A data processing method and related device

By iterative pruning and rebirth processing of the initial click-through rate estimate model, the model performance bias caused by the pruning strategy is solved, and a faster and more accurate click-through rate estimate is achieved.

CN114611690BActive Publication Date: 2024-11-08TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210223399.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-09
Publication Date
2024-11-08
Estimated Expiration
2042-03-09

AI Technical Summary

Technical Problem

In the prior art, pruning strategies may lead to the inaccurate deletion of important weight parameters during model training, affecting the performance of the model and inaccurate prediction click-through rate.

Method used

The initial click-to-rebirth model is used to prune and rebirth the initial click-to-rate estimate model. The model parameters are sparse through pruning, and some pruned parameters are restored during the iteration process to retain important connections and form a lightweight Target click-through rate estimate model.

Benefits of technology

It improves the speed and accuracy of model training, reduces the memory occupied by the model, and achieves better estimates when using fewer parameters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114611690B_ABST
    Figure CN114611690B_ABST
Patent Text Reader

Abstract

The present application relates to the field of computer technology, and provides a data processing method and related devices for improving the model training speed. The method includes: based on a sample data set, an initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model, wherein, in one iteration, the current click rate prediction model is pruned, and the click rate prediction model obtained after the pruning process is trained, and the current click rate prediction model is regenerated, and the click rate prediction model obtained after the regeneration process is trained. In this way, a lightweight target click rate prediction model can be obtained, the memory occupied by the model can be reduced, and the model training speed can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Background Art

[0002] With the continuous development of network and computer technology, more and more target objects can obtain multimedia information through the Internet, including pictures, texts, public accounts, videos, news, etc. In order to more accurately display the corresponding multimedia information to each target object, machine learning models are usually used to estimate the click rate of multimedia information.

[0003] In related technologies, in order to solve the problems of excessive training consumption and low training efficiency caused by complex model structures, a lightweight click-through rate prediction model is usually obtained through a pruning strategy. The pruning strategy refers to the process of continuously pruning certain parameters contained in the initial click-through rate prediction model according to a set parameter pruning ratio during the model training process.

[0004] However, the use of the above pruning strategy may cause some important weight parameters to be erroneously reduced, affecting model performance, causing the trained model to have a biased perception of the target object's preferences, and making the predicted click rate output by the model inaccurate. Summary of the invention

[0005] The embodiments of the present application provide a data processing method and related devices to increase the model training speed while improving the model accuracy.

[0006] In a first aspect, an embodiment of the present application provides a data processing method, including:

[0007] Obtain a sample data set and an initial click rate prediction model;

[0008] Based on the sample data set, the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model; wherein, in one iteration, the following operations are performed:

[0009] Determining at least one model parameter to be deleted from the currently retained model parameters, and pruning the current click-through rate prediction model based on the at least one model parameter to be deleted, and training the click-through rate prediction model obtained after the pruning based on the sample data set;

[0010] At least one model parameter to be restored is determined from the currently pruned deleted model parameters, and based on the at least one model parameter to be restored, the current click-through rate prediction model is regenerated, and based on the sample data set, the click-through rate prediction model obtained after the regeneration is trained.

[0011] In a second aspect, an embodiment of the present application provides a data processing device, including:

[0012] An acquisition unit, used to acquire a sample data set and an initial click rate prediction model;

[0013] The training unit is used to perform pruning-regeneration processing on the initial click rate prediction model in an iterative manner based on the sample data set to obtain a target click rate prediction model; wherein, in one iteration process, the following operations are performed:

[0014] Determining at least one model parameter to be deleted from the currently retained model parameters, and pruning the current click-through rate prediction model based on the at least one model parameter to be deleted, and training the click-through rate prediction model obtained after the pruning based on the sample data set;

[0015] At least one model parameter to be restored is determined from the currently pruned deleted model parameters, and based on the at least one model parameter to be restored, the current click-through rate prediction model is regenerated, and based on the sample data set, the click-through rate prediction model obtained after the regeneration is trained.

[0016] As a possible implementation method, the initial click rate prediction model at least includes a deep neural network component;

[0017] Then, based on the sample data set, the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model, the training unit is used to:

[0018] Based on the sample data set, the deep neural network component included in the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model.

[0019] As a possible implementation method, the initial click rate prediction model also includes a non-deep neural network component;

[0020] Then, based on the sample data set, the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model. The training unit is further used to:

[0021] Building a new estimation model based on the non-deep neural network components included in the target click rate estimation model;

[0022] Based on the sample data set, the new estimation model is trained, and the trained new estimation model is used as a new target click rate estimation model.

[0023] As a possible implementation manner, when determining at least one model parameter to be deleted from the currently retained model parameters, the training unit is used to:

[0024] Based on the parameter values ​​corresponding to the currently reserved model parameters, select a reserved model parameter whose corresponding parameter value is not greater than a preset first value threshold from the reserved model parameters, and use the selected at least one reserved model parameter as the at least one model parameter to be deleted; or

[0025] The retained model parameters are sorted based on parameter values ​​corresponding to the currently retained model parameters, and the at least one model parameter to be deleted is determined from the retained model parameters based on a corresponding first sorting result and a preset first selection ratio.

[0026] As a possible implementation manner, when pruning the current click rate prediction model based on the at least one model parameter to be deleted, the training unit is used to:

[0027] In the current click rate prediction model, the parameter value of the at least one model parameter to be deleted is set to a first set value.

[0028] As a possible implementation manner, when the click rate prediction model obtained after pruning is trained based on the sample data set, the training unit is used to:

[0029] Inputting the sample data set into the click rate prediction model obtained after pruning processing to obtain the click rate prediction value corresponding to each sample data contained in the sample data set;

[0030] Based on the obtained estimated click-through rates and the actual click-through rates corresponding to the sample data, the parameter adjustment values ​​corresponding to the deletion model parameters are determined.

[0031] As a possible implementation manner, when determining at least one model parameter to be restored from the currently pruned deleted model parameters, the training unit is used to:

[0032] Based on the parameter adjustment values ​​corresponding to the deletion model parameters, a deletion model parameter whose corresponding parameter adjustment value is not less than a preset second value threshold is selected from the deletion model parameters, and the selected at least one deletion model parameter is used as the at least one model parameter to be restored; or

[0033] The deleted model parameters are sorted based on the parameter adjustment values ​​corresponding to the deleted model parameters, and the at least one model parameter to be restored is determined from the deleted model parameters based on the corresponding second sorting result and a preset second selection ratio.

[0034] As a possible implementation, when regenerating the current click rate prediction model based on the model parameters to be restored, the training unit is used to:

[0035] In the current click rate prediction model, the parameter value of the at least one model parameter to be restored is adjusted to the corresponding parameter adjustment value.

[0036] As a possible implementation manner, after obtaining the sample data set and the initial click rate prediction model, and before training the click rate prediction model obtained after pruning based on the sample data set, the training unit is further used to:

[0037] Acquire each newly added sample data, and add each newly added sample data to the sample data set;

[0038] When the click rate prediction model obtained after pruning is trained based on the sample data set, the training unit is used to:

[0039] Based on the newly added sample data included in the sample data set, the click rate prediction model obtained after pruning is trained.

[0040] As a possible implementation manner, the training unit obtains the initial click rate prediction model in the following manner:

[0041] Get the click rate estimation model to be trained;

[0042] Based on the sample data set, the click rate prediction model to be trained is iteratively trained until a model convergence condition is met, thereby obtaining the initial click rate prediction model.

[0043] As a possible implementation method, satisfying the model convergence condition includes:

[0044] For N consecutive times, the first difference between the model evaluation indicators of two adjacent iterations is not greater than a preset first difference threshold, where the value of N is a positive integer; or

[0045] A second difference between the model evaluation index of the click-through rate prediction model to be trained and the model evaluation index of the specified model is not greater than a preset second difference threshold.

[0046] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above-mentioned data processing method.

[0047] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which includes a computer program. When the computer program is run on an electronic device, the computer program is used to enable the electronic device to execute the steps of the above-mentioned data processing method.

[0048] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program, and the computer program is stored in a computer-readable storage medium. The processor of an electronic device reads and executes the computer program from the computer-readable storage medium, so that the electronic device performs the steps of the above-mentioned data processing method.

[0049] In an embodiment of the present application, a sample data set and an initial click-through rate prediction model are obtained, and based on the sample data set, the initial click-through rate prediction model is pruned and regenerated in an iterative manner to obtain a target click-through rate prediction model, wherein, in one iteration, the current click-through rate prediction model is pruned, and then the click-through rate prediction model obtained after the pruning is trained, and the current click-through rate prediction model is regenerated, and then the click-through rate prediction model obtained after the regeneration is trained.

[0050] In this way, during the training process of the click-through rate prediction model, the model is pruned to make the originally dense model parameters sparse, thereby improving the model processing efficiency. At the same time, during the iteration process, some of the pruned model parameters are restored, so that the more important connections in the click-through rate prediction model are retained. In addition, by continuously iteratively pruning and restoring the model, a lightweight target click-through rate prediction model can be obtained, which reduces the memory occupied by the model and can achieve better model prediction effects while using fewer model parameters.

[0051] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0053] Figure 1 A schematic diagram of an application scenario provided in an embodiment of the present application;

[0054] Figure 2 A flowchart of a data processing method provided in an embodiment of the present application;

[0055] Figure 3 A schematic diagram of an initial click rate prediction model provided in an embodiment of the present application;

[0056] Figure 4A A schematic diagram of another initial click rate prediction model provided in an embodiment of the present application;

[0057] Figure 4B A logical diagram of a pruning-regeneration process provided in an embodiment of the present application;

[0058] Figure 5A A schematic diagram of a target click rate prediction model provided in an embodiment of the present application;

[0059] Figure 5B A schematic diagram of another target click rate prediction model provided in an embodiment of the present application;

[0060] Figure 6 A schematic diagram of a process for determining model parameters to be deleted provided in an embodiment of the present application;

[0061] Figure 7 A logical diagram of another pruning-regeneration process provided in an embodiment of the present application;

[0062] Figure 8 A logical schematic diagram of a model training process provided in an embodiment of the present application;

[0063] Fig. 9 A schematic diagram of another application scenario provided in an embodiment of the present application;

[0064] Fig.10 A schematic diagram of the structure of a data processing device provided in an embodiment of the present application;

[0065] Fig.11 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the technical solution of the present application, rather than all of the embodiments. Based on the embodiments recorded in the application documents, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the technical solution of the present application.

[0067] The following is an introduction to some concepts involved in the embodiments of the present application.

[0068] Neural network: A widely interconnected network of simple adaptive units, organized to simulate the interactive responses of biological nervous systems to real-world objects. Neural networks can simulate the nervous system's response to inputs and are a type of neural network that simulates the human brain in order to achieve artificial intelligence-like machine learning technology.

[0069] Pruning: Neural networks have many parameters, but some of them do not contribute much to the final click-through rate estimate and appear redundant. Pruning, as the name suggests, is to cut off these redundant parameters.

[0070] 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 comprehensive technology in 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.

[0071] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Basic artificial intelligence technologies generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, mechatronics and other technologies. Artificial intelligence software technologies mainly include computer vision technology, speech processing technology, natural language processing technology, as well as machine learning / deep learning, autonomous driving, smart transportation and other major directions.

[0072] Among them, machine learning (ML) is a multi-disciplinary interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0073] The solution provided in the embodiment of the present application involves machine learning technology of artificial intelligence. In the embodiment of the present application, it mainly involves the training process of the click-through rate prediction model. Specifically, based on the sample data set, the initial click-through rate prediction model is pruned and reborn in an iterative manner to obtain the target click-through rate prediction model. The detailed pruning and rebirth process is described below and will not be repeated here.

[0074] With the continuous development of network and computer technology, more and more target objects can obtain multimedia information through the Internet, including pictures, texts, public accounts, videos, news, etc. In order to more accurately display the corresponding multimedia information to each target object, machine learning models are usually used to estimate the click rate of multimedia information.

[0075] In related technologies, in order to solve the problems of excessive training consumption and low training efficiency caused by complex model structures, a lightweight click-through rate prediction model is usually obtained through a pruning strategy. The pruning strategy refers to the process of continuously pruning certain parameters contained in the initial click-through rate prediction model according to a set parameter pruning ratio during the model training process.

[0076] However, the use of the above pruning strategy may cause some important weight parameters to be erroneously reduced, affecting model performance, causing the trained model to have a biased perception of the target object's preferences, and making the predicted click rate output by the model inaccurate.

[0077] In an embodiment of the present application, a sample data set and an initial click-through rate prediction model are obtained, and based on the sample data set, the initial click-through rate prediction model is pruned and regenerated in an iterative manner to obtain a target click-through rate prediction model, wherein, in one iteration, the current click-through rate prediction model is pruned, and then the click-through rate prediction model obtained after the pruning is trained, and the current click-through rate prediction model is regenerated, and then the click-through rate prediction model obtained after the regeneration is trained.

[0078] In this way, during the training process of the click-through rate prediction model, the model is pruned to make the originally dense model parameters sparse, thereby improving the model processing efficiency. At the same time, during the iteration process, some of the pruned model parameters are restored, so that the more important connections in the click-through rate prediction model are retained. In addition, by continuously iteratively pruning and restoring the model, a lightweight target click-through rate prediction model can be obtained, which reduces the memory occupied by the model and can achieve better model prediction effects while using fewer model parameters.

[0079] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application. In addition, the embodiments of the present application and the features in the embodiments may be combined with each other if there is no conflict.

[0080] See also Figure 1 As shown, it is a schematic diagram of an application scenario provided in an embodiment of the present application. The application scenario includes at least a terminal device 110 and a server 120. The number of terminal devices 110 can be one or more, and the number of servers 120 can also be one or more. The present application does not specifically limit the number of terminal devices 110 and servers 120. The terminal device 110 in the embodiment of the present application can be installed with a client related to data processing, and the server 120 can be a server related to data processing. In addition, the client in the present application can be software, or a web page, a small program, etc., and the server is a background server corresponding to the software or web page, a small program, etc., or a server specifically used for data processing, etc., which is not specifically limited in the present application.

[0081] In the embodiment of the present application, the terminal device 110 can be a mobile phone, an intelligent voice interaction device, a smart home appliance, a vehicle terminal, an aircraft, etc., but is not limited thereto.

[0082] The server 120 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The terminal device 110 and the server 120 may be directly or indirectly connected via wired or wireless communication, which is not limited in this application.

[0083] It should be noted that the data processing method in the embodiment of the present application can be executed by a server or a terminal device alone, or can be executed by a server and a terminal device together.

[0084] For example, a terminal device obtains a sample data set and an initial click rate prediction model, and then, based on the sample data set, an iterative method is used to prune and regenerate the initial click rate prediction model to obtain a target click rate prediction model; wherein, in one iteration, the following operations are performed: at least one model parameter to be deleted is determined from each currently retained model parameter, and the current click rate prediction model is pruned based on the at least one model parameter to be deleted, and the click rate prediction model obtained after the pruning process is trained based on the sample data set; at least one model parameter to be restored is determined from each currently pruned deleted model parameter, and the current click rate prediction model is regenerated based on the at least one model parameter to be restored, and the click rate prediction model obtained after the regeneration process is trained based on the sample data set. Alternatively, the above data processing process is performed by a server. Alternatively, the server obtains a target click-through rate prediction model based on the above data processing process, and the terminal device obtains a click-through rate prediction value corresponding to the information to be processed based on the target click-through rate prediction model, and presents the click-through rate prediction value in the operation interface, etc. This application does not make any specific limitations on this.

[0085] The following describes the data processing method provided by the exemplary embodiment of the present application in combination with the application scenarios described above and with reference to the accompanying drawings. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the implementation methods of the present application are not limited in this regard.

[0086] See also Figure 2 As shown, it is a flow chart of a data processing method provided in an embodiment of the present application. The flow of the method can be executed by an electronic device, which can be a terminal device or a server. The specific flow is as follows:

[0087] S201. Obtain a sample data set and an initial click rate prediction model.

[0088] In the embodiment of the present application, each sample data may include exposure data and click data of multimedia content, wherein the exposure data is used to characterize the number of presentations of the multimedia content, and the click data is used to characterize the number of clicks on the multimedia content. Multimedia content includes but is not limited to video, audio, text, etc.

[0089] The exposure data and click data of the multimedia content may be collected according to a preset data collection interval. For example, the data collection interval may be 1 hour.

[0090] After obtaining the exposure data and click data of the multimedia content, the corresponding sample data can be labeled according to the exposure data and click data of the multimedia content. For example, the multimedia information that triggers a click operation after exposure is used as a positive sample, and the multimedia information that does not trigger a click operation after exposure is used as a negative sample. Then, in S202, the click rate prediction model can be trained according to the positive samples and the negative samples.

[0091] It is understandable that in the specific implementation of this application, related data such as exposure data and click data are involved. When the embodiments of this application are applied to specific products or technologies, permission or consent for use is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0092] S202: Based on the sample data set, the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model; wherein, in one iteration, the following operations are performed:

[0093] Determine at least one model parameter to be deleted from the currently retained model parameters, perform pruning on the current click-through rate prediction model based on the at least one model parameter to be deleted, and train the click-through rate prediction model obtained after the pruning based on the sample data set;

[0094] At least one model parameter to be restored is determined from the currently pruned deleted model parameters, and based on the at least one model parameter to be restored, the current click-through rate prediction model is regenerated, and based on the sample data set, the click-through rate prediction model obtained after the regeneration is trained.

[0095] In the embodiment of the present application, the initial click rate prediction model can adopt but is not limited to deep neural networks (DNN).

[0096] The initial click rate prediction model may include an input layer, an embedding layer, a feature interaction layer, and an output layer, wherein the input layer is used to obtain input sample data, the embedding layer is used to map the input discrete features to a lower-dimensional embedding representation, and the interaction layer is used to implement complex feature interaction processing. Exemplarily, the interaction layer may include non-DNN components and DNN components, wherein the non-DNN components are used to implement explicit low-order interactions, and the DNN components are used to implement implicit high-order interactions. Non-DNN components include, but are not limited to, logistic regression (LR), factorization machines (FM), and field-weighted factorization machines (FwFM).

[0097] See also Figure 3 As shown, it is a possible initial click rate prediction model provided in an embodiment of the present application, the initial click rate prediction model includes an input layer, an Embedding layer, a FwFM component, a DNN component and an output layer, wherein the input layer is used to obtain sparse features, the Embedding layer is used to map the sparse features to Embedding representations, the FwFM component includes one or more linear units, and includes one or more quantization units, Figure 3 In the figure, diamonds are used to represent linear units and pentagons are used to represent quantization units.

[0098] In order to ensure the model prediction effect and model generalization ability, the click-through rate prediction model may include one or more DNN components. Since the DNN component has a greater impact on model training, in an embodiment of the present application, the model parameters contained in the DNN component may be pruned and regenerated.

[0099] Specifically, in an embodiment of the present application, the initial click-through rate prediction model includes at least a deep neural network component. When executing S202, based on a sample data set, an iterative method is used to perform pruning and regeneration processing on the DNN component included in the initial click-through rate prediction model to obtain a target click-through rate prediction model.

[0100] Among them, the model parameters included in the DNN component may include one or more of weights and weights.

[0101] It should be noted that in the embodiment of the present application, the number of DNN components included in the initial click-through rate prediction model can be one or more, and there is no limitation on this. If the initial click-through rate prediction model includes one DNN component, when the initial click-through rate prediction model is pruned and regenerated, the one DNN component can be pruned and regenerated. If the initial click-through rate prediction model includes multiple DNN components, any one of the multiple DNN components can be pruned and regenerated according to actual application requirements, or some of the multiple DNN components can be pruned and regenerated, or the multiple DNN components can be pruned and regenerated. In the following, only the example of the initial click-through rate prediction model including one DNN component is used for explanation.

[0102] For example, see Figure 4A As shown in FIG. 1 , the initial click rate prediction model includes a DNN component, a non-DNN component, and an output layer. The non-DNN component may be a FwFM component, which is used to realize the first-order and second-order feature interactions. The initial click rate prediction model also includes an input layer, an embedding layer, etc. ( Figure 4A not shown).

[0103] In the initial click rate prediction model, each neuron in the DNN component is connected in a fully connected manner, that is, neuron 1 is connected to neuron 2, neuron 3, and neuron 4, and neuron 2, neuron 3, and neuron 4 are connected to neurons 5 to 10 respectively.

[0104] Taking the first iteration process as an example, the currently retained model parameters are all model parameters. From the currently retained model parameters, the model parameters to be deleted are determined, and based on the determined model parameters to be deleted, the current click-through rate prediction model is pruned, and the click-through rate prediction model obtained after the pruning process is trained based on the sample data set. The click-through rate prediction model obtained after the pruning process is referred to Figure 4B As shown, Figure 4B , the currently retained model parameters are parameter 1, parameter 2, parameter 3, parameter 4, parameter 5, and parameter 6, wherein parameter 1 represents the weight between neuron 2 and neuron 5, parameter 2 represents the weight between neuron 2 and neuron 8, parameter 3 represents the weight between neuron 2 and neuron 10, parameter 4 represents the weight between neuron 4 and neuron 5, parameter 5 represents the weight between neuron 4 and neuron 7, and parameter 6 represents the weight between neuron 4 and neuron 8. That is to say, in the first iterative pruning process, all model parameters except parameters 1-parameter 6 in the initial click rate prediction model are pruned.

[0105] After pruning, the currently pruned deleted model parameters except parameters 1-6 include parameters 7, 8 and 9, among which parameter 7 represents the weight between neuron 2 and neuron 7, parameter 8 represents the weight between neuron 2 and neuron 9, and parameter 9 represents the weight between neuron 4 and neuron 10. From the currently pruned deleted model parameters, the model parameters to be restored are determined, and the model parameters to be restored include parameters 7, 8 and 9. Then, based on the determined model parameters to be restored, the current click-through rate prediction model is regenerated, and then based on the sample data set, the click-through rate prediction model obtained after the regeneration is trained.

[0106] See also Figure 5A As shown, after multiple iterations, the target click rate prediction model is obtained. In the target click rate prediction model, the retained model parameters include: the weight between neuron 2 and neuron 8, the weight between neuron 4 and neuron 5, and the weight between neuron 4 and neuron 7, that is, retained parameters 2, parameter 4 and parameter 5.

[0107] Since sparse DNN can pass gradient information to embedding during reverse gradient propagation, sparse DNN has a positive effect on the overall model during training and assists in embedding learning. The redundant parameters of the DNN part of the pruned model are removed.

[0108] It should be noted that Figure 4A The first-order, second-order, and DNN parts of the model shown can be added after obtaining their respective logistic regressions (logit), or they can be merged (concat) and then processed linearly (linear). The two are theoretically consistent, that is, Linear([a;b])=Linear(a)+Linear(b), where a and b represent two outputs, and Linear() is used for linear processing. The former makes it easier to observe the effectiveness of the DNN part.

[0109] It should be noted that in an embodiment of the present application, in an iteration process, after training the click-through rate prediction model obtained after the pruning process, it can be determined whether the model convergence conditions are met, and if so, the target click-through rate prediction model is output; it can also be determined whether the model convergence conditions are met after training the click-through rate prediction model obtained after the rebirth process, and if so, the target click-through rate prediction model is output; it can also be determined whether the model convergence conditions are met after training the click-through rate prediction model obtained after the pruning process and after training the click-through rate prediction model obtained after the rebirth process. For specific model convergence conditions, please refer to below.

[0110] As the pruning-regeneration process continues, in extreme cases, this mechanism can completely prune DNN components. For example, see Figure 5B As shown in the figure, after multiple pruning-rebirth processes, all the model parameters of the DNN component in the target click rate prediction model are pruned.

[0111] Assume that the number of features in the input layer is 30, the embedding size is 128, the DNN contains 4 layers of neurons, the number of neurons in each layer is 512-256-128-64, the number of ID features (i.e. discrete features, features with a dimension of 1) of each sample data is 400, and before pruning, the number of model parameters (excluding the embedding mapping table) is: 30*30+30+128*30+30 +256*400+256*30+128*30*512+512*256+256*128+128*64+64=2253156, and after the DNN component is completely pruned, the number of model parameters is 30*30+30+128*30+30+256*400+256*30+128*30=118720. Obviously, through the pruning operation, the number of parameters is reduced by 94%, which greatly reduces the amount of calculation.

[0112] Obviously, through the above implementation method, during the model training process, the DNN component can be pruned and reborn to improve the model training speed and ensure low latency. In addition, after the pruning and rebirth process, a large number of model parameters contained in the DNN component are pruned, thereby reducing the memory occupied by the model and further improving the model training speed.

[0113] In some embodiments, based on at least one model parameter to be deleted, pruning the current click rate prediction model includes:

[0114] In the current click rate prediction model, a parameter value of at least one model parameter to be deleted is set to a first set value.

[0115] For example, assuming that the first set value is 0, the model parameters to be deleted include parameter 3, ..., parameter 9, and in the current click rate prediction model, the parameter values ​​of parameter 3, ..., parameter 9 are all set to 0.

[0116] Through the above implementation, since the parameter value of the model parameter is set to 0, it is not actually deleted from the model. Therefore, the deleted model parameter can still participate in the back propagation process, thereby improving the model accuracy.

[0117] In some embodiments, when determining at least one model parameter to be deleted from the currently reserved model parameters, any one of the following operations may be used:

[0118] Operation A1: Based on the parameter values ​​corresponding to each of the currently retained model parameters, select a retained model parameter whose corresponding parameter value is not greater than a preset first value threshold from the retained model parameters, and use at least one selected retained model parameter as at least one model parameter to be deleted.

[0119] For example, assuming that the currently retained model parameters include parameter 1, parameter 2, and parameter 3, wherein the parameter values ​​corresponding to parameter 1, parameter 2, and parameter 3 are 0.1, 0.2, and 0.3, respectively, and the preset first value threshold is 0.15, based on the parameter values ​​corresponding to the currently retained model parameters, a retained model parameter whose corresponding parameter value is not greater than 0.15 is selected from the retained model parameters, that is, parameter 1 is selected, and the selected parameter 1 is used as a model parameter to be deleted.

[0120] For another example, assuming that the currently retained model parameters include parameter 1, parameter 2, and parameter 3, wherein the parameter values ​​corresponding to parameter 1, parameter 2, and parameter 3 are 0.1, 0.2, and 0.3, respectively, and the preset first value threshold is 0.25, based on the parameter values ​​corresponding to the currently retained model parameters, the retained model parameters whose corresponding parameter values ​​are not greater than 0.15 are selected from the retained model parameters, that is, parameter 1 and parameter 2 are selected, and the selected parameter 1 and parameter 2 are used as the two model parameters to be deleted.

[0121] Operation A2: sorting the retained model parameters based on the parameter values ​​corresponding to the currently retained model parameters, and determining at least one model parameter to be deleted from the retained model parameters based on the corresponding first sorting result and the preset first selection ratio.

[0122] It should be noted that in the embodiment of the present application, when sorting the retained model parameters, they can be sorted in order from large to small, or in order from small to large, and there is no limitation on this.

[0123] For example, see Figure 6As shown, it is assumed that the currently retained model parameters include parameter 1, parameter 2, parameter 3, and parameter 4, wherein the parameter values ​​corresponding to parameter 1, parameter 2, parameter 3, and parameter 4 are 0.1, 0.2, 0.3, and 0.1, respectively, and the preset first selection ratio is 50%. Based on the parameter values ​​corresponding to the currently retained model parameters, the retained model parameters are sorted, and the sorting results are parameter 3, parameter 2, parameter 1, and parameter 4 in descending order. Then, based on the corresponding first sorting results and the preset first selection ratio, the model parameters to be deleted are determined to be parameter 1 and parameter 4 from the retained model parameters.

[0124] Through the above implementation, the model parameters to be deleted can be determined from the retained model parameters through the parameter values ​​of each model parameter, thereby reducing the possibility of parameter being erroneously reduced and improving the accuracy of model training.

[0125] In some embodiments, based on the sample data set, the click rate prediction model obtained after pruning is trained, including:

[0126] Inputting the sample data set into the click rate prediction model obtained after pruning processing to obtain the click rate prediction value corresponding to each sample data contained in the sample data set;

[0127] Based on the obtained estimated values ​​of the click rates and the actual values ​​of the click rates corresponding to the sample data, the parameter adjustment values ​​corresponding to the deletion model parameters are determined.

[0128] It should be noted that, in the embodiment of the present application, the parameter adjustment value can be determined based on the back propagation of the deep neural network. The click rate estimate can be expressed in numerical values ​​or in levels, and there is no limitation on this. The following only uses numerical values ​​as an example for explanation.

[0129] For example, assuming that the currently pruned deletion model parameter is parameter 3, the sample data set includes sample data 1, and the true value of the click-through rate of sample data 1 is 100%. Sample data 1 is input into the click-through rate prediction model obtained after the pruning process to obtain the click-through rate estimation value corresponding to sample data 1, wherein the click-through rate estimation value corresponding to sample data 1 is 50%. Then, based on the obtained click-through rate estimation value and the true value of the click-through rate corresponding to sample data 1, the parameter adjustment value corresponding to parameter 3 is determined to be 0.12.

[0130] Through the above implementation method, in the back propagation process, the parameter adjustment value of the deleted model parameter is determined by the true value and the estimated value, which improves the model convergence effect and further improves the model performance.

[0131] In some embodiments, determining at least one model parameter to be restored from the currently pruned deleted model parameters may be performed by any one of the following operations:

[0132] Operation B1: based on the parameter adjustment values ​​corresponding to the respective deletion model parameters, select from the deletion model parameters a deletion model parameter whose corresponding parameter adjustment value is not less than a preset second value threshold, and use at least one selected deletion model parameter as at least one model parameter to be restored.

[0133] For example, assuming that the deletion model parameters include parameter 2 and parameter 3, the parameter adjustment value corresponding to parameter 2 is 0.01, the parameter adjustment value corresponding to parameter 3 is 0.12, and the second value threshold is 0.1, then, based on the parameter adjustment values ​​corresponding to each deletion model parameter, a deletion model parameter whose corresponding parameter adjustment value is not less than 0.1 is selected from each deletion model parameter, that is, parameter 3 is selected, and then the selected parameter 3 is used as the model parameter to be restored.

[0134] Operation B2: sorting the deleted model parameters based on the parameter adjustment values ​​corresponding to the deleted model parameters, and determining at least one model parameter to be restored from the deleted model parameters based on the corresponding second sorting result and a preset second selection ratio.

[0135] It should be noted that, in the embodiment of the present application, when sorting the deletion model parameters, they can be sorted in order from large to small, or in order from small to large, and there is no limitation on this.

[0136] For example, assuming that each deletion model parameter includes parameter 1, parameter 2, and parameter 3, wherein the parameter adjustment values ​​corresponding to parameter 1, parameter 2, and parameter 3 are 0.02, 0.01, and 0.12, respectively, and the preset first selection ratio is 60%. Based on the parameter adjustment values ​​corresponding to each deletion model parameter, each deletion model parameter is sorted, and the sorting results are parameter 3, parameter 1, and parameter 2 in descending order. Then, based on the corresponding second sorting result and the preset second selection ratio, the model parameter to be restored is determined to be parameter 3 from each deletion model parameter.

[0137] Through the above implementation method, the model parameters to be restored can be determined from the deleted model parameters through the parameter adjustment values ​​of each deleted model parameter, thereby reducing the possibility of parameters being erroneously reduced, so that important connections in the deep neural network are retained, and the model can achieve better click-through rate estimation benefits while using fewer parameters.

[0138] In some embodiments, based on the model parameters to be restored, the current click rate prediction model is regenerated, including:

[0139] In the current click rate prediction model, a parameter value of at least one model parameter to be restored is adjusted to a corresponding parameter adjustment value.

[0140] For example, assuming that the model parameter to be restored is parameter 3, and the parameter adjustment value corresponding to parameter 3 is 0.12, in the current click rate prediction model, the parameter value of parameter 3 is adjusted to 0.12.

[0141] Through the above implementation method, some pruned model parameters in the click rate prediction model can be restored, so that important connections in the deep neural network are retained, thereby improving the model performance and improving the prediction accuracy of the model.

[0142] In some embodiments, in order to improve the model training efficiency, the initial click rate prediction model may be a relatively stable model obtained through training. Specifically, the initial click rate prediction model may be obtained by:

[0143] Get the click rate estimation model to be trained;

[0144] Based on the sample data set, the click rate prediction model to be trained is iteratively trained until the model convergence condition is met, and an initial click rate prediction model is obtained.

[0145] Specifically, the model convergence condition may be at least one of the following conditions but is not limited to:

[0146] Condition 1: For N consecutive times, the first difference between the model evaluation indicators of two adjacent iterations of the click rate prediction model to be trained is not greater than a preset first difference threshold, where the value of N is a positive integer.

[0147] It should be noted that in the embodiments of the present application, the model evaluation index may be, but is not limited to, the area under the ROC curve (AUC) and the coordinate axis, the click-through rate (CTR), and the like.

[0148] Among them, AUC can be calculated in the following way: first, for each sample data contained in the sample data set, sort them from large to small according to the click rate estimate, then, according to the sorting result, let the sample data with the largest click rate estimate be ranked as n, let the sample data with the second largest click rate estimate be ranked as n-1, and so on, where n is M+Q, M is the number of positive samples, Q is the number of negative samples, and positiveClass represents the positive samples. After that, add up the ranks of all positive samples, subtract the counts of the two combinations of M-1 positive samples, and divide by M×Q.

[0149] Specifically, the AUC can be calculated using the following formula (1):

[0150]

[0151] CTR can be calculated using the following formula (2), where #Click represents click data and #Expose represents exposure data:

[0152]

[0153] For example, assuming that the value of N is 2, the preset first difference threshold is 0.05, in the first iteration, the AUC value of the click-through rate prediction model to be trained is 0.6, in the second iteration, the AUC value of the click-through rate prediction model to be trained is 0.65, in the third iteration, the AUC value of the click-through rate prediction model to be trained is 0.7, the first difference between the AUCs of the first iteration and the second iteration is 0.05, and the first difference between the AUCs of the second iteration and the third iteration is 0.05. For two consecutive times, the first difference between the model evaluation indicators of two adjacent iterations of the click-through rate prediction model to be trained is not greater than 0.05. Therefore, in the third iteration, it is determined that the model convergence condition is met and the initial click-through rate prediction model is output.

[0154] Condition 2: The second difference between the model evaluation index of the click rate estimation model to be trained and the model evaluation index of the specified model is not greater than a preset second difference threshold.

[0155] For example, suppose that the preset second difference threshold is 0.1, the CTR prediction model to be trained is a deep neural network model based on FM, the specified model is based on the Wide&Deep model, and the Wide&Deep model is divided into a wide part and a deep part, wherein the wide part is a linear network with memory, and the deep part is a nonlinear network with generalizability, such as a deep neural network. If the second difference between the AUC of the CTR prediction model to be trained and the AUC of the specified model is 0.05, then the second difference between the model evaluation index of the CTR prediction model to be trained and the model evaluation index of the specified model is not greater than 0.1. At this time, it is determined that the model convergence condition is met and the initial CTR prediction model is output.

[0156] Through the above implementation, when the offline AUC indicator or the online CTR indicator of the model no longer changes significantly, it is determined that the model has converged, thereby improving the model performance and the estimation accuracy of the model.

[0157] In some embodiments, when executing S202, the following two methods may be used but are not limited to:

[0158] Method 1: During the incremental training process, pruning and regeneration are performed. That is, the sample data set contains newly added sample data.

[0159] Specifically, before training the click rate prediction model obtained after the pruning process or the click rate prediction model obtained after the regeneration process, each newly added sample data is obtained and added to the sample data set.

[0160] Accordingly, based on the sample data set, the click rate prediction model obtained after pruning is trained, including:

[0161] Based on the newly added sample data included in the sample data set, the click rate prediction model obtained after pruning is trained.

[0162] It should be noted that each newly added sample data may be sample data collected within a preset time period, for example, sample data within one hour.

[0163] For example, see Figure 7 As shown, in one iteration, after the current click-through rate prediction model is pruned based on at least one model parameter to be deleted, the click-through rate prediction model obtained after the pruning process is trained using each newly added sample data obtained between 16:00 and 17:00, and after the current click-through rate prediction model is regenerated based on at least one model parameter to be restored, the click-through rate prediction model obtained after the regeneration process is trained using each newly added sample data obtained between 17:00 and 18:00. Similarly, in the next iteration, the click-through rate prediction model obtained after the pruning process can be trained using each newly added sample data obtained between 18:00 and 19:00, and the click-through rate prediction model obtained after the regeneration process can be trained using each newly added sample data obtained between 19:00 and 20:00.

[0164] Through the above implementation method, it can be ensured that the model can self-iterate over time, and then the trained click-through rate prediction model can be updated online in real time, thereby improving the efficiency and accuracy of click-through rate prediction.

[0165] Method 2: After obtaining the target CTR estimation model, the non-DNN part of the target CTR estimation model is transplanted to a new model with the same structure but without DNN as the initial parameters of the new model, and then the stabilized model is obtained through incremental training.

[0166] Specifically, a new estimation model is constructed based on the non-deep neural network components included in the target click rate estimation model;

[0167] Based on the sample data set, the new estimation model is trained, and the trained new estimation model is used as the new target click rate estimation model.

[0168] It should be noted that in the embodiment of the present application, the new estimation model has the same structure as the target click-through rate estimation model and does not include a DNN component. The initial values ​​of the model parameters of the new estimation model adopt the model parameters corresponding to the non-deep neural network components in the target click-through rate estimation model.

[0169] For example, see Figure 8 As shown, after obtaining the target click-through rate prediction model, a new prediction model is constructed based on the FwFM component included in the target click-through rate prediction model. The new prediction model includes the FwFM component. The initial values ​​of the model parameters of the FwFM component use the corresponding model parameters in the target click-through rate prediction model. Then, based on the sample data set, the new prediction model is trained, and the trained new prediction model is used as the new target click-through rate prediction model.

[0170] Through the above implementation method, the effective gain of DNN for embedding training can be retained, and the overall model is simpler, achieving faster speed during online reasoning.

[0171] In actual application, an AB experiment was conducted in the scenario of rough sorting of videos recommended by subscription accounts. The experimental results are shown in Table 1:

[0172] Table 1 Experimental results of AB experiment

[0173]

[0174] Among them, B1 is the online model Wide+Deep+Attentional FactorizationMachine (AFM) model, B2 is the pruned DeepFM model, tlfeeds and mainsort are used to characterize two different business scenarios. For example, tlfeeds is used to characterize the subscription account recommendation flow scenario, and mainsort is used to characterize the subscription account main scenario. In a week of online AB experiments, taking the click-through rate indicator as an example, the click-through rate indicator increased by 4.910%, and the benefits of multiple other indicators increased significantly. It can be seen that the pruning-rebirth process can significantly improve the model benefits.

[0175] At the same time, a statistical analysis was conducted on the average inference time of a single sample, and the results are shown in Table 2 below:

[0176] Table 2 Average inference time for a single sample

[0177]

[0178] It can be seen that for the unpruned DeepFM model, the average single-sample inference time of the pruned model is greatly reduced, down 41%; in addition, for the Wide+Deep+AFM model, although the pruned model uses higher-dimensional embeddings for training and inference, its model is simpler, without the calculation time of the DNN part and the second-order cross attention part, and the overall time consumption is reduced by 58%.

[0179] The following is an explanation based on two specific application scenarios.

[0180] Application scenario 1: Subscription account recommendation.

[0181] See also Fig. 9 As shown, an instant messaging application client is installed in the terminal device, the instant messaging application client can provide a subscription account recommendation service, and the server is a server corresponding to the instant messaging application client.

[0182] Based on the sample data set, the server uses an iterative method to prune and regenerate the initial click-through rate prediction model. After obtaining the target click-through rate prediction model, the server obtains information to be processed 1 and information to be processed 2, where information to be processed 1 is the emoticon-related message published by subscription account A, and information to be processed 2 is the consulting video published by subscription account B.

[0183] The server determines the target recommendation information based on the target click-through rate prediction model. Specifically, the server can obtain the click-through rate estimation value corresponding to each piece of information to be processed based on the target click-through rate prediction model. Assume that the click-through rate estimation value corresponding to the information public account is 90%, and the click-through rate estimation value corresponding to the sports public account is 80%.

[0184] Afterwards, the server determines that both the information to be processed 1 and the information to be processed 2 are target recommended information based on the estimated click rates corresponding to the respective information to be processed, and the recommendation order of the information to be processed 1 and the information to be processed 2 is the information to be processed 1 and the information to be processed 2. Next, the server generates a corresponding recommendation interface based on the recommendation order, so that the terminal device presents the target recommended information in the operation interface.

[0185] Application scenario 2: video recommendation.

[0186] A video application client is installed in the terminal device, and the video application client can provide a video recommendation service. The server is a server corresponding to the video application client.

[0187] Based on the sample data set, the server uses an iterative method to prune and regenerate the initial click rate prediction model. After obtaining the target click rate prediction model, the server obtains video 1 to be recommended, video 2 to be recommended, and video 3 to be recommended. Among them, video 1 to be recommended is a food video, video 2 to be recommended is a game video, and video 3 to be recommended is a news video.

[0188] Based on the target click rate prediction model, the server obtains the click rate estimation value corresponding to each video to be recommended. Assume that the click rate estimation value corresponding to the video to be recommended 1 is 90%, the click rate estimation value corresponding to the video to be recommended 2 is 95%, and the click rate estimation value corresponding to the video to be recommended 3 is 96%.

[0189] Afterwards, the server determines that video 2 and video 3 to be recommended are target recommendation information based on the estimated click rates corresponding to each piece of information to be processed. Afterwards, the server generates a corresponding recommendation interface so that the terminal device presents video 2 and video 3 to be recommended in the operation interface.

[0190] Based on the same inventive concept, the present application embodiment provides a data processing device. Fig.10 As shown, it is a schematic diagram of the structure of the data processing device 1000, which may include:

[0191] An acquisition unit 1001 is used to acquire a sample data set and an initial click rate prediction model;

[0192] The training unit 1002 is used to perform pruning-regeneration processing on the initial click rate prediction model in an iterative manner based on the sample data set to obtain a target click rate prediction model; wherein, in one iteration process, the following operations are performed:

[0193] Determining at least one model parameter to be deleted from the currently retained model parameters, and pruning the current click-through rate prediction model based on the at least one model parameter to be deleted, and training the click-through rate prediction model obtained after the pruning based on the sample data set;

[0194] At least one model parameter to be restored is determined from the currently pruned deleted model parameters, and based on the at least one model parameter to be restored, the current click-through rate prediction model is regenerated, and based on the sample data set, the click-through rate prediction model obtained after the regeneration is trained.

[0195] As a possible implementation method, the initial click rate prediction model at least includes a deep neural network component;

[0196] Then, based on the sample data set, the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model, the training unit 1002 is used to:

[0197] Based on the sample data set, the deep neural network component included in the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model.

[0198] As a possible implementation method, the initial click rate prediction model also includes a non-deep neural network component;

[0199] Then, based on the sample data set, the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model. Then, the training unit 1002 is further used to:

[0200] Building a new estimation model based on the non-deep neural network components included in the target click rate estimation model;

[0201] Based on the sample data set, the new estimation model is trained, and the trained new estimation model is used as a new target click rate estimation model.

[0202] As a possible implementation manner, when determining at least one model parameter to be deleted from the currently retained model parameters, the training unit 1002 is used to:

[0203] Based on the parameter values ​​corresponding to the currently reserved model parameters, select a reserved model parameter whose corresponding parameter value is not greater than a preset first value threshold from the reserved model parameters, and use the selected at least one reserved model parameter as the at least one model parameter to be deleted; or

[0204] The retained model parameters are sorted based on parameter values ​​corresponding to the currently retained model parameters, and the at least one model parameter to be deleted is determined from the retained model parameters based on a corresponding first sorting result and a preset first selection ratio.

[0205] As a possible implementation manner, when pruning the current click rate prediction model based on the at least one model parameter to be deleted, the training unit 1002 is used to:

[0206] In the current click rate prediction model, the parameter value of the at least one model parameter to be deleted is set to a first set value.

[0207] As a possible implementation, when the click rate prediction model obtained after pruning is trained based on the sample data set, the training unit 1002 is used to:

[0208] Inputting the sample data set into the click rate prediction model obtained after pruning processing to obtain the click rate prediction value corresponding to each sample data contained in the sample data set;

[0209] Based on the obtained estimated click-through rates and the actual click-through rates corresponding to the sample data, the parameter adjustment values ​​corresponding to the deletion model parameters are determined.

[0210] As a possible implementation, when determining at least one model parameter to be restored from the currently pruned deleted model parameters, the training unit 1002 is used to:

[0211] Based on the parameter adjustment values ​​corresponding to the deletion model parameters, a deletion model parameter whose corresponding parameter adjustment value is not less than a preset second value threshold is selected from the deletion model parameters, and the selected at least one deletion model parameter is used as the at least one model parameter to be restored; or

[0212] The deleted model parameters are sorted based on the parameter adjustment values ​​corresponding to the deleted model parameters, and the at least one model parameter to be restored is determined from the deleted model parameters based on the corresponding second sorting result and a preset second selection ratio.

[0213] As a possible implementation, when regenerating the current click rate prediction model based on the model parameters to be restored, the training unit 1002 is used to:

[0214] In the current click rate prediction model, the parameter value of the at least one model parameter to be restored is adjusted to the corresponding parameter adjustment value.

[0215] As a possible implementation, after obtaining the sample data set and the initial click rate prediction model, and before training the click rate prediction model obtained after pruning based on the sample data set, the training unit 1002 is further used to:

[0216] Acquire each newly added sample data, and add each newly added sample data to the sample data set;

[0217] When the click rate prediction model obtained after pruning is trained based on the sample data set, the training unit 1002 is used to:

[0218] Based on the newly added sample data included in the sample data set, the click rate prediction model obtained after pruning is trained.

[0219] As a possible implementation, the training unit 1002 obtains the initial click rate prediction model in the following manner:

[0220] Get the click rate estimation model to be trained;

[0221] Based on the sample data set, the click rate prediction model to be trained is iteratively trained until a model convergence condition is met, thereby obtaining the initial click rate prediction model.

[0222] As a possible implementation method, satisfying the model convergence condition includes:

[0223] For N consecutive times, the first difference between the model evaluation indicators of two adjacent iterations is not greater than a preset first difference threshold, where the value of N is a positive integer; or

[0224] A second difference between the model evaluation index of the click-through rate prediction model to be trained and the model evaluation index of the specified model is not greater than a preset second difference threshold.

[0225] For the convenience of description, the above parts are divided into modules (or units) according to their functions and described separately. Of course, when implementing this application, the functions of each module (or unit) can be implemented in the same or multiple software or hardware.

[0226] Regarding the device in the above embodiment, the specific manner in which each unit executes the request has been described in detail in the embodiment of the method, and will not be elaborated here.

[0227] Those skilled in the art will appreciate that various aspects of the present application may be implemented as a system, method or program product. Therefore, various aspects of the present application may be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software, which may be collectively referred to as "circuit", "module" or "system" herein.

[0228] Based on the same inventive concept, the embodiment of the present application also provides an electronic device. In one embodiment, the electronic device can be a server or a terminal device. Fig.11 As shown, it is a schematic diagram of the structure of a possible electronic device provided in an embodiment of the present application, Fig.11 In the embodiment, the electronic device 1100 includes: a processor 1110 and a memory 1120.

[0229] The memory 1120 stores a computer program that can be executed by the processor 1110 , and the processor 1110 can perform the steps of the above-mentioned data processing method by executing the instructions stored in the memory 1120 .

[0230] The memory 1120 may be a volatile memory, such as a random-access memory (RAM); the memory 1120 may also be a non-volatile memory, such as a read-only memory (ROM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD); or the memory 1120 may be any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 1120 may also be a combination of the above memories.

[0231] The processor 1110 may include one or more central processing units (CPU) or a digital processing unit, etc. The processor 1110 is configured to implement the above data processing method when executing the computer program stored in the memory 1120 .

[0232] In some embodiments, the processor 1110 and the memory 1120 may be implemented on the same chip. In some embodiments, they may also be implemented on separate chips.

[0233] The specific connection medium between the processor 1110 and the memory 1120 is not limited in the embodiment of the present application. In the embodiment of the present application, the processor 1110 and the memory 1120 are connected via a bus as an example. Fig.11 The connections between the other components are only for illustration and are not intended to be limiting. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of description, Fig.11 The diagram shows that only one thick line is used, but this does not mean that there is only one bus or only one type of bus.

[0234] Based on the same inventive concept, an embodiment of the present application provides a computer-readable storage medium, which includes a computer program. When the computer program is run on an electronic device, the computer program is used to enable the electronic device to perform the steps of the above-mentioned data processing method. In some possible implementations, various aspects of the data processing method provided by the present application can also be implemented in the form of a program product, which includes a computer program. When the program product is run on an electronic device, the computer program is used to enable the electronic device to perform the steps in the above-mentioned data processing method. For example, the electronic device can perform the following steps: Figure 2 Follow the steps shown in .

[0235] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a RAM, a ROM, an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (Compact Disk Read Only Memory, CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0236] The program product of the embodiment of the present application may adopt a CD-ROM and include a computer program, and can be run on an electronic device. However, the program product of the present application is not limited thereto, and in this document, a readable storage medium may be any tangible medium containing or storing a computer program, which may be used by or in combination with a command execution system, apparatus or device.

[0237] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, wherein a readable computer program is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A readable signal medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a computer program for use by or in conjunction with a command execution system, apparatus, or device.

[0238] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0239] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A data processing method, characterized in that: The method comprises: Obtain a sample data set and an initial click rate prediction model; the initial click rate prediction model includes a deep neural network component and a non-deep neural network component; Based on the sample data set, the initial click rate prediction model is pruned and regenerated in an iterative manner to obtain a target click rate prediction model; wherein, in one iteration, the following operations are performed: Determine at least one model parameter to be deleted from the currently retained model parameters, and based on the at least one model parameter to be deleted, prune the deep neural network component in the current click-through rate prediction model, and train the click-through rate prediction model obtained after the pruning process based on the newly added sample data, wherein the click-through rate prediction model obtained after the pruning process includes a deep neural network component and a non-deep neural network component; Determine at least one model parameter to be restored from the currently pruned deleted model parameters, and based on the at least one model parameter to be restored, regenerate the deep neural network component in the current click-through rate prediction model, and train the click-through rate prediction model obtained after the regeneration process based on the newly added sample data, wherein the click-through rate prediction model obtained after the regeneration process includes a deep neural network component and a non-deep neural network component; Based on the non-deep neural network component in the target click-through rate prediction model, a new prediction model is constructed, and based on the newly added sample data, the new prediction model is trained, and the trained new prediction model is used as a new target click-through rate prediction model.

2. The method according to claim 1, characterized in that The step of determining at least one model parameter to be deleted from the currently reserved model parameters includes: Based on the parameter values ​​corresponding to the currently reserved model parameters, select a reserved model parameter whose corresponding parameter value is not greater than a preset first value threshold from the reserved model parameters, and use the selected at least one reserved model parameter as the at least one model parameter to be deleted; or The retained model parameters are sorted based on parameter values ​​corresponding to the currently retained model parameters, and the at least one model parameter to be deleted is determined from the retained model parameters based on a corresponding first sorting result and a preset first selection ratio.

3. The method according to claim 1, characterized in that The pruning of the current click rate prediction model based on the at least one model parameter to be deleted includes: In the current click rate prediction model, the parameter value of the at least one model parameter to be deleted is set to a first set value.

4. The method according to claim 1, characterized in that The training of the click rate prediction model obtained after pruning based on the newly added sample data includes: Inputting the sample data set into the click rate prediction model obtained after pruning processing to obtain the click rate prediction value corresponding to each sample data contained in the sample data set; Based on the obtained estimated click-through rates and the actual click-through rates corresponding to the sample data, the parameter adjustment values ​​corresponding to the deletion model parameters are determined.

5. The method according to claim 4, characterized in that The step of determining at least one model parameter to be restored from the currently pruned deleted model parameters includes: Based on the parameter adjustment values ​​corresponding to the deletion model parameters, a deletion model parameter whose corresponding parameter adjustment value is not less than a preset second value threshold is selected from the deletion model parameters, and the selected at least one deletion model parameter is used as the at least one model parameter to be restored; or The deleted model parameters are sorted based on the parameter adjustment values ​​corresponding to the deleted model parameters, and the at least one model parameter to be restored is determined from the deleted model parameters based on the corresponding second sorting result and a preset second selection ratio.

6. The method according to claim 4, characterized in that The regenerating process of the current click rate prediction model based on the model parameters to be restored includes: In the current click rate prediction model, the parameter value of the at least one model parameter to be restored is adjusted to the corresponding parameter adjustment value.

7. The method according to any one of claims 1 to 6, characterized in that The initial click rate estimation model is obtained in the following way: Get the click rate estimation model to be trained; Based on the sample data set, the click rate prediction model to be trained is iteratively trained until a model convergence condition is met, thereby obtaining the initial click rate prediction model.

8. The method according to claim 7, characterized in that The model convergence condition is satisfied, including: For N consecutive times, the first difference between the model evaluation indicators of two adjacent iterations is not greater than a preset first difference threshold, where the value of N is a positive integer; or A second difference between the model evaluation index of the click-through rate prediction model to be trained and the model evaluation index of the specified model is not greater than a preset second difference threshold.

9. A data processing device, characterized in that: include: An acquisition unit is used to acquire a sample data set and an initial click rate prediction model; the initial click rate prediction model includes a deep neural network component and a non-deep neural network component; The training unit is used to perform pruning-regeneration processing on the initial click rate prediction model in an iterative manner based on the sample data set to obtain a target click rate prediction model; wherein, in one iteration process, the following operations are performed: Determine at least one model parameter to be deleted from the currently retained model parameters, and based on the at least one model parameter to be deleted, prune the deep neural network component in the current click-through rate prediction model, and train the click-through rate prediction model obtained after the pruning process based on the newly added sample data, wherein the click-through rate prediction model obtained after the pruning process includes a deep neural network component and a non-deep neural network component; Determine at least one model parameter to be restored from the currently pruned deleted model parameters, and based on the at least one model parameter to be restored, regenerate the deep neural network component in the current click-through rate prediction model, and train the click-through rate prediction model obtained after the regeneration process based on the newly added sample data, wherein the click-through rate prediction model obtained after the regeneration process includes a deep neural network component and a non-deep neural network component; The training unit is also used to construct a new estimation model based on the non-deep neural network component in the target click-through rate estimation model, train the new estimation model based on the newly added sample data, and use the trained new estimation model as the new target click-through rate estimation model.

10. An electronic device, characterized in that: It comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: It includes a computer program. When the computer program is run on an electronic device, the computer program is used to enable the electronic device to execute the steps of the method according to any one of claims 1 to 8.

12. A computer program product, characterized in that It includes a computer program, which is stored in a computer-readable storage medium. The processor of the electronic device reads and executes the computer program from the computer-readable storage medium, so that the electronic device performs the steps of the method in any one of claims 1 to 8.