Method, device and recommendation system for training models for online prediction of click-through rate

By integrating the click-through rate prediction values ​​of multiple teacher models to train student models, the problem of high online prediction delay in the prior art is solved, and the prediction accuracy similar to that of the teacher model and the effect of reducing prediction delay is achieved.

CN112182362BActive Publication Date: 2024-11-26HUAWEI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010901702.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-08-31
Publication Date
2024-11-26
Estimated Expiration
2040-08-31

AI Technical Summary

Technical Problem

When the prior art recommends information to users in massive information, the online prediction delay based on multiple models is high, which affects the user experience.

Method used

By obtaining the click-through rate prediction values ​​of the sample set by multiple teacher models, fusion processing is performed, fusion prediction values ​​are generated, and student models are trained based on this until the student model meets the convergence conditions to achieve knowledge transfer and reduce prediction delay.

Benefits of technology

The trained student model can achieve a prediction accuracy similar to that of the teacher model, or even exceed the accuracy of the teacher model. At the same time, a single student model can complete the click-through rate prediction, reducing the prediction delay.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112182362B_ABST
    Figure CN112182362B_ABST
Patent Text Reader

Abstract

The embodiment of the present application discloses a method, device and recommendation system for training a model for online prediction of click-through rate in the field of artificial intelligence technology, which is used to reduce the delay of online prediction. The method of the embodiment of the present application includes: obtaining the predicted value of the click-through rate of each sample in a sample set by each teacher model in multiple teacher models, and fusing the predicted values ​​of the click-through rate of the same sample by multiple teacher models into a fused predicted value, and then determining the loss function based on the fused predicted value of the click-through rate corresponding to each sample in the sample set, and finally training the student model based on the loss function and the sample set until the student model meets the first convergence condition, and the student model can be used to predict the click-through rate online.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a method, device and recommendation system for training a model for online prediction of click-through rate. Background Art

[0002] Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a branch of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making. Research in the field of artificial intelligence includes robotics, natural language processing, computer vision, decision-making and reasoning, human-computer interaction, recommendation and search, basic AI theory, etc.

[0003] How to recommend information to users based on their operation behaviors and characteristics in a vast amount of information is also one of the important research directions of artificial intelligence. As information continues to become richer and more abundant, how to recommend information to users in a more targeted, efficient and accurate manner is the research focus of recommendation technology.

[0004] In order to improve the accuracy of recommending information to users, the current main method is to train two different models based on the user's historical click behavior data, and then use these two different models to predict the user's click rate for various information online, and then recommend information with high click rate to users.

[0005] However, this approach increases the complexity of online prediction and leads to higher prediction latency. Summary of the invention

[0006] Embodiments of the present application provide a method, device, and recommendation system for training a model for online prediction of click-through rate, so as to reduce the delay of online prediction.

[0007] A first aspect of an embodiment of the present application provides a method for training a model for online prediction of click-through rate, comprising:

[0008] Obtain a predicted value of the click rate of each sample in a sample set by each teacher model in a plurality of teacher models, wherein the plurality of teacher models are obtained by training based on the sample set, and the sample set contains a plurality of samples; the plurality of teacher models may contain models of the same type or models of different types; for example, the plurality of teacher models may include at least one of a decision tree model and a deep neural network model; fuse the predicted values ​​of the click rate of the same sample in the sample set by the plurality of teacher models to obtain a fused predicted value of the click rate, wherein each sample in the sample set corresponds to a fused predicted value of the click rate; determine a loss function based on the fused predicted value of the click rate corresponding to each sample in the sample set, and the loss function may be expressed in a variety of forms; train a student model based on the loss function and the sample set until the student model meets a first convergence condition, which may be of a variety of forms, and the student model that meets the first convergence condition is used to predict the click rate online.

[0009] The predicted values ​​of the click rate of the same sample in the sample set by multiple teacher models are fused into a fused predicted value of the click rate, the loss function is determined according to the fused predicted value of the click rate corresponding to each sample in the sample set, and the student model is trained based on the loss function and the sample set, thereby realizing the transfer of knowledge from multiple teacher models to a single student model; the trained student model can achieve a prediction accuracy similar to that of the teacher model, or even exceed that of the teacher model; at the same time, a single student model can complete the prediction of the click rate without the need for multiple models to predict online at the same time, thereby reducing the prediction delay.

[0010] In one implementation, the predicted values ​​of the click rate of the same sample in the sample set by multiple teacher models are fused to obtain a fused predicted value of the click rate, including: calculating the weight of each teacher model in the multiple teacher models according to the predicted values ​​of the click rate of the target sample by the multiple teacher models, the weight corresponding to the target sample, and the target sample is any sample in the sample set; according to the weight of each teacher model in the multiple teacher models, performing weighted averaging processing on the predicted values ​​of the click rate of the target sample by the multiple teacher models to obtain a fused predicted value of the click rate.

[0011] According to the predicted values ​​of the click rates of target samples by multiple teacher models, the weight of each teacher model in the multiple teacher models is calculated, and then the predicted values ​​of the click rates of the target samples by the multiple teacher models are fused according to the weight of each teacher model in the multiple teacher models, thereby realizing the adaptive fusion of multiple teacher models. Moreover, the fused predicted value of the click rate after fusion can reduce the interference caused by the erroneous prediction values ​​of one or more teacher models. Using the fused predicted value of the click rate to train the student model can improve the prediction accuracy of the student model, so that the prediction accuracy of the student model is close to that of the teacher model, or even exceeds that of the teacher model.

[0012] In one implementation, the loss function includes a distillation loss term, which is determined based on the fused prediction value of the click rate corresponding to each sample in the sample set; the first convergence condition is that the value of the distillation loss term no longer decreases in multiple consecutive iterative calculations. In the first convergence condition, the number of iterative calculations can be set according to actual needs.

[0013] Since the first convergence condition is that the value of the distillation loss term no longer decreases in multiple consecutive iterative calculations, there is no need to select an additional portion of samples from the sample set as verification samples during the training of the student model, so the samples in the sample set can be fully utilized.

[0014] In one implementation, the loss function also includes a non-distillation loss term, which is determined based on the sample label of each sample in the sample set, where the sample label represents whether the user has clicked on the sample or not. For example, when the sample label y is 1, it represents that the user has clicked on the sample; when the sample label y is 0, it represents that the user has not clicked on the sample.

[0015] The existence of the non-distillation loss term makes the predicted value of the click rate of a certain sample output by the student model continuously approach the fused predicted value of the click rate corresponding to the sample, and continuously approach the sample label of the sample, thereby improving the prediction accuracy of the student model, and even making the prediction accuracy of the student model exceed that of the teacher model.

[0016] In one implementation, before obtaining the predicted value of the click rate of each sample in the sample set by each teacher model in multiple teacher models, the method also includes: training the multiple teacher models based on the sample set until the multiple teacher models all meet the second convergence condition, and there can be multiple second convergence conditions.

[0017] In one implementation, training multiple teacher models based on a sample set until multiple teacher models all meet the second convergence condition includes: training a target teacher model based on a training sample set, the training sample set contains at least one sample in the sample set, and the target teacher model is any one of the multiple teacher models; testing the target teacher model based on a verification sample set to obtain a test result, and the training stops when the test result indicates that the target teacher model meets the second convergence condition, and the verification sample set contains at least one sample in the sample set; wherein the AUC indicator can be used as a test result for evaluating the performance of the target teacher model, and accordingly, the second convergence condition is that the AUC indicator no longer increases in multiple consecutive iterative calculations.

[0018] This implementation provides a feasible solution for training multiple teacher models.

[0019] In one implementation, the collection time of any sample in the training sample set is earlier than the collection time of any sample in the verification sample set.

[0020] Since the characteristic information of the environment in the samples with late collection time (i.e., the latest samples) is closer to the characteristic information of the environment during online prediction, the embodiment of the present application uses the samples with late collection time (i.e., the latest samples) to verify the target teacher model, which can ensure that the prediction accuracy of the target teacher model that meets the second convergence condition is high.

[0021] In one implementation, each sample in the sample set contains user feature information, feature information of the object to be recommended, feature information of the environment, and indication information, wherein the indication information indicates whether the user has clicked on the object to be recommended or whether the user has not clicked on the object to be recommended. The user feature information may include the user's age group, the city where the user is located, the user's purchase history, etc.; the object feature information may include the object's category, the object's description, the object's attribute label, etc.; the environment feature information may include time, network status, whether it is a weekend, whether it is a holiday, etc.

[0022] This implementation provides a specific example of sample content.

[0023] In one implementation, the object to be recommended is at least one of an application, music, video, article, and merchandise.

[0024] This implementation provides multiple possible examples of objects to be recommended.

[0025] A second aspect of an embodiment of the present application provides a device for training a model for online prediction of click-through rate, comprising:

[0026] An acquisition unit is used to obtain a predicted value of the click rate of each sample in a sample set by each teacher model in a plurality of teacher models, wherein the plurality of teacher models are obtained by training based on the sample set, and the sample set contains a plurality of samples; a fusion unit is used to fuse the predicted values ​​of the click rate of the same sample in the sample set by the plurality of teacher models to obtain a fused predicted value of the click rate, wherein each sample in the sample set corresponds to a fused predicted value of the click rate; a loss function determination unit is used to determine a loss function based on the fused predicted value of the click rate corresponding to each sample in the sample set; a training unit is used to train a student model based on the loss function and the sample set until the student model meets a first convergence condition, and the student model meeting the first convergence condition is used to predict the click rate online.

[0027] In one implementation, the fusion unit is used to calculate the weight of each teacher model in multiple teacher models based on the predicted values ​​of the click rate of the target sample by multiple teacher models, where the weight corresponds to the target sample, and the target sample is any sample in the sample set; according to the weight of each teacher model in the multiple teacher models, the predicted values ​​of the click rate of the target sample by multiple teacher models are weighted averaged to obtain a fused predicted value of the click rate.

[0028] In one implementation, the loss function includes a distillation loss term, which is determined based on a fusion prediction value of the click rate corresponding to each sample in the sample set; the first convergence condition is that the value of the distillation loss term no longer decreases in multiple consecutive iterative calculations.

[0029] In one implementation, the loss function also includes a non-distillation loss term, which is determined based on a sample label of each sample in the sample set, where the sample label represents whether the user has clicked on it or not.

[0030] In one implementation, the training unit is further used to train multiple teacher models based on the sample set until the multiple teacher models all meet the second convergence condition.

[0031] In one implementation, a training unit is used to train a target teacher model based on a training sample set, where the training sample set contains at least one sample in the sample set, and the target teacher model is any one of multiple teacher models; the target teacher model is tested based on a verification sample set to obtain a test result, and the training stops when the test result indicates that the target teacher model satisfies a second convergence condition, and the verification sample set contains at least one sample in the sample set.

[0032] In one implementation, the collection time of any sample in the training sample set is earlier than the collection time of any sample in the verification sample set.

[0033] In one implementation, each sample in the sample set includes characteristic information of the user, characteristic information of the object to be recommended, characteristic information of the environment, and indication information, wherein the indication information indicates that the user has clicked on the object to be recommended or the user has not clicked on the object to be recommended.

[0034] In one implementation, the object to be recommended is at least one of an application, music, video, article, and merchandise.

[0035] Among them, for the specific implementation, relevant instructions and technical effects of the above units, please refer to the description of the first aspect of the embodiment of this application.

[0036] A third aspect of an embodiment of the present application provides a computer device, comprising: at least one processor and a memory, the memory storing computer execution instructions that can be run on the processor, when the computer execution instructions are executed by the processor, the processor executes the method described in any one of the embodiments of the first aspect above.

[0037] A fourth aspect of an embodiment of the present application provides a chip or a chip system, which includes at least one processor and a communication interface, the communication interface and at least one processor are interconnected by lines, and the at least one processor is used to run computer programs or instructions to execute a method as described in any one of the embodiments of the first aspect above.

[0038] A fifth aspect of an embodiment of the present application provides a computer-readable storage medium storing one or more computer-executable instructions. When the computer-executable instructions are executed by a processor, the processor executes a method as described in any one of the implementations of the first aspect above.

[0039] A sixth aspect of the embodiments of the present application provides a computer program product, which includes computer software instructions, and the computer software instructions can be loaded by a processor to implement the method described in any one of the implementation methods of the first aspect above.

[0040] A seventh aspect of the embodiments of the present application provides a recommendation system, including a first server and a second server;

[0041] The first server is used to execute the method described in any one of the embodiments of the first aspect above;

[0042] The second server is used to deploy the student model trained by the first server.

[0043] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0044] The predicted values ​​of the click rate of the same sample in the sample set by multiple teacher models are fused into a fused predicted value of the click rate, the loss function is determined according to the fused predicted value of the click rate corresponding to each sample in the sample set, and the student model is trained based on the loss function and the sample set, thereby realizing the transfer of knowledge from multiple teacher models to a single student model; the trained student model can achieve a prediction accuracy similar to that of the teacher model, or even exceed that of the teacher model; at the same time, the structure of the student model is simpler than that of the teacher model, and a single student model can complete the prediction of the click rate without the need for multiple models to predict online at the same time, thereby reducing the prediction delay. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1A schematic diagram of an artificial intelligence main framework provided for an embodiment of the present application;

[0046] Figure 2 This is a schematic diagram of an application scenario of the recommendation system in an embodiment of the present application;

[0047] Figure 3 This is a schematic diagram of the architecture of the recommendation system in the embodiment of the present application;

[0048] Figure 4 A schematic diagram of an example of the working process of the recommendation system in the embodiment of the present application;

[0049] Figure 5 This is a schematic diagram of an embodiment of a method for training a model for online prediction of click-through rate in an embodiment of the present application;

[0050] Figure 6 This is a schematic diagram of an embodiment of obtaining a fusion prediction value in an embodiment of the present application;

[0051] Figure 7 Another schematic diagram of the working process of the recommendation system in the embodiment of the present application;

[0052] Figure 8 This is a schematic diagram of the recommended interface in the embodiment of the present application;

[0053] Fig. 9 A schematic diagram of an embodiment of a device for training a model for online prediction of click-through rate in an embodiment of the present application;

[0054] Fig.10 It is a schematic diagram of the structure of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0055] Embodiments of the present application provide a method, device, and recommendation system for training a model for online prediction of click-through rate, so as to reduce the delay of online prediction.

[0056] The technical solutions in the embodiments of the present invention will be described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] See first Figure 1 , Figure 1 A schematic diagram of an artificial intelligence main framework is shown, which describes the overall workflow of the artificial intelligence system and is applicable to general artificial intelligence field requirements.

[0058] The following is an explanation of the above artificial intelligence theme framework from two dimensions: "intelligent information chain" (horizontal axis) and "IT value chain" (vertical axis).

[0059] The "intelligent information chain" reflects a series of processes from data acquisition to processing. For example, it can be a general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, intelligent execution and output. In this process, data undergoes a condensed process of "data-information-knowledge-wisdom".

[0060] The "IT value chain" reflects the value that artificial intelligence brings to the information technology industry, from the underlying infrastructure of human intelligence, information (providing and processing technology implementation) to the system's industrial ecological process.

[0061] (1) Infrastructure:

[0062] The infrastructure provides computing power support for the AI ​​system, enables communication with the outside world, and supports it through the basic platform. It communicates with the outside world through sensors; computing power is provided by smart chips (CPU, NPU, GPU, ASIC, FPGA and other hardware acceleration chips); the basic platform includes distributed computing frameworks and networks and other related platform guarantees and support, which can include cloud storage and computing, interconnected networks, etc. For example, sensors communicate with the outside world to obtain data, and these data are provided to the smart chips in the distributed computing system provided by the basic platform for calculation.

[0063] (2) Data

[0064] The data on the upper layer of the infrastructure is used to represent the data sources in the field of artificial intelligence. The data involves graphics, images, voice, text, and IoT data of traditional devices, including business data of existing systems and perception data such as force, displacement, liquid level, temperature, and humidity.

[0065] (3) Data processing

[0066] Data processing usually includes data training, machine learning, deep learning, search, reasoning, decision-making and other methods.

[0067] Among them, machine learning and deep learning can symbolize and formalize data for intelligent information modeling, extraction, preprocessing, and training.

[0068] Reasoning refers to the process of simulating human intelligent reasoning in computers or intelligent systems, using formalized information to perform machine thinking and solve problems based on reasoning control strategies. Typical functions are search and matching.

[0069] Decision-making refers to the process of making decisions after intelligent information is reasoned, usually providing functions such as classification, sorting, and prediction.

[0070] (4) General ability

[0071] After the data has undergone the data processing mentioned above, some general capabilities can be further formed based on the results of the data processing, such as an algorithm or a general system, for example, translation, text analysis, computer vision processing, speech recognition, image recognition, etc.

[0072] (5) Smart products and industry applications

[0073] Smart products and industry applications refer to the products and applications of artificial intelligence systems in various fields. They are the encapsulation of the overall artificial intelligence solution, which productizes intelligent information decision-making and realizes practical application. Its application areas mainly include: smart manufacturing, smart transportation, smart home, smart medical care, smart security, autonomous driving, safe cities, smart terminals, etc.

[0074] based on Figure 1 In the artificial intelligence system shown in the embodiment of the present application, the data acquired by the infrastructure is the user's historical click behavior data, and then the historical click behavior data is analyzed and learned.

[0075] As a type of artificial intelligence system, the recommendation system is introduced below. The recommendation system refers to the use of machine learning algorithms to analyze and learn based on the user's historical click behavior data to obtain learning results; then predict the user's needs based on the learning results, and recommend corresponding information based on the prediction results. The embodiments of this application do not specifically limit the type of recommended information. The type of recommended information is related to the application scenario of the recommendation system.

[0076] Specifically, Figure 2 Several common scenarios of recommendation systems are shown, including but not limited to scenarios involving e-commerce product recommendations, search engine result recommendations, application market recommendations, music recommendations, article recommendations, video recommendations, etc. The recommended information in various application scenarios is hereinafter referred to as "objects" to facilitate subsequent descriptions, that is, in different recommendation scenarios, the recommended object can be an application APP, or a video, or music, or an article (such as news), or an advertisement, or a certain product (such as the presentation interface of an online shopping platform, which will display different products for presentation according to different users, which can actually be presented through the recommendation results of the recommendation system).

[0077] Combine the following Figure 3 This paper introduces the architecture and working process of the recommendation system in detail. Figure 3 As shown in Figure 3, the recommendation system mainly includes an offline training module and an online prediction module.

[0078] The offline training module collects user behavior logs, and then extracts the interaction information between users and objects (which can be application APPs, videos, articles, music, advertisements or certain products) from the user behavior logs. The interaction information can be browsing, clicking or downloading, etc. Then the interaction information is combined with the user's characteristic information, the object's characteristic information, and the environment's characteristic information to construct a sample. The sample may include the user's characteristic information, the object's characteristic information, the environment's characteristic information, and whether the user has clicked on the object. Among them, the user's characteristic information may include the user's age group, the user's city, the user's purchase history, etc.; the object's characteristic information may include the object's category, the object's description, the object's attribute label, etc.; the environment's characteristic information may include time, network status, whether it is a weekend, whether it is a holiday, etc. Finally, the offline training module is trained based on the sample and the pre-set machine learning algorithm to obtain a model for online prediction of click-through rate. Among them, the click-through rate refers to the probability of a user clicking on an object in a specific environment.

[0079] Then, the offline training module sends the model file corresponding to the model used for online prediction of click-through rate to the online prediction module, and the online prediction module receives the model file. When a user initiates an access request, the online prediction module predicts the click-through rates of multiple objects by running the model corresponding to the model file, and then sorts the multiple objects according to the predicted click-through rates, and finally displays at least one of the top ranked objects on the user interface.

[0080] Among them, in the embodiment of the present application, online prediction refers to real-time prediction based on user requests.

[0081] Based on the above description, we can see that the working process of the recommendation system is as follows: Figure 4 As shown, it may specifically include: performing offline training based on user behavior logs to obtain a model for predicting click rate; then using the model to perform online prediction to obtain an object list containing at least one object, and finally displaying the objects in the object list to the user. The object list includes object a, object b, object c, object d, object e, and object f.

[0082] It should be understood that if the object displayed to the user is more in line with the user's needs, the user's click-through rate and download rate will be higher, which will increase the revenue of the recommendation system. Therefore, in order to make the objects displayed to the user fit the user's needs as much as possible, it is necessary to improve the accuracy of the prediction of the online prediction module. However, if multiple different models are used to predict the click-through rate online, although the accuracy of the prediction can be improved, it will also cause a higher prediction delay, thereby affecting the user experience. To this end, an embodiment of the present application provides a method for training a model for online prediction of click-through rate, in which a given multiple complex teacher models are first trained based on the acquired samples, and then multiple teacher models are used to guide the training of the student model, that is, the knowledge of multiple teacher models is transferred to a single student model by the method of knowledge distillation. The student model trained by this method can achieve a prediction accuracy similar to that of the teacher model, or even exceed the prediction accuracy of the teacher model; at the same time, a single student model can complete the prediction of the click-through rate without the need for multiple models to predict online at the same time, so the prediction delay can be reduced.

[0083] Specifically, see Figure 5 , a schematic diagram of an embodiment of a method for training a model for online prediction of click-through rate provided in an embodiment of the present application. Figure 5 As shown, the embodiment of the present application provides an embodiment of a method for training a model for online prediction of click-through rate, which can be applied to any computer device that can train a model, and the computing device is usually a server; the embodiment includes:

[0084] Step 101, training multiple teacher models based on a sample set until the multiple teacher models all meet a second convergence condition, and the sample set contains multiple samples.

[0085] It should be noted that the content of each sample in the sample set can be adjusted according to actual needs. For example, each sample in the sample set contains the user's characteristic information, the characteristic information of the object to be recommended, the characteristic information of the environment, and the indication information, and the indication information indicates that the user has clicked on the object to be recommended or the user has not clicked on the object to be recommended. Among them, the user's characteristic information, the characteristic information of the object to be recommended, and the characteristic information of the environment can be understood by referring to the above-mentioned relevant instructions.

[0086] The object to be recommended can be any content that can be recommended; illustratively, the object to be recommended is at least one of an application, music, video, article, advertisement, and commodity. The characteristic information of the object to be recommended may include the category of the object to be recommended, the description of the object to be recommended, and the attribute label of the object to be recommended. For example, if the object to be recommended is an application, the category of the object to be recommended may include social, game, sports, financial management, and audio-visual, etc.; the description of the object to be recommended may be that the application is a team fair competition mobile game, and the game modes include classic mode, border breakout mode, and brawl mode; the attribute label of the object to be recommended may be team competition, gunfight, and puzzle, etc.

[0087] The embodiment of the present application does not specifically limit the type of teacher model, that is, multiple teacher models may include models of the same type or models of different types; for example, multiple teacher models may include at least one of a decision tree model and a deep neural network model.

[0088] The embodiment of the present application does not specifically limit the number of teacher models of the teacher model. For example, the number of teacher models can be 2, 3, or more than 3.

[0089] It should be understood that in order to reduce training time, a parallel training method can be used to train multiple teacher models; the training method of each teacher model can be the same or different, and accordingly, the second convergence condition corresponding to each teacher model can be the same or different.

[0090] The following takes one of the multiple teacher models as an example to illustrate the training process of the teacher model.

[0091] Before training, the sample set can be divided into a training sample set and a verification sample set, wherein the training sample set contains at least one sample in the sample set, and the verification sample set contains at least one sample in the sample set; the training sample set is used to train the teacher model, and the verification sample set is used to verify whether the teacher model meets the second convergence condition.

[0092] It should be noted that there are many methods for dividing the sample set into a training sample set and a verification sample set, and the embodiments of the present application do not specifically limit this.

[0093] Exemplarily, the sample set can be divided into a training sample set and a verification sample set according to the sample collection time. As an achievable method, samples with an earlier collection time are divided into the training sample set, and samples with a later collection time (i.e., the latest samples) are divided into the verification sample set. In this case, the collection time of any sample in the training sample set is earlier than the collection time of any sample in the verification sample set.

[0094] For example, if the samples in the sample set are collected from August 1 to August 7, the samples collected from August 1 to August 6 can be divided into the training sample set, and the samples collected on August 7 can be divided into the verification sample set.

[0095] Since the characteristic information of the environment in the samples with late collection time (i.e., the latest samples) is closer to the characteristic information of the environment during online prediction, the embodiment of the present application uses the samples with late collection time (i.e., the latest samples) to verify the target teacher model, which can ensure that the prediction accuracy of the target teacher model that meets the second convergence condition is high.

[0096] Based on the training sample set and the validation sample set, training multiple teacher models based on the sample set until the multiple teacher models all meet the second convergence condition may include:

[0097] The target teacher model is trained based on the training sample set, and the target teacher model is any one of multiple teacher models.

[0098] It should be noted that the objective function of the training can be set first, and then the objective function can be optimized through iterative calculation to achieve the training of the target teacher model. Among them, the objective function can be set according to actual needs.

[0099] For example, in the embodiment of the present application, assuming that the target teacher model is the i-th teacher model among multiple teacher models, the objective function can be in, represents the loss function corresponding to sample j, y j is the sample label of sample j, is the sigmoid function, which is used to Convert to a value between 0 and 1. It is the predicted value of the click rate of the jth sample output by the i-th teacher model, i is the number of the teacher model in multiple teacher models, and j is the number of the sample in the sample set.

[0100] In the present embodiment, y j The value of can be 1 or 0; usually, when y j When the value of is 1, it means the user has clicked. j When the value of is 0, it means that the user has not clicked it.

[0101] Since the objective function is a loss function, in the process of optimizing the objective function, the predicted value of the click rate output by the target teacher model is Continuously approaching the sample label y j .

[0102] The target teacher model is tested based on the validation sample set to obtain a test result, and the training stops when the test result indicates that the target teacher model satisfies the second convergence condition.

[0103] It should be noted that there are many methods for testing the target teacher model, which can be selected according to actual needs, and accordingly, the second convergence condition can also be set according to actual needs. For example, the AUC index can be used as a test result for evaluating the performance of the target teacher model, and accordingly, the second convergence condition is that the AUC index no longer increases in multiple consecutive iterative calculations, for example, the AUC index no longer increases in two or three consecutive iterative calculations.

[0104] In the embodiment of the present application, step 101 is optional.

[0105] Step 102, obtaining a predicted value of the click rate of each sample in the sample set by each teacher model in the multiple teacher models, wherein the multiple teacher models are trained based on the sample set, and the sample set contains multiple samples.

[0106] After multiple teacher models satisfy their respective second convergence conditions, the samples in the sample set can be re-input into the teacher model, and accordingly, the teacher model will output the predicted value of the click rate.

[0107] Taking the target teacher model as an example, each sample in the sample set is input into the target teacher model respectively to obtain the predicted value of the click rate of each sample in the sample set by the target teacher model.

[0108] Step 103, the predicted values ​​of the click rate of the same sample in the sample set by multiple teacher models are merged to obtain a fused predicted value of the click rate, and each sample in the sample set corresponds to a fused predicted value of the click rate.

[0109] It is understandable that when the types of multiple teacher models are different, the predicted values ​​of the click rate of the same sample in the sample set by the multiple teacher models may be different; even if the types of multiple teacher models are the same, the predicted values ​​of the click rate of the same sample in the sample set by the multiple teacher models may also be different. Therefore, in order to use the predicted value of the click rate of each teacher model in the multiple teacher models to guide the training of the student model, the embodiment of the present application fuses the predicted values ​​of the click rate of the same sample in the sample set by the multiple teacher models to obtain a fused predicted value of the click rate.

[0110] For example, if there are three teacher models, for the sample numbered 2 in the sample set, the predicted values ​​of the click rate of multiple teacher models can be expressed as Then and Fusion, to obtain the fusion prediction value corresponding to the sample numbered 2.

[0111] There are many methods for fusing the predicted values ​​of the click rate of the same sample in the sample set by multiple teacher models, and the embodiments of the present application do not make specific limitations on this.

[0112] For example, the average of the predicted values ​​of the click rate of the same sample in the sample set by multiple teacher models can be used as the fusion predicted value of the click rate.

[0113] It is understandable that the same teacher model may have different prediction accuracy rates for different samples. For example, the sample label of sample number 1 indicates that the user has clicked on it, and the sample label of sample number 2 also indicates that the user has clicked on it; the predicted value of the click rate of sample number 1 by the teacher model number 1 is 0.98, and the predicted value of the click rate of sample number 2 is 0.93; while the predicted value of the click rate of sample number 1 by the teacher model number 2 is 0.93, and the predicted value of the click rate of sample number 2 is 0.98. It can be seen from this that the prediction accuracy of the teacher model number 1 is higher for the sample number 1, and the prediction accuracy of the teacher model number 2 is higher for the sample number 2.

[0114] Therefore, the embodiment of the present application determines the weights of multiple teacher models relative to the sample based on the predicted values ​​of the click rate of the same sample by multiple teacher models, and then fuses the predicted values ​​of the click rate of the sample by multiple teacher models based on the weights of the multiple teacher models relative to the sample.

[0115] Specifically, Figure 6 As shown, the predicted values ​​of the click rate of the same sample in the sample set by multiple teacher models are fused to obtain the fused predicted value of the click rate, including:

[0116] Step 201, based on the predicted values ​​of the click rate of the target sample by the multiple teacher models, the weight of each teacher model in the multiple teacher models is calculated, and the weight corresponds to the target sample, and the target sample is any sample in the sample set.

[0117] It should be noted that there are many methods for calculating the weight of each teacher model in multiple teacher models, and this embodiment of the application does not specifically limit this. For example, the formula Determine the weight of each teacher model in multiple teacher models for the target sample.

[0118] Among them, i is the number of the teacher model in multiple teacher models, j is the number of the sample in the sample set; α ij represents the weight of the i-th teacher model relative to the j-th sample, is the predicted value of the click rate of the jth sample output by the i-th teacher model, w ij and b ijare all parameters to be learned, and n represents the number of teacher models. Before training begins, w is set ij and b ij The initial value of .

[0119] Step 202, according to the weight of each teacher model in the multiple teacher models, weighted average processing is performed on the predicted values ​​of the click rate of the target sample by the multiple teacher models to obtain a fusion predicted value of the click rate.

[0120] It can be understood that the fusion prediction value of click rate can be expressed as

[0121] For example, if there are three teacher models, for the sample numbered 2 in the sample set, the predicted values ​​of the click rate of multiple teacher models can be expressed as Then the fusion prediction value corresponding to the sample numbered 2 can be expressed as

[0122] In an embodiment of the present application, the weight of each teacher model in the multiple teacher models is calculated based on the predicted values ​​of the click rate of the target sample by multiple teacher models, and then the predicted values ​​of the click rate of the target sample by the multiple teacher models are fused based on the weight of each teacher model in the multiple teacher models, thereby realizing adaptive fusion of the multiple teacher models, so that the fused predicted value of the click rate after fusion is closer to the sample label, thereby improving the accuracy of the fused predicted value of the click rate; and, by training the student model with the fused predicted value of the click rate calculated by steps 201 and 202, the prediction accuracy of the student model can be improved, so that the prediction accuracy of the student model is close to the prediction accuracy of the teacher model, or even exceeds the prediction accuracy of the teacher model.

[0123] Step 104: determine a loss function based on the fused prediction value of the click rate corresponding to each sample in the sample set.

[0124] It should be noted that there may be multiple methods for determining the loss function, which is not limited in the embodiments of the present application; accordingly, there may also be multiple loss functions.

[0125] For example, the loss function of the student model can be expressed as Among them, i is the number of the teacher model in multiple teacher models, j is the number of the sample in the sample set; α ij represents the weight of the i-th teacher model relative to the j-th sample, is the predicted value of the click rate of the jth sample output by the i-th teacher model, represents a loss function, m represents the number of samples in the sample set, n represents the number of teacher models, represents the predicted value of the click rate output by the student model, L SRepresents the loss function of the student model.

[0126] In the embodiment of the present application, knowledge distillation refers to the transfer of knowledge from multiple teacher models to a student model with a simple structure, and the knowledge of the teacher model can be represented by the predicted value of the click rate of the sample by the teacher model. It is constructed based on the predicted value of the click rate of each sample in the sample set by each teacher model in multiple teacher models, so It can also be called the distillation loss term.

[0127] Based on the above description, it can be seen that the distillation loss term is determined based on the fused prediction value of the click rate corresponding to each sample in the sample set.

[0128] For example, the loss function can also be expressed as Among them, y j represents the sample label of the j-th sample, γ is the weight value, and the value of γ can be set according to actual needs, for example, it can be 1; the description of other parameters can refer to the previous text.

[0129] because It can be called the distillation loss term, so It can be called the non-distillation loss term.

[0130] Based on the above description, we can know that the non-distillation loss term is determined based on the sample label of each sample in the sample set, and the sample label represents whether the user has clicked or not. For example, the sample label y j When it is 1, it indicates that the user has clicked it; sample label y j When it is 0, it indicates that the user has not clicked it.

[0131] Step 105: Train the student model based on the loss function and the sample set until the student model meets the first convergence condition. The student model that meets the first convergence condition is used to predict click-through rate online.

[0132] The process of training the student model can be understood as using the samples in the sample set to perform iterative calculations to optimize the loss function until the student model meets the first convergence condition.

[0133] When the loss function is , as the loss function is continuously optimized, the predicted value of the click rate of a certain sample output by the student model continues to approach the fusion predicted value of the click rate corresponding to the sample; when the loss function is , the predicted value of the click rate of a certain sample output by the student model keeps approaching the fusion predicted value of the click rate corresponding to the sample, and keeps approaching the sample label of the sample.

[0134] Among them, in the process of iterative calculation, the stochastic gradient descent method can be used to update the loss function LS .

[0135] It should be noted that the first convergence condition can be set according to actual needs; illustratively, the loss function includes a distillation loss term, and the distillation loss term is determined based on the fused prediction value of the click rate corresponding to each sample in the sample set; the first convergence condition is that the value of the distillation loss term no longer decreases in multiple consecutive iterative calculations.

[0136] Among them, in the first convergence condition, the number of iterative calculations can be set according to actual needs. For example, the value of the distillation loss term no longer decreases in two consecutive iterative calculations, or no longer decreases in three consecutive iterative calculations.

[0137] The fact that the value of the distillation loss term no longer decreases in consecutive multiple iterative calculations can be understood as the value of the distillation loss term in consecutive multiple iterative calculations is greater than or equal to the value of the distillation loss term in the previous iterative calculation.

[0138] Since the first convergence condition is that the value of the distillation loss term no longer decreases in multiple consecutive iterative calculations, there is no need to select an additional portion of samples from the sample set as verification samples during the training of the student model, so the samples in the sample set can be fully utilized.

[0139] In an embodiment of the present application, the predicted values ​​of the click-through rate of the same sample in the sample set by multiple teacher models are fused into a fused predicted value of the click-through rate, the loss function is determined according to the fused predicted value of the click-through rate corresponding to each sample in the sample set, and the student model is trained based on the loss function and the sample set, thereby realizing the transfer of knowledge from multiple teacher models to a single student model; the trained student model can achieve a prediction accuracy similar to that of the teacher model, or even exceed the prediction accuracy of the teacher model; at the same time, a single student model can complete the prediction of the click-through rate without the need for multiple models to predict online at the same time, thereby reducing the prediction delay.

[0140] In addition, in the embodiment of the present application, multiple teacher models are trained first, and then the student model is trained by the multiple teacher models. Only the student model is used to predict the click rate online, and the student model and the teacher model are independent of each other. Therefore, without increasing the complexity of the student model, the number of teacher models can be flexibly adjusted as needed. Therefore, the recommendation system trained by the method provided in the embodiment of the present application has strong scalability and versatility.

[0141] At this time, if Figure 7 As shown, Figure 4The offline training process in may include: training multiple teacher models in parallel to obtain multiple converged teacher models, then adaptively fusing the multiple converged teacher models to obtain a fused teacher model, and then using the fused teacher model to guide the student model to train to obtain a student model. The fused teacher model refers to the fused prediction value corresponding to each sample.

[0142] After the student model training is completed, the student model can be used for online prediction, and then the objects that the user is interested in are recommended based on the results of the online prediction of the student model. The following is an example to illustrate the recommendation process.

[0143] Specifically, Figure 8 As shown in the figure, when a user enters the keyword "novel" in the application market of the terminal device, the student model deployed in the server will predict the probability of the current user clicking on all reading applications (i.e., click-through rate) based on the characteristic information of the current user, the characteristic information of the current environment, and the characteristic information of the reading applications to which the novel belongs, and then sort all reading applications based on the probability of being clicked, and recommend the top three reading applications to the application market interface. Figure 8 It can be seen that three applications, "Novel aaa", "Novel bbb" and "Novel ccc", are finally recommended.

[0144] It should be noted that in the above examples, terminal devices include but are not limited to mobile phones, tablet computers, laptops, desktop computers, smart screens, car terminals, etc. Figure 8 Only the mobile phone terminal device is shown.

[0145] See also Fig. 9 , a schematic diagram of an embodiment of a device for training a model for online prediction of click-through rate in an embodiment of the present application. Fig. 9 As shown, the embodiment of the present application provides an embodiment of a device for training a model for online prediction of click-through rate, including:

[0146] An acquisition unit 301 is used to acquire a predicted value of the click rate of each sample in a sample set by each teacher model in a plurality of teacher models, wherein the plurality of teacher models are obtained by training based on the sample set, and the sample set includes a plurality of samples;

[0147] A fusion unit 302 is used to fuse the prediction values ​​of the click rate of the same sample in the sample set by multiple teacher models to obtain a fusion prediction value of the click rate, and each sample in the sample set corresponds to a fusion prediction value of the click rate;

[0148] A loss function determination unit 303, configured to determine a loss function according to a fusion prediction value of a click rate corresponding to each sample in the sample set;

[0149] The training unit 304 is used to train the student model based on the loss function and the sample set until the student model meets the first convergence condition, and the student model that meets the first convergence condition is used to predict the click rate online.

[0150] In one implementation, the fusion unit 302 is used to calculate the weight of each teacher model in the multiple teacher models based on the predicted values ​​of the click rate of the target sample by the multiple teacher models, where the weight corresponds to the target sample, and the target sample is any sample in the sample set; according to the weight of each teacher model in the multiple teacher models, the predicted values ​​of the click rate of the target sample by the multiple teacher models are weighted averaged to obtain a fused predicted value of the click rate.

[0151] In one implementation, the loss function includes a distillation loss term, which is determined based on a fusion prediction value of the click rate corresponding to each sample in the sample set; the first convergence condition is that the value of the distillation loss term no longer decreases in multiple consecutive iterative calculations.

[0152] In one implementation, the loss function also includes a non-distillation loss term, which is determined based on a sample label of each sample in the sample set, where the sample label represents whether the user has clicked on it or not.

[0153] In one implementation, the training unit 304 is further configured to train multiple teacher models based on the sample set until the multiple teacher models all meet the second convergence condition.

[0154] In one implementation, the training unit 304 is used to train the target teacher model based on the training sample set, the training sample set contains at least one sample in the sample set, and the target teacher model is any one of multiple teacher models; the target teacher model is tested based on the verification sample set to obtain a test result, and the training stops when the test result indicates that the target teacher model meets the second convergence condition, and the verification sample set contains at least one sample in the sample set.

[0155] In one implementation, the collection time of any sample in the training sample set is earlier than the collection time of any sample in the verification sample set.

[0156] In one implementation, each sample in the sample set includes characteristic information of the user, characteristic information of the object to be recommended, characteristic information of the environment, and indication information, wherein the indication information indicates that the user has clicked on the object to be recommended or the user has not clicked on the object to be recommended.

[0157] In one implementation, the object to be recommended is at least one of an application, music, video, article, and merchandise.

[0158] For the specific implementation, related description and technical effects of the above units, please refer to the embodiments of this application. Figures 2 to 8 Description of the illustrated embodiment.

[0159] See also Fig.10 , Fig.10 It is a schematic diagram of the structure of a computer device in an embodiment of the present application.

[0160] In the embodiment of the present application, a computer device may include one or more processors 401 , a memory 402 , and a communication interface 403 .

[0161] The memory 402 may be a temporary storage or a permanent storage. Furthermore, the processor 401 may be configured to communicate with the memory 402 and execute a series of instruction operations in the memory 402 on the computer device.

[0162] In this embodiment, the processor 401 may execute the aforementioned Figure 5 and Figure 6 The steps in the illustrated embodiment will not be described in detail here.

[0163] Exemplarily, the processor 401 may perform the following steps:

[0164] Obtain a predicted value of the click rate of each sample in a sample set by each teacher model in multiple teacher models, where the multiple teacher models are trained based on the sample set, and the sample set contains multiple samples;

[0165] The predicted values ​​of the click rate of the same sample in the sample set by multiple teacher models are merged to obtain the fused prediction value of the click rate. Each sample in the sample set corresponds to a fused prediction value of the click rate.

[0166] Determine the loss function based on the fused prediction value of the click rate corresponding to each sample in the sample set;

[0167] The student model is trained based on the loss function and the sample set until the student model meets the first convergence condition, and the student model that meets the first convergence condition is used to predict the click rate online.

[0168] At this time, the specific functional module division in the processor 401 can be the same as the above Fig. 9 The division of the functional modules such as the acquisition unit, fusion unit, loss function determination unit and training unit described in is similar and will not be repeated here.

[0169] The embodiment of the present application also provides a chip or a chip system, which includes at least one processor and a communication interface, the communication interface and the at least one processor are interconnected through a line, and the at least one processor is used to run a computer program or instruction to execute the aforementioned Figure 5 and Figure 6 The steps in the illustrated embodiment will not be described in detail here.

[0170] The communication interface in the chip may be an input / output interface, a pin or a circuit, etc.

[0171] The embodiment of the present application also provides a first implementation of a chip or a chip system, wherein the chip or chip system described above in the present application further includes at least one memory, and the at least one memory stores instructions. The memory may be a storage unit inside the chip, such as a register, a cache, etc., or may be a storage unit of the chip (e.g., a read-only memory, a random access memory, etc.).

[0172] The embodiment of the present application also provides a computer storage medium, which is used to store computer software instructions used by the above-mentioned computer device, including a program designed for executing the computer device.

[0173] The computer device can be as described above Fig. 9 The described apparatus is a model for online prediction of click-through rate.

[0174] The present application also provides a computer program product, which includes computer software instructions, which can be loaded by a processor to implement the aforementioned Figure 5 and Figure 6 The process in the method shown.

[0175] The embodiment of the present application also provides a recommendation system, including a first server and a second server;

[0176] The first server is used to execute the above Figure 5 and Figure 6 The method shown;

[0177] The second server is used to deploy the student model trained by the first server.

[0178] Exemplarily, the second server is further used to predict the click rate of the object to be recommended through the student model.

[0179] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0180] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0181] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0182] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0183] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

Claims

1. A method for training a model for online prediction of click-through rate, It is characterized in that include: Obtain a predicted value of the click rate of each sample in a sample set by each teacher model in a plurality of teacher models, wherein the plurality of teacher models are trained based on the sample set, and the sample set includes a plurality of samples; Fusion the predicted values ​​of the click rate of the same sample in the sample set by the multiple teacher models to obtain a fused predicted value of the click rate, wherein each sample in the sample set corresponds to a fused predicted value of the click rate; A loss function is determined based on a fused prediction value of a click rate corresponding to each sample in the sample set, wherein the loss function includes a distillation loss term and a non-distillation loss term, wherein the distillation loss term is determined based on a fused prediction value of a click rate corresponding to each sample in the sample set; and the non-distillation loss term is determined based on a sample label of each sample in the sample set, wherein the sample label represents whether a user has clicked or not clicked; The student model is trained based on the loss function and the sample set until the student model meets a first convergence condition. The student model that meets the first convergence condition is used to predict click-through rate online. The first convergence condition is that the value of the distillation loss term no longer decreases in multiple consecutive iterative calculations.

2. The method according to claim 1, It is characterized in that The step of fusing the predicted values ​​of the click rate of the same sample in the sample set by the multiple teacher models to obtain a fused predicted value of the click rate includes: Calculate the weight of each teacher model in the multiple teacher models according to the predicted values ​​of the click rates of the target samples by the multiple teacher models, where the weight corresponds to the target sample, and the target sample is any sample in the sample set; According to the weight of each teacher model in the multiple teacher models, weighted average processing is performed on the predicted values ​​of the click rate of the target sample by the multiple teacher models to obtain a fused predicted value of the click rate.

3. The method according to any one of claims 1 to 2, It is characterized in that Before obtaining the predicted value of the click rate of each sample in the sample set by each teacher model in the multiple teacher models, the method further includes: The multiple teacher models are trained based on the sample set until the multiple teacher models all meet a second convergence condition.

4. The method according to claim 3, It is characterized in that The training of the multiple teacher models based on the sample set until the multiple teacher models all meet the second convergence condition includes: Training a target teacher model based on a training sample set, wherein the training sample set includes at least one sample in the sample set, and the target teacher model is any one of the multiple teacher models; The target teacher model is tested based on a validation sample set to obtain a test result, and training stops when the test result indicates that the target teacher model satisfies a second convergence condition, wherein the validation sample set includes at least one sample in the sample set.

5. The method according to claim 4, It is characterized in that The collection time of any sample in the training sample set is earlier than the collection time of any sample in the verification sample set.

6. The method according to any one of claims 1 to 2, It is characterized in that Each sample in the sample set contains characteristic information of the user, characteristic information of the object to be recommended, characteristic information of the environment and indication information, wherein the indication information indicates that the user has clicked on the object to be recommended or the user has not clicked on the object to be recommended.

7. The method according to claim 6, It is characterized in that The object to be recommended is at least one of an application, music, video, article and commodity.

8. A device for training a model for online prediction of click-through rate, It is characterized in that include: an acquisition unit, configured to acquire a predicted value of a click rate of each sample in a sample set by each teacher model in a plurality of teacher models, wherein the plurality of teacher models are obtained by training based on the sample set, and the sample set includes a plurality of samples; A fusion unit, used to fuse the prediction values ​​of the click rate of the same sample in the sample set by the multiple teacher models to obtain a fusion prediction value of the click rate, each sample in the sample set corresponds to a fusion prediction value of the click rate; A loss function determination unit, configured to determine a loss function based on a fused prediction value of a click rate corresponding to each sample in the sample set, wherein the loss function includes a distillation loss term and a non-distillation loss term, wherein the distillation loss term is determined based on a fused prediction value of a click rate corresponding to each sample in the sample set; and the non-distillation loss term is determined based on a sample label of each sample in the sample set, wherein the sample label represents whether the user has clicked or has not clicked. A training unit is used to train the student model based on the loss function and the sample set until the student model meets a first convergence condition. The student model that meets the first convergence condition is used to predict the click rate online. The first convergence condition is that the value of the distillation loss term no longer decreases in multiple consecutive iterative calculations.

9. A computer device, It is characterized in that include: At least one processor and a memory, wherein the memory stores computer-executable instructions that can be run on the processor, and when the computer-executable instructions are executed by the processor, the processor executes the method as described in any one of claims 1 to 7.

10. A computer storage medium, It is characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer is caused to perform the method according to any one of claims 1 to 7.

11. A recommendation system, It is characterized in that comprising a first server and a second server; The first server is used to execute the method according to any one of claims 1 to 7; The second server is used to deploy the student model trained by the first server.