risk model predicts the probability that the player to be identified is a suspicious underage player. If the predicted probability is greater than or equal to 0.5, the player to be identified is determined to be a suspicious underage player.

[0188] The model's training data consists of positive samples representing underage players and negative samples representing adult players. Player characteristics are derived from game-reported logs, specifically including basic attribute data (i.e., player profile attributes), match data, match duration data, suspicious device data, underage friend data, registration data, and game category data. Basic attribute data includes pre-real-name age, pre-real-name gender, pre-real-name device, pre-real-name model, pre-real-name network, pre-real-name regional classification (first-tier, second-tier, third-tier cities, etc.), and pre-real-name operating system. Match data includes win rate, draw rate, loss rate, win rate for each mode, total number of matches, total number of days with matches, average daily number of matches (total number of matches / number of days with matches) for July and August, average daily number of matches on weekdays (total number of matches on weekdays / number of weekdays with matches), average daily number of matches on weekends & holidays (total number of matches on weekends & holidays / number of weekends & holidays with matches), percentage of average daily number of matches on weekends & holidays, average number of matches per hour on weekends & holidays, and percentage of average daily number of matches per hour on weekends & holidays for July and August. Duration data includes average daily duration on weekdays, average daily duration on weekends & holidays, duration for different time slots on weekdays, and duration for weekends & holidays. The data includes the duration of each time period during the holiday, and the percentage of time spent in the morning, noon, afternoon, evening, and late night each day. Suspicious device data includes the number of devices that have logged in historically, the number of natural persons who have been linked historically, the number of accounts logged in by devices from July to August, the number of accounts linked by natural persons from July to August, and whether there were suspicious devices (devices with a history of underage logins) logged in during July and August. Data on underage friends includes the number of underage friends, the proportion of underage friends, and the total intimacy level of underage friends. Registration and game category data includes the registration year, the earliest registration year, the number of active days for PC games in the past month, the active duration for PC games in the past month, the number of active days for mobile games in the past month, and the active duration for mobile games in the past month.

[0189] The training process of the ensemble model: The ensemble model is an additive model that sequentially trains a set of CART (Classification and Regression Tree) decision trees, resulting in a strong learner based on all the decision trees. During decision tree training, each decision tree fits the negative gradient direction of the current loss function and learns the residuals of the previous decision tree predictions (the residuals refer to the differences between the predicted and actual values). The output data of the ensemble model is obtained by fusing the predictions from each decision tree.

[0190] Key parameters used during model training: Random number seed (needs to specify the starting point of iterations during multiple iterations): 45; Number of decision trees: 100; Purity (represents the probability that a randomly selected sample will be correctly classified in the model): Gini coefficient; Maximum tree depth (represents the distance between a leaf node and the root node; the maximum tree depth is the critical point at which decision tree iterations stop; when the decision tree depth reaches the maximum, it will stop splitting): 6; Maximum number of feature bins: 32; Validation set ratio (when constructing the model, the dataset is split into training and validation sets; the training set data is used to construct the model, and the validation set data is used to test the model's accuracy): 20%.

[0191] The validation results of the high-risk model based on the validation set are shown in Table 7.

[0192] Table 7

[0193] AUC value 0.961 Accuracy 0.703 Recall rate 0.598 F1 value 0.613 False positive rate 4.20％ Proportion of simulation delivery of minors 13.84％

[0194] The simulated underage ratio refers to the proportion of underage individuals among those who have undergone real-name authentication.

[0195] Furthermore, the effects of deploying the high-risk model online are shown in Table 8.

[0196] Table 8

[0197]

[0198] As shown in Table 8, in the high-risk model's face verification pop-up effect from December 1st to 7th, the daily average proportion of minors verifying their identities was 11.5%, which is 9.58 times that of the overall market, and the pop-up interception rate was 1.25 times that of the overall market. The pop-up interception rate is calculated as follows: (Number of objects whose faces failed verification in the pop-up / Number of objects in the pop-up) * 100%.

[0199] In the above embodiments, by combining object profile attributes, in-game data during holidays, characteristics of holiday game periods, recent game duration features, active behavior in PC games, suspicious devices, and natural persons, an ensemble model algorithm is used for modeling, and the model training parameters are optimized to train an ensemble model. This ensemble model can accurately and efficiently identify underage gamers among those playing games during holidays (such as summer vacation and National Day holidays), thereby restricting underage players from playing games and preventing minors from becoming addicted to games.

[0200] It is understood that the object classification method of this application can also be applied to other applications, such as video applications, e-commerce applications, etc.

[0201] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0202] Based on the same inventive concept, this application also provides an object classification apparatus for implementing the object classification method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more object classification apparatus embodiments provided below can be found in the limitations of the object classification method described above, and will not be repeated here.

[0203] In one embodiment, such as Figure 7 As shown, an object classification device 700 is provided, including: a training data acquisition module 702, an initial decision tree generation module 704, a prediction error calculation module 706, a backward decision tree generation module 708, and an object classification model generation module 710, wherein:

[0204] The training data acquisition module 702 is used to acquire the training operation data and training labels corresponding to the training object; the training operation data is obtained based on the operation data of the training object in the target application, and the training labels are used to determine the operation permissions of the training object in the target application.

[0205] The initial decision tree generation module 704 is used to generate an initial decision tree based on the training operation data and training labels, and to use the initial decision tree as the current decision tree.

[0206] The prediction error calculation module 706 is used to input the training operation data into the current decision tree to obtain the prediction result of the current round, and generate the prediction error based on the training label, the prediction result of the current round and the prediction results of the previous round.

[0207] The backward decision tree generation module 708 is used to generate a backward decision tree of the current decision tree based on the training operation data and the prediction error. The backward decision tree is used as the current decision tree. The step of inputting the training operation data into the current decision tree to obtain the prediction result of this round is returned to the execution until the first convergence condition is met, and multiple backward decision trees are obtained.

[0208] The object classification model generation module 710 is used to generate an object classification model based on the initial decision tree and each backward decision tree. The object classification model is used to input the target operation data corresponding to the target object into the initial decision tree and each backward decision tree, and to fuse the prediction results of the initial decision tree and each backward decision tree to obtain the prediction label corresponding to the target object.

[0209] The aforementioned object classification device can train an object classification model based on training data and training labels. Classifying objects using this model improves classification accuracy and efficiency. Furthermore, the object classification model includes multiple decision trees. Each backward decision tree is obtained by learning the error between the prediction results of the forward decision trees and the training labels. The backward decision trees can improve the model's accuracy based on the forward decision trees. Therefore, by fusing the prediction results of each decision tree, the object classification model can obtain more accurate predicted labels, further improving object classification accuracy.

[0210] In one embodiment, the training operation data includes at least one of the following: object attribute features of the training object, operation interaction features, operation duration features, device login features, registration time features, operating platform features, and target association features between the training object and the target associated object. The target associated object is the associated object among the various associated objects of the training object that has the target operation permission.

[0211] In one embodiment, the training data acquisition module is further configured to perform statistical analysis on the operation data of the training object in the target application to obtain initial operation data; the initial operation data includes multiple initial operation features; multiple target operation features are determined from each initial operation feature; the target operation features are cross-referenced to obtain cross operation data; and training operation data is obtained based on the initial operation data and the cross operation data.

[0212] In one embodiment, the initial decision tree generation module includes:

[0213] The basic decision tree generation unit is used to generate the first basic decision tree based on the training operation data and training labels;

[0214] The data prediction unit is used to input the training operation data into the first basic decision tree to obtain the initial prediction results;

[0215] The parameter adjustment unit is used to adjust the decision tree parameters of the first basic decision tree based on the initial prediction results and training labels until the second convergence condition is met, thus obtaining the initial decision tree.

[0216] In one embodiment, the training objects include multiple training operation features, and the training operation data includes multiple training operation features. The basic decision tree generation unit is further configured to bin the training operation features corresponding to the same category, obtaining feature bin sets corresponding to each category; the feature bin sets include multiple feature bins; calculate the first split coefficient corresponding to each current feature bin based on the training labels corresponding to each current feature bin; determine the second split coefficient based on the first split coefficient of each current feature bin corresponding to the same category, obtaining the second split coefficient corresponding to each category; determine the target category from each category based on the second split coefficient; determine the target feature bin from each current feature bin corresponding to the target category based on the first split coefficient of each current feature bin corresponding to the target category; generate decision tree nodes based on the target feature bins; update the feature bins corresponding to the remaining categories based on the target feature bins, obtaining updated feature bins corresponding to each category; use the updated feature bins as the current feature bins; return to the step of calculating the first split coefficient corresponding to each current feature bin based on the number of features and training labels corresponding to each current feature bin, until a first preset condition is met, obtaining each decision tree node; and generate a first basic decision tree based on each decision tree node.

[0217] In one embodiment, the basic decision tree generation unit is further configured to perform initial binning on each training operation feature corresponding to the current category, obtain multiple candidate bins, and use the candidate bins as the current bins; calculate the binning merging coefficients corresponding to adjacent current bins based on the training labels corresponding to each current bin; merge adjacent current bins based on the binning merging coefficients to obtain multiple merged bins; use the merged bins as the current bins, and return to the step of calculating the binning merging coefficients corresponding to adjacent current bins based on the training labels corresponding to each current bin, until the second preset condition is met, resulting in multiple feature bins corresponding to the current category; and obtain a feature bin set corresponding to the current category based on each feature bin corresponding to the current category.

[0218] In one embodiment, the basic decision tree generation unit is further configured to determine a reference feature bin from each current feature bin corresponding to the same category, and divide each current feature bin corresponding to the same category into a first category bin and a second category bin based on the reference feature bin; obtain the label ratios corresponding to the first category bin and the second category bin based on the total number of labels corresponding to the first category bin and the second category bin respectively; obtain the label distribution coefficient corresponding to the current category bin based on the number of labels corresponding to various training labels in the current category bin and the total number of labels corresponding to the current category bin; and obtain the first split coefficient corresponding to the reference feature bin based on the label ratios and label distribution coefficients corresponding to the first category bin and the second category bin.

[0219] In one embodiment, the basic decision tree generation unit is further configured to obtain the first split coefficient with the smallest value from each first split coefficient corresponding to the same category as the second split coefficient, thereby obtaining the second split coefficients corresponding to each category; obtain the category corresponding to the second split coefficient with the smallest value as the target category; and obtain the current feature bin corresponding to the first split coefficient with the smallest value from each first split coefficient corresponding to the target category as the target feature bin.

[0220] In one embodiment, the prediction error calculation module is further configured to fuse the prediction results of the current round and the prediction results of the previous round to obtain a fused prediction result, perform mapping processing on the fused prediction result to obtain the target prediction label; generate a prediction loss based on the difference between the training label and the target prediction label; calculate the loss gradient of the prediction loss, and obtain the prediction error based on the loss gradient.

[0221] In one embodiment, the backward decision tree generation module is further configured to obtain a second basic decision tree; input training operation data into the second basic decision tree to obtain basic prediction results; generate loss information based on the difference between the basic prediction results and the prediction error; adjust the model parameters of the second basic decision tree based on the loss information until the third convergence condition is met, and obtain the backward decision tree of the current decision tree.

[0222] In one embodiment, such as Figure 8 As shown, an object classification device 800 is provided, including: a data acquisition module 802 and a label prediction module 804, wherein:

[0223] The data acquisition module 802 is used to acquire the target operation data corresponding to the target object; the target operation data is obtained based on the operation data of the target object in the target application.

[0224] The label prediction module 804 is used to input the target operation data into the object classification model to obtain the predicted label corresponding to the target object; the predicted label is used to determine the operation permission of the target object in the target application, and the predicted label is obtained by fusing the prediction results of each decision tree in the object classification model;

[0225] The training process of the object classification model includes the following steps: obtaining training operation data and training labels corresponding to the training objects; generating an initial decision tree based on the training operation data and training labels, and using the initial decision tree as the current decision tree; inputting the training operation data into the current decision tree to obtain the prediction result for this round, and generating a prediction error based on the training labels, the prediction result for this round, and the prediction results for previous rounds; generating a backward decision tree based on the training operation data and prediction error, using the backward decision tree as the current decision tree, and returning to the step of inputting the training operation data into the current decision tree to obtain the prediction result for this round, until the first convergence condition is met, resulting in multiple backward decision trees; and generating an object classification model based on the initial decision tree and each backward decision tree.

[0226] The aforementioned object classification device inputs the target operation data corresponding to the target object into an object classification model generated by multiple decision trees. The object classification model outputs a predicted label for the target object by fusing the prediction results of the initial decision tree and each backward decision tree. The object classification model can be trained based on training operation data and training labels. Classifying objects based on this model improves the accuracy and efficiency of object classification. Furthermore, the object classification model includes multiple decision trees. Each backward decision tree is obtained by learning the error between the prediction results of the forward decision tree and the training labels. The backward decision trees can improve the accuracy of the model based on the forward decision trees. Therefore, by fusing the prediction results of each decision tree, the object classification model can obtain a more accurate predicted label, further improving the accuracy of object classification.

[0227] In one embodiment, such as Figure 9 As shown, the object classification device 800 also includes:

[0228] The object authentication module 806 is used to generate an object authentication request when the predicted tag is the target tag, and send the object authentication request to the target terminal corresponding to the target object; obtain the authentication information returned by the target terminal based on the object authentication request, and determine the authentication result of the target object based on the authentication information; when the authentication result is authentication failure, restrict the operation of the target object in the target application.

[0229] Each module in the aforementioned object classification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0230] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data such as object classification models and training operation features. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements an object classification method.

[0231] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 11 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements an object classification method. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0232] Those skilled in the art will understand that Figure 10 , 11The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0233] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0234] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0235] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0236] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0237] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0238] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0239] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method of object classification, characterized by, The method comprises: obtaining training operation data corresponding to a training object and a training label; the training operation data is obtained based on operation data of the training object in a target application, and the training label is used to determine operation permission of the training object in the target application; the training operation data comprises at least one of object attribute features, operation interaction features, operation time length features, device login features, registration time features, running platform features and target association features between the training object and a target associated object; the target associated object is an associated object of the training object that has target operation permission; the object attribute features comprise at least one of object age, object gender, terminal identification of a terminal running the target application, terminal model, terminal running network, terminal running area and terminal running operating system; the operation interaction features are obtained based on interaction operations of the object and other objects in the target application; the operation time length features are used to represent operation time length of the object in the target application; the device login features are obtained based on a terminal and an account used by the object to log in the target application; the registration time features are used to represent registration time and registration time length of the object in the target application; and the running platform features are used to represent activity of the object in operating the target application on different running platforms; the associated object of the object refers to a friend of the object in the target application; generating an initial decision tree based on the training operation data and the training label, and taking the initial decision tree as a current decision tree; inputting the training operation data into the current decision tree to obtain a current round prediction result, generating a prediction error based on the training label, the current round prediction result and historical round prediction results; generating a backward decision tree of the current decision tree based on the training operation data and the prediction error, taking the backward decision tree as the current decision tree, returning to the step of inputting the training operation data into the current decision tree to obtain the current round prediction result, and executing until a first convergence condition is met, to obtain a plurality of backward decision trees; generating an object classification model based on the initial decision tree and each backward decision tree; the object classification model is used to input target operation data corresponding to a target object into the initial decision tree and each backward decision tree, and fuse prediction results of the initial decision tree and each backward decision tree to obtain a prediction label corresponding to the target object.

2. The method of claim 1, wherein, The method comprises: statistically analyzing operation data of the training object in the target application to obtain initial operation data; the initial operation data comprises a plurality of initial operation features; determining a plurality of target operation features from each initial operation feature; crossing each target operation feature to obtain cross operation data; obtaining the training operation data based on the initial operation data and the cross operation data.

3. The method of claim 1, wherein, The method comprises: generating a first basic decision tree based on the training operation data and the training label; inputting the training operation data into the first basic decision tree to obtain an initial prediction result; Adjusting a decision tree parameter of the first base decision tree based on the initial prediction result and the training label until a second convergence condition is met, to obtain the initial decision tree.

4. The method of claim 3, wherein, The training objects include a plurality of training operation features. The training operation data includes a plurality of training operation features. The method comprises the following steps: Each training operation feature corresponding to the same category is subjected to binning processing to obtain a feature binning set corresponding to each category respectively; the feature binning set includes a plurality of feature bins. A first split coefficient corresponding to each current feature bin is calculated based on the training label corresponding to each current feature bin. A second split coefficient is determined based on the first split coefficient of each current feature bin corresponding to the same category, to obtain a second split coefficient corresponding to each category respectively, a target category is determined from each category based on the second split coefficient, and a target feature bin is determined from each current feature bin corresponding to the target category based on the first split coefficient of each current feature bin corresponding to the target category. A decision tree node is generated based on the target feature bin. The feature bins corresponding to the remaining categories are updated based on the target feature bin to obtain updated feature bins corresponding to each category, and the updated feature bins are taken as current feature bins, and the step of calculating the first split coefficient corresponding to each current feature bin based on the number of features and the training label corresponding to each current feature bin is executed until a first preset condition is met, to obtain each decision tree node.

5. The method of claim 4, wherein, The first base decision tree is generated based on each decision tree node. The method comprises the following steps: Each training operation feature corresponding to the same category is subjected to binning processing to obtain a feature binning set corresponding to each category respectively; the feature binning set includes a plurality of feature bins. Each training operation feature corresponding to the current category is subjected to initial binning to obtain a plurality of candidate bins, and the candidate bins are taken as current bins. A bin merging coefficient corresponding to adjacent current bins is calculated based on the training label corresponding to each current bin. Adjacent current bins are subjected to bin merging based on the bin merging coefficient to obtain a plurality of merged bins.

6. The method of claim 4, wherein, The merged bins are taken as current bins, and the step of calculating the bin merging coefficient corresponding to adjacent current bins based on the training label corresponding to each current bin is executed until a second preset condition is met, to obtain a plurality of feature bins corresponding to the current category. The feature binning set corresponding to the current category is obtained based on each feature bin corresponding to the current category. The method comprises the following steps: A reference feature bin is determined from each current feature bin corresponding to the same category, and each current feature bin corresponding to the same category is divided into a first type of bin and a second type of bin based on the reference feature bin; A label proportion corresponding to the first type of bin and the second type of bin is obtained based on the total number of labels corresponding to the first type of bin and the second type of bin respectively. obtaining a label distribution coefficient corresponding to the current class bin based on the number of labels corresponding to various training labels in the current class bin and the total number of labels corresponding to the current class bin; obtaining a first split coefficient corresponding to the reference feature bin based on the label proportion and the label distribution coefficient corresponding to the first class bin and the second class bin.

7. The method of claim 4, wherein, The method comprises the following steps: obtaining a second split coefficient corresponding to each category based on the first split coefficient corresponding to each current feature bin of the same category, determining a target category from each category based on the second split coefficient, and determining a target feature bin from each current feature bin corresponding to the target category based on the first split coefficient corresponding to each current feature bin of the target category, comprising: obtaining the minimum first split coefficient as the second split coefficient from the first split coefficient corresponding to each category, and obtaining the second split coefficient corresponding to each category; obtaining the category corresponding to the minimum second split coefficient as the target category; 8. The method of claim 1, wherein, obtaining the current feature bin corresponding to the minimum first split coefficient as the target feature bin from the first split coefficient corresponding to each category of the target category. The method comprises the following steps: obtaining a fusion prediction result by fusing the current round prediction result and the historical round prediction result, performing mapping processing on the fusion prediction result to obtain a target prediction label; generating a prediction loss based on the difference between the training label and the target prediction label; 9. The method of claim 1, wherein, calculating the loss gradient of the prediction loss to obtain the prediction error based on the loss gradient. The method comprises the following steps: obtaining a second basic decision tree; inputting the training operation data into the second basic decision tree to obtain a basic prediction result; 10. A method of object classification characterized by, generating loss information based on the difference between the basic prediction result and the prediction error, adjusting the model parameters of the second basic decision tree based on the loss information until a third convergence condition is met to obtain a backward decision tree of the current decision tree. The method comprises the following steps: obtaining target operation data corresponding to a target object; the target operation data is obtained based on the operation data of the target object in a target application; inputting the target operation data into an object classification model to obtain a prediction label corresponding to the target object; the prediction label is used to determine the operation permission of the target object in the target application, and the prediction label is obtained by fusing the prediction results of each decision tree in the object classification model; The training process of the object classification model comprises the following steps: The method comprises the following steps: obtaining training operation data and training labels corresponding to a training object, wherein the training operation data comprises at least one of object attribute features, operation interaction features, operation time length features, device login features, registration time features, running platform features, and target association features between the training object and a target associated object, the target associated object being an associated object of the training object that has target operation permissions; the object attribute features comprise at least one of object age, object gender, terminal identification of a terminal running a target application, terminal model, terminal running network, terminal running area, and terminal running operating system; the operation interaction features are obtained based on interactions between the object and other objects in the target application; the operation time length features are used to represent operation time length of the object in the target application; the device login features are obtained based on a terminal and an account used by the object to log in to the target application; the registration time features are used to represent registration time and registration time length of the object in the target application; and the running platform features are used to represent activity of the object in operating the target application on different running platforms; the associated object of the object refers to a friend of the object in the target application; generating an initial decision tree based on the training operation data and the training labels, and taking the initial decision tree as a current decision tree; inputting the training operation data into the current decision tree to obtain a current round of prediction results, generating prediction errors based on the training labels, the current round of prediction results, and historical round of prediction results; generating a backward decision tree of the current decision tree based on the training operation data and the prediction errors, taking the backward decision tree as the current decision tree, and returning to the step of inputting the training operation data into the current decision tree to obtain the current round of prediction results until a first convergence condition is met, and obtaining a plurality of backward decision trees; 11. The method of claim 10, wherein, generating the object classification model based on the initial decision tree and each backward decision tree. The method further comprises the following steps: when the prediction label is a target label, generating an object authentication request, and sending the object authentication request to a target terminal corresponding to the target object; obtaining authentication information returned by the target terminal according to the object authentication request, and determining an authentication result of the target object based on the authentication information; 12. An object classification apparatus characterized by comprising: when the authentication result is authentication failure, limiting operation of the target object in the target application. The device comprises the following components: The training data acquisition module is configured to acquire training operation data corresponding to a training object and a training label. The training operation data is obtained based on operation data of the training object in a target application. The training label is used to determine an operation permission of the training object in the target application. The training operation data includes at least one of an object attribute feature of the training object, an operation interaction feature, an operation time length feature, a device login feature, a registration time feature, a running platform feature, and a target association feature between the training object and a target associated object. The target associated object is an associated object of the training object that has a target operation permission. The object attribute feature includes at least one attribute information of an object age, an object gender, a terminal identifier of a terminal running the target application, a terminal model, a terminal running network, a terminal running area, and a terminal running operating system. The operation interaction feature is obtained based on an interaction operation of the object in the target application and other objects. The operation time length feature is used to represent an operation time length of the object in the target application. The device login feature is obtained based on a terminal and an account used by the object to log in to the target application. The registration time feature is used to represent a registration time and a registration time length of the object in the target application. The running platform feature is used to represent an activity level of the object in operating the target application on different running platforms. The associated object of the object refers to a friend of the object in the target application. The initial decision tree generation module is configured to generate an initial decision tree based on the training operation data and the training label, and to use the initial decision tree as a current decision tree. The prediction error calculation module is configured to input the training operation data into the current decision tree to obtain a current round prediction result, and to generate a prediction error based on the training label, the current round prediction result, and a historical round prediction result. The backward decision tree generation module is configured to generate a backward decision tree of the current decision tree based on the training operation data and the prediction error, to use the backward decision tree as the current decision tree, and to return to the step of inputting the training operation data into the current decision tree to obtain the current round prediction result until a first convergence condition is met, and to obtain a plurality of backward decision trees. The object classification model generation module is configured to generate an object classification model based on the initial decision tree and each backward decision tree. The object classification model is used to input target operation data corresponding to a target object into the initial decision tree and each backward decision tree, and to fuse prediction results of the initial decision tree and each backward decision tree to obtain a prediction label corresponding to the target object.

13. The object classification apparatus according to claim 12, characterized by The training data acquisition module is further configured to statistically analyze operation data of the training object in the target application to obtain initial operation data. The initial operation data includes a plurality of initial operation features. A plurality of target operation features are determined from each initial operation feature. Each target operation feature is cross-featured to obtain cross-operation data. The training operation data is obtained based on the initial operation data and the cross-operation data.

14. The object classification apparatus according to claim 12, wherein The initial decision tree generation module includes: a basic decision tree generation unit, configured to generate a first basic decision tree based on the training operation data and the training label; a data prediction unit, configured to input the training operation data into the first basic decision tree to obtain an initial prediction result; a parameter adjustment unit, configured to adjust a decision tree parameter of the first basic decision tree based on the initial prediction result and the training label until a second convergence condition is met, to obtain the initial decision tree.

15. The object classification apparatus according to claim 14, wherein The training objects include a plurality of training operation features; the basic decision tree generation unit is further configured to perform binning processing on each training operation feature corresponding to a same category to obtain a feature binning set corresponding to each category respectively; the feature binning set includes a plurality of feature bins; a first split coefficient corresponding to each current feature bin is calculated based on a training label corresponding to the current feature bin; a second split coefficient is determined based on the first split coefficients of each current feature bin corresponding to a same category, to obtain a second split coefficient corresponding to each category respectively, a target category is determined from each category based on the second split coefficient, a target feature bin is determined from each current feature bin corresponding to the target category based on the first split coefficients of the current feature bins, and a decision tree node is generated based on the target feature bin; updated feature bins corresponding to each category are obtained by updating the feature bins corresponding to the remaining categories based on the target feature bin, and the updated feature bins are taken as current feature bins, and the step of calculating the first split coefficient corresponding to each current feature bin based on the feature quantity and the training label corresponding to the current feature bin is returned to be executed until a first preset condition is met, to obtain each decision tree node; and the first basic decision tree is generated based on the decision tree nodes.

16. The object classification apparatus according to claim 15, wherein The basic decision tree generation unit is further configured to perform initial binning on each training operation feature corresponding to a current category to obtain a plurality of candidate bins, and the candidate bins are taken as current bins; a bin merging coefficient corresponding to adjacent current bins is calculated based on a training label corresponding to each current bin; bin merging is performed on the adjacent current bins based on the bin merging coefficient, to obtain a plurality of merged bins; the merged bins are taken as current bins, and the step of calculating the bin merging coefficient corresponding to adjacent current bins based on a training label corresponding to each current bin is returned to be executed until a second preset condition is met, to obtain a plurality of feature bins corresponding to the current category; a feature binning set corresponding to the current category is obtained based on each feature bin corresponding to the current category.

17. The object classification apparatus according to claim 15, wherein The basic decision tree generation unit is further configured to determine a reference feature bin from each current feature bin corresponding to the same category, divide each current feature bin corresponding to the same category into a first type of bin and a second type of bin based on the reference feature bin, obtain label proportions of the first type of bin and the second type of bin corresponding to the label total numbers of the first type of bin and the second type of bin respectively, obtain a label distribution coefficient of a current category bin based on label quantities of various training labels corresponding to the current category bin and a label total number corresponding to the current category bin, and obtain a first split coefficient corresponding to the reference feature bin based on the label proportions and the label distribution coefficients of the first type of bin and the second type of bin.

18. The object classification apparatus according to claim 15, wherein The basic decision tree generation unit is further configured to obtain, as a second split coefficient, a first split coefficient with the smallest value from each first split coefficient corresponding to the various categories, obtain a category corresponding to the second split coefficient with the smallest value as the target category, obtain, as the target feature bin, a current feature bin corresponding to the first split coefficient with the smallest value from each first split coefficient corresponding to the target category. The prediction error calculation module is further configured to obtain a fusion prediction result by fusing a current round prediction result and a historical round prediction result, perform mapping processing on the fusion prediction result to obtain a target prediction label, generate a prediction loss based on a difference between the training label and the target prediction label, calculate a loss gradient of the prediction loss, and obtain the prediction error based on the loss gradient.

19. The object classification apparatus according to claim 12, wherein, The backward decision tree generation module is further configured to obtain a second basic decision tree, input the training operation data into the second basic decision tree to obtain a basic prediction result, generate loss information based on a difference between the basic prediction result and the prediction error, adjust model parameters of the second basic decision tree based on the loss information, and obtain a backward decision tree of a current decision tree until a third convergence condition is met.

20. The object classification apparatus according to claim 12, wherein, The apparatus comprises:

21. An object classification apparatus characterized by comprising: a data acquisition module configured to acquire target operation data corresponding to a target object, the target operation data being obtained based on operation data of the target object in a target application; a label prediction module configured to input the target operation data into an object classification model to obtain a prediction label corresponding to the target object, the prediction label being used to determine an operation permission of the target object in the target application, and the prediction label being obtained by fusing prediction results of various decision trees in the object classification model; a training process of the object classification model comprises the following steps: ​ The training operation data of the training object includes at least one of object attribute features, operation interaction features, operation time length features, device login features, registration time features, running platform features, and target association features between the training object and a target associated object, the target associated object being an associated object of the training object that has target operation permission; the object attribute features include at least one of object age, object gender, terminal identifier of a terminal running a target application, terminal model, terminal running network, terminal running area, and terminal running operating system, the operation interaction features being obtained based on interaction operations of the object and other objects in the target application, the operation time length features being used to represent operation time length of the object in the target application, the device login features being obtained based on a terminal and an account used by the object to log in the target application, the registration time features being used to represent registration time and registration time length of the object in the target application, and the running platform features being used to represent activity of the object in operating the target application on different running platforms; the associated object of the object refers to a friend of the object in the target application. The initial decision tree is generated based on the training operation data and the training label, and the initial decision tree is used as a current decision tree. The training operation data is input into the current decision tree to obtain a current round prediction result, a prediction error is generated based on the training label, the current round prediction result, and historical round prediction results. A backward decision tree of the current decision tree is generated based on the training operation data and the prediction error, the backward decision tree is used as the current decision tree, and the step of inputting the training operation data into the current decision tree to obtain the current round prediction result is executed until a first convergence condition is met, and a plurality of backward decision trees are obtained. The object classification model is generated based on the initial decision tree and each backward decision tree.

22. The object classification apparatus according to claim 21, wherein The apparatus further includes: An object authentication module configured to generate an object authentication request when the prediction label is a target label, send the object authentication request to a target terminal corresponding to the target object, obtain authentication information returned by the target terminal according to the object authentication request, determine an authentication result of the target object based on the authentication information, and limit operation of the target object in the target application when the authentication result is authentication failure. 23.A computer device, comprising a memory and a processor, wherein the memory stores a computer program. The processor executes the computer program to implement the method in any one of claims 1 to 11.

24. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 11.

25. A computer program product comprising a computer program, characterised in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 11. The computer program is executed by the processor to implement the method in any one of claims 1 to 11.

Images