Data processing method and apparatus, storage medium, and device

Abstract

Embodiments of the present application disclose a data processing method and device, a storage medium and equipment. The method comprises: obtaining N candidate recommended objects in a sample recommended object set, object features corresponding to each candidate recommended object, and user features corresponding to a recommended user, inputting the features into an initial prediction model, and predicting a target recommended object by using the initial prediction model; determining candidate model parameters corresponding to the initial prediction model and an adjustment degree corresponding to the candidate model parameters according to a user feedback label corresponding to the target recommended object and predicted recommendation features corresponding to the target recommended object; determining target model parameters corresponding to the initial prediction model according to the candidate model parameters and the adjustment degree; and performing parameter updating on the initial prediction model according to the target model parameters, and determining a target prediction model by using the initial prediction model that meets a convergence condition. By using the present application, the prediction accuracy of object recommendation by the target prediction model can be improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a data processing method, apparatus, storage medium and device. Background Technology

[0002] With the development of internet technology, various online content interaction behaviors, such as reading text content (e.g., books, WeChat articles), listening to music, watching videos, and online shopping, have gradually become part of people's daily lives. Therefore, it is possible to recommend objects (e.g., advertisements, books, articles, or videos) to users to promote relevant objects.

[0003] Currently, neural network models can predict which items to recommend to users. However, in training these models, the process involves selecting the item with the highest expected return based on the model's predictions and then training the model based on user feedback. This method, which focuses solely on the item with the highest expected return, may repeatedly recommend the same item to the user without exploring other options. Consequently, it fails to accurately identify items that the user is truly interested in, resulting in a low accuracy rate in predicting recommended items. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of this application is to provide a data processing method, apparatus, storage medium and device that can improve the prediction accuracy of the target prediction model for object recommendation.

[0005] One embodiment of this application provides a data processing method, including:

[0006] Obtain the sample recommendation object set; the sample recommendation object set includes N candidate recommendation objects, where N is a positive integer;

[0007] The object features corresponding to each of the N candidate recommendation objects, and the user features corresponding to the recommended user, are input into the initial prediction model. The initial prediction model then predicts the target recommendation object to be recommended to the recommended user. The N candidate recommendation objects include the target recommendation object.

[0008] Based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object, determine the candidate model parameters and the adjustment degree corresponding to the candidate model parameters for the initial prediction model; the adjustment degree has randomness.

[0009] Based on the candidate model parameters and the adjustment degree, determine the parameter random sampling interval corresponding to the initial prediction model, and determine the target model parameters corresponding to the initial prediction model from the parameter random sampling interval;

[0010] Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

[0011] Specifically, the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the recommended user are input into the initial prediction model. The initial prediction model then predicts the target recommendation objects to be recommended to the recommended user, including:

[0012] Obtain the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the target user;

[0013] The N object features are combined with the user features to obtain the combined features corresponding to each candidate recommendation object;

[0014] Input the combined features corresponding to each candidate recommendation object into the initial prediction model, and output the predicted recommendation probability corresponding to each candidate recommendation object through the initial prediction model;

[0015] The candidate recommendation object corresponding to the highest predicted recommendation probability is determined as the target recommendation object.

[0016] The process involves inputting the combined features corresponding to each candidate recommendation object into the initial prediction model, and then outputting the predicted recommendation probability for each candidate recommendation object, including:

[0017] The combined features corresponding to each candidate recommendation object are input into the initial prediction model. The combined features corresponding to each candidate recommendation object are encoded through the embedding layer in the initial prediction model to obtain the embedded feature vector corresponding to each candidate recommendation object.

[0018] Based on the connection layer in the initial prediction model, the embedded feature vectors corresponding to each candidate recommendation object are concatenated to obtain the concatenated feature vector.

[0019] The concatenated feature vector is input into the fully connected hidden layer in the initial prediction model, and the predicted recommendation features corresponding to N candidate recommendation objects are output through the fully connected hidden layer.

[0020] Based on the classifier in the initial prediction model, the predicted recommendation features corresponding to N candidate recommendation objects are classified to obtain the predicted recommendation probability for each candidate recommendation object.

[0021] The fully connected hidden layer includes a first fully connected layer, a second fully connected layer, and an activation layer.

[0022] Based on the fully connected hidden layer in the initial prediction model, the predicted recommendation features corresponding to N candidate recommendation objects are output through the fully connected hidden layer, including:

[0023] The concatenated feature vector is input into the first fully connected layer. The concatenated feature vector is then convolved with the weight matrix of the first fully connected layer to obtain the first recommended feature vector.

[0024] The first recommended feature vector is input into the second fully connected layer. The first recommended feature vector is then convolved with the weight matrix of the second fully connected layer to obtain the second recommended feature vector.

[0025] The second recommendation feature vector is input into the activation layer, and the activation layer activates the second recommendation feature vector to obtain the predicted recommendation features corresponding to N candidate recommendation objects.

[0026] The data processing methods also include:

[0027] Send the target recommendation object to the user terminal so that the user terminal can obtain user conversion feedback information for the target recommendation object;

[0028] Receive user conversion feedback information returned by the user terminal, and generate user feedback tags corresponding to the target recommendation object based on the user conversion feedback information.

[0029] Specifically, based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object, the candidate model parameters corresponding to the initial prediction model and the adjustment degree corresponding to the candidate model parameters are determined, including:

[0030] Based on user feedback tags and the predicted recommendation features corresponding to the target recommendation object, determine the loss function corresponding to the initial prediction model;

[0031] Based on the loss function, candidate model parameters are determined for adjusting the initial model parameters in the initial prediction model;

[0032] The adjustment degree corresponding to the candidate model parameters is determined based on the candidate model parameters and the initial model parameters in the initial prediction model.

[0033] Specifically, the adjustment degree corresponding to the candidate model parameters is determined based on the candidate model parameters and the initial model parameters in the initial prediction model, including:

[0034] Randomly select the target update coefficient from the update coefficient range;

[0035] Obtain the first product of the target update coefficients and the initial model parameters in the initial prediction model;

[0036] Based on the first product and the candidate model parameters, determine the degree of adjustment corresponding to the candidate model parameters.

[0037] Specifically, the adjustment degree corresponding to the candidate model parameters is determined based on the first product and the candidate model parameters, including:

[0038] Based on the first product, the candidate model parameters are adjusted to obtain the updated candidate model parameters;

[0039] Obtain the parameter difference between the updated candidate model parameters and the original candidate model parameters;

[0040] The degree of adjustment corresponding to the parameters of the candidate model is determined based on the parameter difference.

[0041] Among them, determining the degree of adjustment corresponding to the candidate model parameters based on the parameter difference includes:

[0042] The target randomness coefficient is randomly obtained from the randomness coefficient range;

[0043] Obtain the second product between the target randomness coefficient and the parameter difference, and determine the second product as the adjustment degree corresponding to the candidate model parameters.

[0044] Specifically, based on the candidate model parameters and the adjustment degree, a random sampling interval for the parameters corresponding to the initial prediction model is determined. From this random sampling interval, the target model parameters corresponding to the initial prediction model are determined, including:

[0045] The candidate model parameters are determined as the parameter mean, the adjustment degree is determined as the standard deviation, and the normal distribution associated with the candidate model parameters is determined based on the parameter mean and standard deviation. The normal distribution is then determined as the random sampling interval of the parameters corresponding to the initial prediction model.

[0046] Random sampling is performed within the parameter random sampling interval to obtain the target model parameters corresponding to the initial prediction model.

[0047] Specifically, based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model, including:

[0048] Based on the target model parameters, the initial model parameters in the initial prediction model are updated to obtain the parameter-updated initial prediction model;

[0049] If the adjustment degree corresponding to the initial prediction model after parameter update is the target adjustment threshold, then the initial prediction model after parameter update is determined to meet the convergence condition, and the initial prediction model after parameter update that meets the convergence condition is determined as the target prediction model.

[0050] The data processing methods also include:

[0051] Obtain the target user characteristics corresponding to the target user, and obtain the object characteristics corresponding to each object to be recommended in the set of objects to be recommended;

[0052] Input the target user features and the object features corresponding to each object to be recommended into the target prediction model, and output the recommendation probability corresponding to each object to be recommended through the target prediction model;

[0053] Based on the recommendation probability corresponding to each object to be recommended, a user recommendation object is determined from the set of objects to be recommended to the target user, and the user recommendation object is sent to the target user terminal to which the target user belongs, so that the target user terminal outputs the user recommendation object.

[0054] One embodiment of this application provides a data processing apparatus, including:

[0055] The first acquisition module is used to acquire a sample recommendation object set; the sample recommendation object set includes N candidate recommendation objects, where N is a positive integer;

[0056] The prediction module is used to input the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the recommended user into the initial prediction model, and predict the target recommendation object to be recommended to the recommended user through the initial prediction model; the N candidate recommendation objects include the target recommendation object;

[0057] The first determining module is used to determine the candidate model parameters and the adjustment degree corresponding to the candidate model parameters based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object; the adjustment degree has randomness.

[0058] The second determining module is used to determine the random sampling interval of the parameters corresponding to the initial prediction model based on the candidate model parameters and the adjustment degree, and to determine the target model parameters corresponding to the initial prediction model from the random sampling interval of the parameters.

[0059] The parameter update module is used to update the parameters of the initial prediction model based on the target model parameters. When the initial prediction model after parameter update meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

[0060] The prediction module includes:

[0061] The first acquisition unit is used to acquire the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the target user.

[0062] The combination unit is used to combine N object features with user features to obtain the combined features corresponding to each candidate recommendation object;

[0063] The output unit is used to input the combined features corresponding to each candidate recommendation object into the initial prediction model, and output the predicted recommendation probability corresponding to each candidate recommendation object through the initial prediction model.

[0064] The first determining unit is used to determine the candidate recommendation object corresponding to the highest predicted recommendation probability as the target recommendation object.

[0065] The output unit includes:

[0066] The feature encoding subunit is used to input the combined features corresponding to each candidate recommendation object into the initial prediction model. Through the embedding layer in the initial prediction model, the combined features corresponding to each candidate recommendation object are encoded to obtain the embedded feature vector corresponding to each candidate recommendation object.

[0067] The concatenation processing subunit is used to concatenate the embedded feature vectors corresponding to each candidate recommendation object based on the connection layer in the initial prediction model, so as to obtain the concatenated feature vector.

[0068] The output sub-unit is used to input the concatenated feature vector into the fully connected hidden layer in the initial prediction model, and output the predicted recommendation features corresponding to N candidate recommendation objects through the fully connected hidden layer.

[0069] The classification processing subunit is used to classify the predicted recommendation features corresponding to N candidate recommendation objects based on the classifier in the initial prediction model, and obtain the predicted recommendation probability corresponding to each candidate recommendation object.

[0070] The fully connected hidden layer includes a first fully connected layer, a second fully connected layer, and an activation layer.

[0071] The output subunit is specifically used for:

[0072] The concatenated feature vector is input into the first fully connected layer. The concatenated feature vector is then convolved with the weight matrix of the first fully connected layer to obtain the first recommended feature vector.

[0073] The first recommended feature vector is input into the second fully connected layer. The first recommended feature vector is then convolved with the weight matrix of the second fully connected layer to obtain the second recommended feature vector.

[0074] The second recommendation feature vector is input into the activation layer, and the activation layer activates the second recommendation feature vector to obtain the predicted recommendation features corresponding to N candidate recommendation objects.

[0075] The data processing device also includes:

[0076] The sending module is used to send the target recommendation object to the user terminal so that the user terminal can obtain user conversion feedback information for the target recommendation object;

[0077] The receiving module is used to receive user conversion feedback information returned by the user terminal and generate user feedback tags corresponding to the target recommendation object based on the user conversion feedback information.

[0078] The first determining module includes:

[0079] The second determining unit is used to determine the loss function corresponding to the initial prediction model based on the user feedback tags and the predicted recommendation features corresponding to the target recommendation object.

[0080] The third determining unit is used to determine, based on the loss function, candidate model parameters for adjusting the initial model parameters in the initial prediction model;

[0081] The fourth determining unit is used to determine the adjustment degree corresponding to the candidate model parameters based on the candidate model parameters and the initial model parameters in the initial prediction model.

[0082] The fourth determining unit includes:

[0083] Select sub-units to randomly select target update coefficients from the update coefficient range;

[0084] Obtain the sub-unit, which is used to obtain the first product of the target update coefficient and the initial model parameters in the initial prediction model;

[0085] The sub-unit is determined based on the first product and the candidate model parameters to determine the adjustment degree corresponding to the candidate model parameters.

[0086] Specifically, determining the sub-unit is used for:

[0087] Based on the first product, the candidate model parameters are adjusted to obtain the updated candidate model parameters;

[0088] Obtain the parameter difference between the updated candidate model parameters and the original candidate model parameters;

[0089] The degree of adjustment corresponding to the parameters of the candidate model is determined based on the parameter difference.

[0090] Specifically, determining the subunit is also used for:

[0091] The target randomness coefficient is randomly obtained from the randomness coefficient range;

[0092] Obtain the second product between the target randomness coefficient and the parameter difference, and determine the second product as the adjustment degree corresponding to the candidate model parameters.

[0093] The second determining module includes:

[0094] The fifth determining unit is used to determine the candidate model parameters as the parameter mean, the adjustment degree as the standard deviation, and to determine the normal distribution associated with the candidate model parameters based on the parameter mean and standard deviation, and to determine the normal distribution as the random sampling interval of the parameters corresponding to the initial prediction model.

[0095] The random sampling unit is used to perform random sampling within the parameter random sampling interval to obtain the target model parameters corresponding to the initial prediction model.

[0096] The parameter update module includes:

[0097] The parameter update unit is used to update the initial model parameters in the initial prediction model according to the target model parameters, so as to obtain the parameter-updated initial prediction model.

[0098] The sixth determining unit is used to determine if the adjustment degree corresponding to the initial prediction model after parameter update is the target adjustment threshold, and to determine the initial prediction model after parameter update that satisfies the convergence condition as the target prediction model.

[0099] The data processing device also includes:

[0100] The second acquisition module is used to acquire the target user characteristics corresponding to the target user and to acquire the object characteristics corresponding to each object to be recommended in the set of objects to be recommended.

[0101] The output module is used to input the target user features and the object features corresponding to each object to be recommended into the target prediction model, and output the recommendation probability corresponding to each object to be recommended through the target prediction model.

[0102] The third determining module is used to determine the user recommendation object to be recommended to the target user from the set of objects to be recommended based on the recommendation probability corresponding to each object to be recommended, and send the user recommendation object to the target user terminal to which the target user belongs, so that the target user terminal outputs the user recommendation object.

[0103] This application provides a computer device, including: a processor and a memory;

[0104] The memory is used to store computer programs, and the processor is used to call the computer programs to perform the following steps:

[0105] Obtain the sample recommendation object set; the sample recommendation object set includes N candidate recommendation objects, where N is a positive integer;

[0106] The object features corresponding to each of the N candidate recommendation objects, and the user features corresponding to the recommended user, are input into the initial prediction model. The initial prediction model then predicts the target recommendation object to be recommended to the recommended user. The N candidate recommendation objects include the target recommendation object.

[0107] Based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object, determine the candidate model parameters and the adjustment degree corresponding to the candidate model parameters for the initial prediction model; the adjustment degree has randomness.

[0108] Based on the candidate model parameters and the adjustment degree, determine the parameter random sampling interval corresponding to the initial prediction model, and determine the target model parameters corresponding to the initial prediction model from the parameter random sampling interval;

[0109] Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

[0110] One embodiment of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the following steps:

[0111] Obtain the sample recommendation object set; the sample recommendation object set includes N candidate recommendation objects, where N is a positive integer;

[0112] The object features corresponding to each of the N candidate recommendation objects, and the user features corresponding to the recommended user, are input into the initial prediction model. The initial prediction model then predicts the target recommendation object to be recommended to the recommended user. The N candidate recommendation objects include the target recommendation object.

[0113] Based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object, determine the candidate model parameters and the adjustment degree corresponding to the candidate model parameters for the initial prediction model; the adjustment degree has randomness.

[0114] Based on the candidate model parameters and the adjustment degree, determine the parameter random sampling interval corresponding to the initial prediction model, and determine the target model parameters corresponding to the initial prediction model from the parameter random sampling interval;

[0115] Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

[0116] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method described in the preceding aspect.

[0117] In this embodiment, a sample recommendation object set is obtained. The object features corresponding to each of the N candidate recommendation objects in the sample recommendation object set, along with the user features corresponding to the recommended user, are input into an initial prediction model. The initial prediction model then predicts the target recommendation object to be recommended to the recommended user. Based on the user feedback tags and predicted recommendation features corresponding to the target recommendation object, candidate model parameters and their adjustment degrees are determined for the initial prediction model. A random sampling interval for the parameters of the initial prediction model is then determined based on the candidate model parameters and their adjustment degrees. The target model parameters are then determined from this random sampling interval. Since the candidate model parameters for the initial prediction model have uncertainty, and this uncertainty is related to the uncertainty of the prediction results output by the initial prediction model (i.e., different target model parameters update the initial prediction model, resulting in different prediction results when the updated initial prediction model predicts recommendation objects), the target model parameters can be generated based on the candidate model parameters and their adjustment degrees, thus increasing the randomness of the target model parameters. Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, it is determined as the target prediction model. The uncertainty of the candidate model parameters guides the generation of target model parameters with uncertainty. This increases the randomness of the target model parameters, allowing for targeted exploration of recommended objects that users may be interested in. It also ensures that the initial prediction model converges in a precise convergence direction, improving the accuracy of the target prediction model in recommending objects. Attached Figure Description

[0118] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0119] Figure 1 This is a schematic diagram of the architecture of a data processing system provided in an embodiment of this application;

[0120] Figure 2 This is a schematic diagram illustrating an application scenario of data processing provided in an embodiment of this application;

[0121] Figure 3 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0122] Figure 4 This is a schematic diagram of the network structure of an initial prediction model provided in an embodiment of this application;

[0123] Figure 5 This is a schematic diagram of an initial prediction model training provided in an embodiment of this application;

[0124] Figure 6 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0125] Figure 7 This is a schematic diagram of a random sampling interval for parameters provided in an embodiment of this application;

[0126] Figure 8 This is a schematic diagram illustrating how to update the parameters of an initial prediction model based on the target model parameters, as provided in an embodiment of this application.

[0127] Figure 9 This is an application scenario diagram of book recommendation provided in an embodiment of this application;

[0128] Figure 10 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0129] Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0130] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0131] See Figure 1 , Figure 1 This is a schematic diagram of the structure of a data processing system provided in an embodiment of this application. For example... Figure 1 As shown, the data processing system may include server 10 and a user terminal cluster. The user terminal cluster may include one or more user terminals; the number of user terminals is not limited here. Figure 1 As shown, it can specifically include user terminal 100a, user terminal 100b, user terminal 100c, ..., user terminal 100n. Figure 1 As shown, user terminals 100a, 100b, 100c, ..., 100n can each connect to the server 10 via a network, so that each user terminal can interact with the server 10 through the network connection.

[0132] Each user terminal in this user terminal cluster can include: smartphones, tablets, laptops, desktop computers, wearable devices, smart home devices, head-mounted devices, and other intelligent terminals with trajectory recognition capabilities. It should be understood that, for example... Figure 1 Each user terminal in the user terminal cluster shown can have the target application (i.e., the application client) installed. When the application client runs on each user terminal, it can interact with the aforementioned... Figure 1 Data interaction occurs between the servers 10 shown.

[0133] Among them, such as Figure 1 As shown, the server 10 can be an independent physical server, 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 communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0134] For ease of understanding, the embodiments of this application may be described in detail below. Figure 1 From the plurality of user terminals shown, one user terminal is selected as the target user terminal. The target user terminal may include: a smartphone, tablet computer, laptop computer, desktop computer, smart TV, or other smart terminal with data processing capabilities. For example, for ease of understanding, embodiments of this application may... Figure 1The user terminal 100a shown serves as the target user terminal. User terminal 100a can send the user features corresponding to the user (i.e., the recommended user) to server 10. Server 10 can receive the user features corresponding to the recommended object sent by target user terminal 100a, and obtain N candidate recommended objects from the sample recommended object set. It inputs the object features corresponding to each of the N candidate recommended objects, along with the user features corresponding to the recommended user, into an initial prediction model. The initial prediction model then predicts the target recommended object for recommendation to the recommended user. After obtaining the target recommended object, server 10 can send it to target user terminal 100a, enabling target user terminal 100a to output the target recommended object to the recommended user and obtain user conversion feedback information (such as click feedback or purchase feedback) from the recommended user regarding the target recommended object. Target user terminal 100a can send user conversion feedback information to server 10. After receiving the user conversion feedback information sent by target user terminal, server 10 can generate user feedback tags for the target recommended object based on the user conversion feedback information. Server 10 can determine the random sampling interval of parameters corresponding to the initial prediction model based on the user feedback tags corresponding to the target recommendation object and the prediction recommendation features corresponding to the target recommendation object. Server 10 can obtain the target model parameters corresponding to the initial prediction model from the parameter sampling interval, update the parameters of the initial prediction model based on the target model parameters, and determine the target prediction model based on the initial prediction model that meets the convergence condition when the updated initial prediction model meets the convergence condition. The target prediction model is used to predict recommendation models for users.

[0135] For example, embodiments of this application can... Figure 1The user terminal 100a shown serves as the target user terminal. After the server 10 trains the target prediction model, the user terminal 100a can send the target user features (i.e., the user currently requiring recommendation object prediction) and the object features corresponding to each object in the set of objects to be recommended to the server 10. Upon receiving the target user features and the object features corresponding to each object to be recommended, the server 10 can input these features into the target prediction model, output the recommendation probability for each object to be recommended, and determine the recommended user object from the set of objects to be recommended to the target user based on this probability. The target prediction model can be pre-trained by the server 10. After obtaining the recommended user object, the server 10 can send it to the target user terminal 100a. The target user terminal 100a can receive the recommended user object returned by the server 10 and output the recommended user object to the target user. Through the target prediction model in this scheme, recommended user objects can be more accurately recommended to the target user, improving the prediction accuracy of object recommendations.

[0136] like Figure 2 As shown, Figure 2 This is a schematic diagram illustrating a data processing application scenario provided in an embodiment of this application, such as... Figure 2 As shown, the objects to be recommended can refer to product advertisements, and the set of objects to be recommended can refer to a set of product advertisements corresponding to multiple products. If we need to deliver product advertisements to user A, then user A can be referred to as the target user. If we need to display product advertisements on the browsing interface of user terminal 1, that is, deliver product advertisements to user A, as shown... Figure 2The user terminal 1 shown can identify user A, who currently needs product ad recommendations, as the target user, and obtain the target user's corresponding target user features. It then uploads these target user features, along with the product ad features corresponding to each product ad in the set of recommended product ads, to server 1. Server 1 receives the target user features and the product ad features corresponding to each product ad in the set of recommended product ads uploaded by user terminal 1. It combines each ad feature with the target user features to obtain a combined feature for each product ad. A combined feature is obtained by combining a product ad with the target user features: <product ad feature, target user feature>. Server 1 can input the combined features corresponding to each product ad into a pre-trained target prediction model. This model extracts features from the combined features corresponding to each product ad, obtaining the recommendation probability for each product ad. Based on the recommendation probability, server 1 determines the target product ad from the set of recommended product ads to recommend to the target user and sends this target product ad to the user terminal 1 belonging to the target user. User terminal 1 can receive the target product advertisement sent by server 1 and output the target product advertisement to the target user, that is, display the target product advertisement in the browsing interface of user terminal 1.

[0137] Please see Figure 3 , Figure 3 This is a flowchart illustrating a data processing method provided in an embodiment of this application. This data processing method can be executed by a computer device, which can be a server (as described above). Figure 1 Server 10 in the middle), or user terminal (as mentioned above) Figure 1 This application does not limit the scope to any user terminal in a user terminal cluster, or a system consisting of a server and user terminals. Figure 2 As shown, the data processing method may include steps S101-S105.

[0138] S101, Obtain the sample recommendation object set; the sample recommendation object set includes N candidate recommendation objects.

[0139] Specifically, the computer device can train an initial prediction model to obtain a target prediction model for predicting recommended objects for the user. This target prediction model is used to predict recommended objects, such as advertisements, products, videos, music, and books. The computer device can obtain a sample set of recommended objects, which includes N candidate recommended objects, where N is a positive integer (e.g., N can take values ​​of 1, 2, 3, ...). This sample set of recommended objects is used to train the initial prediction model to obtain the target prediction model. For example, the sample set of recommended objects could be a book library, and the candidate recommended objects could be books from that library.

[0140] S102, input the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the recommended user into the initial prediction model, and predict the target recommendation object to be recommended to the recommended user through the initial prediction model.

[0141] Specifically, the computer device can acquire the object features corresponding to each of the N candidate recommendation objects. These object features can refer to characteristics such as the category, location, price, and usage permissions of the recommended object. For example, when the recommended object is a product advertisement, the object features corresponding to the product advertisement can refer to the product category, product price, and product location, etc. When the recommended object is music, the object features corresponding to the music can refer to the music category and music location, etc. The computer device can also acquire the user features corresponding to the recommended user. The recommended user can refer to the user to whom the recommended object is pushed. For example, if a cosmetic product advertisement needs to be recommended to user A, then user A can be called the recommended user. The user features corresponding to the recommended user can refer to information such as the user's age, user gender, and user location, etc. The computer device can input the object features corresponding to each of the N candidate recommendation objects, as well as the user features corresponding to the recommended user, into the initial prediction model. The initial prediction model predicts the target recommendation object to be recommended to the recommended user through the initial prediction model. The N candidate recommendation objects include the target recommendation object. That is, the initial prediction model can predict the target recommendation object to be recommended to the recommended user from the N candidate recommendation objects based on the object features of the recommended object and the user features corresponding to the recommended user. The number of target recommended objects can be one or more, and this embodiment of the application does not impose any restrictions.

[0142] Optionally, the specific method by which the computer device predicts the target recommendation object for recommendation to the recommended user through the initial prediction model may include: obtaining the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the target user; combining the N object features with the user features to obtain the combined features corresponding to each candidate recommendation object; inputting the combined features corresponding to each candidate recommendation object into the initial prediction model; outputting the predicted recommendation probability corresponding to each candidate recommendation object through the initial prediction model; and determining the candidate recommendation object corresponding to the highest predicted recommendation probability as the target recommendation object.

[0143] Specifically, the computer device can acquire the object features corresponding to each candidate recommendation object and the user features corresponding to the recommended user. The N object features corresponding to N candidate recommendation objects are combined with the user features to obtain combined features for each candidate recommendation object. Similarly, the object features of a candidate recommendation object are combined with the user features of the recommended user to obtain a combined feature. For example, combining object feature 1 of candidate recommendation object 1 with the user features yields the combined feature <object feature 1, user feature>, and combining object feature 2 of candidate recommendation object 2 with the user features yields the combined feature <object feature 2, user feature>. The combined features corresponding to each candidate recommendation object are input into the initial prediction model. The initial prediction model performs feature analysis on the combined features corresponding to each candidate recommendation object and outputs the predicted recommendation probability for each candidate recommendation object. This predicted recommendation probability refers to the probability that the corresponding candidate recommendation object will be pushed to the recommended user or the degree of interest the recommended user has in the candidate recommendation object. The higher the predicted recommendation probability, the greater the likelihood of being pushed to the recommended user or the greater the degree of interest the recommended user has in the object; conversely, the lower the predicted recommendation probability, the less likely the object will be pushed to the recommended user or the less interest the recommended user has in the object. After the computer device obtains the predicted recommendation probability for each candidate recommendation object, when the number of target recommendation objects is 1, the candidate recommendation object with the highest predicted recommendation probability can be determined as the target recommendation object. When the number of target recommendation objects is multiple, the candidate recommendation object with the highest predicted recommendation probability can be determined as the target recommendation object.

[0144] Optionally, the specific method by which the computer device outputs the predicted recommendation probability for each candidate recommendation object through the initial prediction model may include: inputting the combined features corresponding to each candidate recommendation object into the initial prediction model; encoding the combined features corresponding to each candidate recommendation object through the embedding layer in the initial prediction model to obtain the embedded feature vector corresponding to each candidate recommendation object; concatenating the embedded feature vectors corresponding to each candidate recommendation object based on the connection layer in the initial prediction model to obtain the concatenated feature vector; inputting the concatenated feature vector into the fully connected hidden layer in the initial prediction model; and outputting the predicted recommendation features corresponding to N candidate recommendation objects through the fully connected hidden layer. Finally, classifying the predicted recommendation features corresponding to the N candidate recommendation objects based on the classifier in the initial prediction model to obtain the predicted recommendation probability corresponding to each candidate recommendation object.

[0145] Specifically, the initial prediction model can include an embedding layer, a connection layer, a fully connected hidden layer, and a classifier. The computer device can output the combined features corresponding to each candidate recommendation object into the initial prediction model. Through the embedding layer in the initial prediction model, the combined features corresponding to each candidate recommendation object are encoded to obtain the embedded feature vector corresponding to each candidate recommendation object. The embedding layer in the initial prediction model can transform the sparse matrix corresponding to each combined feature into a dense matrix through linear transformation (such as through a lookup table operation to convert the combined features into feature codes, generating a specific encoding matrix (such as an encoding matrix containing 0 or 1)). This dense matrix is ​​used to represent the combined features corresponding to each recommendation object. This dense matrix contains a large amount of feature information and their interrelationships. After obtaining the embedded feature vector corresponding to each candidate recommendation object, the computer device can concatenate the embedded feature vectors corresponding to each candidate recommendation object based on the connection layer in the initial prediction model to obtain a concatenated feature vector corresponding to N embedded feature vectors. That is, concatenating N embedded feature vectors yields a single concatenated feature vector. The concatenated feature vectors corresponding to N embedding vectors are input into the fully connected hidden layer of the initial prediction model. The fully connected hidden layer performs convolution on the concatenated feature vectors, outputting predicted recommendation features for N candidate recommendation objects. The role of the fully connected hidden layer is to abstract the concatenated feature vectors to another dimension, representing their more abstract features for better linear classification. After obtaining the predicted recommendation features, the computer device can classify the predicted recommendation features corresponding to the N candidate recommendation objects based on the classifier in the initial prediction model, obtaining the predicted recommendation probability for each candidate recommendation object.

[0146] Optionally, the fully connected hidden layer includes a first fully connected layer, a second fully connected layer, and an activation layer. The specific method by which the computer device outputs predicted recommendation features corresponding to N candidate recommendation objects through the fully connected network layer can include: inputting a concatenated feature vector into the first fully connected layer, performing convolution processing on the concatenated feature vector using the weight matrix corresponding to the first fully connected layer to obtain a first recommendation feature vector; inputting the first recommendation feature vector into the second fully connected layer, performing convolution processing on the first recommendation feature vector using the weight matrix corresponding to the second fully connected layer to obtain a second recommendation feature vector; and inputting the second recommendation feature vector into the activation layer, performing activation processing on the second recommendation feature vector to obtain the predicted recommendation features corresponding to the N candidate recommendation objects.

[0147] Specifically, since a single feature vector is input into the fully connected hidden layer at a time, the embedded feature vectors corresponding to the N candidate recommendation objects can be concatenated to obtain a single concatenated feature vector. This reduces the computational cost of feature extraction for the N candidate recommendation objects, thereby improving the efficiency of feature extraction. The computer device can input the concatenated feature vector corresponding to the N candidate recommendation objects into the first fully connected layer. The weight matrix in the first fully connected layer performs convolution on the concatenated feature vector to obtain the first recommendation feature vector. The computer device can then input the first recommendation feature vector into the second fully connected layer. The weight matrix in the second fully connected layer performs convolution on the first recommendation feature vector to obtain the second recommendation feature vector. The second recommendation feature vector is then input into the activation layer, where it is activated to obtain the predicted recommendation features corresponding to the N candidate recommendation objects. Essentially, the fully connected layer in the fully connected hidden layer linearly transforms one feature space to another, mapping the learned concatenated feature vector (i.e., distributed feature representation) to the sample label space for subsequent classification. The number of fully connected layers in a fully connected hidden layer can be one or more, and can be set according to specific needs. In this embodiment, the number of fully connected layers in a fully connected hidden layer is not limited.

[0148] like Figure 4 As shown, Figure 4 This is a schematic diagram of the network structure of an initial prediction model provided in an embodiment of this application, as shown below. Figure 4As shown, the initial prediction model can include an embedding layer, a connection layer, a fully connected hidden layer, and a classifier. The computer device can input the combined features corresponding to each candidate recommendation object into the embedding layer. This embedding layer includes embedding layer 1, embedding layer 2, ..., embedding layer n-1, embedding layer n. Each embedding layer is used to encode the feature combination corresponding to a recommendation object. For example, embedding layer 1 can be used to encode the features of recommendation object 1 to obtain the embedded feature vector corresponding to recommendation object 1, and embedding layer 2 can be used to encode the features of recommendation object 2 to obtain the embedded feature vector corresponding to recommendation object 2. Through the embedding layer in the initial prediction model, one-hot encoding (one-hot encoding uses an N-bit state register to encode N states, each state has its own independent register bit, and only one bit is valid at any given time) is used to linearly transform the feature matrix corresponding to each feature combination to obtain the embedded feature vector corresponding to each candidate recommendation object. After obtaining the embedded feature vectors corresponding to each candidate recommendation object, the computer device can concatenate these vectors based on the connected layers in the initial prediction model, resulting in a concatenated feature vector of N embedded feature vectors. Since a fully connected layer can only input one feature vector at a time, concatenating N embedded feature vectors into a single concatenated feature vector reduces computation and improves feature extraction efficiency. The concatenated feature vector is then input into the fully connected hidden layer of the initial prediction model. This fully connected hidden layer performs convolution on the concatenated feature vector, outputting the predicted recommendation features corresponding to the N candidate recommendation objects. This fully connected hidden layer can include a first fully connected layer and a second fully connected layer, through which the concatenated feature vector is processed. After obtaining the predicted recommendation features, the computer device can classify the predicted recommendation features corresponding to the N candidate recommendation objects based on the classifier in the initial prediction model, obtaining the predicted recommendation probability for each candidate recommendation object.

[0149] The initial prediction model can be any intermediate deep neural network model. The prediction effect of the initial prediction model can be improved by modifying the network structure, adding or removing network layers, changing the size of hidden layers and embedding layers, etc., according to specific needs. For example, the initial prediction model can include convolutional network layers or recurrent network layers, and the predicted recommendation probability corresponding to each candidate recommendation object can be obtained. This application embodiment does not limit the network structure in the initial prediction model and the number of network layers in each network structure (such as the number of fully connected layers in the fully connected hidden layer), and can be set according to specific needs.

[0150] Optionally, after the computer device determines the target recommendation object for recommendation to the recommended user from the sample recommendation objects, it can also send the target recommendation object to the user terminal so that the user terminal can obtain user conversion feedback information for the target recommendation object, receive the user conversion feedback information returned by the user terminal, and generate user feedback tags corresponding to the target recommendation object based on the user conversion feedback information.

[0151] Specifically, the computer device can send the target recommendation object to the user terminal, i.e., the user terminal to which the recommended user belongs, so that the user terminal outputs the target recommendation object to the recommended user and obtains the user conversion feedback information of the recommended user regarding the target recommendation object. For example, when the target recommendation object is a target product advertisement, the user terminal can display the target product advertisement to the recommended user, and the user conversion feedback information of the recommended user regarding the target recommendation object can be the recommended user's click behavior (such as click behavior or non-click behavior). For example, when the target recommendation object is a book, the user conversion feedback information of the recommended user regarding the target recommendation object can be the purchase behavior (such as purchase behavior or non-purchase behavior). After obtaining the user conversion feedback behavior of the recommended user regarding the target recommendation object, the user terminal can send the user conversion feedback information to the computer device in this embodiment. The computer device in this embodiment can receive the user conversion feedback information returned by the user terminal and generate user feedback tags corresponding to the target recommendation object based on the user conversion feedback information. If the target recommendation is a product advertisement, and the user conversion feedback information for that target recommendation is whether the user clicked the product, then the user feedback tag for that target recommendation can be set to 1 if the recommended user clicked the product, and to 0 if the recommended user did not click the product. Similarly, if the target recommendation is a book, and the user conversion feedback information for that target recommendation is whether the user read the book, then the user feedback tag for that target recommendation can be set to 1 if the recommended user read the book, and to 0 if the recommended user did not read the book.

[0152] Optionally, the computer device in this embodiment can be a system composed of a server and a user terminal. After the server in the computer device determines the target recommendation object from the sample recommendation objects, it can send the target recommendation object to the user terminal in the computer device. The user terminal can receive the target recommendation object sent by the server, output the target recommendation object to the recommended user, obtain the user conversion feedback information of the recommended user regarding the target recommendation object, and send the user conversion feedback information to the server. The server can receive the user conversion feedback information sent by the user terminal and generate a user feedback tag corresponding to the target recommendation object based on the user conversion feedback information.

[0153] S103, based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object, determine the candidate model parameters and the adjustment degree corresponding to the candidate model parameters for the initial prediction model.

[0154] Specifically, after obtaining user feedback tags corresponding to the target recommendation object, the computer device can determine the loss function corresponding to the initial prediction model based on the user feedback tags and the predicted recommendation features corresponding to the target recommendation object. Based on this loss function, the initial adjusted model parameters are determined, and these initial adjusted model parameters are designated as candidate model parameters for the initial prediction model. The adjustment degree of the candidate model parameters is determined based on the variation between the candidate model parameters and the initial model parameters in the initial prediction model. This adjustment degree, which has randomness, is used to adjust the candidate model parameters to obtain the target model parameters. These target model parameters are then used by the user to update the initial model parameters.

[0155] Optionally, the computer device may determine the candidate model parameters corresponding to the initial prediction model and the degree of adjustment corresponding to the candidate model parameters by: determining the loss function corresponding to the initial prediction model based on user feedback tags and the prediction recommendation features corresponding to the target recommendation object; determining the candidate model parameters for adjusting the parameters of the initial prediction model based on the loss function; and determining the degree of adjustment corresponding to the candidate model parameters based on the candidate model parameters and the initial model parameters in the initial prediction model.

[0156] Specifically, the computer device can determine the loss function corresponding to the initial prediction model based on the user feedback tags corresponding to the target recommendation object (these user feedback tags are obtained through user conversion feedback information of the recommended user) and the predicted recommendation features of the target recommendation object (these predicted recommendation features are obtained through feature extraction from the initial prediction model). Based on this loss function, the current loss value of the initial prediction model is calculated (i.e., determining the difference between the predicted recommendation features corresponding to the initial prediction model and the user feedback tags). Based on this loss value, candidate model parameters corresponding to the current initial prediction model are determined. These candidate model parameters are used to adjust the model parameters in the initial prediction model. For example, the candidate model parameters corresponding to the current initial prediction model can be calculated using gradient descent, i.e., the error is calculated using the predicted recommendation features output by the initial prediction model and the user feedback tags, the error is backpropagated to each node in the network of the initial prediction model, and the parameters of each node are updated. Based on the candidate model parameters and the initial model parameters in the initial prediction model, the adjustment degree corresponding to the candidate model parameters is determined.

[0157] Optionally, the specific method by which the computer device determines the degree of adjustment corresponding to the candidate model parameters based on the candidate model parameters and the initial model parameters in the initial prediction model may include: randomly selecting a target update coefficient from the update coefficient range, obtaining a first product between the target update coefficient and the initial model parameters in the initial prediction model, and determining the degree of adjustment corresponding to the candidate model parameters based on the first product and the candidate model parameters.

[0158] Specifically, the computer equipment can randomly select a target update coefficient from the update coefficient range. For example, if the update coefficient range is (0, 1), the target update coefficient can be any value within the range (0, 1), such as 0.1, 0.3, or 0.7. The target update coefficient, also known as the exponential moving average coefficient (decay), is used to determine the adjustment degree corresponding to the candidate model parameters. The first product between the target update coefficient and the initial model parameters in the initial prediction model is obtained. Based on this first product, the candidate model parameters are adjusted to determine the adjustment degree corresponding to the candidate model parameters.

[0159] Optionally, the specific method by which the computer device determines the degree of adjustment corresponding to the candidate model parameters based on the first product and the candidate model parameters may include: adjusting the candidate model parameters based on the first product to obtain updated candidate model parameters, obtaining the parameter difference between the updated candidate model parameters and the candidate model parameters, and determining the degree of adjustment corresponding to the candidate model parameters based on the parameter difference.

[0160] Specifically, the computer equipment can obtain the difference between the target threshold and the target update coefficient, and then obtain the third product of this difference with the candidate model parameters. The target threshold can be the natural number 1 or other thresholds. The sum of the first product (i.e., the product between the target update coefficient and the initial model parameters in the initial prediction model) and the third product (i.e., the product of the difference between the target threshold and the target update coefficient, and the candidate model parameters) is obtained, and this sum is used to determine the updated candidate model parameters. Finally, the parameter difference between the candidate model parameters and the updated candidate model parameters, i.e., the magnitude of change between the candidate model parameters and the updated candidate model parameters, is obtained, and the adjustment degree corresponding to the candidate model parameters is determined based on this parameter difference.

[0161] The calculation method for obtaining updated candidate model parameters by computer equipment can be as follows: (1)

[0162] (1)

[0163] Among them, in formula (1) This refers to updating the candidate model parameters. Decay refers to the exponential moving average coefficient (i.e., the target randomness coefficient), and the value range of Decay is (0, 1). This refers to the initial model parameters in the initial prediction model. This refers to the parameters of the candidate model.

[0164] The calculation method for the parameter difference between the candidate model parameters obtained by the computer equipment and the updated candidate model parameters can be as follows (2):

[0165] (2)

[0166] Among them, in formula (2) This refers to the parameter difference between the candidate model parameters and the updated candidate model parameters. `abs()` is used to calculate the absolute value. This refers to the candidate model parameters. This refers to updating the parameters of the candidate model.

[0167] Optionally, the specific method by which the computer device determines the degree of adjustment corresponding to the candidate model parameters based on the parameter difference may include: randomly obtaining the target randomness coefficient from the randomness coefficient range, obtaining the second product between the target randomness coefficient and the parameter difference, and determining the second product as the degree of adjustment corresponding to the candidate model parameters.

[0168] Specifically, the computer equipment can randomly obtain the target randomness coefficient from the randomness coefficient interval. The target randomness coefficient can also be called the linear coefficient between the parameter uncertainty and the parameter change. The range of the randomness coefficient interval can be (0, +∞). The target randomness coefficient can be any value in the interval (0, +∞). The second product between the target randomness coefficient and the parameter difference is obtained. That is, the parameter difference is adjusted by the target randomness coefficient. The second product is determined as the adjustment degree corresponding to the candidate model parameter.

[0169] The calculation method for determining the adjustment degree corresponding to the candidate model parameters based on the parameter difference by computer equipment can be as follows (3):

[0170] (3)

[0171] Among them, in formula (3) This refers to the degree of adjustment. This refers to the target randomness coefficient. This refers to the parameter difference.

[0172] It should be noted that when obtaining the uncertainty of model parameter changes in the initial prediction model, i.e., obtaining the adjustment degree corresponding to the candidate model parameters, other methods besides those mentioned above can be used, including but not limited to changing the target randomness coefficient decay, changing the value of the target randomness coefficient α, and obtaining the adjustment degree using other forms of expressions. The value of is not limited in this embodiment of the application.

[0173] S104. Based on the candidate model parameters and the adjustment degree, determine the random sampling interval of the parameters corresponding to the initial prediction model, and determine the target model parameters corresponding to the initial prediction model from the random sampling interval of the parameters.

[0174] Specifically, after obtaining the candidate model parameters and their corresponding adjustment degrees, the computer equipment can determine the random sampling interval for the parameters of the initial prediction model based on these parameters and adjustment degrees. If the candidate model parameters can be adjusted based on the adjustment degrees to obtain adjusted candidate model parameters, the random sampling interval for the parameters of the initial prediction model can be determined based on the unadjusted candidate model parameters and the adjusted candidate model parameters. The target model parameters for the initial prediction model are then determined from this random sampling interval and used to update the parameters of the initial prediction model.

[0175] S105, based on the target model parameters, update the parameters of the initial prediction model. When the initial prediction model after parameter update meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

[0176] Specifically, after obtaining the target model parameters, the computer equipment can update the parameters of the initial prediction model based on these parameters. When the updated initial prediction model meets the convergence condition, it can be identified as the target prediction model. If the updated initial prediction model does not meet the convergence condition, the updated initial prediction model is used to predict the target recommendation object from the sample recommendation objects. The updated initial prediction model is then trained based on the user feedback tags and prediction recommendation features corresponding to the target recommendation object, resulting in a re-updated initial prediction model. Training of the initial prediction model stops when the updated initial prediction model meets the convergence condition.

[0177] The target model parameters are randomly selected from a random sampling interval generated based on the adjustment degree and candidate model parameters. This randomness, utilizing the uncertainty of the candidate model parameters to guide the generation of uncertain target model parameters, increases the randomness of the updated initial prediction model. This allows the updated initial prediction model to explore other types of recommended objects to deliver to users, thereby acquiring more training data for model training and improving the accuracy of the target prediction model, rather than simply updating the initial model parameters to those of the candidate models. Different target model parameters randomly selected from the random sampling interval will result in different target recommended objects predicted by the updated initial prediction model in the next round; that is, the uncertainty of parameter sampling affects the uncertainty of the prediction results of the initial prediction model. During the training of the initial prediction model, as the prediction results of the initial prediction model become more accurate, the adjustment degree of the parameters of the candidate model corresponding to the initial prediction model will gradually decrease, and the random sampling interval of the parameters of the initial prediction model will also gradually decrease. In this way, the uncertainty of the target model parameters will also decrease, making the training process of the initial prediction model controllable, purposeful, and efficient. This can reduce the waste of resources in the exploration process and avoid recommending objects that users are not interested in, thus affecting the user experience and improving the prediction accuracy of the target prediction model in object recommendation.

[0178] This application can be applied to advertising recommendation scenarios, which are also application scenarios of exploration and utilization. Exploration refers to random exploration, that is, randomly selecting candidate advertisements to be delivered to users, exploring users' preferences for different advertisements, collecting more data to train the initial prediction model, so that the obtained target prediction model can more accurately predict recommended advertisements. Utilization refers to greedy utilization, that is, using the prediction results of the initial prediction model, selecting the candidate advertisement with the highest expected revenue (expected revenue refers to the mathematical expectation of the revenue generated after an advertisement is delivered, which can be equal to the predicted value of the click-through rate of that advertisement) to deliver to users, in order to maximize the current revenue. By using the initial prediction model to predict the expected revenue that a specific advertisement can bring to a specific user, when there is enough sample training data to train the initial prediction model, the initial prediction model can relatively accurately predict the expected revenue of the candidate advertisement (such as the advertisement click-through rate, which is equal to the ratio of the number of ad clicks to the number of ad impressions in terms of frequency, and can represent the probability of the ad being clicked after being exposed in terms of probability). By using the above greedy utilization method, the revenue of each recommendation can be maximized. However, in real-world advertising recommendation scenarios, new ads and users constantly emerge, and user interests may also change. Simply using the greedy exploitation method described above will lead to inaccurate predictions of the expected revenue of new <ad, user> combinations by the initial prediction model. Therefore, when making advertising recommendations, the initial prediction model needs to consider appropriate exploration (i.e., randomly selecting candidate ads and delivering them to users to explore user preferences for different ads). By proactively delivering new ads and different types of ads to users, feedback information can be collected to provide more sample training data for the initial prediction model, ensuring that the model converges in an accurate convergence direction. This will allow for more accurate predictions of ads that users are interested in. This application allows for the combination of advertising features corresponding to all candidate ads in the ad pool with user features corresponding to the current user (i.e., the user for whom ads are recommended) to obtain a feature combination <ad features, user features> for each candidate ad. This feature combination is then input into an initial prediction model, which outputs the predicted recommendation probability for each candidate ad. Based on a certain delivery strategy (e.g., selecting the candidate ad with the highest predicted recommendation probability), a target recommendation ad for recommending to the current user is determined. In other words, the initial prediction model identifies target recommendation ads that the current user may be interested in from the ad pool and recommends these target recommendation ads to the current user.The computer device then receives user conversion feedback information from the current user. Based on this feedback, it determines the user feedback tags for the target recommended ad. It then uses the user characteristics of the current user, the ad characteristics of the target recommended ad, and the user feedback tags as training data to train the initial prediction model. This trains candidate model parameters (derived from the error between the outputs of the initial prediction model) and obtains the adjustment degree (e.g., the magnitude of parameter updates) corresponding to these candidate model parameters. Based on this adjustment degree and the candidate model parameters, the target model parameters are determined, and the initial model parameters in the initial prediction model are updated to the target model parameters. These target model parameters are selected from a randomly generated parameter sampling interval based on the adjustment degree and the candidate model parameters, thus exhibiting randomness. Utilizing the uncertainty of the candidate model parameters to guide the generation of uncertain target model parameters increases the randomness of the updated initial prediction model. This allows the updated initial prediction model to explore other types of ad recommendations to the user, thereby acquiring more training data for model training and improving the accuracy of the target prediction model, rather than simply updating the initial model parameters in the initial prediction model to the candidate model parameters. If the target model parameters randomly selected within the random sampling interval are different, the target recommended advertisements predicted by the initial prediction model after updating based on the target model parameters will also be different in the next round. That is, the uncertainty of parameter sampling will affect the uncertainty of the prediction results of the initial prediction model. During the training process of the initial prediction model, as the prediction results corresponding to the initial prediction model gradually become more accurate, the adjustment degree of the candidate model parameters corresponding to the initial prediction model will also gradually decrease, and the random sampling interval of the parameters corresponding to the initial prediction model will also gradually decrease. In this way, the uncertainty of the target model parameters will also decrease, making the training process of the initial prediction model controllable, purposeful, and efficient. This can reduce the waste of resources in the exploration process and avoid recommending advertisements that users are not interested in, thus affecting the user experience and improving the prediction accuracy of the target prediction model in recommending advertisements.

[0179] Specifically, given the massive number of ads in the ad pool and the vast number of users requiring ad recommendations, predicting the revenue (e.g., click-through rate) for each combination <ad, user> can be done using a combination feature <ad features, user features> obtained by combining the ad features corresponding to each ad and the user features corresponding to each user. User features can refer to factors like age, gender, and location, while ad features can refer to ad type, product price, etc. After obtaining the feature combination for each ad, the expected revenue (e.g., click-through rate) for each ad can be obtained through the functional relationship between the feature combination and the final revenue (e.g., click-through rate). A deep neural network acts as a general function approximator; by using the feature combination for each ad as input and the final revenue (e.g., click-through rate) as output, the initial prediction model can be trained to obtain a target prediction model for recommending ads to users. Figure 4 As shown, Figure 4 A structure diagram of an initial prediction model based on a deep neural network is provided. For details on the specific structure of the initial prediction model, please refer to [link / reference needed]. Figure 4 The described content will not be repeated here in the embodiments of this application. For example, when the final revenue output by the initial prediction model is determined as the click-through rate (CTR) of an advertisement, the target prediction model can predict the CTR for each advertisement; that is, the target prediction model can be a CTR prediction model. During the training of the initial prediction model, the error (i.e., the loss value) corresponding to the initial prediction model can be determined through the input and output of the initial prediction model. The error is backpropagated to each node in the network of the initial prediction model, and the parameters corresponding to each node are updated, i.e., the initial model parameters in the initial prediction model are updated. The initial prediction model is trained in this way until it meets the convergence condition. The initial prediction model that meets the convergence condition is then determined as the target prediction model. The convergence condition can be that the error corresponding to the initial prediction model is within the target error threshold range, or that the number of training iterations of the initial prediction model reaches the target number of training iterations, etc.

[0180] Specifically, such as Figure 5 As shown, Figure 5 This is a schematic diagram of an initial prediction model training provided in an embodiment of this application, as shown below. Figure 5As shown, the sample recommendation object can be an advertisement. The training model A, parameter smoothing model B, and prediction model C all have the same model parameters. The computer device can combine the advertisement features corresponding to each advertisement in the advertisement pool 501 (i.e., the sample recommendation object set) with the user features corresponding to user B502 (i.e., the recommended user) to obtain the advertisement feature combination corresponding to each advertisement, i.e., advertisement feature + user feature 503. The advertisement feature combination corresponding to each advertisement is input into the prediction model C. The prediction model C performs feature extraction on the advertisement feature combination corresponding to each advertisement and outputs a click-through rate list 505 corresponding to all advertisements in the advertisement pool. This click-through rate list 505 contains the click-through rate (i.e., the predicted recommendation probability) corresponding to all advertisements. The advertisement with the highest click-through rate in the click-through rate list is selected as the target recommendation advertisement 506 for this round. The target recommendation advertisement 506 is delivered to user B502, and the user conversion feedback information of user B502 regarding the target recommendation advertisement (such as whether the user clicked on the target recommendation advertisement) is obtained. Based on the user conversion feedback information of user B502 regarding the target recommendation advertisement 506, a user feedback tag corresponding to the target recommendation advertisement 506 is generated. For example, if user B502 clicks on the target recommended ad 506, the user feedback tag corresponding to the target recommended ad 506 can be set to 1; if user B502 does not click on the target recommended ad 506, the user feedback tag corresponding to the target recommended ad 506 can be set to 0. After obtaining the user feedback tag corresponding to the target recommended ad 506, the computer device can use the ad features corresponding to the target recommended ad 506, the user feedback tag corresponding to the target recommended ad 506, and the user features corresponding to user B502 as a set of training samples 507, and input the samples 507 into the training model A508. Through the training model A508, feature extraction is performed on the ad features corresponding to the target recommended ad 506 and the user features corresponding to user B502 to obtain the predicted recommendation features corresponding to the target recommended ad 506. Based on the predicted recommendation features corresponding to the target recommended ad 506 and the user feedback tag corresponding to the target recommended ad 506, the loss function corresponding to the training model A508 is determined, and the loss value corresponding to the training model A508 is determined based on the loss function. Based on the loss value, determine the adjusted model parameters corresponding to the training model A508 (equivalent to the candidate model parameters corresponding to the initial prediction model). Update the parameters of the training model A508 according to these adjusted model parameters to obtain the parameter-updated training model A. Based on the parameter-updated training model A, update the model parameters in the parameter smoothing model B509 using an exponential moving average method to obtain the parameter-updated parameter smoothing model B (equivalent to updating the candidate model parameters). Obtain the parameter difference between the parameter-updated parameter smoothing model B and the parameter-updated training model A, and determine the adjustment degree corresponding to the model parameters (i.e., candidate model parameters) in the parameter-updated training model A based on this parameter difference.Based on the adjustment degree and the updated training model A, the random sampling interval for the parameters corresponding to the prediction model C504 is determined. From this interval, the target model parameters for updating prediction model C are identified, and the prediction model C is updated accordingly. When the updated prediction model C meets the convergence condition, it is designated as the target prediction model, used to predict recommended advertisements for users. If the updated prediction model C does not meet the convergence condition, the above steps are repeated to train the updated prediction model C until it meets the convergence condition.

[0181] In this embodiment, a sample recommendation object set is obtained. The object features corresponding to each of the N candidate recommendation objects in the sample recommendation object set, along with the user features corresponding to the recommended user, are input into an initial prediction model. The initial prediction model then predicts the target recommendation object to be recommended to the recommended user. Based on the user feedback tags and predicted recommendation features corresponding to the target recommendation object, candidate model parameters and their adjustment degrees are determined for the initial prediction model. A random sampling interval for the parameters of the initial prediction model is then determined based on the candidate model parameters and their adjustment degrees. The target model parameters are then determined from this random sampling interval. Since the candidate model parameters for the initial prediction model have uncertainty, and this uncertainty is related to the uncertainty of the prediction results output by the initial prediction model (i.e., different target model parameters update the initial prediction model, resulting in different prediction results when the updated initial prediction model predicts recommendation objects), the target model parameters can be generated based on the candidate model parameters and their adjustment degrees, thus increasing the randomness of the target model parameters. Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, it is determined as the target prediction model. The uncertainty of the candidate model parameters guides the generation of target model parameters with uncertainty. This increases the randomness of the target model parameters, allowing for targeted exploration of recommended objects that users might be interested in. This ensures the initial prediction model converges in an accurate convergence direction, improving the accuracy of the target prediction model in predicting recommended objects. During the training process of the initial prediction model, as the prediction results become more accurate, the adjustment degree of the candidate model parameters and the random sampling interval of the initial prediction model parameters gradually decrease. This reduces the uncertainty of the target model parameters, making the training process controllable, purposeful, and efficient. It reduces resource waste during the exploration process and avoids recommending objects that users are not interested in, thus improving the accuracy of the target prediction model in recommending objects.

[0182] Please see Figure 6 , Figure 6 This is a flowchart illustrating a data processing method provided in an embodiment of this application. This data processing method can be executed by a computer device, which can be a server (as described above). Figure 1 Server 10 in the middle), or user terminal (as mentioned above) Figure 1 This application does not limit the scope to any user terminal in a user terminal cluster, or a system consisting of a server and user terminals. Figure 2 As shown, the data processing method may include steps S201-S206.

[0183] S201, Obtain the set of sample recommendation objects.

[0184] S202, input the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the recommended user into the initial prediction model, and predict the target recommendation object to be recommended to the recommended user through the initial prediction model.

[0185] S203, based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object, determine the candidate model parameters and the adjustment degree corresponding to the candidate model parameters for the initial prediction model.

[0186] The contents of steps S201-S203 in the embodiments of this application can be found in [reference]. Figure 3 The specific details of steps S101-S103 will not be repeated here in the embodiments of this application.

[0187] S204. The candidate model parameters are determined as the parameter mean and the adjustment degree is determined as the standard deviation. Based on the parameter mean and standard deviation, the normal distribution associated with the candidate model parameters is determined, and the normal distribution is determined as the random sampling interval of the parameters corresponding to the initial prediction model.

[0188] Specifically, after obtaining the candidate model parameters and their corresponding adjustments, the computer device can determine the mean of the candidate model parameters and the standard deviation of the adjustments. Based on the mean and standard deviation, a normal distribution associated with the candidate model parameters is generated, and this normal distribution is used as the random sampling interval for the parameters corresponding to the initial prediction model. Alternatively, other distributions can be used to generate the random sampling interval for the parameters corresponding to the initial prediction model, such as the triangular distribution (i.e., determining the upper and lower bounds of the triangular distribution based on the candidate model parameters and adjustments, and generating the random sampling interval based on these upper and lower bounds), the arcsine distribution (i.e., determining the expected value of the candidate model parameters, the standard deviation of the adjustments, determining the upper and lower bounds of the arcsine distribution, and generating the random sampling interval based on these upper and lower bounds), etc. The specific method for generating the random sampling interval for the parameters corresponding to the initial prediction model can be selected according to specific needs. This application embodiment does not limit the specific method for generating the random sampling interval for the parameters corresponding to the initial prediction model.

[0189] like Figure 7 As shown, Figure 7 This is a schematic diagram of a random sampling interval for parameters provided in an embodiment of this application, such as... Figure 7As shown, it can be used to generate random sampling intervals for parameters corresponding to the initial prediction model using the normal distribution, and the candidate model parameters can be used as a normal distribution. mean The adjustment degree corresponding to the candidate model parameters is taken as a normal distribution. standard deviation Generate random sampling intervals of parameters associated with candidate model parameters, with mean... and standard deviation The values ​​of are all positive integers. The size of the random sampling interval for parameters depends on the adjustment degree of the candidate model parameters. When the loss value corresponding to the initial prediction model is large, the uncertainty of the candidate model parameters is greater, and the corresponding adjustment degree is larger, thus the random sampling interval for the initial prediction model parameters is larger; when the loss value corresponding to the initial prediction model is small, the uncertainty of the candidate model parameters is smaller, and the corresponding adjustment degree is smaller, thus the random sampling interval for the initial prediction model parameters is smaller. For example... Figure 7 As shown in the normal distribution 1, when the uncertainty of the candidate model parameters is greater (e.g., in the early stage of training the initial prediction model, the uncertainty of the candidate model parameters used to adjust the initial prediction model is relatively large), the adjustment degree of the candidate model parameters is greater (which can be obtained according to formula (3)), and therefore the random sampling interval of the parameters corresponding to the initial prediction model is larger. Figure 7 As shown in the normal distribution 2, when the uncertainty of the candidate model parameters is smaller (such as in the later stage of the initial prediction model training, the uncertainty of the candidate model parameters used to adjust the initial prediction model is smaller), the adjustment degree of the candidate model parameters is smaller (which can be obtained according to formula (3)), and therefore the random sampling interval of the parameters corresponding to the initial prediction model is smaller.

[0190] S205, random sampling is performed in the parameter random sampling interval to obtain the target model parameters corresponding to the initial prediction model.

[0191] Specifically, after the computer equipment determines the random sampling interval for the parameters corresponding to the initial prediction model, it can randomly sample within this interval, selecting a model parameter as the target model parameter. Since the candidate model parameters used for updating the initial prediction model are uncertain during training, simply using the loss value between the predicted recommendation features output by the initial prediction model and the user feedback labels to determine the candidate model parameters and then directly using these as the target model parameters for updating the initial prediction model would lead to the final target prediction model repeatedly recommending similar categories of objects to the user, lacking exploration of other types of objects and failing to find recommendations that the user is truly interested in, resulting in low prediction accuracy. This solution, however, uses the loss value between the predicted recommendation features output by the initial prediction model and the user feedback labels to determine the candidate model parameters. Based on the uncertainty of the candidate model parameters, it then determines the random sampling interval for the parameters corresponding to the initial prediction model and randomly selects the target model parameter from this interval for updating the initial prediction model. When determining the target model parameters, the uncertainty (i.e., adjustment degree) of the candidate model parameters is introduced. Therefore, the prediction results of the initial prediction model after updating the parameters based on the target model parameters will also change with the uncertainty of the candidate parameters and random sampling. As the initial prediction model is trained for multiple rounds, the distance between the initial prediction model and the convergence condition will gradually decrease, that is, the loss value corresponding to the initial prediction model will gradually decrease, and the uncertainty (i.e., adjustment degree) of the candidate model parameters corresponding to the initial prediction model will also gradually decrease, that is, gradually approach 0, thus gradually approaching the convergence condition corresponding to the initial prediction model. In this way, after the target model parameters are generated by random sampling, the prediction results obtained by the initial prediction model updated based on the target model parameters are also random. Therefore, it can reduce the repeated recommendation of similar categories of recommended objects to users, and can randomly explore other categories of recommended objects to more accurately determine the recommended objects that users are truly interested in, thereby improving the accuracy of the prediction of recommended objects.

[0192] Among them, the formula for determining the random sampling interval of the parameters corresponding to the initial prediction model based on the candidate model parameters and the corresponding adjustment degree can be as follows (4):

[0193] (4)

[0194] Among them, in formula (4) This refers to the target model parameters. (D) refers to a single sampling from the data distribution D. This refers to the candidate model parameters The mean is the adjustment degree corresponding to the candidate model parameters. It is a normal distribution with standard deviation.

[0195] S206. Based on the target model parameters, update the parameters of the initial prediction model. When the initial prediction model after parameter update meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model.

[0196] Specifically, the computer equipment updates the parameters of the initial prediction model based on the target model parameters. When the updated initial prediction model meets the convergence condition, the method for determining the convergent initial prediction model as the target prediction model can include: updating the model parameters in the initial prediction model based on the target model parameters to obtain the updated initial prediction model. If the adjustment degree corresponding to the updated initial prediction model is the target adjustment threshold, then the updated initial prediction model is determined to meet the convergence condition, and the updated initial prediction model that meets the convergence condition is determined as the target prediction model.

[0197] Specifically, after the computer equipment determines the target model parameters corresponding to the initial prediction model, it can update the initial model parameters in the initial prediction model according to the target model parameters to obtain the parameter-updated initial prediction model. For example, the initial model parameters in the initial prediction model can be replaced with the target model parameters to obtain the parameter-updated initial prediction model. If the adjustment degree corresponding to the parameter-updated initial prediction model meets the target adjustment threshold, it can be determined that the parameter-updated initial prediction model meets the convergence condition, and the parameter-updated initial prediction model that meets the convergence condition is determined as the target prediction model. The target adjustment threshold can be a threshold such as 0, 0.06, or 0.02, which can be set according to specific needs, and this application embodiment does not impose any restrictions. If the target adjustment threshold is 0, it means that the uncertainty of the candidate model parameters corresponding to the initial prediction model is 0, that is, the candidate model parameters determined in this round of training are the same as the candidate model parameters determined in the previous round of training, indicating that the loss value corresponding to the initial prediction model is 0, and the initial model parameters in the initial prediction model do not need to be adjusted. Therefore, when the adjustment degree corresponding to the parameter-updated initial prediction model meets the target adjustment threshold, it can be determined that the parameter-updated initial prediction model meets the convergence condition. Of course, if the accuracy requirement for the initial prediction model is not high, the target adjustment threshold can be set to other thresholds such as 0.06 or 0.02.

[0198] The initial prediction model is trained as described above to obtain the target prediction model, which can then be applied to scenarios such as advertising recommendation, video recommendation, and text content recommendation.

[0199] Optionally, after training the initial prediction model as described above to obtain the trained target prediction model, the computer device can acquire the target user features corresponding to the target user, acquire the object features corresponding to each object to be recommended in the set of objects to be recommended, input the target user features and the object features corresponding to each object to be recommended into the target prediction model, and output the recommendation probability corresponding to each object to be recommended through the target prediction model. Based on the recommendation probability corresponding to each object to be recommended, the user recommendation object is determined from the set of objects to be recommended to the target user, and the user recommendation object is sent to the target user terminal to which the target user belongs, so that the target user terminal outputs the user recommendation object.

[0200] Specifically, the computer device acquires the target user characteristics corresponding to the target user, and the object characteristics corresponding to each object in the set of objects to be recommended. The target user can refer to the user for whom recommendation object prediction is currently needed, and the object to be recommended can refer to the object to be recommended to the target user. For example, if we need to determine which novel to recommend to user C from the novel library, then user C can be called the target user, and the object to be recommended can be the novel. The target user characteristics and the object characteristics corresponding to each object to be recommended are combined, that is, the object characteristics corresponding to each object to be recommended are combined with the target user characteristics to obtain the combined characteristics corresponding to each object to be recommended. The combined characteristics corresponding to each object to be recommended are input into the target prediction model, and the target prediction model outputs the recommendation probability corresponding to each object to be recommended. Based on the recommendation probability corresponding to each object to be recommended, the user recommendation object to be recommended to the target user is determined from the set of objects to be recommended. The user can choose the recommended object with the highest recommendation probability as the recommended object, or the top three recommended objects with the highest recommendation probability ranking (e.g., sorting the recommendation probabilities of each recommended object in descending order to obtain the recommendation probability ranking) can be chosen as the recommended objects. The number of recommended objects can be one or more, and can be set according to specific needs. This application embodiment does not impose any restrictions on this.

[0201] like Figure 8 As shown, Figure 8 This is a schematic diagram illustrating how to update the parameters of an initial prediction model based on the target model parameters, as provided in an embodiment of this application. Figure 8As shown, after the computer device obtains the user feedback label 80a corresponding to the target recommendation object, it can determine the loss function corresponding to the initial prediction model based on the predicted recommendation feature 80b corresponding to the target recommendation object and the user feedback label 80a corresponding to the target recommendation object. Based on this loss function, it determines the loss value (i.e., the prediction error) corresponding to the initial prediction model. Based on this loss value, it determines the candidate model parameters 80c corresponding to the initial prediction model. Based on the uncertainty of the candidate model parameters (i.e., the magnitude of change compared to the initial model parameters in the initial prediction model), it determines the adjustment degree 80d corresponding to the candidate model parameters. For details on how to determine the adjustment degree of the candidate model parameters, please refer to [link to documentation / reference]. Figure 3 The content of step S103 will not be repeated here in this embodiment. After the computer device obtains the candidate model parameters corresponding to the initial prediction model and the adjustment degree corresponding to the candidate model parameters, it can use the candidate model parameters as the mean and the adjustment degree as the standard deviation to generate a normal distribution associated with the candidate model parameters. This normal distribution is determined as the random sampling interval 80e of the parameters corresponding to the initial prediction model. Figure 8 As shown, the lower bounds c1 and c2 in the random sampling interval of the parameters can refer to the negative and positive values ​​of the adjustment degree, respectively, and the y-values ​​in this normal distribution can refer to the model parameters. After obtaining the random sampling interval of the parameters corresponding to the initial prediction model, one model parameter can be randomly selected from the multiple model parameters in this random sampling interval as the target model parameter 80f. The initial model parameters in the initial prediction model are then updated according to the target model parameters, that is, the initial model parameters in the initial prediction model are replaced with the target model parameters, resulting in the parameter-updated initial prediction model.

[0202] like Figure 9 As shown, Figure 9 This is an application scenario diagram of book recommendation provided in an embodiment of this application, such as... Figure 9As shown, the recommended object can be a book. User 90b can read books on the reading platform in user terminal 90a, and the user terminal can also recommend books that user 90b might be interested in. When user 90b is reading a book on user terminal 90a and needs book recommendations from the reading platform, they can click the book recommendation button in the display interface 90c on user terminal 90a. The user terminal can then redirect to display interface 90d, informing user 90b that book recommendations are being generated. The user terminal can respond to user 90b's trigger operation on the book recommendation button, obtain the user characteristics corresponding to user 90b and the book characteristics corresponding to each book to be recommended in the book library, generate a request for the target recommended book (i.e., the recommended book recommended to user 90b) based on the user characteristics corresponding to user 90b and the book characteristics corresponding to each book to be recommended in the book library, and send the request to server 90e. After receiving the retrieval request from user terminal 90a, server 90e can retrieve the user features corresponding to user 90b and the book features corresponding to each book to be recommended in the book library from the retrieval request. It then inputs these user features and book features into a pre-trained target prediction model 90f. The target prediction model 90f outputs the recommendation probability for each book to be recommended. For details, please refer to [link / reference]. Figure 3 The content described in step S102 will not be repeated here in this embodiment. Server 90e can determine target recommended books from the book library based on the recommendation probability corresponding to each book to be recommended. The number of target recommended books can be one or more, and the target recommended books are sent to user terminal 90a. After receiving the target recommendations sent by server 90e, user terminal 90a can update the display interface 90d, generate a display interface 90g, and output the target recommended books, that is, display the target recommended books to user 90b.

[0203] In this embodiment, a sample recommendation object set is obtained. The object features corresponding to each of the N candidate recommendation objects in the sample recommendation object set, along with the user features corresponding to the recommended user, are input into an initial prediction model. The initial prediction model then predicts the target recommendation object to be recommended to the recommended user. Based on the user feedback tags and predicted recommendation features corresponding to the target recommendation object, candidate model parameters and their adjustment degrees are determined for the initial prediction model. A random sampling interval for the parameters of the initial prediction model is then determined based on the candidate model parameters and their adjustment degrees. The target model parameters are then determined from this random sampling interval. Since the candidate model parameters for the initial prediction model have uncertainty, and this uncertainty is related to the uncertainty of the prediction results output by the initial prediction model (i.e., different target model parameters update the initial prediction model, resulting in different prediction results when the updated initial prediction model predicts recommendation objects), the target model parameters can be generated based on the candidate model parameters and their adjustment degrees, thus increasing the randomness of the target model parameters. Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, it is determined as the target prediction model. The uncertainty of the candidate model parameters guides the generation of target model parameters with uncertainty. This increases the randomness of the target model parameters, allowing for targeted exploration of recommended objects that the user might be interested in. This ensures the initial prediction model converges in the accurate convergence direction and improves the prediction accuracy of the target prediction model in object recommendation. Furthermore, by training the initial prediction model with randomized target model parameters, the training process is more controlled, purposeful, and efficient. This reduces resource waste during the exploration process (i.e., exploring recommended objects that the user is interested in) and minimizes the recommendation of objects that the user is not interested in.

[0204] Please see Figure 10 , Figure 10 This is a schematic diagram of the structure of a data processing apparatus provided in an embodiment of this application. The aforementioned data processing apparatus can be a computer program (including program code) running on a computer device; for example, the data processing apparatus is an application software. The data processing apparatus can be used to execute corresponding steps in the data processing method provided in the embodiments of this application. Figure 10 As shown, the data processing device may include: a first acquisition module 11, a prediction module 12, a first determination module 13, a second determination module 14, a parameter update module 15, a sending module 16, a receiving module 17, a second acquisition module 18, an output module 19, and a third determination module 20.

[0205] The first acquisition module 11 is used to acquire a sample recommendation object set; the sample recommendation object set includes N candidate recommendation objects; N is a positive integer;

[0206] Prediction module 12 is used to input the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the recommended user into the initial prediction model, and predict the target recommendation object to be recommended to the recommended user through the initial prediction model; the N candidate recommendation objects include the target recommendation object;

[0207] The first determining module 13 is used to determine the candidate model parameters and the adjustment degree of the candidate model parameters based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object; the adjustment degree has randomness.

[0208] The second determining module 14 is used to determine the random sampling interval of the parameters corresponding to the initial prediction model based on the candidate model parameters and the adjustment degree, and to determine the target model parameters corresponding to the initial prediction model from the random sampling interval of the parameters.

[0209] The parameter update module 15 is used to update the parameters of the initial prediction model according to the target model parameters. When the initial prediction model after parameter update meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

[0210] The prediction module 12 includes:

[0211] The first acquisition unit 1201 is used to acquire the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the target user.

[0212] Combination unit 1202 is used to combine N object features with user features respectively to obtain the combined features corresponding to each candidate recommendation object;

[0213] The output unit 1203 is used to input the combined features corresponding to each candidate recommendation object into the initial prediction model, and output the predicted recommendation probability corresponding to each candidate recommendation object through the initial prediction model.

[0214] The first determining unit 1204 is used to determine the candidate recommendation object corresponding to the highest predicted recommendation probability as the target recommendation object.

[0215] The output unit 1203 includes:

[0216] The feature encoding subunit 120301 is used to input the combined features corresponding to each candidate recommendation object into the initial prediction model. Through the embedding layer in the initial prediction model, the combined features corresponding to each candidate recommendation object are encoded to obtain the embedded feature vector corresponding to each candidate recommendation object.

[0217] The splicing processing subunit 120302 is used to splice the embedded feature vectors corresponding to each candidate recommendation object based on the connection layer in the initial prediction model to obtain the spliced ​​feature vector.

[0218] The output subunit 120303 is used to input the concatenated feature vector into the fully connected hidden layer in the initial prediction model, and output the predicted recommendation features corresponding to N candidate recommendation objects through the fully connected hidden layer.

[0219] The classification processing subunit 120304 is used to classify the predicted recommendation features corresponding to N candidate recommendation objects based on the classifier in the initial prediction model, and obtain the predicted recommendation probability corresponding to each candidate recommendation object.

[0220] The fully connected hidden layer includes a first fully connected layer, a second fully connected layer, and an activation layer.

[0221] Output subunit 120303 is specifically used for:

[0222] The concatenated feature vector is input into the first fully connected layer. The concatenated feature vector is then convolved with the weight matrix of the first fully connected layer to obtain the first recommended feature vector.

[0223] The first recommended feature vector is input into the second fully connected layer. The first recommended feature vector is then convolved with the weight matrix of the second fully connected layer to obtain the second recommended feature vector.

[0224] The second recommendation feature vector is input into the activation layer, and the activation layer activates the second recommendation feature vector to obtain the predicted recommendation features corresponding to N candidate recommendation objects.

[0225] The data processing device also includes:

[0226] Sending module 16 is used to send the target recommendation object to the user terminal so that the user terminal can obtain user conversion feedback information for the target recommendation object;

[0227] The receiving module 17 is used to receive user conversion feedback information returned by the user terminal and generate user feedback tags corresponding to the target recommendation object based on the user conversion feedback information.

[0228] The first determining module 13 includes:

[0229] The second determining unit 1301 is used to determine the loss function corresponding to the initial prediction model based on the user feedback tags and the prediction recommendation features corresponding to the target recommendation object.

[0230] The third determining unit 1302 is used to determine, based on the loss function, candidate model parameters for adjusting the initial model parameters in the initial prediction model.

[0231] The fourth determining unit 1303 is used to determine the adjustment degree corresponding to the candidate model parameters based on the candidate model parameters and the initial model parameters in the initial prediction model.

[0232] The fourth determining unit 1303 includes:

[0233] Select subunit 130301 to randomly select the target update coefficient from the update coefficient range;

[0234] Obtain subunit 130302, which is used to obtain the first product of the target update coefficient and the initial model parameters in the initial prediction model;

[0235] Subunit 130303 is defined to determine the adjustment degree corresponding to the candidate model parameters based on the first product and the candidate model parameters.

[0236] Specifically, subunit 130303 is used for:

[0237] Based on the first product, the candidate model parameters are adjusted to obtain the updated candidate model parameters;

[0238] Obtain the parameter difference between the updated candidate model parameters and the original candidate model parameters;

[0239] The degree of adjustment corresponding to the parameters of the candidate model is determined based on the parameter difference.

[0240] Specifically, the determined subunit 130303 is also used for:

[0241] The target randomness coefficient is randomly obtained from the randomness coefficient range;

[0242] Obtain the second product between the target randomness coefficient and the parameter difference, and determine the second product as the adjustment degree corresponding to the candidate model parameters.

[0243] The second determining module 14 includes:

[0244] The fifth determining unit 1401 is used to determine the candidate model parameters as the parameter mean, the adjustment degree as the standard deviation, determine the normal distribution associated with the candidate model parameters based on the parameter mean and standard deviation, and determine the normal distribution as the random sampling interval of the parameters corresponding to the initial prediction model.

[0245] The random sampling unit 1402 is used to perform random sampling in the parameter random sampling interval to obtain the target model parameters corresponding to the initial prediction model.

[0246] The parameter update module 15 includes:

[0247] The parameter update unit 1501 is used to update the initial model parameters in the initial prediction model according to the target model parameters, so as to obtain the parameter-updated initial prediction model.

[0248] The sixth determining unit 1502 is used to determine that the initial prediction model after parameter update satisfies the convergence condition if the adjustment degree corresponding to the initial prediction model after parameter update is the target adjustment threshold, and to determine the initial prediction model after parameter update that satisfies the convergence condition as the target prediction model.

[0249] The data processing device also includes:

[0250] The second acquisition module 18 is used to acquire the target user characteristics corresponding to the target user and to acquire the object characteristics corresponding to each object to be recommended in the set of objects to be recommended.

[0251] Output module 19 is used to input the target user features and the object features corresponding to each object to be recommended into the target prediction model, and output the recommendation probability corresponding to each object to be recommended through the target prediction model.

[0252] The third determining module 20 is used to determine the user recommendation object to be recommended to the target user from the set of objects to be recommended according to the recommendation probability corresponding to each object to be recommended, and send the user recommendation object to the target user terminal to which the target user belongs, so that the target user terminal outputs the user recommendation object.

[0253] According to one embodiment of this application, Figure 3 The steps involved in the data processing method shown can be derived from... Figure 10 The data processing device shown is executed by various modules. For example, Figure 3 Step S101 shown can be performed by Figure 10 The first acquisition module 11 in the process is executed. Figure 3 Step S102 shown can be performed by Figure 10 The prediction module 12 in the middle is used to perform the operation. Figure 3 Step S103 shown can be performed by Figure 10 The first determined module 13 in the process is executed. Figure 3 Step S104 shown can be derived from Figure 10 The second determining module 14 in the middle is used to execute, Figure 3 Step S105 shown can be performed by Figure 10The parameter update module 15 in the middle is used to perform, etc.

[0254] According to one embodiment of this application, Figure 10 The modules in the data processing device shown can be individually or entirely combined into one or more units, or one or more of these units can be further divided into multiple functionally smaller sub-units to achieve the same operation without affecting the technical effects of the embodiments of this application. The above modules are based on logical functional division. In practical applications, the function of one module can be implemented by multiple units, or the function of multiple modules can be implemented by one unit. In other embodiments of this application, the testing device may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0255] In this embodiment, a sample recommendation object set is obtained. The object features corresponding to each of the N candidate recommendation objects in the sample recommendation object set, along with the user features corresponding to the recommended user, are input into an initial prediction model. The initial prediction model then predicts the target recommendation object to be recommended to the recommended user. Based on the user feedback tags and predicted recommendation features corresponding to the target recommendation object, candidate model parameters and their adjustment degrees are determined for the initial prediction model. A random sampling interval for the parameters of the initial prediction model is then determined based on the candidate model parameters and their adjustment degrees. The target model parameters are then determined from this random sampling interval. Since the candidate model parameters for the initial prediction model have uncertainty, and this uncertainty is related to the uncertainty of the prediction results output by the initial prediction model (i.e., different target model parameters update the initial prediction model, resulting in different prediction results when the updated initial prediction model predicts recommendation objects), the target model parameters can be generated based on the candidate model parameters and their adjustment degrees, thus increasing the randomness of the target model parameters. Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, it is determined as the target prediction model. The uncertainty of the candidate model parameters guides the generation of target model parameters with uncertainty. This increases the randomness of the target model parameters, allowing for targeted exploration of recommended objects that the user might be interested in. This ensures the initial prediction model converges in the accurate convergence direction and improves the prediction accuracy of the target prediction model in object recommendation. Furthermore, by training the initial prediction model with randomized target model parameters, the training process is more controlled, purposeful, and efficient. This reduces resource waste during the exploration process (i.e., exploring recommended objects that the user is interested in) and minimizes the recommendation of uninteresting objects to the user. During the training of the initial prediction model, as the prediction results of the initial prediction model become more accurate, the adjustment degree of the parameters of the candidate model corresponding to the initial prediction model will gradually decrease, and the random sampling interval of the parameters of the initial prediction model will also gradually decrease. In this way, the uncertainty of the target model parameters will also decrease, making the training process of the initial prediction model controllable, purposeful, and efficient. This can reduce the waste of resources in the exploration process and avoid recommending objects that users are not interested in, thus affecting the user experience and improving the prediction accuracy of the target prediction model in object recommendation.

[0256] Please see Figure 11 , Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 11 As shown, the aforementioned computer device 1000 may include: a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the computer device 1000 may also include: a target user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The target user interface 1003 may include a display screen and a keyboard; optionally, the target 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 non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 11 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a target user interface module, and a device control application.

[0257] exist Figure 11 In the computer device 1000 shown, the network interface 1004 provides network communication functionality; the target user interface 1003 is mainly used to provide an input interface for the target user; and the processor 1001 can be used to call the device control application stored in the memory 1005 to achieve:

[0258] Obtain the sample recommendation object set; the sample recommendation object set includes N candidate recommendation objects, where N is a positive integer;

[0259] The object features corresponding to each of the N candidate recommendation objects, and the user features corresponding to the recommended user, are input into the initial prediction model. The initial prediction model then predicts the target recommendation object to be recommended to the recommended user. The N candidate recommendation objects include the target recommendation object.

[0260] Based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object, determine the candidate model parameters and the adjustment degree corresponding to the candidate model parameters for the initial prediction model; the adjustment degree has randomness.

[0261] Based on the candidate model parameters and the adjustment degree, determine the parameter random sampling interval corresponding to the initial prediction model, and determine the target model parameters corresponding to the initial prediction model from the parameter random sampling interval;

[0262] Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

[0263] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned... Figure 3 or Figure 6 The description of the data processing method in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0264] As an example, the above program instructions can be deployed and executed on a single computer device, or deployed and executed on multiple computer devices located in one location, or executed on multiple computer devices distributed across multiple locations and interconnected via a communication network. Multiple computer devices distributed across multiple locations and interconnected via a communication network can form a blockchain network.

[0265] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0266] It should be understood that in the specific implementation of this application, user characteristics and other related data are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0267] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A data processing method, characterized in that, include: Obtain a sample recommendation object set; the sample recommendation object set includes N candidate recommendation objects, where N is a positive integer; The object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the recommended user are input into the initial prediction model. The initial prediction model is used to predict the target recommendation object to be recommended to the recommended user. The N candidate recommendation objects include the target recommendation object. Based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object, the candidate model parameters corresponding to the initial prediction model are determined; Randomly select the target update coefficient from the update coefficient range; Obtain the first product of the target update coefficient and the initial model parameters in the initial prediction model; Based on the first product, the candidate model parameters are adjusted to obtain updated candidate model parameters; Obtain the parameter difference between the updated candidate model parameters and the candidate model parameters; The adjustment degree corresponding to the candidate model parameters is determined based on the parameter difference; the adjustment degree has randomness. Based on the candidate model parameters and the adjustment degree, determine the parameter random sampling interval corresponding to the initial prediction model, and determine the target model parameters corresponding to the initial prediction model from the parameter random sampling interval; Based on the target model parameters, the initial prediction model is updated. When the updated initial prediction model meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

2. The method according to claim 1, characterized in that, The step of inputting the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the recommended user into the initial prediction model, and predicting the target recommendation object to be recommended to the recommended user through the initial prediction model, includes: Obtain the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the target user; The N object features are combined with the user features respectively to obtain the combined features corresponding to each candidate recommendation object; The combined features corresponding to each candidate recommendation object are input into the initial prediction model, and the predicted recommendation probability corresponding to each candidate recommendation object is output through the initial prediction model. The candidate recommendation object corresponding to the highest predicted recommendation probability is determined as the target recommendation object.

3. The method according to claim 2, characterized in that, The step of inputting the combined features corresponding to each candidate recommendation object into the initial prediction model, and outputting the predicted recommendation probability corresponding to each candidate recommendation object through the initial prediction model, includes: The combined features corresponding to each candidate recommendation object are input into the initial prediction model. The combined features corresponding to each candidate recommendation object are encoded through the embedding layer in the initial prediction model to obtain the embedded feature vector corresponding to each candidate recommendation object. Based on the connection layer in the initial prediction model, the embedded feature vectors corresponding to each candidate recommendation object are concatenated to obtain a concatenated feature vector. The concatenated feature vector is input into the fully connected hidden layer in the initial prediction model, and the predicted recommendation features corresponding to the N candidate recommendation objects are output through the fully connected hidden layer. Based on the classifier in the initial prediction model, the predicted recommendation features corresponding to the N candidate recommendation objects are classified to obtain the predicted recommendation probability corresponding to each candidate recommendation object.

4. The method according to claim 3, characterized in that, The fully connected hidden layer includes a first fully connected layer, a second fully connected layer, and an activation layer; The step of outputting predicted recommendation features corresponding to the N candidate recommendation objects through the fully connected hidden layer in the initial prediction model includes: The concatenated feature vector is input into the first fully connected layer, and the concatenated feature vector is convolved through the weight matrix corresponding to the first fully connected layer to obtain the first recommended feature vector. The first recommended feature vector is input into the second fully connected layer, and the first recommended feature vector is convolved through the weight matrix corresponding to the second fully connected layer to obtain the second recommended feature vector. The second recommendation feature vector is input into the activation layer, and the activation layer activates the second recommendation feature vector to obtain the predicted recommendation features corresponding to the N candidate recommendation objects.

5. The method according to claim 1, characterized in that, The method further includes: The target recommendation object is sent to the user terminal so that the user terminal can obtain user conversion feedback information for the target recommendation object; The system receives the user conversion feedback information returned by the user terminal and generates user feedback tags corresponding to the target recommendation object based on the user conversion feedback information.

6. The method according to claim 1, characterized in that, The step of determining the candidate model parameters corresponding to the initial prediction model based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object includes: Based on the user feedback tags and the predicted recommendation features corresponding to the target recommendation object, determine the loss function corresponding to the initial prediction model; Based on the loss function, candidate model parameters are determined for adjusting the initial model parameters in the initial prediction model.

7. The method according to claim 1, characterized in that, The step of determining the adjustment degree corresponding to the candidate model parameters based on the parameter difference includes: The target randomness coefficient is randomly obtained from the randomness coefficient range; Obtain the second product between the target randomness coefficient and the parameter difference, and determine the second product as the adjustment degree corresponding to the candidate model parameters.

8. The method according to claim 1, characterized in that, The step of determining the random sampling interval of parameters corresponding to the initial prediction model based on the candidate model parameters and the adjustment degree, and determining the target model parameters corresponding to the initial prediction model from the random sampling interval of parameters, includes: The candidate model parameters are determined as the parameter mean, the adjustment degree is determined as the standard deviation, and a normal distribution associated with the candidate model parameters is determined based on the parameter mean and the standard deviation. The normal distribution is then determined as the random sampling interval of the parameters corresponding to the initial prediction model. Random sampling is performed within the parameter random sampling interval to obtain the target model parameters corresponding to the initial prediction model.

9. The method according to claim 1, characterized in that, The step of updating the parameters of the initial prediction model according to the target model parameters, and determining the initial prediction model that satisfies the convergence condition as the target prediction model when the updated initial prediction model meets the convergence condition, includes: Based on the target model parameters, the initial model parameters in the initial prediction model are updated to obtain the parameter-updated initial prediction model; If the adjustment degree corresponding to the initial prediction model after parameter update is the target adjustment threshold, then the initial prediction model after parameter update is determined to meet the convergence condition, and the initial prediction model after parameter update that meets the convergence condition is determined as the target prediction model.

10. The method according to claim 1, characterized in that, The method further includes: Obtain the target user characteristics corresponding to the target user, and obtain the object characteristics corresponding to each object to be recommended in the set of objects to be recommended; The target user features and the object features corresponding to each object to be recommended are input into the target prediction model, and the recommendation probability corresponding to each object to be recommended is output by the target prediction model. Based on the recommendation probability corresponding to each object to be recommended, a user recommendation object is determined from the set of objects to be recommended to the target user, and the user recommendation object is sent to the target user terminal to which the target user belongs, so that the target user terminal outputs the user recommendation object.

11. A data processing apparatus, characterized in that, include: The first acquisition module is used to acquire a set of sample recommendation objects; The sample recommendation object set includes N candidate recommendation objects, where N is a positive integer; The prediction module is used to input the object features corresponding to each of the N candidate recommendation objects and the user features corresponding to the recommended user into the initial prediction model, and predict the target recommendation object to be recommended to the recommended user through the initial prediction model; the N candidate recommendation objects include the target recommendation object; The first determining module is used to determine the candidate model parameters and the adjustment degree corresponding to the candidate model parameters based on the user feedback tags corresponding to the target recommendation object and the predicted recommendation features corresponding to the target recommendation object. The degree of adjustment is random; The first determining module includes: The fourth determining unit is used to determine the adjustment degree corresponding to the candidate model parameters based on the candidate model parameters and the initial model parameters in the initial prediction model; The fourth determining unit includes: Select sub-units to randomly select target update coefficients from the update coefficient range; Obtain a sub-unit, used to obtain the first product of the target update coefficient and the initial model parameters; A subunit is defined for adjusting the candidate model parameters based on the first product to obtain updated candidate model parameters; obtaining the parameter difference between the updated candidate model parameters and the candidate model parameters; and determining the adjustment degree corresponding to the candidate model parameters based on the parameter difference. The second determining module is used to determine the parameter random sampling interval corresponding to the initial prediction model based on the candidate model parameters and the adjustment degree, and to determine the target model parameters corresponding to the initial prediction model from the parameter random sampling interval; The parameter update module is used to update the parameters of the initial prediction model according to the target model parameters. When the initial prediction model after parameter update meets the convergence condition, the initial prediction model that meets the convergence condition is determined as the target prediction model. The target prediction model is used to predict recommended objects for users.

12. A computer device, characterized in that, include: Processor and memory; The memory stores a computer program, which, when executed by a processor, performs the method of any one of claims 1 to 10.

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

14. A computer program product characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium, which, when read and executed by a processor of a computer device, cause the computer device to perform the method of any one of claims 1 to 10.

Images