Network model training method and device, equipment and computer readable storage medium

Abstract

This invention discloses a network model training method, comprising the following steps: acquiring training data and validation data corresponding to a preset message; generating multiple candidate subnetworks through a deep neural subnetwork generator, and determining the training error corresponding to each candidate subnetwork based on the training data and validation data; determining the optimal subnetwork among the candidate subnetworks based on each training error, and determining the target prediction model based on the optimal subnetwork and the main network. This invention also discloses a network model training device, apparatus, and computer-readable storage medium. This invention enables automatic training of network models through machine learning. For different preset messages, it eliminates the need for manual training and optimization of different network models, saving significant human resources, reducing the workload of model training and tuning, and improving the training efficiency of network models.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a network model training method, apparatus, device, and computer-readable storage medium. Background Technology

[0002] With the rapid development of artificial intelligence technology, neural network models (which can be simply referred to as network models) have broad application prospects in fields such as system identification, pattern recognition, and intelligent control. Typically, a pre-set network model is trained based on a training sample set to obtain a trained network model. Then, the sample to be detected is input into the trained network model to obtain the actual output result of the network model, which is the prediction result for the sample to be detected.

[0003] For example, in programmatic advertising, the characteristics of each user can be input into the network model, and a decision on whether to run ads can be made quickly based on the actual output of the network model.

[0004] However, the current network models require professional algorithm engineers to train and optimize them specifically for different advertising businesses. Different advertising businesses cannot use the same network model. Algorithm engineers need to train and optimize different network models for different advertising businesses, which takes up a lot of their time. Furthermore, the model structure used for each training session needs to be manually specified by the algorithm engineers, resulting in a lack of intelligence in the training and optimization of different network models.

[0005] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The main objective of this invention is to provide a network model training method, apparatus, device, and computer-readable storage medium, aiming to solve the technical problem of unintelligent training and optimization of different network models.

[0007] To achieve the above objectives, the present invention provides a network model training method, which includes the following steps:

[0008] Obtain the training and validation data corresponding to the message;

[0009] Multiple candidate subnetworks are generated using a deep neural subnetwork generator, and the training error corresponding to each candidate subnetwork is determined based on the training data and validation data.

[0010] Based on the training errors, the optimal subnetwork is determined among the candidate subnetworks, and the target prediction model is determined based on the optimal subnetwork and the main network.

[0011] Furthermore, the step of determining the training error corresponding to each candidate sub-network based on the training data and validation data includes:

[0012] The target candidate subnetwork is obtained sequentially from each candidate subnetwork.

[0013] The target candidate subnetwork is embedded into the main network to obtain the updated main network, and the training data and validation data are input into the updated main network to obtain the training error corresponding to the target candidate subnetwork.

[0014] Further, the step of inputting the training data and validation data into the updated main network to obtain the training error corresponding to the target candidate sub-network includes:

[0015] The training data and validation data are input into the updated main network, and the number of training iterations is updated.

[0016] When the number of training iterations reaches the preset number of training iterations, the training error of the updated main network after training is used as the training error of the target candidate sub-network.

[0017] If the number of training iterations does not reach the preset number of training iterations, the updated main network after training is used as the updated main network, and the process returns to the step of inputting the training data and validation data into the updated main network and updating the number of training iterations.

[0018] Furthermore, the step of determining the optimal subnetwork among the candidate subnetworks based on each of the training errors includes:

[0019] Obtain the penalty coefficient corresponding to the network layer number of each candidate subnetwork;

[0020] Based on each penalty coefficient and training error, the weight score corresponding to each candidate subnetwork is determined respectively;

[0021] The candidate subnetwork with the smallest weight score among all candidate subnetworks is selected as the optimal subnetwork.

[0022] Furthermore, the network model training method also includes:

[0023] Update the search count when generating multiple candidate subnetworks using the deep neural subnetwork generator;

[0024] The step of determining the target prediction model based on the optimal sub-network and the main network includes:

[0025] Determine whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network;

[0026] If the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, then it is determined whether the number of search attempts has reached the preset number of search attempts.

[0027] When the number of searches reaches the preset number of searches, the optimal subnetwork is embedded into the main network to obtain the target prediction model.

[0028] Further, after determining whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, the method further includes:

[0029] If the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, then return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator.

[0030] Furthermore, after determining whether the number of searches has reached the preset number of searches, the method further includes:

[0031] If the number of searches does not reach the preset number of searches, the optimal subnetwork is embedded into the main network;

[0032] The main network after embedding the optimal subnetwork is used as the main network, and the process returns to the step of generating multiple candidate subnetworks through the deep neural subnetwork generator.

[0033] Furthermore, after the step of determining the target prediction model based on the optimal sub-network and the main network, the network model training method further includes:

[0034] When the duration after determining the target prediction model reaches the preset duration, return to the steps of obtaining the training data and validation data corresponding to the preset message.

[0035] Furthermore, the training data is user feature data, the validation data is exposure and click data, and after the step of determining the target prediction model based on the optimal sub-network and the main network, the network model training further includes:

[0036] Obtain the user feature data to be predicted, and input the user data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message;

[0037] When the click probability is greater than a preset threshold, the user corresponding to the user feature data to be predicted is determined as the user to be targeted.

[0038] Furthermore, to achieve the above objectives, the present invention also provides a network model training apparatus, the network model training apparatus comprising:

[0039] The acquisition module is used to acquire training data and validation data corresponding to preset messages;

[0040] The generation module is used to generate multiple candidate subnetworks through a deep neural subnetwork generator, and to determine the training error corresponding to each candidate subnetwork based on the training data and validation data.

[0041] The determination module is used to determine the optimal subnetwork among the candidate subnetworks based on the training errors, and then determine the target prediction model based on the optimal subnetwork and the main network.

[0042] In addition, to achieve the above objectives, the present invention also provides a network model training device, the network model training device comprising: a memory, a processor, and a network model training program stored in the memory and executable on the processor, wherein the network model training program, when executed by the processor, implements the steps of the aforementioned network model training method.

[0043] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a network model training program, which, when executed by a processor, implements the steps of the aforementioned network model training method.

[0044] This invention acquires training and validation data corresponding to a preset message; then generates multiple candidate subnetworks using a deep neural subnetwork generator; and determines the training error of each candidate subnetwork based on the training and validation data; subsequently, based on each training error, determines the optimal subnetwork among the candidate subnetworks; and determines the target prediction model based on the optimal subnetwork and the main network. This enables automatic training of network models through machine learning, achieving intelligent training and optimization of different network models. For different advertising businesses, it eliminates the need for manual training and optimization of different network models, saving significant human resources, reducing the workload of model training and tuning, and improving the training efficiency of network models. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the structure of the network model training device in the hardware operating environment involved in the embodiments of the present invention;

[0046] Figure 2 This is a flowchart illustrating the first embodiment of the network model training method of the present invention;

[0047] Figure 3 This is a schematic diagram of the functional modules of an embodiment of the network model training device of the present invention.

[0048] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0049] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0050] like Figure 1 As shown, Figure 1 This is a schematic diagram of the structure of a network model training device in the hardware operating environment involved in the embodiments of the present invention.

[0051] The terminal in this invention embodiment can be a PC, or a smartphone, tablet computer, e-book reader, MP3 (Moving Picture Experts Group Audio Layer III) player, MP4 (Moving Picture Experts Group Audio Layer IV) player, portable computer, or other portable terminal devices with display functions.

[0052] like Figure 1 As shown, the network model training device may include: a processor 1001, such as a CPU; a network interface 1004; a user interface 1003; a memory 1005; and a communication bus 1002. The communication bus 1002 is used to establish communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or stable non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0053] Optionally, the network model training device may also include a camera, RF (Radio Frequency) circuitry, sensors, audio circuitry, a WiFi module, and so on. These sensors may include light sensors, motion sensors, and other sensors. Specifically, light sensors may include ambient light sensors and proximity sensors; of course, the network model training device may also be equipped with other sensors such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, which will not be elaborated upon here.

[0054] Those skilled in the art will understand that Figure 1The network model training device structure shown does not constitute a limitation on the network model training device. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0055] like Figure 1 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a network model training program.

[0056] exist Figure 1 In the network model training device shown, the network interface 1004 is mainly used to connect to the backend server and communicate data with the backend server; the user interface 1003 is mainly used to connect to the client (user end) and communicate data with the client; and the processor 1001 can be used to call the network model training program stored in the memory 1005.

[0057] In this embodiment, the network model training device includes: a memory 1005, a processor 1001, and a network model training program stored in the memory 1005 and executable on the processor 1001. When the processor 1001 calls the network model training program stored in the memory 1005, it executes the steps of the network model training methods in the following embodiments.

[0058] This invention also provides a network model training method, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the network model training method of the present invention.

[0059] In this embodiment, the network model training method includes the following steps:

[0060] Step S100: Obtain the training data and validation data corresponding to the preset message;

[0061] In this embodiment, the preset message can be a preset advertisement, a preset product, a questionnaire, etc., and the verification data can be the click data, exposure data, and viewing data of the preset message. The training data can be the user feature information corresponding to the click data, exposure data, or viewing data. When training the model for the preset message, the training data and verification data corresponding to the preset message are first obtained, and the obtained training data and verification data are used as the training data, i.e., the sample data for subsequent model training. Among them, the user feature data can include the gender, age, occupation, and hobbies (e.g., several network links that the user clicks most frequently, or the network link that the user clicks most frequently) of each user who clicks the preset message.

[0062] For example, taking a preset advertisement as an example, the preset advertisement can be any advertisement placed online (such as on the Internet). After the preset advertisement has been placed for a period of time, that is, when there is exposure and click data corresponding to the preset advertisement, a model can be trained for the preset advertisement to obtain the target prediction model of the preset advertisement. For example, the model can be trained for the preset advertisement when the duration after the preset advertisement is placed reaches a preset duration, or when the exposure and click data corresponding to the preset advertisement is detected to be greater than the preset exposure and click value after the preset advertisement is placed, the model can be trained for the preset advertisement.

[0063] Step S200: Generate multiple candidate subnetworks using a deep neural subnetwork generator, and determine the training error corresponding to each candidate subnetwork based on the training data and validation data.

[0064] In this embodiment, after obtaining the training data and verification data corresponding to the preset message, a deep neural subnetwork generator generates multiple candidate subnetworks. The number of layers in each candidate subnetwork is not limited, that is, the number of network layers in each candidate subnetwork is random. Furthermore, the number of candidate subnetworks generated by the deep neural subnetwork generator each time is also random. Of course, in other embodiments, the number of candidate subnetworks can also be limited. For example, a preset number of candidate subnetworks can be generated each time by the deep neural subnetwork generator.

[0065] Then, based on the training and validation data, the training error corresponding to each candidate sub-network is determined. Specifically, each candidate sub-network is embedded into the main network in sequence, and the training and validation data are input into the main network embedded with the candidate sub-networks to train the model and obtain the training error corresponding to each candidate sub-network.

[0066] It should be noted that the deep neural subnetwork generator can be specified by the algorithm developer, or multiple deep neural subnetwork generators can be pre-set, and one of them can be randomly selected to generate multiple candidate subnetworks.

[0067] Step S300: Based on each of the training errors, determine the optimal subnetwork among the candidate subnetworks, and determine the target prediction model based on the optimal subnetwork and the main network.

[0068] In this embodiment, when the training error is obtained, the optimal subnetwork is determined among the candidate subnetworks based on the training error. For example, the candidate subnetwork with the smallest training error is selected as the optimal subnetwork, or the training error and the number of layers of each candidate subnetwork are weighted and the candidate subnetwork with the smallest weight score after weight calculation is selected as the optimal subnetwork.

[0069] Then, based on the optimal subnetwork and the main network, a target prediction model is determined. If only one search / training is required, the neural network structure obtained by embedding the optimal subnetwork into the main network is the target prediction model. If multiple searches / trainings are required, it is determined whether the number of searches / trainings has reached the specified number. If yes, the neural network structure obtained by embedding the optimal subnetwork into the main network is the target prediction model. If not, the optimal subnetwork is embedded into the main network, and the main network after embedding the optimal subnetwork is used as the new main network. Then, the process returns to step S200. This enables automatic training of the network model through machine learning. For different preset messages, there is no need for manual training and optimization of different network models. Only a small amount of time is required from algorithm engineers to obtain a machine learning model for online preset message delivery, saving a lot of human resources and improving the training efficiency of the network model.

[0070] Furthermore, in one embodiment, step S300 is followed by:

[0071] When the duration after determining the target prediction model reaches the preset duration, return to the steps of obtaining the training data and validation data corresponding to the preset message.

[0072] In this embodiment, when the duration of the target prediction model reaches the preset duration, the execution step S100 is returned to achieve the timed update of the target prediction model. Then, based on automatic machine learning technology, the target prediction model is updated cyclically. The model can be automatically optimized according to changes in user preferences (user feature data) to ensure the delivery effect of the preset message.

[0073] It should be noted that in other embodiments, the verification data corresponding to the preset message can be stored as the first verification data, and the click data, exposure data or viewing data corresponding to the preset message can be counted in real time to obtain the current second verification data. Then, the difference between the second verification data and the first verification data is calculated. When the difference is greater than the preset difference, the execution step S100 is returned to realize the updating of the target prediction model according to the number of clicks (verification data) of the user, so as to ensure the delivery effect of the preset message.

[0074] The network model training method proposed in this embodiment obtains training data and validation data corresponding to a preset message; then, it generates multiple candidate sub-networks through a deep neural sub-network generator, and determines the training error corresponding to each candidate sub-network based on the training data and validation data; then, based on each training error, it determines the optimal sub-network among the candidate sub-networks, and determines the target prediction model based on the optimal sub-network and the main network. This method can achieve automatic training of the network model through machine learning, realize intelligent training and optimization of different network models, and eliminate the need for manual training and optimization of different network models for different message services, saving a lot of human resources, reducing the workload of model training and tuning, and improving the training efficiency of the network model.

[0075] Based on the first embodiment, a second embodiment of the network model training method of the present invention is proposed. In this embodiment, step S200 includes:

[0076] Step S210: Sequentially obtain the target candidate subnetwork in each candidate subnetwork;

[0077] Step S220: Embed the target candidate sub-network into the main network to obtain the updated main network, and input the training data and validation data into the updated main network to obtain the training error corresponding to the target candidate sub-network.

[0078] In this embodiment, after obtaining multiple candidate subnetworks, target candidate subnetworks are obtained sequentially from each candidate subnetwork. The target candidate subnetwork is then embedded into the main network to obtain an updated main network. The training data and validation data are then input into the updated main network to train the model and obtain the training error corresponding to the target candidate subnetwork, thereby accurately obtaining the training error corresponding to each candidate subnetwork.

[0079] It should be noted that when inputting the training data and validation data into the updated main network for model training, multiple model training sessions can be performed. Specifically, the training data and validation data are input into the updated main network, and the training count is updated. If the training count is less than the preset training count, the training data and validation data are input into the updated main network again. The updated main network is the main network that has been trained with the training data and validation data to embed the target candidate subnetwork. If the training count reaches the preset training count, the loss function value of the currently updated main network is used as the training error. The preset training count can be set reasonably.

[0080] In this embodiment, after training with the updated main network, the click probability of each user clicking / viewing the preset message corresponding to the training data can be obtained. Users whose click probability is less than the preset click probability are regarded as users who will not click / view the preset message. The training error is obtained by dividing the number of users who will not click the preset message by the number of verification data, where the number of verification data can be the number of users in the training data.

[0081] The network model training method proposed in this embodiment obtains target candidate subnetworks sequentially from each candidate subnetwork; then, the target candidate subnetworks are embedded into the main network to obtain an updated main network, and the training data and validation data are input into the updated main network to obtain the training error corresponding to the target candidate subnetwork. This method can accurately obtain the training error corresponding to each candidate subnetwork, thereby improving the accuracy of the target prediction model in predicting the preset message, ensuring the delivery effect of the preset message, and further improving the training efficiency of the network model.

[0082] Based on the second embodiment, a third embodiment of the network model training method of the present invention is proposed. In this embodiment, step S220 includes:

[0083] Step S221: Input the training data and validation data into the updated main network, and update the training count;

[0084] Step S222: When the number of training iterations reaches the preset number of training iterations, the training error of the updated main network after training is used as the training error of the target candidate sub-network.

[0085] Step S223: When the number of training iterations has not reached the preset number of training iterations, the updated main network after training is used as the updated main network, and the process returns to the step of inputting the trained data and verification data into the updated main network and updating the number of training iterations.

[0086] In this embodiment, in order to improve the efficiency of model training, the updated main network needs to be trained multiple times. Then, when the training data and validation data are input into the updated main network for training, the number of training iterations is the current number of training iterations of the updated main network.

[0087] Then, it is determined whether the training count has reached the preset number of training counts. If the training count has reached the preset number of training counts, the training error of the updated main network after training is used as the training error of the target candidate sub-network, and the training count is reset. Otherwise, the updated main network after training is used as the updated main network, and the steps of inputting the training data and verification data into the updated main network and updating the training count are returned. This ensures that the updated main network corresponding to each target candidate sub-network is trained for the preset number of times, thereby improving the accuracy of the training error.

[0088] The preset number of training sessions can be set appropriately.

[0089] The network model training method proposed in this embodiment involves inputting the training data and validation data into the updated main network and updating the training iterations. Then, when the training iterations reach a preset number, the training error of the updated main network is used as the training error of the target candidate sub-network. If the training iterations do not reach the preset number, the updated main network is used as the updated main network, and the process returns to inputting the training data and validation data into the updated main network and updating the training iterations. This improves the accuracy of the training errors of each candidate sub-network by increasing the number of training iterations of the updated main network, thereby enhancing the accuracy of the target prediction model in predicting the preset message and improving the training efficiency of the network model.

[0090] Based on the first embodiment, a fourth embodiment of the network model training method of the present invention is proposed. In this embodiment, step S300 includes:

[0091] Step S310: Obtain the penalty coefficient corresponding to the network layer number of each candidate sub-network;

[0092] Step S320: Based on each penalty coefficient and training error, determine the weight score corresponding to each candidate sub-network.

[0093] Step S330: Select the candidate subnetwork with the smallest weight score among all candidate subnetworks as the optimal subnetwork.

[0094] In this embodiment, a preset penalty coefficient corresponding to the number of network layers or a mapping relationship between the number of network layers and the preset penalty coefficient can be preset. After obtaining the training error corresponding to each candidate sub-network, the penalty coefficient corresponding to the number of network layers of each candidate sub-network is determined according to the number of network layers of each candidate sub-network and the mapping relationship.

[0095] Then, based on each penalty coefficient and the training error, the weight score corresponding to each candidate sub-network is determined. Specifically, the weight of the penalty coefficient and the weight of the training error are obtained respectively. The weight score corresponding to each candidate sub-network is determined based on the penalty coefficient, the training error, the weight of the penalty coefficient, and the weight of the training error. For example, the weight score = the weight of the penalty coefficient * the penalty coefficient + the weight of the training error * the training error.

[0096] The network model training method proposed in this embodiment obtains the penalty coefficients corresponding to the network layers of each candidate sub-network; then, based on each penalty coefficient and the training error, determines the weight score corresponding to each candidate sub-network; and then selects the candidate sub-network with the smallest weight score as the optimal sub-network. By comprehensively considering the penalty coefficients and the training error, the determined candidate sub-network is a candidate sub-network with a simple network and a small training error, which improves the accuracy of the target prediction model in predicting the preset message while reducing the complexity of the network model.

[0097] Based on the first embodiment, a fifth embodiment of the network model training method of the present invention is proposed. In this embodiment, the network model training method further includes:

[0098] Step S100: When generating multiple candidate subnetworks through the deep neural subnetwork generator, update the search count;

[0099] Step S300 includes:

[0100] Step S340: Determine whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network;

[0101] Step S350: If the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, then determine whether the number of search attempts has reached the preset number of search attempts.

[0102] Step S360: When the number of searches reaches the preset number of searches, the optimal sub-network is embedded into the main network to obtain the target prediction model.

[0103] In this embodiment, in order to improve the prediction accuracy of the target prediction model, it is necessary to stack multiple optimal subnetworks. Therefore, when generating multiple candidate subnetworks through the deep neural subnetwork generator, the search count is updated. Specifically, the search count is incremented by 1 to obtain the updated search count.

[0104] Then, it is determined whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network. When no sub-network is embedded in the main network, the training error of the main network is greater than the training error corresponding to any candidate sub-network. When a sub-network has been embedded in the main network, the training error of the main network can be obtained based on the previous training data.

[0105] If the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, then it is determined whether the search count has reached the preset search count, so as to determine whether the search / training of the preset search count has been completed. When the search count reaches the preset search count, the optimal sub-network is embedded into the main network to obtain the target prediction model, and the search count is reset at the same time.

[0106] Furthermore, in one embodiment, after step S340, the method further includes:

[0107] If the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, then return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator.

[0108] In this embodiment, if the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, then the update of the current main network is abandoned, and the process returns to the step of generating multiple candidate sub-networks through the deep neural sub-network generator. In this step of generating multiple candidate sub-networks, the search count may or may not be updated; or, if the search count has reached the preset search count, then the search count is not updated when generating multiple candidate sub-networks; if the search count has not reached the preset search count, then the search count is updated.

[0109] Furthermore, in yet another embodiment, after step S350, the method further includes:

[0110] Step S360: If the number of searches has not reached the preset number of searches, then the optimal sub-network is embedded into the main network;

[0111] Step S370: The main network after embedding the optimal subnetwork is taken as the main network, and the process returns to the step of generating multiple candidate subnetworks through the deep neural subnetwork generator.

[0112] In this embodiment, the number of searches has not reached the preset number of searches, and the model training has not reached the specified number of times. Therefore, the optimal subnetwork is embedded into the main network, and the main network after embedding the optimal subnetwork is used as the main network. Then, the step of generating multiple candidate subnetworks through the deep neural subnetwork generator is returned to achieve iterative training.

[0113] The network model training method proposed in this embodiment determines whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network; then, if the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, it determines whether the number of searches has reached a preset number of searches; then, when the number of searches reaches the preset number of searches, the optimal sub-network is embedded into the main network to obtain the target prediction model. By obtaining the target prediction model through multiple searches, the accuracy of the target prediction model in predicting the preset message is further improved, and the training efficiency of the model is enhanced.

[0114] Based on the above embodiments, a sixth embodiment of the network model training method of the present invention is proposed. In this embodiment, the training data is user feature data, the verification data is exposure and click data, and after step S300, the network model training method further includes:

[0115] Step S400: Obtain the user feature data to be predicted, and input the user data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message;

[0116] Step S500: When the click probability is greater than a preset threshold, determine the user corresponding to the user feature data to be predicted as the user to be targeted.

[0117] In this embodiment, after obtaining the target prediction model, the user feature data to be predicted can be obtained at any time, and the user feature data to be predicted can be input into the target prediction model to train the model and obtain the click probability corresponding to the preset message, that is, the click probability of the user feature data to be predicted clicking the preset message.

[0118] Then, it is determined whether the click probability is greater than a preset threshold. If it is, the user corresponding to the user feature data to be predicted is determined to be the user to be delivered to. The preset message can be delivered to the user terminal corresponding to the user feature data to be predicted, or the user terminal information of the user to be delivered to is pushed to the delivery server of the preset message, so that the delivery server delivers the preset message to the user terminal corresponding to the user feature data to be predicted, so as to achieve accurate delivery of the preset message, ensure that the delivered preset message can be clicked and viewed by the user with a high probability, and thus ensure the delivery effect of the preset message.

[0119] For example, when a user is watching a video, the user's feature data can be obtained as the user feature data to be predicted. The model is then used to train the model to obtain the probability that the user will click on the preset message. If the click probability is greater than the preset threshold, the preset message will be pushed to the user's terminal when the user finishes watching the video.

[0120] The network model training method proposed in this embodiment obtains the user feature data to be predicted and inputs the user feature data to be predicted into the target prediction model for model training to obtain the click probability corresponding to the preset message; then, when the click probability is greater than a preset threshold, the user corresponding to the user feature data to be predicted is determined as the user to be delivered to. By accurately predicting the click probability of the user, the preset message is delivered, thereby improving the delivery effect of the preset message.

[0121] This invention also provides a network model training device, referring to... Figure 3 The network model training device includes:

[0122] The acquisition module 100 is used to acquire the training data and validation data corresponding to the preset message;

[0123] The generation module 200 is used to generate multiple candidate subnetworks through a deep neural subnetwork generator, and determine the training error corresponding to each candidate subnetwork based on the training data and validation data.

[0124] The determination module 300 is used to determine the optimal subnetwork among the candidate subnetworks based on the training errors, and then determine the target prediction model based on the optimal subnetwork and the main network.

[0125] Optionally, the generation module 200 is also used for:

[0126] The target candidate subnetwork is obtained sequentially from each candidate subnetwork.

[0127] The target candidate subnetwork is embedded into the main network to obtain the updated main network, and the training data and validation data are input into the updated main network to obtain the training error corresponding to the target candidate subnetwork.

[0128] Optionally, the generation module 200 is also used for:

[0129] The training data and validation data are input into the updated main network, and the number of training iterations is updated.

[0130] When the number of training iterations reaches the preset number of training iterations, the training error of the updated main network after training is used as the training error of the target candidate sub-network.

[0131] If the number of training iterations does not reach the preset number of training iterations, the updated main network after training is used as the updated main network, and the process returns to the step of inputting the training data and validation data into the updated main network and updating the number of training iterations.

[0132] Optionally, the determining module 300 is also used for:

[0133] Obtain the penalty coefficient corresponding to the network layer number of each candidate subnetwork;

[0134] Based on each penalty coefficient and training error, the weight score corresponding to each candidate subnetwork is determined respectively;

[0135] The candidate subnetwork with the smallest weight score among all candidate subnetworks is selected as the optimal subnetwork.

[0136] Optionally, the network model training device also includes:

[0137] Update the search count when generating multiple candidate subnetworks using the deep neural subnetwork generator;

[0138] The determination module 300 is also used for:

[0139] Determine whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network;

[0140] If the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, then it is determined whether the number of search attempts has reached the preset number of search attempts.

[0141] When the number of searches reaches the preset number of searches, the optimal subnetwork is embedded into the main network to obtain the target prediction model.

[0142] Optionally, the determining module 300 is also used for:

[0143] If the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, then return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator.

[0144] Optionally, the determining module 300 is also used for:

[0145] If the number of searches does not reach the preset number of searches, the optimal subnetwork is embedded into the main network;

[0146] The main network after embedding the optimal subnetwork is used as the main network, and the process returns to the step of generating multiple candidate subnetworks through the deep neural subnetwork generator.

[0147] Optionally, the network model training device also includes:

[0148] When the duration after determining the target prediction model reaches the preset duration, return to the steps of obtaining the training data and validation data corresponding to the preset message.

[0149] Optionally, the network model training device also includes:

[0150] Obtain the user feature data to be predicted, and input the user data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message;

[0151] When the click probability is greater than a preset threshold, the user corresponding to the user feature data to be predicted is determined as the user to be targeted.

[0152] The methods executed by the above-mentioned program modules can be referred to in the various embodiments of the network model training method of the present invention, and will not be repeated here.

[0153] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a network model training program, which, when executed by a processor, implements the steps of the network model training method described above.

[0154] The method implemented when the network model training program running on the processor is executed can be referred to in various embodiments of the network model training method of the present invention, and will not be repeated here.

[0155] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0156] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0157] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0158] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A network model training method, characterized in that, The network model training method includes the following steps: Obtain training data and verification data corresponding to a preset message, wherein the training data is user feature data and the verification data is exposure and click data; Multiple candidate subnetworks are generated using a deep neural subnetwork generator. Based on the training data and validation data, the training error of each candidate subnetwork is determined. The candidate subnetwork with the smallest training error is selected as the target candidate subnetwork. The target candidate subnetwork is embedded into the main network to obtain an updated main network. The training data and validation data are then input into the updated main network to obtain the training error corresponding to the target candidate subnetwork. Based on the training error corresponding to each of the target candidate sub-networks, the optimal sub-network is determined among the target candidate sub-networks, and the target prediction model is determined based on the optimal sub-network and the main network. Obtain the user feature data to be predicted, and input the user feature data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message; When the click probability is greater than a preset threshold, the user corresponding to the user feature data to be predicted is determined as the user to be targeted.

2. The network model training method as described in claim 1, characterized in that, The step of inputting the training data and validation data into the updated main network to obtain the training error corresponding to the target candidate sub-network includes: The training data and validation data are input into the updated main network, and the number of training iterations is updated. When the number of training iterations reaches the preset number of training iterations, the training error of the updated main network after training is used as the training error of the target candidate sub-network. If the number of training iterations does not reach the preset number of training iterations, the updated main network after training is used as the updated main network, and the process returns to the step of inputting the training data and validation data into the updated main network and updating the number of training iterations.

3. The network model training method as described in claim 1, characterized in that, The step of determining the optimal subnetwork among the candidate subnetworks based on each of the training errors includes: Obtain the penalty coefficient corresponding to the network layer number of each candidate subnetwork; Based on each penalty coefficient and training error, the weight score corresponding to each candidate subnetwork is determined respectively; The candidate subnetwork with the smallest weight score among all candidate subnetworks is selected as the optimal subnetwork. The step of determining the weight score corresponding to each candidate sub-network based on each penalty coefficient and training error includes: Obtain the weights of the penalty coefficient and the training error respectively. Based on the penalty coefficient, the training error, the weights of the penalty coefficient and the training error, determine the weight score corresponding to each candidate subnetwork, including: weight score = weight of penalty coefficient * penalty coefficient + weight of training error * training error.

4. The network model training method as described in claim 1, characterized in that, The network model training method also includes: Update the search count when generating multiple candidate subnetworks using the deep neural subnetwork generator; The step of determining the target prediction model based on the optimal sub-network and the main network includes: Determine whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network; If the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, then it is determined whether the number of search attempts has reached the preset number of search attempts. When the number of searches reaches the preset number of searches, the optimal subnetwork is embedded into the main network to obtain the target prediction model.

5. The network model training method as described in claim 4, characterized in that, After determining whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, the method further includes: If the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, then return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator.

6. The network model training method as described in claim 4, characterized in that, After determining whether the number of searches has reached the preset number of searches, the method further includes: If the number of searches does not reach the preset number of searches, the optimal subnetwork is embedded into the main network; The main network after embedding the optimal subnetwork is used as the main network, and the process returns to the step of generating multiple candidate subnetworks through the deep neural subnetwork generator.

7. The network model training method as described in claim 1, characterized in that, After the step of determining the target prediction model based on the optimal sub-network and the main network, the network model training method further includes: When the duration after determining the target prediction model reaches the preset duration, return to the steps of obtaining the training data and validation data corresponding to the preset message.

8. A network model training device, characterized in that, The network model training device includes: The acquisition module is used to acquire training data and verification data corresponding to a preset message, wherein the training data is user feature data and the verification data is exposure and click data; The generation module is used to generate multiple candidate subnetworks through a deep neural subnetwork generator, and determine the training error corresponding to each candidate subnetwork based on the training data and validation data. The candidate subnetwork with the smallest training error is selected as the target candidate subnetwork. The target candidate subnetwork is embedded into the main network to obtain an updated main network, and the training data and validation data are input into the updated main network to obtain the training error corresponding to the target candidate subnetwork. The determination module is used to determine the optimal subnetwork among the target candidate subnetworks based on the training error corresponding to each target candidate subnetwork, and then determine the target prediction model based on the optimal subnetwork and the main network; obtain the user feature data to be predicted, and input the user feature data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message; when the click probability is greater than the preset threshold, determine the user corresponding to the user feature data to be predicted as the user to be delivered.

9. A network model training device, characterized in that, The network model training device includes: a memory, a processor, and a network model training program stored in the memory and executable on the processor. When the network model training program is executed by the processor, it implements the steps of the network model training method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a network model training program, which, when executed by a processor, implements the steps of the network model training method as described in any one of claims 1 to 7.

Images