A data processing method and device, computer equipment and storage medium

Abstract

Embodiments of the present application disclose a data processing method and device, computer equipment and a storage medium, comprising: obtaining N sample advertisements and N sample data pairs corresponding to the N sample advertisements; determining a root node based on the N sample advertisements, taking a sample advertisement attribute feature of each sample advertisement in the root node as an initial attribute feature, determining a first split point set corresponding to a feature type of the initial attribute feature, determining a first best split point from the first split point set, dividing the N sample advertisements according to a split condition indicated by the first best split point, and obtaining an initial probability boosting tree; and iteratively training the initial probability boosting tree based on a tree convergence condition and sample advertisement conversion data of each sample advertisement in the N sample advertisements, to obtain a target probability boosting tree for predicting a conversion data probability distribution of a target advertisement. The embodiments of the present application can improve the prediction speed of the probability distribution.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and particularly relates to a data processing method and device, computer equipment and storage medium. BACKGROUND

[0002] Traditional probability estimation methods (for example, Gaussian process regression) are mainly used for time series estimation. Since the covariance matrix between all data points in the entire sequence needs to be calculated when the distribution process of the data points in the estimated time sequence is estimated, a large amount of calculation is required when the probability distribution is predicted, so that the traditional probability estimation method is not suitable for a large number of application scenarios. Therefore, the traditional probability estimation method has the disadvantages of large amount of calculation and poor flexibility, and the computer equipment needs to spend a large amount of time when predicting the probability distribution, so that the prediction speed of the probability distribution is reduced. SUMMARY

[0003] The embodiments of the present application provide a data processing method and device, computer equipment and storage medium, which can improve the prediction speed of the probability distribution.

[0004] In one aspect, the embodiments of the present application provide a data processing method, comprising:

[0005] N sample advertisements and N sample data pairs corresponding to the N sample advertisements are obtained; N is a positive integer; one sample advertisement corresponds to one sample data pair, and one sample data pair is determined by a sample advertisement attribute feature and sample advertisement conversion data;

[0006] Based on the N sample advertisements, a root node for constructing an initial probability boosting tree is determined, the sample advertisement attribute feature of each sample advertisement in the root node is taken as an initial attribute feature, a first split point set corresponding to the feature type of the initial attribute feature is determined, a first best split point is determined from the first split point set, the N sample advertisements are divided according to the split condition indicated by the first best split point, and an initial probability boosting tree is obtained;

[0007] Based on the tree convergence condition associated with the initial probability boosting tree and the sample advertisement conversion data of each sample advertisement in the N sample advertisements, the initial probability boosting tree is iteratively trained to obtain a target probability boosting tree for predicting the conversion data probability distribution of the target advertisement.

[0008] In one aspect, the embodiments of the present application provide a data processing method, comprising:

[0009] A delivery feature request for a target advertisement is obtained; the delivery feature request includes a target advertisement attribute feature of the target advertisement;

[0010] obtain a target probability boosting tree associated with the target advertisement, input the target advertisement attribute feature into the target probability boosting tree, and output a conversion data probability distribution of the target advertisement by the target probability boosting tree; the target probability boosting tree is obtained by iteratively training an initial probability boosting tree based on a tree convergence condition and sample advertisement conversion data of each sample advertisement in N sample advertisements; the N sample advertisements are used to construct a root node of the initial probability boosting tree; the initial probability boosting tree is obtained by dividing the N sample advertisements based on a split condition indicated by a first best split point; the first best split point is determined based on sample advertisement attribute features of each sample advertisement in the root node; N is a positive integer;

[0011] determine target advertisement conversion data of the target advertisement based on the conversion data probability distribution of the target advertisement.

[0012] The embodiment of the application provides a data processing device, comprising:

[0013] a sample advertisement obtaining module configured to obtain N sample advertisements and N sample data pairs corresponding to the N sample advertisements; N is a positive integer; one sample advertisement corresponds to one sample data pair, and one sample data pair is determined by sample advertisement attribute features and sample advertisement conversion data;

[0014] an initial tree determining module configured to determine a root node used to construct an initial probability boosting tree based on the N sample advertisements, determine a first split point set corresponding to feature types of initial attribute features by taking sample advertisement attribute features of each sample advertisement in the root node as the initial attribute features, determine a first best split point from the first split point set, divide the N sample advertisements according to a split condition indicated by the first best split point, and obtain the initial probability boosting tree;

[0015] an iterative training module configured to iteratively train the initial probability boosting tree based on a tree convergence condition associated with the initial probability boosting tree and sample advertisement conversion data of each sample advertisement in the N sample advertisements, and obtain a target probability boosting tree used to predict a conversion data probability distribution of a target advertisement.

[0016] The initial tree determining module comprises:

[0017] a first split point determining unit configured to determine a root node used to construct an initial probability boosting tree based on the N sample advertisements, determine a first split point set corresponding to feature types of initial attribute features by taking sample advertisement attribute features of each sample advertisement in the root node as the initial attribute features, and determine a first best split point from the first split point set;

[0018] The first dividing unit is configured to divide the sample advertisements in the root node according to a split condition indicated by the first best split point, to obtain a first type node and a second type node;

[0019] The stop condition obtaining unit is configured to take the first type node and the second type node as to-be-split nodes respectively, and obtain split stop conditions associated with the initial probability boosting tree;

[0020] The second split point determining unit is configured to, when the to-be-split node does not satisfy the split stop condition, take a sample advertisement attribute feature of each sample advertisement in the to-be-split node as a target attribute feature, determine a second split point set consistent with a feature type of the target attribute feature, and determine a second best split point from the second split point set.

[0021] The second dividing unit is configured to divide the sample advertisements in the to-be-split node according to a split condition indicated by the second best split point, and construct the initial probability boosting tree based on the root node and the divided nodes when the divided nodes satisfy the split stop condition.

[0022] The first split point determining unit includes:

[0023] The initial feature determining subunit is configured to determine, based on the N sample advertisements, a root node for constructing the initial probability boosting tree, and take a sample advertisement attribute feature of each sample advertisement in the root node as an initial attribute feature.

[0024] The de-duplication processing subunit is configured to take a feature type of the initial attribute feature as a split point, perform de-duplication processing on the split point, and add the de-duplication processed split point to the first split point set.

[0025] The split point screening subunit is configured to screen, based on a split point determining condition for the root node, a best split point satisfying the split point determining condition from the first split point set, and take the screened best split point as the first best split point.

[0026] The first split point set includes M split points; M is a positive integer; and the M split points include a split point F a ; a is less than or equal to M.

[0027] The split point screening subunit includes:

[0028] The loss reduction parameter determining subunit is configured to obtain the split point determining condition for the root node, and obtain a node loss reduction parameter corresponding to the root node.

[0029] The first split point obtaining subunit is configured to obtain the split point F a from the first split point set, and obtain, based on the split point F aThe split condition indicated splits the sample advertisements in the root node to obtain initial child nodes of the root node; the initial child nodes include a first initial node and a second initial node;

[0030] The first parameter obtaining subunit is configured to obtain a first feature loss reduction parameter associated with the split point F a The first feature loss reduction parameter is associated with the split point F

[0031] The first comparison subunit is configured to compare the first feature loss reduction parameter and the node loss reduction parameter.

[0032] The first determination subunit is configured to determine that the split point F a satisfies a split point determination condition, and determine the split point F a as a first optimal split point.

[0033] The total probability distribution indicated by the root node includes a total shape parameter and a total slope parameter; the first sub-probability distribution indicated by the first initial node includes a first shape parameter and a first slope parameter; and the second sub-probability distribution indicated by the second initial node includes a second shape parameter and a second slope parameter.

[0034] The first parameter obtaining subunit is further configured to:

[0035] The first parameter obtaining subunit is further configured to:

[0036] The first parameter obtaining subunit is further configured to:

[0037] The first parameter obtaining subunit is further configured to:

[0038] The first parameter obtaining subunit is further configured to:

[0039] The feature loss reduction rule is obtained based on the total number of samples, the first number of samples, the second number of samples, the total node loss, the first node loss, the second node loss, and the feature loss reduction rule. a The first feature loss reduction parameter is associated with the split point F

[0040] The split point screening subunit further includes:

[0041] The second split point obtaining subunit is configured to determine the split point Fa does not satisfy the split point determination condition, obtaining a split point F from the first split point set a+1 ;

[0042] the second determination subunit is configured to, if a+1 is equal to M, determine the split point F as the first optimal split point. a+1

[0043] The split point screening subunit comprises:

[0044] The second parameter obtaining subunit is configured to, if a+1 is less than M, obtain a second feature loss reduction parameter associated with the split point F based on the split condition indicated by the split point F. a+1 a+1 The second feature loss reduction parameter is used for dividing the sample advertisement in the root node and obtaining a second feature loss reduction parameter associated with the split point F.

[0045] The second comparison subunit is configured to compare the second feature loss reduction parameter and the node loss reduction parameter.

[0046] The third determination subunit is configured to, if the node loss reduction parameter is less than or equal to the second feature loss reduction parameter, determine that the split point F satisfies the split point determination condition. a+1 a+1 The third determination subunit is configured to, if the node loss reduction parameter is less than or equal to the second feature loss reduction parameter, determine that the split point F satisfies the split point determination condition.

[0047] The first division unit comprises:

[0048] The sample advertisement obtaining subunit is configured to obtain a sample advertisement i from the root node; i is a positive integer less than or equal to N.

[0049] The first division subunit is configured to, if the sample advertisement i satisfies the split condition indicated by the first optimal split point, divide the sample advertisement i to a first type node.

[0050] The second division subunit is configured to, if the sample advertisement i does not satisfy the split condition indicated by the first optimal split point, divide the sample advertisement i to a second type node; the first type node and the second type node are both child nodes of the root node.

[0051] The iteration training module comprises:

[0052] The tree convergence condition obtaining unit is configured to obtain a tree convergence condition associated with the initial probability boosting tree; the tree convergence condition comprises a distribution error threshold.

[0053] The sample probability distribution determining unit is configured to determine a sample conversion data probability distribution based on sample advertisement conversion data of each sample advertisement in the N sample advertisements.

[0054] ​​​The prediction probability distribution determination unit is configured to, when the prediction conversion data probability distribution output by the initial probability boosting tree is obtained, determine a distribution error of the initial probability boosting tree based on the sample conversion data probability distribution and the prediction conversion data probability distribution.

[0055] The first target tree determination unit is configured to, when the distribution error is less than or equal to a distribution error threshold, determine that the initial probability boosting tree satisfies a tree convergence condition, and determine the initial probability boosting tree satisfying the tree convergence condition as a target probability boosting tree for predicting the conversion data probability distribution of the target advertisement.

[0056] The iteration training module further includes:

[0057] The condition non-satisfaction unit is configured to, when the distribution error is greater than the distribution error threshold, determine that the initial probability boosting tree does not satisfy the tree convergence condition.

[0058] The second target tree determination unit is configured to adjust the tree parameters of the initial probability boosting tree, and determine the adjusted initial probability boosting tree as a transition probability boosting tree, until the transition probability boosting tree satisfies the tree convergence condition, and determine the transition probability boosting tree satisfying the tree convergence condition as the target probability boosting tree for predicting the conversion data probability distribution of the target advertisement.

[0059] The application embodiment provides a data processing apparatus, including:

[0060] The delivery request acquisition module is configured to acquire a delivery feature request for the target advertisement; the delivery feature request includes a target advertisement attribute feature of the target advertisement.

[0061] The probability distribution determination module is configured to acquire a target probability boosting tree associated with the target advertisement, input the target advertisement attribute feature into the target probability boosting tree, and output a conversion data probability distribution of the target advertisement by the target probability boosting tree; the target probability boosting tree is obtained by iteratively training an initial probability boosting tree based on a tree convergence condition and sample advertisement conversion data of each sample advertisement in N sample advertisements; the N sample advertisements are used to construct a root node of the initial probability boosting tree; the initial probability boosting tree is obtained by dividing the N sample advertisements based on a split condition indicated by a first best split point; the first best split point is determined based on sample advertisement attribute features of each sample advertisement in the root node; N is a positive integer.

[0062] The target conversion data determination module is configured to determine target advertisement conversion data of the target advertisement based on the conversion data probability distribution of the target advertisement.

[0063] The application embodiment provides a computer device, including a processor and a memory.

[0064] The processor is connected with the memory, wherein the memory is configured to store a computer program, and the computer program is configured to be executed by the processor, so that the computer device executes the method provided by the embodiments of the present application.

[0065] The embodiments of the present application provide a computer readable storage medium, which stores a computer program. The computer program is suitable for being loaded and executed by a processor, so that a computer device having the processor executes the method provided by the embodiments of the present application.

[0066] The embodiments of the present application provide a computer program product or a computer program, which comprises 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 the processor executes the computer instructions, so that the computer device executes the method provided by the embodiments of the present application.

[0067] In the embodiments of the present application, the computer device can determine a root node for constructing an initial probability boosting tree according to N pairs of sample data corresponding to N sample advertisements, and then can determine a first best split point through a sample advertisement attribute feature of each sample advertisement in the root node, divide the sample advertisements in the root node, and thus obtain the initial probability boosting tree. Further, the computer device can iteratively train the initial probability boosting tree, and thus obtain a target probability boosting tree for predicting a conversion data probability distribution (for example, a conversion cost probability distribution) of a target advertisement. The entire process does not need to consume a large amount of computing resources to calculate the correlation parameters (for example, a covariance matrix) between all sample data as in the traditional probability estimation method, but uses the target probability boosting tree to quickly predict the conversion data probability distribution of the target advertisement, so as to reduce the calculation time and improve the prediction speed of the probability distribution. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0069] Figure 1 is a structural schematic diagram of a network architecture provided by the embodiments of the present application;

[0070] Figure 2 is a scene diagram for training a probability boosting tree provided by the embodiments of the present application;

[0071] Figure 3 is a flowchart of a data processing method provided by an embodiment of the present application;

[0072] Figure 4 is a scenario diagram of constructing an initial probability boosting tree provided by an embodiment of the present application;

[0073] Figure 5 is a scenario diagram of iteratively training an initial probability boosting tree provided by an embodiment of the present application;

[0074] Figure 6 is a flowchart of a data processing method provided by an embodiment of the present application;

[0075] Figure 7a is a comparison diagram of a predicted conversion data probability distribution determined by a target probability boosting tree and an actual conversion data probability distribution provided by an embodiment of the present application;

[0076] Figure 7b is a comparison diagram of a predicted conversion data probability distribution determined by a target probability boosting tree and an actual conversion data probability distribution provided by an embodiment of the present application;

[0077] Figure 8 is a scenario diagram of determining target conversion data provided by an embodiment of the present application;

[0078] Figure 9 is a structural diagram of a data processing apparatus provided by an embodiment of the present application;

[0079] Figure 10 is a structural diagram of a data processing apparatus provided by an embodiment of the present application;

[0080] Figure 11 is a diagram of a computer device provided by an embodiment of the present application;

[0081] Figure 12 is a structural diagram of a data processing system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0083] Please refer to Figure 1 , Figure 1 is a structural diagram of a network architecture provided by an embodiment of the present application. As shown inFigure 1 As shown, this network architecture may include a server 10F and a user terminal cluster. The user terminal cluster may include one or more user terminals. For example... Figure 1 As shown, the user terminal cluster may specifically include user terminal 100a, user terminal 100b, user terminal 100c, ..., user terminal 100n. Figure 1 As shown, user terminals 100a, 100b, 100c, ..., 100n can each connect to the aforementioned server 10F via a network, enabling each user terminal to interact with the server 10F through this network connection. The network connection method is not limited; it can be a direct or indirect connection via wired communication, a direct or indirect connection via wireless communication, or other methods. This application does not impose any restrictions on this method.

[0084] Each user terminal in this user terminal cluster can include: smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, in-vehicle terminals, smart TVs, and other smart terminals with boosting tree training capabilities. It should be understood that, for example... Figure 1 Each user terminal in the user terminal cluster shown can have the target application (i.e., the application client) installed. When the application client runs on each user terminal, it can interact with the aforementioned... Figure 1 Data interaction occurs between the servers 10F shown. The application client can include social clients, multimedia clients (e.g., video clients), entertainment clients (e.g., game clients), educational clients, live streaming clients, etc. This application client can be a standalone client or an embedded sub-client integrated into another client (e.g., social clients, educational clients, and multimedia clients), and this is not limited here.

[0085] like Figure 1 As shown, in this embodiment, server 10F can be the server corresponding to the application client. Server 10F can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. This embodiment does not limit the number of user terminals and servers.

[0086] For ease of understanding, the embodiments of this application may be described in detail below. Figure 1 From the multiple user terminals shown, one user terminal is selected as the target user terminal. For example, in the embodiments of this application, a user terminal can be selected as the target user terminal. Figure 1The user terminal 100a shown as a target user terminal can integrate a target application (i.e., an application client). At this time, the target user terminal can realize data interaction between the application client corresponding service data platform and the server 10F. The target application can run a trained probability boosting tree (i.e., a target probability boosting tree), which can be used to predict the conversion data probability distribution of the target advertisement (i.e., the advertisement to be predicted). It should be understood that the conversion data probability distribution can be used to provide a conversion data control reference for the advertiser, so as to subsequently determine the final conversion data of the target advertisement based on the conversion data probability distribution, which is of great significance to advertising marketing and financial risk control. The conversion data probability distribution refers to the probability law of the random variable (i.e., the conversion data) value. The probability of an event represents the likelihood of a certain result occurring in a trial.

[0087] It should be understood that the embodiments of the present application propose a distribution estimation method based on a probability boosting tree, which can involve machine learning in the field of artificial intelligence. It can be understood that artificial intelligence (AI) is to use digital computers or digital computer-controlled computing simulation, extension and expansion of human intelligence, perception of the environment, acquisition of knowledge and use of knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machine has the functions of perception, reasoning and decision-making.

[0088] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, automatic driving, intelligent transportation, etc.

[0089] Machine Learning (ML) is a multi-disciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithmic complexity theory, etc. It is a specialized study of how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure, and continuously improve performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent. It is applied in various fields of artificial intelligence. Machine learning and deep learning generally include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rule-based learning.

[0090] It can be understood that the boosting method based on the decision tree in the embodiments of the present application can be referred to as a boosting tree. The boosting tree is considered as one of the best methods in statistical learning. The decision tree here is a decision analysis method for obtaining the probability that the expected value of net present value is greater than or equal to zero by constructing a decision tree based on the probability of various situations, evaluating project risk, and determining its feasibility. It is a graphical method for intuitive probability analysis.

[0091] It should be understood that the application scenarios involved in the target probability boosting tree can include application scenario 1 (for example, a conversion cost distribution estimation scenario) and application scenario 2 (for example, a default asset distribution estimation scenario). When training the target probability boosting tree, the computer device needs to obtain N sample advertisements and sample data pairs corresponding to the N sample advertisements, respectively. N is a positive integer, one sample advertisement corresponds to one sample data pair, and one sample data pair is determined by sample advertisement attribute features and sample advertisement conversion data. The advertisement attribute features of the sample advertisement (for example, optimization target, advertisement type, and advertiser industry) can be referred to as sample advertisement attribute features, and the conversion data of the sample advertisement (for example, advertisement conversion cost or advertisement default asset) can be referred to as sample conversion data.

[0092] Further, the computer device can determine a root node for constructing an initial probability boosting tree based on the N sample advertisements, and then determine a first best split point through the sample advertisement attribute features of each sample advertisement in the root node, and divide the sample advertisements in the root node according to the first best split point, so as to obtain the initial probability boosting tree. At this time, the computer device can iteratively train the initial probability boosting tree, and then obtain the target probability boosting tree when the training is completed. This means that the computer device does not need to consume a large amount of computing resources, and can directly use the target probability boosting tree to quickly predict the conversion data probability distribution of the target advertisement, so as to improve the prediction speed of the probability distribution.

[0093] For ease of understanding, further, please refer to Figure 2 , Figure 2 is an embodiment of the present application provides a scene for training probability boosting tree. As Figure 2 indicated, the computer device in the embodiment of the present application can be a computer device with the function of training the tree, which can be any one of the user terminals in the user terminal cluster as shown in Figure 1 , for example, user terminal 100a, the computer device can also be the server 10F as shown in Figure 1 , here will not be limited.

[0094] It should be understood that the computer device can obtain N sample advertisements and N sample data pairs corresponding to the N sample advertisements; N is a positive integer; one sample advertisement corresponds to one sample data pair, and one sample data pair is determined by sample advertisement attribute features and sample advertisement conversion data. As Figure 2 indicated, the number of sample advertisements in the embodiment of the present application is taken as an example, which can include sample advertisement 1, sample advertisement 2, sample advertisement 3, sample advertisement 4 and sample advertisement 5. Among them, the sample data pair corresponding to the sample advertisement 1 is the sample data pair S1; the sample data pair corresponding to the sample advertisement 2 is the sample data pair S2; the sample data pair corresponding to the sample advertisement 3 is the sample data pair S3; the sample data pair corresponding to the sample advertisement 4 is the sample data pair S4; the sample data pair corresponding to the sample advertisement 5 is the sample data pair S5.

[0095] Among them, the computer device can determine the root node (for example, the initial value as shown in Figure 2 ) for constructing the initial probability boosting tree based on the five sample advertisements. Further, the computer device can determine a first set of split points consistent with the feature type of the initial attribute feature by taking the sample advertisement attribute feature of each sample advertisement in the root node as the initial attribute feature, determine a first best split point from the first set of split points, divide the five sample advertisements according to the split condition indicated by the first best split point, and obtain the initial probability boosting tree (for example, tree 1 as shown in Figure 2 ). Among them, the sample advertisements divided into the first leaf node of tree 1 can include sample advertisement 1, sample advertisement 2 and sample advertisement 3, and the sample advertisements divided into the second leaf node of the tree 1 can include sample advertisement 4 and sample advertisement 5.

[0096] For ease of understanding, further, please refer to Table 1, which is a sample probability distribution parameter calculation process table before simplification provided by the embodiment of the present application. As shown in Table 1:

[0097] Table 1

[0098]

[0099] wherein p j,l may represent the probability distribution of the lth leaf node in the jth tree; for example, p 1,1 may represent the probability distribution of the 1st leaf node in the 1st tree; C can represent a normalization parameter, wherein the first subscript of C is used to represent the current tree (i.e., the number of iterations) in which the sample advertisement i is located, the second subscript of C is used to represent the position of the sample advertisement i in the leaf node of the previous tree, and the third subscript of C is used to represent the position of the sample advertisement i in the leaf node of the current tree.

[0100] Further, the computer device can obtain a tree convergence condition associated with the initial probability boosting tree. Here, the tree convergence condition can be used to indicate the number of iterations of the iterative training, i.e., the number of trees included in the target probability boosting tree; optionally, the tree convergence condition can also be used to indicate that the distribution error of the target probability boosting tree is less than a pre-set distribution error threshold (e.g., 0.3), and the tree convergence condition will not be limited here.

[0101] Further, the computer device can perform iterative training on the initial probability boosting tree (e.g., the probability boosting tree 20T including tree 1) based on the tree convergence condition and the sample advertisement conversion data of each of the five sample advertisements, to obtain a target probability boosting tree for predicting the conversion data probability distribution of the target advertisement.

[0102] For example, the computer device needs to determine whether the probability boosting tree 20T including tree 1 is the target probability boosting tree for predicting the conversion data probability distribution of the target advertisement based on the tree convergence condition. If the probability boosting tree 20T including tree 1 does not satisfy the tree convergence condition, i.e., it is determined that the probability boosting tree 20T including tree 1 is not the target probability boosting tree, the computer device needs to perform a second round of iteration on the probability boosting tree 20T, i.e., to construct tree 2 based on tree 1, to obtain a probability boosting tree (e.g., a new probability boosting tree 20T) including tree 1 and tree 2. At this time, the computer device needs to further determine whether the new probability boosting tree 20T is the target probability boosting tree for predicting the conversion data probability distribution of the target advertisement based on the tree convergence condition, and so on, until the probability boosting tree 20T after the iterative training satisfies the tree convergence condition, and the probability boosting tree 20T after the iterative training is taken as the target probability boosting tree.

[0103] Therefore, the computer device in the embodiment of the present application does not need to consume a large amount of computing resources when training the initial probability boosting tree model, so that the calculation amount of probability prediction can be greatly reduced. In addition, the target probability boosting tree trained by the computer device can quickly predict the conversion data probability distribution of the target advertisement, so as to improve the prediction speed of the probability distribution.

[0104] The specific implementation of the computer device with the boosting tree training function training the initial probability boosting tree through N sample data pairs corresponding to N sample advertisements to obtain the target probability boosting tree can be seen from the following Figures 3-8 corresponding embodiment.

[0105] Further, please refer to Figure 3 , Figure 3 is a flowchart of a data processing method provided by an embodiment of the present application. As Figure 3 shown, the method can be executed by a computer device with a boosting tree training function. The computer device can be a user terminal (for example, any one of the user terminals in the user terminal cluster shown in the above Figure 1 , for example, the user terminal 100a), or a server (for example, the server 10F shown in the above Figure 1 , and the present application is not limited thereto. For ease of understanding, the embodiment of the present application takes the method executed by the server with the boosting tree training function as an example, and the method can at least include the following steps S101-S103:

[0106] Step S101, obtaining N sample advertisements and N sample data pairs corresponding to the N sample advertisements.

[0107] Specifically, in order to quickly and flexibly estimate the complete probability distribution, the computer device with the boosting tree training function can obtain N sample advertisements and N sample data pairs corresponding to the N sample advertisements from the historically launched advertisements. Here, N can be a positive integer, one sample advertisement corresponds to one sample data pair, and the sample data pair is determined by the sample advertisement attribute feature and the sample advertisement conversion data.

[0108] For ease of understanding, further, please refer to Table 2, which is a sample data table provided by an embodiment of the present application. The table 2 can include N sample advertisements, and one sample data pair corresponding to one sample advertisement. For example, the N sample advertisements can include sample advertisement i, and the sample data pair corresponding to the sample advertisement i can be (x i , y i ). Wherein, i is a positive integer less than or equal to N; x i may refer to the sample advertisement attribute feature of the sample advertisement i, yi may refer to sample ad conversion data (e.g., conversion cost) of the sample ad i. As shown in Table 2:

[0109] Table 2

[0110]

[0111] The ad attribute features of the ad can specifically include the following attribute features: a first attribute feature (e.g., an optimization target), a second attribute feature (e.g., an ad type), and a third attribute feature (e.g., an advertiser industry). For example, the optimization target can include feature 1 (e.g., targeting download), feature 2 (e.g., targeting activation), feature 3 (e.g., targeting secondary stay), feature 4 (e.g., targeting payment assets), and feature 5 (e.g., targeting credit), etc. The ad type can include feature 6 (e.g., operating system A download), feature 7 (e.g., operating system B download), feature 8 (e.g., commodity promotion), and feature 9 (small game promotion), etc. The operating system A and the operating system B belong to two different operating systems. The advertiser industry can include feature 10 (e.g., e-commerce industry), feature 11 (e.g., education industry), feature 12 (e.g., financial industry), and feature 13 (e.g., tourism industry), etc.

[0112] For example, the sample data pair of the sample ad 1 shown in Table 2 can be (x1, y1), that is, the sample ad attribute features x1 of the sample ad 1 can include feature 1 (e.g., targeting download), feature 6 (e.g., operating system A download), and feature 13 (e.g., tourism industry), and the sample ad conversion data y1 of the sample ad 1 can be 1000.

[0113] In step S102, based on the N sample ads, a root node for constructing an initial probability boosting tree is determined, the sample ad attribute features of each sample ad in the root node are taken as initial attribute features, a first set of split points consistent with the feature types of the initial attribute features is determined, a first best split point is determined from the first set of split points, the N sample ads are divided according to the split condition indicated by the first best split point, and the initial probability boosting tree is obtained.

[0114] Specifically, the computer device can determine a root node for constructing an initial probability boosting tree based on N sample advertisements, and then can take the sample advertisement attribute features of each sample advertisement in the root node as initial attribute features, and can determine a first set of split points consistent with the feature types of the initial attribute features, and determine a first best split point from the first set of split points. Further, the computer device can divide the sample advertisements in the root node according to the split condition indicated by the first best split point, thereby obtaining a first type node (e.g., a left child node) and a second type node (e.g., a right child node). Further, the computer device can take the first type node and the second type node as nodes to be split, and at the same time, the computer device can also obtain a split stop condition associated with the initial probability boosting tree. Here, the split stop condition is used to indicate the depth (e.g., 5) of the initial probability boosting tree. Optionally, the split stop condition can also be used to indicate that the number of sample advertisements in the node to be split (i.e., the current node) is 1, i.e., the node to be split cannot be further divided, and the split stop condition will not be limited here. When the node to be split satisfies the split stop condition, the computer device can construct the initial probability boosting tree based on the root node and the divided nodes. Optionally, when the node to be split does not satisfy the split stop condition, the computer device can take the sample advertisement attribute features of each sample advertisement in the node to be split as target attribute features, and then can determine a second set of split points consistent with the feature types of the target attribute features, and determine a second best split point from the second set of split points. Further, the computer device can divide the sample advertisements in the node to be split according to the split condition indicated by the second best split point, until the divided node satisfies the split stop condition, and then can construct the initial probability boosting tree based on the root node and the divided nodes.

[0115] It should be understood that the embodiments of the present application can use the idea of probability boosting tree to propose a systematic scheme for modeling probability distribution, wherein a single probability boosting tree is basically consistent with an ordinary decision tree in structure: it is a binary tree, and a non-leaf node can be divided based on the split condition indicated by the split point. For example, sample advertisements satisfying the split condition can be divided into a first type node (e.g., a left child node), and sample advertisements not satisfying the split condition can be divided into a second type node (e.g., a right child node). Multiple probability boosting trees can use the idea of ensemble learning to take the output of the leaf node in each round of training as the likelihood probability, and then update it by applying the Bayes formula.

[0116] Specifically, if the sample set obtained by the computer device is {(x i ,y i| i = 1, 2, …, N}, then the probability density function p(y i |x i ) of the sample ad i in the distribution estimation model (e.g., a model indicated by a target probability boosting tree containing T trees) can be seen from the following formula (1), and the related formula of ensemble learning can be seen from formula (2):

[0117]

[0118]

[0119] wherein p j,l is the probability distribution of the lth leaf node in the jth tree; T trees can represent that T rounds of iterative training are performed, where T is a positive integer; and C is a normalization parameter. may represent that the sample data i is divided into the lth node in the jth tree. may represent the sample set on the lth node in the jth tree.

[0120] wherein it is assumed that the output of the leaf node of the probability boosting tree obeys a certain parameter distribution, for example, a Gaussian distribution or a Gamma distribution. Since in the advertisement conversion scenario, the advertisement conversion data (e.g., advertisement conversion cost or advertisement default asset) is generally a value greater than 0, and the definition domain of the Gamma distribution is 0 to positive infinity, the parameter distribution to which the output of the leaf node of the probability boosting tree in the embodiment of the present application can be taken as the Gamma distribution.

[0121] It can be understood that the Gamma distribution with two parameters can be uniquely determined by the shape parameter (k) and the slope parameter (θ). Wherein, the shape parameter herein can be used to describe the distribution shape of the probability distribution (i.e., to describe the shape of the curve), and the slope parameter can be used to describe the slope of the probability distribution (i.e., to describe the steepness of the curve). Therefore, the probability density function of a certain node can be written as p(y|k, θ) in the embodiment of the present application, where y can be the advertisement conversion data.

[0122] Specifically, when the computer device performs the jth round (i.e., the jth tree) of iterative training on the initial probability boosting tree, the node loss of node l can be seen from the following formula (3) to formula (11):

[0123]

[0124]

[0125]

[0126] Apply the Gamma distribution formula And after simple arrangement, we get:

[0127]

[0128]

[0129]

[0130]

[0131]

[0132]

[0133] wherein, Likelihood j-1 is the likelihood function of the population at the j-1th iteration training; p(y i |k j-1,i ,θ j-1,i ) is the probability density function of the sample advertisement i at the j-1th iteration training, and N is the total number of sample advertisements. Loss j-1 (L j-1 ) is the loss function of the population at the j-1th iteration training; Loss j-1 (S l ) is the loss function of node l at the j-1th iteration training; s is the total number of sample advertisements in node l. is called the double Gamma function, which is monotonically increasing when k>0; η1 is the learning rate for the shape parameter k; η2 is the learning rate for the shape parameter is the learning rate for the slope parameter θ; Loss j (S l ) is the node loss of node l at the jth iteration training.

[0134] Specifically, the formula for determining the node loss reduction parameter of any one node (for example, node l) by the computer device can refer to the following formula (12):

[0135] loss_reduce=Loss j (S l )-Loss j-1 (S l ) (12)

[0136] wherein, loss_reduce herein refers to the node loss reduction parameter of node l, and the initial value of the node loss reduction parameter is 0.

[0137] Specifically, the feature loss reduction rule (i.e., used to determine the split point Fa The formula of the associated current feature loss reduction parameter (tmp_loss_reduce) can be seen from the following formula (13):

[0138]

[0139] Wherein, |S l | can represent the total number of samples of the sample advertisement in the node l; Can represent the first sample number of the sample advertisement divided into the first type node (for example, based on the split point F a The left child node of the node l divided by the indicated split condition); Can represent the second sample number of the sample advertisement divided into the second type node (for example, based on the split point F a The right child node of the node l divided by the indicated split condition); Can represent the node loss of the left child node of the node l at the jth round of iterative training; Can represent the node loss of the right child node of the node l at the jth round of iterative training.

[0140] It should be understood that when determining the first optimal split point (i.e. the optimal split point for dividing the sample advertisements in the root node), the computer device can determine the root node for constructing the initial probability boosting tree based on the N sample advertisements, and then the sample advertisement attribute features of each sample advertisement in the root node can be used as the initial attribute features. At this time, the computer device can take the feature type of the initial attribute features as the split point, and perform deduplication processing on the split point, and add the deduplicated split point to the first split point set. For example, the split point set associated with the root node (i.e. the first split point set) C = {F1, F2, F3, …, F M}, wherein M here can be a positive integer.

[0141] Further, the computer device can filter out the optimal split point satisfying the split point determination condition from the first split point set based on the split point determination condition for the root node, and then the filtered out optimal split point can be taken as the first optimal split point. The split point determination condition here can be used to indicate that the current split point is the optimal split point when the node loss reduction parameter is less than or equal to the current feature loss reduction parameter (i.e. loss_reduce≤tmp_loss_reduce). Wherein, the node loss reduction parameter of the root node is the initial value (for example, 0) at initialization, and the optimal split point sequence number of the current root node is the initial value (for example, -1).

[0142] It can be understood that the computer device can further obtain the node loss reduction parameter (i.e., initial value 0) corresponding to the root node when obtaining the split point determination condition for the root node. At this time, the computer device can obtain the first split point (for example, split point F a , where a is less than or equal to M) from the first split point set associated with the root node, and can further divide the sample advertisements in the root node based on the split point F a indicated split condition to obtain the initial child nodes of the root node. Wherein, the initial child nodes can include a first initial node (for example, the left child node of the root node divided based on the split point F a indicated split condition) and a second initial node (for example, the right child node of the root node divided based on the split point F a indicated split condition).

[0143] Further, the computer device obtains the current feature loss reduction parameter (i.e., the first feature loss reduction parameter) associated with the split point F a . Wherein, the shape parameter in the probability distribution indicated by the root node (i.e., the total probability distribution) can be referred to as the total shape parameter, and the slope parameter in the probability distribution indicated by the root node can be referred to as the total slope parameter; the shape parameter in the probability distribution indicated by the first initial node of the root node (i.e., the first sub-probability distribution) can be referred to as the first shape parameter, and the slope parameter in the first sub-probability distribution can be referred to as the first slope parameter; the shape parameter in the probability distribution indicated by the second initial node of the root node (i.e., the second sub-probability distribution) can be referred to as the second shape parameter, and the slope parameter in the second sub-probability distribution can be referred to as the second slope parameter.

[0144] At this time, the computer device can respectively count the total number of sample advertisements in the root node, the first number of sample advertisements in the first initial node, and the second number of sample advertisements in the second initial node. Further, the computer device can determine the total node loss of the root node based on the total shape parameter, the total slope parameter, and the node loss determination rule shown in the above formula (11). At the same time, the computer device can also determine the first node loss of the first initial node based on the first shape parameter, the first slope parameter, and the node loss determination rule shown in the above formula (11). Meanwhile, the computer device can determine the second node loss of the second initial node based on the second shape parameter, the second slope parameter, and the node loss determination rule shown in the above formula (11). Further, the computer device can obtain the feature loss reduction rule shown in the above formula (13), and determine the feature loss reduction parameter associated with the split point Fa the associated first feature loss reduction parameter.

[0145] It should be understood that the computer device can compare the first feature loss reduction parameter with the node loss reduction parameter. If the node loss reduction parameter is less than or equal to the first feature loss reduction parameter, the computer device can determine that the split point F a satisfies the split point determination condition, and the split point F a is determined as the first optimal split point, i.e., the serial number of the current optimal split point is a. Alternatively, if the node loss reduction parameter is greater than the first feature loss reduction parameter, the computer device can determine that the split point F a does not satisfy the split point determination condition, and the computer device can obtain the next split point (e.g., the split point F a+1 ) of the split point F a from the first split point set. Since there is a split point in the first split point set that can effectively reduce the node loss of the root node, in order to improve the determination efficiency of the optimal split point, the computer device can quickly determine the first optimal split point for dividing the sample advertisements in the root node by determining whether the split point F a+1 belongs to the last split point in the first split point set, i.e., whether a+1 is equal to M.

[0146] If a+1 is equal to M, the computer device can determine that the split point F a+1 belongs to the last split point in the first split point set, and the computer device does not need to determine the split point F a+1 corresponding to the current feature loss reduction parameter. Instead, the computer device can directly determine the split point F a+1 as the first optimal split point, i.e., the serial number of the current optimal split point is a+1. Alternatively, if a+1 is less than M, the computer device can re-divide the sample advertisements in the root node based on the split condition indicated by the split point F a+1 to obtain new initial sub-nodes, where the new initial sub-nodes can include a new first initial node (e.g., the left child node of the root node divided based on the split condition indicated by the split point F a+1 ) and a new second initial node (e.g., the right child node of the root node divided based on the split condition indicated by the split point F a+1 ). Further, the computer device can obtain the node loss reduction parameter of the new initial sub-nodes according to the above formula (13). a+1The associated current feature loss reduction parameter (i.e., the second current feature loss reduction parameter) can be compared with the node loss reduction parameter. If the node loss reduction parameter is less than or equal to the second current feature loss reduction parameter, the computer device can determine that the split point F a+1 satisfies the split point determination condition, and then the split point F a+1 is the first optimal split point, i.e., the sequence number of the current optimal split point is a+1. The specific implementation of the computer device determining the second feature loss reduction parameter can refer to the specific implementation of the computer device determining the first feature loss reduction parameter described above, which will not be described here.

[0147] When the first optimal split point is determined, the computer device can divide the sample advertisements in the root node according to the split condition indicated by the first optimal split point, so as to obtain a first type node and a second type node. The first type node refers to the left child node of the root node divided based on the split condition indicated by the first optimal split point; the second type node refers to the right child node of the root node divided based on the split condition indicated by the first optimal split point.

[0148] The computer device obtains a sample advertisement i from the root node; where i can be a positive integer less than or equal to N. If the sample advertisement i satisfies the split condition indicated by the first optimal split point, the computer device can divide the sample advertisement i to the first type node. Optionally, if the sample advertisement i does not satisfy the split condition indicated by the first optimal split point, the computer device can divide the sample advertisement i to the second type node; where the first type node and the second type node are both child nodes of the root node.

[0149] Further, the computer device takes the first type node and the second type node as a to-be-split node, and then obtains a split cutoff condition associated with the initial probability boosting tree. When the to-be-split node satisfies the split cutoff condition, the computer device can directly construct the initial probability boosting tree based on the root node and the divided nodes. Optionally, when the to-be-split node does not satisfy the split cutoff condition, the computer device can take the sample advertisement attribute feature of each sample advertisement in the to-be-split node as a target attribute feature, and then determine a second split point set corresponding to the feature type of the target attribute feature, and determine a second optimal split point from the second split point set. Further, the computer device can divide the sample advertisements in the to-be-split node according to the split condition indicated by the second optimal split point, until the divided node satisfies the split cutoff condition, at which time the computer device can construct the initial probability boosting tree based on the root node and the divided nodes.

[0150] For ease of understanding, further, please refer to Figure 4 , Figure 4 is a scene schematic diagram provided by an embodiment of the present application for constructing an initial probability boosting tree. As shown in Figure 4 , the computer device in the embodiment of the present application can be a computer device with a boosting tree training function. The computer device can be any one of the user terminals in the user terminal cluster shown in Figure 1 , for example, the user terminal 100a. The computer device can also be the server 10F shown in Figure 1 . Here, the computer device will not be limited.

[0151] As shown in Figure 4 , for ease of description, the N sample ads obtained by the computer device in the embodiment of the present application when constructing the initial probability boosting tree can be taken as an example of 5 sample ads. The 5 sample ads can specifically include sample ad 1, sample ad 2, sample ad 3, sample ad 4, and sample ad 5. One sample ad corresponds to one sample data pair, and the sample data pair here is determined by the sample ad attribute feature and the sample ad conversion data.

[0152] It can be understood that the computer device can determine a root node (for example, node J1) for constructing the initial probability boosting tree based on the 5 sample ads. For example, the computer device can obtain the sample ad conversion data of each of the 5 sample ads, and then can determine the total probability distribution of the 5 sample ads at the root node based on the obtained 5 sample ad conversion data. At this time, Figure 4 , the computer device can take node J1 as a parent node, and then determine a first best split point according to the ad attribute features of the sample ads in node J1 to obtain a first type node and a second type node of the node J1. The first type node and the second type node are both child nodes of node J1.

[0153] It should be understood that the computer device can take the sample ad attribute features of each sample ad in node J1 as initial attribute features. At this time, the computer device can take the feature types of the initial attribute features as split points, and then can perform a deduplication processing on the split points to add the deduplicated split points to a first split point set. For example, the split point set (i.e., the first split point set) associated with node J1 can be C1={F1, F2, F3, …, F M}, where M can be a positive integer. In this embodiment, four split points can be used as an example. The first set of split points can specifically include split point F1 (e.g., download as the goal), split point F2 (e.g., mini-game promotion), split point F3 (e.g., e-commerce industry), and split point F4 (e.g., operating system A download).

[0154] Furthermore, the computer device can, based on the split point determination conditions for node J1, select the optimal split point from the first set of split points that satisfies the split point determination conditions, and then use the selected optimal split point as the first optimal split point. For example, the computer device can first obtain split point F1 from the first set of split points, and then, based on... Figure 4 The splitting condition indicated by splitting point F1 is used to sequentially divide each sample advertisement in node J1, thereby obtaining the initial child nodes of node J1. For example, if sample advertisement 1 in the root node satisfies the splitting condition indicated by splitting point F1, then sample advertisement 1 is assigned to the first initial node (e.g., the left child node of node J1 divided based on the splitting condition indicated by splitting point F1). Optionally, if sample advertisement 2 in the root node does not satisfy the splitting condition indicated by splitting point F1, then sample advertisement 2 is assigned to the second initial node (e.g., the right child node of node J1 divided based on the splitting condition indicated by splitting point F1), and so on. After the division is completed, the computer device can obtain the current feature loss reduction parameter (i.e., the first feature loss reduction parameter) associated with splitting point F1 based on the feature loss reduction rule shown in the above formula (13).

[0155] If the node loss reduction parameter (e.g., initial value 0) is greater than the first feature loss reduction parameter, the computer device can determine that the split point F1 does not meet the split point determination condition. In this case, the computer device can continue to obtain the next split point (e.g., split point F2) of the split point F1 from the first set of split points. Then, based on the split condition indicated by the split point F2, each sample advertisement in node J1 can be re-divided sequentially to obtain a new initial child node of node J1. The new initial child node of node J1 may include a new first initial node (e.g., the left child node of node J1 divided based on the split condition indicated by the split point F2) and a new second initial node (e.g., the right child node of node J1 divided based on the split condition indicated by the split point F2). Similarly, the computer device can obtain the current feature loss reduction parameter (i.e., the second feature loss reduction parameter) associated with the split point F2 based on the feature loss reduction rule shown in the above formula (13).

[0156] If the node loss reduction parameter is greater than the second feature loss reduction parameter, the computer device can determine that split point F2 also does not meet the split point determination condition. In this case, the computer device still needs to obtain the next split point (e.g., split point F3) of split point F2 from the first split point set, so as to re-divide each sample advertisement in node J1 according to the split condition indicated by split point F3, thereby obtaining the target child node. The target child node of node J1 may include the first target node (e.g., the left child node of node J1 divided according to the split condition indicated by split point F3) and the second target node (e.g., the right child node of node J1 divided according to the split condition indicated by split point F3). Similarly, the computer device can obtain the current feature loss reduction parameter (i.e., the third feature loss reduction parameter) associated with split point F3 based on the feature loss reduction rule shown in the above formula (13).

[0157] When the node loss reduction parameter is less than or equal to the third feature loss reduction parameter, the computer device can determine that split point F3 meets the split point determination condition. In this case, the computer device can use split point F3 as the first optimal split point for node J1. At this point, the computer device can also use the first target node of node J1 (e.g., based on the splitting condition indicated by split point F3) as the first target node of node J1. Figure 4 The node J2 shown is called the first type node of node J1, and the second target node of node J1 (e.g., based on the splitting condition indicated by the splitting point F3) is the node divided based on the splitting condition indicated by the splitting point F3. Figure 4 The node J3 shown is called the second type node of node J1.

[0158] Furthermore, the computer device uses nodes J2 and J3 as nodes to be split, and can then obtain the splitting cutoff condition associated with the probability boosting tree 40T to determine whether to continue splitting the nodes to be split. For example, the splitting cutoff condition can be used to indicate that the number of sample advertisements in the node to be split (i.e., the current node) is 1, meaning that the node to be split can no longer be split.

[0159] like Figure 4 As shown, the sample advertisements in node J2 may include sample advertisement 4 and sample advertisement 5; the sample advertisements in node J3 may include sample advertisement 1, sample advertisement 2, and sample advertisement 3. For nodes J2 and J3, the number of sample advertisements in both nodes is not 1. Therefore, the computer device can determine that neither node J2 nor node J3 meets the splitting cutoff condition, and thus, nodes J2 and J3 are respectively designated as new parent nodes for further splitting.

[0160] For any node to be split (e.g., node J2), the computer device can use the sample advertisement attribute features of each sample advertisement in node J2 as target attribute features, and then determine a second set of split points that matches the feature type of the target attribute features. For example, the set of split points associated with node J2 (i.e., the second set of split points) can be C2 = {F1, F2, F4}, that is, the second set of split points can specifically include split point F1 (e.g., download as the target), split point F2 (e.g., mini-game promotion), and split point F4 (e.g., operating system A download). At this time, the computer device can refer to the specific implementation of determining the first optimal split point described above, determine the second optimal split point (e.g., split point F2) from the second set of split points, and then divide the sample advertisements in node J2 according to the splitting conditions indicated by the second optimal split point of node J2 to obtain a new first type of node (e.g., Figure 4 The node J4 shown) and the new second type of node (e.g., Figure 4 (See node J5). Nodes J4 and J5 are both child nodes of node J2. Since sample ads in node J4 include sample ad 4, and sample ads in node J5 include sample ad 5, it can be determined that when both node J4 and node J5 meet the splitting cutoff condition, there is no need to split nodes J4 and J5.

[0161] Similarly, when the computer device determines that the second optimal split point of node J3 is split point F2, it can divide the sample advertisements in node J3 according to the splitting conditions indicated by the second optimal split point of node J3, and obtain new first-type nodes (e.g., Figure 4 The node J6 shown) and the new second type of node (e.g., Figure 4 (See node J7). Nodes J6 and J7 are both child nodes of node J3. If the splitting cutoff condition is used to indicate the depth of the initial probability boosting tree (e.g., 3), the computer device can determine that both nodes J6 and J7 satisfy the splitting cutoff condition, thus eliminating the need to split nodes J6 and J7. Figure 4 As shown, the sample advertisement in node J6 can include sample advertisement 2 and sample advertisement 3, and the sample advertisement in node J7 can include sample advertisement 1.

[0162] Once it is determined that all the partitioned nodes satisfy the splitting cutoff condition, the computer device can construct a new system based on node J1 and the partitioned nodes (e.g., nodes J2, J3, J4, J5, J6, and J7). Figure 5The probability boosting tree 40T (i.e., the initial probability boosting tree) is shown. It can be understood that the sum of the probabilities of each level of nodes in the initial probability boosting tree is 1.

[0163] In step S103, the initial probability boosting tree is iteratively trained based on a tree convergence condition associated with the initial probability boosting tree and sample advertisement conversion data of each of the N sample advertisements, to obtain a target probability boosting tree for predicting a conversion data probability distribution of a target advertisement.

[0164] Specifically, the computer device can obtain a tree convergence condition associated with the initial probability boosting tree. Here, the tree convergence condition can be used to indicate the number of iterations, i.e., the number of trees included in the target probability boosting tree; optionally, the tree convergence condition can also be used to indicate that the distribution error of the target probability boosting tree is less than a pre-set distribution error threshold (e.g., 0.3), which will not be limited here. In this embodiment of the present application, the tree convergence condition including the distribution error threshold is taken as an example, at this time, the computer device needs to determine a sample conversion data probability distribution (i.e., an actual conversion data probability distribution) based on the sample advertisement conversion data of each of the N sample advertisements. Further, when the predicted conversion data probability distribution output by the initial probability boosting tree is obtained, the computer device can determine the distribution error (e.g., the root mean square error) of the initial probability boosting tree based on the sample conversion data probability distribution and the predicted conversion data probability distribution. Further, the computer device can determine the target probability boosting tree for predicting the conversion data probability distribution of the target advertisement based on the distribution error, the distribution error threshold, and the tree convergence condition.

[0165] It can be understood that when the distribution error is less than or equal to the distribution error threshold, the computer device can determine that the initial probability boosting tree satisfies the tree convergence condition, at this time, the computer device can take the initial probability boosting tree satisfying the tree convergence condition as the target probability boosting tree for predicting the conversion data probability distribution of the target advertisement. Optionally, when the distribution error is greater than the distribution error threshold, the computer device can determine that the initial probability boosting tree does not satisfy the tree convergence condition, at this time, the computer device can adjust the tree parameters of the initial probability boosting tree, and take the adjusted initial probability boosting tree as a transition probability boosting tree, until the transition probability boosting tree satisfies the tree convergence condition, the computer device can take the transition probability boosting tree satisfying the tree convergence condition as the target probability boosting tree for predicting the conversion data probability distribution of the target advertisement.

[0166] For ease of understanding, further, Figure 5 , Figure 5 is a scenario diagram provided by an embodiment of the present application for iteratively training an initial probability boosting tree. As Figure 1As shown, the computer device in the embodiments of the present application can be a computer device with a boosted tree training function. The computer device can be the user terminal 100a in the user terminal cluster as shown above. Figure 1 As shown, the computer device in the embodiments of the present application can be a computer device with a boosted tree training function. The computer device can be the user terminal 100a in the user terminal cluster as shown above. Figure 5 As shown, the computer device in the embodiments of the present application can be a computer device with a boosted tree training function. The computer device can be the user terminal 100a in the user terminal cluster as shown above.

[0167] As shown, the computer device in the embodiments of the present application can be a computer device with a boosted tree training function. The computer device can be the user terminal 100a in the user terminal cluster as shown above. Figure 5 As shown, the computer device in the embodiments of the present application can be a computer device with a boosted tree training function. The computer device can be the user terminal 100a in the user terminal cluster as shown above. Figure 5 As shown, the computer device in the embodiments of the present application can be a computer device with a boosted tree training function. The computer device can be the user terminal 100a in the user terminal cluster as shown above. Figure 5 As shown, the computer device in the embodiments of the present application can be a computer device with a boosted tree training function. The computer device can be the user terminal 100a in the user terminal cluster as shown above.

[0168] For ease of understanding, further, please refer to Table 3, which is a simplified sample probability distribution parameter calculation process table provided by the embodiments of the present application. As shown in Table 3:

[0169] Table 3

[0170]

[0171] In Table 3, Δk j,l may represent the parameter update value of the shape parameter corresponding to the sample advertisement i being divided into the lth leaf node of the jth tree; Δθ j,l may represent the parameter update value of the slope parameter corresponding to the sample advertisement i being divided into the lth leaf node of the jth tree.

[0172] It should be understood that the computer device can acquire tree convergence conditions associated with the initial probability boosting tree (e.g., probability boosting tree 51T). For example, the tree convergence conditions here can be used to indicate that the distribution error of the target probability boosting tree is less than a pre-set distribution error threshold (e.g., 0.3). At this point, the computer device needs to determine the sample conversion data probability distribution (i.e., the actual conversion data probability distribution) based on the sample ad conversion data of each of the five sample ads. Further, when the predicted conversion data probability distribution output by probability boosting tree 51T is obtained based on the sample conversion data probability distribution parameters shown in Table 3, the computer device can determine the distribution error (e.g., root mean square error) of probability boosting tree 51T based on the sample conversion data probability distribution and the predicted conversion data probability distribution.

[0173] Understandably, when the distribution error (e.g., 0.2) is less than or equal to the distribution error threshold (e.g., 0.3), the computer device can determine that the probability boosting tree 51T satisfies the tree convergence condition. In this case, the computer device can directly use the probability boosting tree 51T that satisfies the tree convergence condition as the target probability boosting tree for predicting the probability distribution of conversion data for the target advertisement. At this time, the target probability boosting tree may include one tree, namely tree 1.

[0174] Optionally, when the distribution error (0.5) is greater than the distribution error threshold (e.g., 0.3), the computer device can determine that the probability boosting tree 51T does not meet the tree convergence condition. In this case, the computer device can adjust the tree parameters of the probability boosting tree 51T and perform a second round of iterative training, that is, construct tree 2 based on tree 1. At this time, the computer device can use the adjusted probability boosting tree 51T as an interim probability boosting tree (e.g., Figure 6 The probability boosting tree 52T shown is an example. This probability boosting tree 52T can include tree 1 and tree 2. Further, the computer device needs to determine whether the current probability boosting tree 52T satisfies the tree convergence condition. If the current probability boosting tree 52T satisfies the tree convergence condition, the computer device can use the probability boosting tree 52T including tree 1 and tree 2 as the target probability boosting tree. If the current probability boosting tree 52T does not satisfy the tree convergence condition, the computer device can continue to construct a new tree 3 based on the current probability boosting tree 52T to obtain a new probability boosting tree 52T. At this time, the new probability boosting tree 52T includes tree 1, tree 2, and tree 3, until the new probability boosting tree satisfies the tree convergence condition. Then, the computer device can use the new probability boosting tree that satisfies the tree convergence condition as the target probability boosting tree for predicting the probability distribution of conversion data for the target advertisement.

[0175] In the embodiments of the present application, the computer device can determine the root node for constructing the initial probability boosting tree according to the N sample data pairs corresponding to the N sample advertisements, and then can determine the first best split point through the sample advertisement attribute features of each sample advertisement in the root node, divide the sample advertisements in the root node, and thus obtain the initial probability boosting tree. Further, the computer device can iteratively train the initial probability boosting tree, and thus obtain the target probability boosting tree for predicting the conversion data probability distribution (for example, the conversion cost probability distribution) of the target advertisement. The entire process does not need to consume a large amount of computing resources to calculate the correlation parameters (for example, the covariance matrix) between all sample data as in the traditional probability estimation method, but uses the target probability boosting tree to quickly predict the conversion data probability distribution of the target advertisement, so as to reduce the calculation time and thus improve the prediction speed of the probability distribution.

[0176] Further, please refer to Figure 6 , Figure 6 is a flowchart of a data processing method provided by the embodiments of the present application. As Figure 1 indicated, the method can be executed by a user terminal with a boosting tree training function (for example, any one of the user terminals in the user terminal cluster shown in the above Figure 1 , for example, the user terminal 100a), can be executed by a server with a boosting tree training function (for example, the server 10F shown in the above Figure 3 , and can be executed by interaction between a user terminal with a boosting tree application function and a server with a boosting tree training function. Herein, no limitation is made. The method can at least include the following steps S201-S211:

[0177] Step S201, obtaining N sample advertisements and N sample data pairs corresponding to the N sample advertisements.

[0178] Step S202, determining the root node for constructing the initial probability boosting tree based on the N sample advertisements, taking the sample advertisement attribute features of each sample advertisement in the root node as the initial attribute features, determining the first split point set consistent with the feature type of the initial attribute features, determining the first best split point from the first split point set, dividing the N sample advertisements according to the split condition indicated by the first best split point, and obtaining the initial probability boosting tree.

[0179] Step S203, iteratively training the initial probability boosting tree based on the tree convergence condition associated with the initial probability boosting tree and the sample advertisement conversion data of each sample advertisement in the N sample advertisements, and obtaining the target probability boosting tree for predicting the conversion data probability distribution of the target advertisement.

[0180] In the data processing method in the embodiments of the present application, the boosting tree training process and the boosting tree application process can be included. It can be understood that steps S201-S203 describe the boosting tree application process, and the specific implementation of the boosting tree application process can be referred to the above description. Figure 7a The description of steps S101-S103 in the corresponding embodiments will not be repeated here.

[0181] For better understanding, further, please refer to Figure 7a , Figure 7a is a comparison diagram of the predicted conversion data probability distribution determined by the target probability boosting tree and the actual conversion data probability distribution provided by the embodiments of the present application. As shown in Figure 7a , line L1 refers to the predicted probability distribution (i.e. estimated distribution) of the conversion data of a certain advertisement in scenario 1 by the computer device using the target probability boosting tree, and line L2 refers to the estimated average distribution fitted after predicting the conversion data probability distribution of all advertisements in scenario 1 by the computer device using the target probability boosting tree, as shown in Figure 7a , the estimated distribution in scenario 1 can better fit the real distribution.

[0182] Similarly, line L3 refers to the predicted probability distribution (i.e. estimated distribution) of the conversion data of a certain advertisement in scenario 2 by the computer device using the target probability boosting tree, and line L4 refers to the estimated average distribution fitted after predicting the conversion data probability distribution of all advertisements in scenario 2 by the computer device using the target probability boosting tree, as shown in Figure 7a , the estimated distribution in scenario 2 can also better fit the real distribution.

[0183] In the embodiments of the present application, the distribution difference between the estimated distribution and the real distribution can be measured by the index of JS divergence (Jensen Shannon Divergence, JSD). The value range of JSD is between 0 and 1, and the closer to 0, the smaller the distribution difference, and vice versa. Figure 7b As shown in the figure, the JSD is less than 0.2.

[0184] For better understanding, further, please refer to Figure 7b , Figure 7b is a comparison diagram of the predicted conversion data probability distribution determined by the target probability boosting tree and the actual conversion data probability distribution provided by the embodiments of the present application. As shown in Figure 7aAs shown, line L1 refers to a probability distribution (i.e., an estimated distribution) of conversion data of a certain advertisement in scenario 3 (e.g., a secondary stay scenario) predicted by the computer device using the target probability boosting tree, and line L2 refers to an estimated average distribution fitted after predicting the probability distribution of conversion data of all advertisements in scenario 3 using the target probability boosting tree, as shown in Figure 7b As shown, the estimated distribution in scenario 3 can better fit the real distribution.

[0185] Similarly, line L3 refers to a probability distribution (i.e., an estimated distribution) of conversion data of a certain advertisement in scenario 4 (e.g., an activation scenario) predicted by the computer device using the target probability boosting tree, and line L4 refers to an estimated average distribution fitted after predicting the probability distribution of conversion data of all advertisements in scenario 4 using the target probability boosting tree, as shown in Figure 7b As shown, the estimated distribution in scenario 4 can also better fit the real distribution. Wherein, Figure 8 As shown, the JSD is less than 0.2.

[0186] The boosting tree application process can refer to the descriptions of steps S204-S206 below.

[0187] In step S204, a delivery feature request for a target advertisement is obtained.

[0188] The delivery feature request can include target advertisement attribute features of the target advertisement.

[0189] In step S205, a target probability boosting tree associated with the target advertisement is obtained, and the target advertisement attribute features are input into the target probability boosting tree to output a probability distribution of conversion data of the target advertisement.

[0190] Specifically, the computer device can obtain a target probability boosting tree associated with the target advertisement. The target probability boosting tree can be obtained by iteratively training an initial probability boosting tree based on a tree convergence condition and sample advertisement conversion data of each sample advertisement in N sample advertisements. The N sample advertisements can be used to construct a root node of the initial probability boosting tree. The initial probability boosting tree is obtained by dividing the N sample advertisements based on a split condition indicated by a first best split point. The first best split point is determined based on sample advertisement attribute features of each sample advertisement in the root node. N is a positive integer. Further, the computer device can input the target advertisement attribute features into the target probability boosting tree to output a probability distribution of conversion data of the target advertisement.

[0191] In step S206, the target advertisement conversion data of the target advertisement is determined based on the conversion data probability distribution of the target advertisement.

[0192] Specifically, the computer device can select the maximum conversion data probability from the conversion data probability distribution of the target advertisement, and then the advertisement conversion data corresponding to the selected maximum conversion data probability can be taken as the target advertisement conversion data of the target advertisement.

[0193] For ease of understanding, further, please refer to Figure 8 , Figure 8 is a scene schematic diagram for determining target conversion data provided by an embodiment of the present application. As Figure 1 indicated, the server 8F in the embodiment of the present application can be a computer device with a boosted tree training function, which can be the server 10F as indicated in the above Figure 8 .

[0194] It should be understood that the conversion data in the embodiment of the present application can be taken as an example of advertisement conversion cost, which is used to describe the specific implementation of the computer device using a model based on target probability boosted tree to predict the advertisement data of the target advertisement. It should be understood that the conversion link of the advertisement in the advertisement conversion scenario is long, and the advertisement 8G as indicated in the above Figure 8 (e.g., a game advertisement targeting payment assets) is taken as an example: the entire download link at least goes through exposure, intermediate conversion links (e.g., clicking, downloading, installing, activating), final conversion (e.g., first payment of assets), and the like, and the ultimate goal of the advertiser is often the payment of assets by the user, so the advertiser needs to assess the cost required for the first payment of assets before launching the advertisement. On the one hand, for any advertisement, the need to estimate the conversion data (e.g., cost) distribution before the advertisement is launched. On the other hand, in the long advertisement conversion process, there are many influencing factors, and these factors will affect the final distribution of the conversion cost. Based on this, the embodiment of the present application can collect the conversion cost of the historical advertisement and the advertisement attribute features of the historical advertisement to construct a sample data pair for training an initial probability boosted tree, thereby obtaining a model (e.g., a distribution estimation model based on target boosted tree) capable of estimating the complete conversion data probability distribution, and providing a reference for the advertiser to control the conversion data.

[0195] The placement feature request for the advertisement 8G in the embodiments of the present application can be generated by the server 8F based on the advertisement attribute features of the advertisement 8G when the advertiser creates the advertisement 8G. The placement feature request can also be sent by a user terminal (for example, a terminal device used by the advertiser) having a network connection relationship with the server 8F to the server 8F. Here, the placement feature request will not be limited. At this time, the server 8F can obtain the advertisement attribute features of the advertisement 8G carried in the placement feature request when receiving the placement feature request. Then, the server 8F can input the advertisement attribute features of the advertisement 8G into the pre-trained target probability boosting tree, so as to output the conversion data probability distribution (for example, the conversion cost probability distribution) of the advertisement 8G.

[0196] Further, the server 8F can obtain Figure 8 the confidence intervals of the estimation errors of different positions of the advertisement conversion link to adaptively determine the final actual conversion cost of the advertisement 8G according to the conversion data probability distribution of the advertisement 8G. For example, as shown in Figure 9 the confidence interval of the estimation error of the click-through rate can be [0, 2], the confidence interval of the estimation error of the conversion rate can be [0, 1.3], and the confidence interval of the estimation error of the deep conversion rate can be [0, 0.7]. Then, when the computer device needs to urgently obtain the click rate of the advertisement 8G, the computer device can take the maximum cost in the conversion data probability distribution of the advertisement 8G as the target advertisement conversion data of the advertisement 8G, and return the target advertisement conversion data to the terminal device used by the advertiser, so that the advertiser sets the actual conversion cost of the advertisement 8G by referring to the target advertisement conversion data. Optionally, the server 8F can also return the conversion data probability distribution of the advertisement 8G to the terminal device used by the advertiser, so that the advertiser sets the actual conversion cost of the advertisement 8G by referring to the conversion data probability distribution of the advertisement 8G.

[0197] In the embodiments of the present application, the computer device can determine the root node for constructing the initial probability boosting tree according to the N sample data pairs corresponding to the N sample ads, and then can determine the first best split point through the sample ad attribute features of each sample ad in the root node, divide the sample ads in the root node, and thus obtain the initial probability boosting tree. Further, the computer device can iteratively train the initial probability boosting tree, and thus obtain a target probability boosting tree for predicting the conversion data probability distribution (for example, the conversion cost probability distribution) of the target ad. The entire process does not need to consume a large amount of computing resources to calculate the correlation parameters (for example, the covariance matrix) between all sample data as in the traditional probability estimation method, but uses the target probability boosting tree to quickly and accurately predict the conversion data probability distribution of the target ad, so as to reduce the calculation time and improve the prediction speed of the probability distribution. In addition, since the computer device needs to consume a large amount of computing resources, the target probability boosting tree can also be applicable in a large number of application scenarios, thereby improving the applicability of the probability distribution prediction.

[0198] Further, please refer to Figure 9 , Figure 9 is a structural schematic diagram of a data processing apparatus provided by the embodiments of the present application. As shown in Figure 9 , the data processing apparatus 1 can be a computer program (including program code) running in a computer device, for example, the data processing apparatus 1 is an application software; the data processing apparatus 1 can be used to execute the corresponding steps in the method provided by the embodiments of the present application. As shown in Figure 1 , the data processing apparatus 1 can run in a computer device with a boosting tree training function, which can be the server 10F in the embodiments corresponding to the above Figure 1 , or can be any one of the user terminals in the user terminal cluster in the embodiments corresponding to the above Figure 3 , for example, the user terminal 100a. The data processing apparatus 1 can include: a sample ad acquisition module 10, an initial tree determination module 20, and an iterative training module 30.

[0199] The sample ad acquisition module 10 is configured to acquire N sample ads and N sample data pairs corresponding to the N sample ads; N is a positive integer; one sample ad corresponds to one sample data pair, and one sample data pair is determined by sample ad attribute features and sample ad conversion data;

[0200] The initial tree determination module 20 is configured to determine a root node for constructing an initial probability boosting tree based on the N sample advertisements, take a sample advertisement attribute feature of each sample advertisement in the root node as an initial attribute feature, determine a first split point set corresponding to a feature type of the initial attribute feature, determine a first best split point from the first split point set, divide the N sample advertisements according to a split condition indicated by the first best split point, and obtain the initial probability boosting tree.

[0201] The initial tree determination module 20 includes a first split point determination unit 201, a first division unit 202, a stop condition acquisition unit 203, a second split point determination unit 204, and a second division unit 205.

[0202] The first split point determination unit 201 is configured to determine a root node for constructing an initial probability boosting tree based on the N sample advertisements, take a sample advertisement attribute feature of each sample advertisement in the root node as an initial attribute feature, and determine a first split point set corresponding to a feature type of the initial attribute feature.

[0203] The first split point determination unit 201 includes an initial feature determination subunit 2011, a duplicate processing subunit 2012, and a split point screening subunit 2013.

[0204] The initial feature determination subunit 2011 is configured to determine a root node for constructing an initial probability boosting tree based on the N sample advertisements, and take a sample advertisement attribute feature of each sample advertisement in the root node as an initial attribute feature.

[0205] The duplicate processing subunit 2012 is configured to take a feature type of the initial attribute feature as a split point, perform duplicate processing on the split point, and add the split point after the duplicate processing to the first split point set.

[0206] The split point screening subunit 2013 is configured to screen a best split point that satisfies a split point determination condition from the first split point set based on the split point determination condition for the root node, and take the screened best split point as the first best split point.

[0207] The first split point set includes M split points; M is a positive integer; and the M split points include a split point F a a is less than or equal to M;

[0208] The split point screening unit 2013 includes a loss reduction parameter determination subunit 20130, a first split point acquisition subunit 20131, a first parameter acquisition subunit 20132, a first comparison subunit 20133, a first determination subunit 20134, a second split point acquisition subunit 20135, a second determination subunit 20136, a second parameter acquisition subunit 20137, a second comparison subunit 20138, and a third determination subunit 20139.

[0209] The loss reduction parameter determination subunit 20130 is configured to acquire a split point determination condition for the root node, and acquire a node loss reduction parameter corresponding to the root node.

[0210] The first split point acquisition subunit 20131 is configured to acquire a split point F a from a first split point set. a Based on a split condition indicated by the split point F a , the sample advertisements in the root node are divided to obtain initial child nodes of the root node; the initial child nodes include a first initial node and a second initial node.

[0211] The first parameter acquisition subunit 20132 is configured to acquire a first feature loss reduction parameter associated with the split point F a .

[0212] The total probability distribution indicated by the root node includes a total shape parameter and a total slope parameter; the first sub-probability distribution indicated by the first initial node includes a first shape parameter and a first slope parameter; and the second sub-probability distribution indicated by the second initial node includes a second shape parameter and a second slope parameter.

[0213] The first parameter acquisition subunit 20132 is further configured to:

[0214] respectively, a total number of sample advertisements in the root node, a first number of sample advertisements in the first initial node, and a second number of sample advertisements in the second initial node.

[0215] Based on the total shape parameter, the total slope parameter, and a node loss determination rule, a total node loss of the root node is determined.

[0216] Based on the first shape parameter, the first slope parameter, and the node loss determination rule, a first node loss of the first initial node is determined.

[0217] Based on the second shape parameter, the second slope parameter, and the node loss determination rule, a second node loss of the second initial node is determined.

[0218] obtaining a feature loss reduction rule, determining a split point F a associated with the first feature loss reduction parameter.

[0219] The first comparison sub-unit 20133 is configured to compare the first feature loss reduction parameter and the node loss reduction parameter.

[0220] The first determination sub-unit 20134 is configured to determine the split point F a as the first optimal split point if the node loss reduction parameter is less than or equal to the first feature loss reduction parameter. a as the first optimal split point if the node loss reduction parameter is less than or equal to the first feature loss reduction parameter.

[0221] The second split point obtaining sub-unit 20135 is configured to determine the split point F a from the first split point set if the node loss reduction parameter is greater than the first feature loss reduction parameter. a+1 ;

[0222] The second determination sub-unit 20136 is configured to determine the split point F a+1 as the first optimal split point if a+1 is equal to M.

[0223] The second parameter obtaining sub-unit 20137 is configured to obtain a second feature loss reduction parameter associated with the split point F a+1 based on the split point F a+1 if a+1 is less than M.

[0224] The second comparison sub-unit 20138 is configured to compare the second feature loss reduction parameter and the node loss reduction parameter.

[0225] The third determination sub-unit 20139 is configured to determine the split point F a+1 as the first optimal split point if the node loss reduction parameter is less than or equal to the second feature loss reduction parameter. a+1 as the first optimal split point if the node loss reduction parameter is less than or equal to the second feature loss reduction parameter.

[0226] The specific implementation of the loss reduction parameter determination subunit 20130, the first split point acquisition subunit 20131, the first parameter acquisition subunit 20132, the first comparison subunit 20133, the first determination subunit 20134, the second split point acquisition subunit 20135, the second determination subunit 20136, the second parameter acquisition subunit 20137, the second comparison subunit 20138, and the third determination subunit 20139 can be referred to the related description above. Figure 3 The description of the best split point in the corresponding embodiment will not be repeated here.

[0227] The specific implementation of the initial feature determination subunit 2011, the de-duplication processing subunit 2012, and the split point screening subunit 2013 can be referred to the related description above. Figure 3 The description of the best split point in the corresponding embodiment will not be repeated here.

[0228] The first division unit 202 is configured to divide the sample advertisements in the root node according to the split condition indicated by the first best split point, to obtain a first type node and a second type node.

[0229] The first division unit 202 includes a sample advertisement acquisition subunit 2021, a first division subunit 2022, and a second division subunit 2023.

[0230] The sample advertisement acquisition subunit 2021 is configured to acquire a sample advertisement i from the root node; i is a positive integer less than or equal to N.

[0231] The first division subunit 2022 is configured to divide the sample advertisement i to the first type node if the sample advertisement i satisfies the split condition indicated by the first best split point.

[0232] The second division subunit 2023 is configured to divide the sample advertisement i to the second type node if the sample advertisement i does not satisfy the split condition indicated by the first best split point; the first type node and the second type node are both child nodes of the root node.

[0233] The specific implementation of the sample advertisement acquisition subunit 2021, the first division subunit 2022, and the second division subunit 2023 can be referred to the related description above. Figure 3 The description of the division of the sample advertisements in the root node in the corresponding embodiment will not be repeated here.

[0234] The stop condition acquisition unit 203 is configured to acquire the split stop condition associated with the initial probability boosting tree by taking the first type node and the second type node as to-be-divided nodes.

[0235] The second split point determination unit 204 is configured to, when the node to be split does not satisfy the split stop condition, take a sample advertisement attribute feature of each sample advertisement in the node to be split as a target attribute feature, determine a second split point set corresponding to a feature type of the target attribute feature, and determine a second best split point from the second split point set.

[0236] The second division unit 205 is configured to divide the sample advertisements in the node to be split according to a split condition indicated by the second best split point, and construct an initial probability boosting tree based on the root node and the divided nodes when the divided nodes satisfy the split stop condition.

[0237] The specific implementation of the first split point determination unit 201, the first division unit 202, the stop condition acquisition unit 203, the second split point determination unit 204, and the second division unit 205 can refer to the description of the corresponding embodiments of step S102, which will not be repeated here. Figure 3 The description of step S102 in the corresponding embodiments will not be repeated here.

[0238] The iterative training module 30 is configured to perform iterative training on the initial probability boosting tree based on a tree convergence condition associated with the initial probability boosting tree and sample advertisement conversion data of each sample advertisement in the N sample advertisements, to obtain a target probability boosting tree for predicting a conversion data probability distribution of a target advertisement.

[0239] The iterative training module 30 includes a tree convergence condition acquisition unit 301, a sample probability distribution determination unit 302, a predicted probability distribution determination unit 303, a first target tree determination unit 304, a condition not satisfied unit 305, and a second target tree determination unit 306.

[0240] The tree convergence condition acquisition unit 301 is configured to acquire a tree convergence condition associated with the initial probability boosting tree; the tree convergence condition includes a distribution error threshold;

[0241] The sample probability distribution determination unit 302 is configured to determine a sample conversion data probability distribution based on sample advertisement conversion data of each sample advertisement in the N sample advertisements.

[0242] The predicted probability distribution determination unit 303 is configured to, when a predicted conversion data probability distribution output by the initial probability boosting tree is acquired, determine a distribution error of the initial probability boosting tree based on the sample conversion data probability distribution and the predicted conversion data probability distribution.

[0243] The first target tree determination unit 304 is configured to determine that the initial probability boosting tree satisfies the tree convergence condition when the distribution error is less than or equal to the distribution error threshold, and determine the initial probability boosting tree satisfying the tree convergence condition as the target probability boosting tree for predicting the probability distribution of the conversion data of the target advertisement.

[0244] The condition not satisfied unit 305 is configured to determine that the initial probability boosting tree does not satisfy the tree convergence condition when the distribution error is greater than the distribution error threshold.

[0245] The second target tree determination unit 306 is configured to adjust the tree parameters of the initial probability boosting tree, and determine the initial probability boosting tree after the adjustment as a transition probability boosting tree, until the transition probability boosting tree satisfies the tree convergence condition, and determine the transition probability boosting tree satisfying the tree convergence condition as the target probability boosting tree for predicting the probability distribution of the conversion data of the target advertisement.

[0246] The specific implementation of the tree convergence condition acquisition unit 301, the sample probability distribution determination unit 302, the predicted probability distribution determination unit 303, the first target tree determination unit 304, the condition not satisfied unit 305, and the second target tree determination unit 306 can refer to the descriptions of the corresponding embodiments of the above Figure 3 The descriptions of steps S101-S103 in the corresponding embodiments will not be repeated here.

[0247] The specific implementation of the sample advertisement acquisition module 10, the initial tree determination module 20, and the iterative training module 30 can refer to the descriptions of the corresponding embodiments of the above Figure 10 The descriptions of steps S101-S103 in the corresponding embodiments will not be repeated here. In addition, the descriptions of the beneficial effects of using the same method will not be repeated here.

[0248] Further, please refer to Figure 10 , Figure 10 is a structural schematic diagram of a data processing apparatus provided in an embodiment of the present application. The data processing apparatus 2 can be a computer program (including program code) running in a computer device, for example, the data processing apparatus 2 is an application software; the data processing apparatus 2 can be used to execute the corresponding steps in the method provided in the embodiments of the present application. As shown in Figure 1 , the data processing apparatus 2 can run in a computer device with a boosting tree training function, which can be the server 10F in the above Figure 1 corresponding embodiments, or can be the server 10F in the above Figure 6Any one of the user terminals in the user terminal cluster in the corresponding embodiment runs a user terminal of the target probability boosting tree, for example, the user terminal 100a. The data processing apparatus 2 can include: a delivery request acquisition module 100, a probability distribution determination module 200, and a target conversion data determination module 300.

[0249] The delivery request acquisition module 100 is configured to acquire a delivery feature request for a target advertisement; the delivery feature request includes a target advertisement attribute feature of the target advertisement.

[0250] The probability distribution determination module 200 is configured to acquire a target probability boosting tree associated with the target advertisement, input the target advertisement attribute feature into the target probability boosting tree, and output a conversion data probability distribution of the target advertisement by the target probability boosting tree; the target probability boosting tree is obtained by iteratively training an initial probability boosting tree based on a tree convergence condition and sample advertisement conversion data of each sample advertisement in N sample advertisements; the N sample advertisements are used to construct a root node of the initial probability boosting tree; the initial probability boosting tree is obtained by dividing the N sample advertisements based on a split condition indicated by a first best split point; the first best split point is determined based on sample advertisement attribute features of each sample advertisement in the root node; N is a positive integer.

[0251] The target conversion data determination module 300 is configured to determine target advertisement conversion data of the target advertisement based on the conversion data probability distribution of the target advertisement.

[0252] The specific implementation of the delivery request acquisition module 100, the probability distribution determination module 200, and the target conversion data determination module 300 can be referred to the above Figure 11 The corresponding embodiments describe steps S201-S206, which will not be described again. In addition, the beneficial effects of using the same method will not be described again.

[0253] Further, please refer to Figure 11 , Figure 11 is a schematic diagram of a computer device provided by the embodiments of the present application. As shown in Figure 11As shown, the computer device 1000 can be a computer device with boosting tree training capabilities. The computer device 1000 may include: at least one processor 1001, such as a CPU; at least one network interface 1004; a memory 1005; and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 11 As shown, the memory 1005, as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a device control application program. In some embodiments, the computer device may further include... Figure 1 The user interface 1003 shown, for example, if the computer device is Figure 11 The user terminal shown has a boosting tree training function (e.g., user terminal 100a). The computer device may also include the user interface 1003, which may include a display screen, a keyboard, etc.

[0254] exist Figure 3 In the computer device 1000 shown, the network interface 1004 is mainly used for network communication; the user interface 1003 is mainly used to provide an input interface for the user; and the processor 1001 can be used to call the device control application stored in the memory 1005.

[0255] It should be understood that the computer device 1000 described in the embodiments of this application can execute the foregoing text. Figure 6 and Figure 9 The description of the data processing method in the corresponding embodiment can also be performed as described above. Figure 10 The corresponding embodiment refers to the data processing device 1 or Figure 3 The description of the data processing device 2 in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated here.

[0256] Furthermore, it should be noted that this application embodiment also provides a computer-readable storage medium storing a computer program executed by the aforementioned data processing device 1 or data processing device 2, and the computer program includes program instructions. When the processor executes the program instructions, it can execute the aforementioned... Figure 6 orFigure 3 The description of the data processing method in the corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer-readable storage medium embodiments involved in the present application, please refer to the description of the method embodiments of the present application. As an example, the program instructions can be deployed to execute on one computing device, or on multiple computing devices located in one place, or on multiple computing devices distributed in multiple places and interconnected through a communication network, which can constitute a blockchain system.

[0257] In one aspect, the present application provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device can execute the above Figure 6 Or Figure 12 The description of the data processing method in the corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of using the same method will also not be repeated.

[0258] Further, please refer to Figure 12 , Figure 9 is a structural schematic diagram of a data processing system provided by an embodiment of the present application. The data processing system 3 can include a data processing apparatus 1a and a data processing apparatus 2a. The data processing apparatus 1a can be the above-mentioned Figure 1 The data processing apparatus 1 in the corresponding embodiment can be integrated into a computer device with a boosted tree training function. For example, the computer device can be the above-mentioned Figure 1 The server 10F in the corresponding embodiment can also be the above-mentioned Figure 10 The user terminal running the target probability boosting tree in the user terminal cluster in the corresponding embodiment, for example, the user terminal 100a. Therefore, the description will not be repeated here. The data processing apparatus 2a can be the above-mentioned Figure 1 The data processing apparatus 2 in the corresponding embodiment can be integrated into a computer device with a boosted tree application function. The computer device can be the above-mentioned ​Any one of the user terminals in the user terminal cluster in the corresponding embodiment runs the user terminal of the target probability boosting tree, for example, the user terminal 100a. Therefore, it will not be described here. In addition, the beneficial effect of using the same method will not be described here. For technical details not disclosed in the data processing system embodiments involved in the present application, please refer to the description of the method embodiments of the present application.

[0259] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The above-mentioned program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments of each method. The above-mentioned storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

[0260] The above only discloses the preferred embodiments of the present application, and of course cannot limit the scope of the rights of the present application. Therefore, equivalent changes made in accordance with the claims of the present application are still within the scope of the present application.

Claims

1. A data processing method, characterized in that, include: Obtain N sample advertisements and N sample data pairs corresponding to the N sample advertisements; where N is a positive integer; One sample ad corresponds to one sample data pair, and one sample data pair is determined by the sample ad attribute features and sample ad conversion data; Based on the N sample advertisements, determine the root node for constructing the initial probability boosting tree, take the sample advertisement attribute features of each sample advertisement in the root node as the initial attribute features, determine the first set of split points that matches the feature type of the initial attribute features, and determine the first optimal split point from the first set of split points. Based on the splitting conditions indicated by the first optimal splitting point, the sample advertisements in the root node are divided to obtain a first type of node and a second type of node. The first type of node and the second type of node are respectively taken as nodes to be split, and the splitting cutoff condition associated with the initial probability boosting tree is obtained; When the node to be split does not meet the split cutoff condition, the sample advertisement attribute features of each sample advertisement in the node to be split are taken as target attribute features, a second set of split points that matches the feature type of the target attribute features is determined, and a second optimal split point is determined from the second set of split points. According to the splitting condition indicated by the second optimal splitting point, the sample advertisements in the node to be split are divided until the splitting nodes meet the splitting cutoff condition. Based on the root node and the splitting nodes, the initial probability boosting tree is constructed. Based on the tree convergence condition associated with the initial probability boosting tree and the sample ad conversion data of each of the N sample ads, the initial probability boosting tree is iteratively trained to obtain a target probability boosting tree for predicting the conversion data probability distribution of the target ad.

2. The method according to claim 1, characterized in that, Based on the N sample advertisements, the root node for constructing the initial probability boosting tree is determined. The sample advertisement attribute features of each sample advertisement in the root node are used as initial attribute features. A first set of split points matching the feature type of the initial attribute features is determined. A first optimal split point is determined from the first set of split points, including: Based on the N sample advertisements, determine the root node for constructing the initial probability boosting tree, and use the sample advertisement attribute features of each sample advertisement in the root node as the initial attribute features. The feature type of the initial attribute feature is used as the split point. The split point is deduplicated and then added to the first split point set. Based on the split point determination conditions for the root node, the best split point that satisfies the split point determination conditions is selected from the first set of split points, and the selected best split point is taken as the first best split point.

3. The method according to claim 2, characterized in that, The first set of split points includes M split points; M is a positive integer; the M split points include split point F. a The a is less than or equal to the M; The step of selecting the optimal split point from the first set of split points based on the split point determination conditions for the root node, and taking the selected optimal split point as the first optimal split point, includes: Obtain the split point determination conditions for the root node, and obtain the node loss reduction parameters corresponding to the root node; Obtain the split point F from the first set of split points. a Based on the split point F a The indicated splitting condition divides the sample advertisements in the root node to obtain the initial child nodes of the root node; the initial child nodes include a first initial node and a second initial node; Obtain the split point F a The associated first feature loss reduction parameter; Compare the first feature loss reduction parameter and the node loss reduction parameter; If the node loss reduction parameter is less than or equal to the first feature loss reduction parameter, then the split point F is determined. a If the split point determination condition is met, then the split point F is set. a As the first optimal split point.

4. The method according to claim 3, characterized in that, The total probability distribution indicated by the root node includes a total shape parameter and a total slope parameter; the first sub-probability distribution indicated by the first initial node includes a first shape parameter and a first slope parameter; the second sub-probability distribution indicated by the second initial node includes a second shape parameter and a second slope parameter. The acquisition of the split point F a The associated first feature loss reduction parameters include: The total number of sample ads in the root node, the first number of sample ads in the first initial node, and the second number of sample ads in the second initial node are counted respectively. Based on the total shape parameter, the total slope parameter, and the node loss determination rule, the total node loss of the root node is determined; Based on the first shape parameter, the first slope parameter, and the node loss determination rule, the first node loss of the first initial node is determined. Based on the second shape parameter, the second slope parameter, and the node loss determination rule, the second node loss of the second initial node is determined. Obtain the feature loss reduction rule, and based on the total number of samples, the first number of samples, the second number of samples, the summary point loss, the first node loss, the second node loss, and the feature loss reduction rule, determine the split point F. a The associated first feature loss reduction parameter.

5. The method according to claim 3, characterized in that, The method further includes: If the node loss reduction parameter is greater than the first feature loss reduction parameter, then the split point F is determined. a If the conditions for determining the split point are not met, obtain the split point F from the first set of split points. a+1 ; If a+1 equals M, then the split point F is... a+1 As the first optimal split point.

6. The method according to claim 5, characterized in that, The method further includes: If a+1 is less than M, then based on the split point F a+1 The indicated splitting condition is used to divide the sample advertisements in the root node and obtain the result at the splitting point F. a+1 The associated second feature loss reduction parameter; The second feature loss reduction parameter and the node loss reduction parameter are compared; If the node loss reduction parameter is less than or equal to the second feature loss reduction parameter, then the split point F is determined. a+1 If the split point determination condition is met, then the split point F is set. a+1 As the first optimal split point.

7. The method according to claim 1, characterized in that, The step of dividing the sample advertisements in the root node according to the splitting conditions indicated by the first optimal splitting point to obtain a first type of node and a second type of node includes: Obtain sample advertisement i from the root node; i is a positive integer less than or equal to N; If the sample advertisement i satisfies the splitting condition indicated by the first optimal splitting point, then the sample advertisement i is assigned to the first type of node; If the sample advertisement i does not meet the splitting conditions indicated by the first optimal splitting point, then the sample advertisement i is divided into a second type of node; both the first type of node and the second type of node are child nodes of the root node.

8. The method according to claim 1, characterized in that, The step of iteratively training the initial probability boosting tree based on the tree convergence condition associated with the initial probability boosting tree and the sample ad conversion data of each of the N sample ads to obtain a target probability boosting tree for predicting the conversion data probability distribution of the target ad includes: Obtain the tree convergence condition associated with the initial probability boosting tree; the tree convergence condition includes a distribution error threshold; Determine the probability distribution of sample conversion data based on the sample conversion data of each sample ad in N sample ads; When the predicted conversion data probability distribution output by the initial probability boosting tree is obtained, the distribution error of the initial probability boosting tree is determined based on the sample conversion data probability distribution and the predicted conversion data probability distribution. When the distribution error is less than or equal to the distribution error threshold, the initial probability boosting tree is determined to satisfy the tree convergence condition, and the initial probability boosting tree that satisfies the tree convergence condition is used as the target probability boosting tree for predicting the probability distribution of conversion data for the target advertisement.

9. The method according to claim 8, characterized in that, The method further includes: When the distribution error is greater than the distribution error threshold, it is determined that the initial probability boosting tree does not satisfy the tree convergence condition; The tree parameters of the initial probability boosting tree are adjusted, and the adjusted initial probability boosting tree is used as the transition probability boosting tree. When the transition probability boosting tree meets the tree convergence condition, the transition probability boosting tree that meets the tree convergence condition is used as the target probability boosting tree for predicting the probability distribution of conversion data of the target advertisement.

10. A data processing method, characterized in that, include: Obtain a delivery feature request for the target advertisement; the delivery feature request includes the target advertisement attribute features of the target advertisement; Obtain the target probability boosting tree associated with the target advertisement, input the attribute features of the target advertisement into the target probability boosting tree, and output the conversion data probability distribution of the target advertisement from the target probability boosting tree; the target probability boosting tree is obtained by iteratively training an initial probability boosting tree based on the tree convergence condition and the sample advertisement conversion data of each sample advertisement in N sample advertisements; the N sample advertisements are used to construct the root node of the initial probability boosting tree; the initial probability boosting tree is constructed based on the root node and the partitioned nodes, according to the splitting condition indicated by the second optimal splitting point, dividing the sample advertisements in the node to be split until the partitioned node meets the splitting cutoff condition, wherein the second optimal splitting point is... When the node to be split does not meet the splitting cutoff condition, the sample advertisement attribute features of each sample advertisement in the node to be split are taken as target attribute features and determined from a second set of splitting points that match the feature type of the target attribute features; the splitting cutoff condition is obtained when the first type node and the second type node are taken as nodes to be split respectively, and the first type node and the second type node are obtained by dividing the sample advertisements in the root node according to the splitting condition indicated by the first optimal splitting point, the first optimal splitting point is determined by taking the sample advertisement attribute features of each sample advertisement in the root node as initial attribute features and determining from a first set of splitting points that match the feature type of the initial attribute features; N is a positive integer; Based on the probability distribution of the conversion data of the target advertisement, the target advertisement conversion data is determined.

11. A data processing apparatus, characterized in that, include: The sample ad acquisition module is used to acquire N sample ads and N sample data pairs corresponding to the N sample ads; where N is a positive integer; one sample ad corresponds to one sample data pair, and one sample data pair is determined by the attribute features of the sample ad and the conversion data of the sample ad; The initial tree determination module is used to determine the root node for constructing an initial probability boosting tree based on the N sample advertisements, take the sample advertisement attribute features of each sample advertisement in the root node as the initial attribute features, determine a first set of split points that matches the feature type of the initial attribute features, and determine a first optimal split point from the first set of split points. The initial tree determination module is further configured to divide the sample advertisements in the root node according to the splitting conditions indicated by the first optimal splitting point, to obtain a first type of node and a second type of node. The initial tree determination module is further configured to take the first type of node and the second type of node as nodes to be split, and obtain the splitting cutoff condition associated with the initial probability boosting tree; The initial tree determination module is further configured to, when the node to be split does not meet the split cutoff condition, take the sample advertisement attribute features of each sample advertisement in the node to be split as the target attribute features, determine a second set of split points that matches the feature type of the target attribute features, and determine a second optimal split point from the second set of split points. The initial tree determination module is further configured to divide the sample advertisements in the node to be split according to the splitting condition indicated by the second optimal splitting point, until the splitting nodes meet the splitting cutoff condition, and then construct the initial probability boosting tree based on the root node and the splitting nodes. The iterative training module is used to iteratively train the initial probability boosting tree based on the tree convergence condition associated with the initial probability boosting tree and the sample ad conversion data of each of the N sample ads, to obtain a target probability boosting tree for predicting the conversion data probability distribution of the target ad.

12. A data processing apparatus, characterized in that, include: The delivery request acquisition module is used to acquire delivery feature requests for the target advertisement; the delivery feature requests include the target advertisement attribute features of the target advertisement; A probability distribution determination module is used to obtain a target probability boosting tree associated with the target advertisement, input the attribute features of the target advertisement into the target probability boosting tree, and output the conversion data probability distribution of the target advertisement from the target probability boosting tree. The target probability boosting tree is obtained by iteratively training an initial probability boosting tree based on tree convergence conditions and the conversion data of each sample advertisement in N sample advertisements. The N sample advertisements are used to construct the root node of the initial probability boosting tree. The initial probability boosting tree is constructed based on the root node and the partitioned nodes, according to the splitting conditions indicated by the second optimal splitting point, dividing the sample advertisements in the node to be split until the partitioned nodes meet the splitting cutoff condition. The optimal split point is determined from a second set of split points that match the feature type of the target attribute feature, using the sample ad attribute features of each sample ad in the node to be split as the target attribute feature when the node to be split does not meet the split cutoff condition. The split cutoff condition is obtained when the first type node and the second type node are respectively used as nodes to be split, and the first type node and the second type node are obtained by dividing the sample ads in the root node according to the split condition indicated by the first optimal split point. The first optimal split point is determined from a first set of split points that match the feature type of the initial attribute feature, using the sample ad attribute features of each sample ad in the root node as the initial attribute feature. N is a positive integer. The target conversion data determination module is used to determine the target ad conversion data of the target ad based on the probability distribution of the conversion data of the target ad.

13. A computer device, characterized in that, include: Processor and memory; The processor is connected to a memory, wherein the memory is used to store a computer program, and the processor is used to invoke the computer program to cause the computer device to perform the method according to any one of claims 1-10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1-10.

15. A computer program product or computer program, characterized in that, The computer program product or computer program includes computer instructions stored in a computer-readable storage medium, the computer instructions being adapted to be read and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1-10.

Images