User preference prediction model training method and related device

Abstract

The embodiment of the invention discloses a user preference prediction model training method, a user preference prediction method and related devices. Each group of training sample data comprises feature portrait data of a user account and preference data of the user account, and the preference data comprises a preference label of the user for each function of the application program, and obtaining co-occurrence information among multiple functions in each function combination of the application program, the initial prediction model outputs a preference prediction result of the user account for each function of the application program according to the feature portrait data and the co-occurrence information, and the initial prediction model is trained according to the preference prediction results and the preference labels to obtain a target prediction model. The target prediction model can predict the preference condition of the user for multiple functions of the application program by only reading the portrait data of the user once, so that computing resources are saved; the model can mine the potential preferences of the user for other functions based on the correlation between the learned functions, so that the preference prediction is more accurate and comprehensive.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of data processing, and specifically to a user preference prediction model training method, a user preference prediction method, and related devices. Background Art

[0002] Applications provide users with a variety of functions, and users can use various functions to obtain different user experiences. For example, music software provides users with multiple functions such as listening to songs, configuring sound effects, socializing, and playing games, so that users can configure song sound effects while listening to songs, or interact with music friends, play games, etc.

[0003] The existing solution trains a binary classification model for each function of the application, and the binary classification model identifies whether the user likes this function based on the user portrait data. However, this method requires maintaining a binary classification model for each function. When identifying the user's preference for multiple functions of the application, multiple binary classification models corresponding to multiple functions need to read the user portrait data separately, that is, repeatedly read the user portrait data multiple times, which greatly wastes data computing resources.

[0004] Secondly, the training and application processes of multiple binary classification models are independent of each other, ignoring the correlation between multiple functions. For example, users who like to use the sound effect function often have a strong preference for the sound quality function. However, each binary classification model only performs model training and identification of the model application process for the user portrait data corresponding to a single function, resulting in low accuracy of the training and recognition effects of the binary classification model, which affects user preference prediction. Summary of the invention

[0005] The embodiments of the present application provide a user preference prediction model training method, a user preference prediction method and related devices for improving the accuracy of user preference function prediction.

[0006] A first aspect of an embodiment of the present application provides a method for training a user preference prediction model, the method comprising:

[0007] Acquire multiple sets of training sample data, each set of the training sample data includes feature portrait data of a user account and preference data of the user account, the preference data includes a preference label of the user account for a function of the application program; the preference label is used to indicate the preference of the user account for the function of the application program;

[0008] Obtaining co-occurrence information between multiple functions in each function combination of the application, wherein the co-occurrence information is obtained based on preference data of each user account for multiple functions in the function combination;

[0009] Obtaining an initial prediction model, and inputting the multiple sets of training sample data and the co-occurrence information into the initial prediction model, so that the initial prediction model outputs a user account's preference prediction result for each function of the application according to the feature portrait data and the co-occurrence information;

[0010] The model parameters of the initial prediction model are adjusted according to the preference prediction results and the preference labels, and the model training is stopped until the convergence conditions of the model training are met to obtain a target prediction model; the target prediction model is used to predict the preference of any user account for any function of the application.

[0011] A second aspect of an embodiment of the present application provides a method for predicting user preferences, the method comprising:

[0012] Obtaining a target prediction model, wherein the target prediction model is obtained by training a plurality of sets of training sample data and co-occurrence information between multiple functions in each function combination of the application program using a machine learning algorithm; wherein each set of the training sample data includes feature portrait data and preference data of the user account, wherein the preference data includes a preference label of the user account for each function of the application program; the preference label is used to indicate the preference of the user account for the function of the application program; and the co-occurrence information is used to indicate that the user account prefers each function in the function combination;

[0013] Acquire characteristic portrait data of the target user account, and input the characteristic portrait data of the target user account into the target prediction model, so that the target prediction model outputs the preference of the target user account for each function of the application according to the characteristic portrait data of the target user account.

[0014] A third aspect of an embodiment of the present application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the method of the first aspect or the second aspect when executing the computer program.

[0015] A fourth aspect of an embodiment of the present application provides a computer storage medium, in which instructions are stored. When the instructions are executed on a computer, the computer executes the method of the first aspect or the second aspect.

[0016] A fifth aspect of the embodiments of the present application provides a computer program product. When the computer program product is run on a computer device, the computer device executes the method of the first aspect or the second aspect.

[0017] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0018] Each set of training sample data includes the feature portrait data of the user account and the preference data of the user account. The preference data includes the user's preference label for each function of the application, and obtains the co-occurrence information between multiple functions in each function combination of the application. The co-occurrence information is obtained based on the preference data statistics of each user account for multiple functions in the function combination. Multiple sets of training sample data and the co-occurrence information are input into the initial prediction model, and the initial prediction model outputs the user account's preference prediction results for each function of the application based on the feature portrait data and the co-occurrence information, and trains the initial prediction model based on the preference prediction results and the preference labels to obtain the target prediction model. Compared with using multiple binary classification models to predict user preferences, the target prediction model can predict the user's preferences for multiple functions of the application by reading the user's portrait data only once, saving computing resources, and the model can learn the correlation between multiple functions from the co-occurrence information of multiple functions, so that based on the learned correlation between functions, the user's potential preference for other functions can be mined, making the preference prediction more accurate and comprehensive. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic diagram of the network framework in the embodiment of the present application;

[0020] Figure 2 A flowchart of a method for training a user preference prediction model in an embodiment of the present application;

[0021] Figure 3 This is an exemplary schematic diagram of a co-occurrence frequency matrix corresponding to multiple function combinations in an embodiment of the present application;

[0022] Figure 4 A flowchart of a method for predicting user preferences in an embodiment of the present application;

[0023] Figure 5 This is an exemplary schematic diagram of a display method of a function of displaying a target prediction model predicting a target user account preference on a client interface in an embodiment of the present application;

[0024] Figure 6 It is a structural schematic diagram of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The embodiments of the present application provide a user preference prediction model training method, a user preference prediction method and related devices for improving the accuracy of user preference function prediction.

[0026] See also Figure 1 In the embodiment of the present application, the network framework includes:

[0027] The service server 100 and the terminal cluster; the terminal cluster may include: terminal device 200a, terminal device 200b, terminal device 200c, ..., terminal device 200n and other terminal devices.

[0028] The business server 100 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud database, cloud service, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, CDN, and big data and artificial intelligence platform. The terminal device (including terminal device 200a, terminal device 200b, terminal device 200c, ..., terminal device 200n) may be a smart phone, tablet computer, laptop computer, desktop computer, PDA, mobile internet device (MID), wearable device (such as smart watch, smart bracelet, etc.), smart computer, smart car and other smart terminals.

[0029] Among them, the business server 100 can establish a communication connection with each terminal device in the terminal cluster, and a communication connection can also be established between each terminal device in the terminal cluster. In other words, the business server 100 can establish a communication connection with each terminal device in the terminal device 200a, terminal device 200b, terminal device 200c, ..., terminal device 200n. For example, a communication connection can be established between the terminal device 200a and the business server 100. A communication connection can be established between the terminal device 200a and the terminal device 200b, and a communication connection can also be established between the terminal device 200a and the terminal device 200c. Among them, the above-mentioned communication connection does not limit the connection method, and can be directly or indirectly connected through a wired communication method, or can be directly or indirectly connected through a wireless communication method, etc., which can be determined according to the actual application scenario, and this application is not limited here.

[0030] It should be understood that Figure 1Each terminal device in the terminal cluster shown can be installed with an application client. When the application client runs in each terminal device, it can interact with the business server 100 respectively, so that the business server 100 can receive business data from each terminal device (such as user identity data uploaded by the user through the terminal device). Among them, the application client can be a music playback application, a karaoke software application, a browser application, a social application, an instant messaging application, a live broadcast application, a game application, a short video application, a video application, a shopping application, a novel application, a payment application, etc., which has the function of displaying text, images, audio, and video and other data information. The specific application can be determined according to the actual application scenario requirements, and no limitation is made here. Among them, the application client can be an independent client or an embedded sub-client integrated in a client (such as a music application, a karaoke software application, etc.). The specific application can be determined according to the actual application scenario, and no limitation is made here.

[0031] The following will be combined Figure 1 The network framework of the present invention is used to describe the user preference prediction model training method in the embodiment of the present application:

[0032] The user preference prediction model training method and the user preference prediction method of the embodiments of the present application can be applied to recommendation scenarios of functions of various application programs, such as recommendation scenarios of various functions of music software. The music software can provide multiple functions, and the user's preferences for the multiple functions of the music software can be predicted based on the method of the embodiments of the present application. Then, when the user uses the music software, the user's preferred functions can be recommended on the UI interface of the music software. For example, the user's preferred functions can be displayed in the form of a pop-up window when the user enters the music software, or the points corresponding to the various functions can be displayed when the user searches for songs in the music software, and the user can click the point corresponding to a function to enter the function.

[0033] Of course, the method of the embodiment of the present application can also be applied to any other application program besides music software, and is not limited here.

[0034] See also Figure 2 In the embodiment of the present application, one embodiment of the user preference prediction model training method includes:

[0035] 201. Acquire multiple sets of training sample data, each set of the training sample data includes feature portrait data of a user account and preference data of the user account, the preference data includes a preference label of the user account for a function of an application program; the preference label is used to indicate the preference of the user account for a function of the application program;

[0036] The method of this embodiment can be applied to a computer device, which can be Figure 1The business server 100 or each terminal device in the network framework shown. The computer device can obtain multiple sets of training sample data, each set of training sample data includes feature portrait data of the user account held by the user and preference data of the user account, and the preference data is used to indicate the preference information of the user account for each function of the application, for example, it may include the preference label of the user account for each function of the application, and the preference label is used to indicate the preference of the user account for the function of the application. For example, if the user account prefers function A of the application, that is, the user account is interested in function A, the number "1" can be used to indicate the preference label of the user account for function A; if the user account does not prefer function B of the application, that is, the user account is not interested in function B of the application, the number "0" can be used to indicate the preference label of the user account for function B.

[0037] It should be noted that the preference prediction and analysis of the feature portrait data of the user account and the preference data of the user account described in this embodiment is equivalent to the preference prediction and analysis of the user holding the user account. The same applies to the user account mentioned later. Any reference to the user account refers to the user holding the user account.

[0038] The feature profile data of a user account includes data used to describe any features of the user account, so as to construct a user account profile, such as data describing the behavioral features of the user account, data describing the attributes of the user account such as age, gender, education, occupation, etc. The feature profile data and preference data included in each set of training sample data are for the same user account, such as one set of training sample data includes the feature profile data of user account A and the preference data of user account A, another set of training sample data includes the feature profile data of user account B and the preference data of user account B, and so on.

[0039] 202. Obtaining co-occurrence information between multiple functions in each function combination of the application, wherein the co-occurrence information is obtained based on preference data of each user account for multiple functions in the function combination;

[0040] The co-occurrence information of a function combination refers to the preference of a user account for multiple functions in the function combination. The co-occurrence information can be obtained based on the preference data of each user account for multiple functions in the function combination. For example, the co-occurrence information can indicate that the user account has a preference for multiple functions in the function combination, or indicates that the user account has no preference for multiple functions in the function combination, or indicates that the user account has a preference for some functions in the function combination but not for other functions. Therefore, the co-occurrence information indicates the correlation between multiple functions of the application, and the correlation can characterize the degree of correlation between the user's preferences for multiple functions. For example, for music software, a user who likes to use the sound effect function often has a strong preference for the sound quality function, and the sound effect function has a strong correlation with the sound quality function.

[0041] Therefore, a user account may have a preference for multiple related functions or no preference at all. This correlation information can be used to deduce the user account's potential preference for the application's functions. For example, if a user account prefers function A of the application, and the co-occurrence information indicates that there is a certain correlation between function A and function B of the application, it can be deduced that the user account may also have a potential preference for function B.

[0042] 203. Obtain an initial prediction model, and input the multiple sets of training sample data and the co-occurrence information into the initial prediction model, so that the initial prediction model outputs a user account preference prediction result for each function of the application according to the feature portrait data and the co-occurrence information;

[0043] 204. Adjust the model parameters of the initial prediction model according to the preference prediction result and the preference label, stop the model training until the convergence condition of the model training is met, and obtain a target prediction model; the target prediction model is used to predict the preference of any user account for any function of the application;

[0044] The computer device may obtain an initial prediction model so as to train the initial prediction model and use the trained model to predict the preferences of the user account. In this embodiment, the initial prediction model may be a multi-task learning network model, such as a multi-gate mixture of experts (MMOE).

[0045] The initial prediction model outputs the user account's preference prediction results for each function of the application based on the feature portrait data of the user account and the correlation between multiple functions described by the co-occurrence information, and adjusts the model parameters of the initial prediction model based on the preference prediction results and the preference labels corresponding to each function. The model parameters are updated multiple times after multiple rounds of iterative training. In multiple rounds of iterative training, the model continuously learns the correspondence between the user account feature portrait and the user account's preferred functions, and learns the correlation between multiple functions and applies them to the user account's preference prediction, so that the prediction results output by the model gradually approach or even become the same as the actual results of the preference labels, until the model training is stopped when the convergence conditions are met, and the target prediction model can be obtained. Therefore, the target prediction model can be used to predict the preference of any user account for any function of the application based on the knowledge and prediction ability obtained from model training.

[0046] Among them, the condition for model training convergence can be that the number of model iteration training reaches a preset number, or that the loss value of the loss function constructed by model training meets the preset value or the loss value tends to be stable. This embodiment does not limit the condition for model training convergence.

[0047] In this embodiment, each set of training sample data includes feature portrait data of a user account and preference data of the user account. The preference data includes the user's preference label for each function of the application, and the co-occurrence information between multiple functions in each function combination of the application is obtained. The co-occurrence information is obtained based on the preference data statistics of each user account for multiple functions in the function combination. Multiple sets of training sample data and the co-occurrence information are input into the initial prediction model, and the initial prediction model outputs the user account's preference prediction result for each function of the application based on the feature portrait data and the co-occurrence information, and the initial prediction model is trained based on the preference prediction result and the preference label to obtain the target prediction model. Compared with using multiple binary classification models to predict user preferences, the target prediction model can predict the user's preferences for multiple functions of the application by reading the user's portrait data only once, saving computing resources, and the model can learn the correlation between multiple functions from the co-occurrence information of multiple functions, so that based on the learned correlation between functions, the user's potential preference for other functions can be mined, making the preference prediction more accurate and comprehensive.

[0048] based on Figure 2In the illustrated embodiment, in a preferred implementation manner, the characteristic portrait data of the user account in each set of training sample data includes the access path usage data of the user account to the application function, the operation behavior data of the user account on the application, and the personal information data of the user account. Therefore, when the model outputs the preference prediction result of the user account for each function of the application based on the set of training sample data, the initial prediction model can extract the behavior characteristics of the user account from the access path usage data and the operation behavior data, extract the attribute characteristics of the user account from the personal information data, and output the preference prediction result of the user account for each function of the application based on the behavior characteristics, the attribute characteristics and the co-occurrence information.

[0049] Among them, the access path of the application function refers to the operation mode and operation method of the user account to enter the application function. For example, for the listening function in the music software, it can be entered in a variety of ways, such as clicking the listening button at the bottom of the homepage of the music software to listen to the music, or clicking the playlist to listen to the music, or searching for songs to listen to the music, or clicking the songs recommended on the homepage to listen to the music, etc. The access usage data can indicate the usage of each access path of the application function by the user account, and can reflect some preferences of the user in using the application.

[0050] The user account's operation behavior data on the application includes data on any behavior of the user account using the application, such as the user account's click operation on each control or button of the application, the user account's click to enter various functions, etc. For example, for music software, the user account's operation behavior data on the music software may include data on multiple behaviors such as the user's listening behavior, the user account's search for songs, or the user account's click to enter various functions. These behavior data can also reflect some preferences and habits of the user account in using the application, and can be used as a reference for predicting user preferences.

[0051] The personal information data of a user account may include the personal information of the user account, such as the user's age, gender, education, occupation and other personal attribute data. These data are also related to the user's preferences. For example, for music software, older users may prefer to use its social functions, while younger users may prefer to use its built-in mini-game functions; music creators may prefer to use the music score function of music software. The personal information data of a user account is also related to the user's preference for application functions, and can therefore be used as a reference for predicting user preferences.

[0052] Among them, the network structure used in the initial prediction model to extract the behavioral characteristics and attribute characteristics of the user account from the feature portrait data of the user account can be a model structure such as a convolutional neural network, a recurrent neural network or a Transformer network structure, which is not limited here.

[0053] After extracting the behavior characteristics and attribute characteristics of the user account, the user's preference prediction results for multiple functions of the application can be output in combination with the co-occurrence information. Specifically, the neuron weights used to predict the user account's preference for each function of the application in the initial prediction model can be set based on the co-occurrence information, and the user account's preference prediction results for each function of the application can be output based on the set neuron weights and the behavior characteristics and the attribute characteristics.

[0054] For example, the co-occurrence information of the function combination can be expressed as a co-occurrence frequency. The co-occurrence frequency can be calculated by counting the number of user accounts m whose preference labels for each function of the function combination in the preference data of n user accounts are the same, and calculating m / n as the co-occurrence frequency of the function combination. For example, a function combination is function 1 and function 2 of a music software. In the preference data of 100 user accounts, the number of user accounts whose preference labels for function 1 and function 2 are the same is counted for these 100 user accounts, for example, the number of user accounts whose labels are all "1" (indicating that both function 1 and function 2 are preferred) is counted. Assuming that there are 60, that is, these 60 user accounts all prefer function 1 and function 2, and are all interested in these two functions, then the co-occurrence frequency of this function combination is 60%. The co-occurrence frequencies of multiple function combinations can constitute a co-occurrence frequency matrix.

[0055] Assume that an application has multiple functions such as A, B, C, and D. Among these functions, there are combinations of "A + D" and "B + D". The co-occurrence frequency of the combination of "A + D" is 0.6, and the co-occurrence frequency of the combination of "B + D" is 0.1. Then, the co-occurrence frequency matrix of these multiple functions can be constructed, such as Figure 3As shown, taking the first row of the co-occurrence frequency matrix as an example, the model learns the correlation of the combination of "Function A + Function D" based on the co-occurrence frequency of this row, that is, the model inputs the co-occurrence frequency in the last layer of the network (such as the fully connected layer), which is equivalent to setting the neuron weights used to predict the user's preference for Function B and Function C to 0 in this layer, and setting the neuron weights used to predict the user's preference for Function A and Function D in this layer to 0.6, and then combining the behavioral features and attribute features extracted by the previous layers of the network structure (such as the convolutional layer structure), and based on the set neuron weights, predict the user's preference results for Function A, Function B, Function C and Function D, and output the preference prediction results. Similarly, the model can learn the correlation between multiple functions in other function combinations based on other co-occurrence frequencies in the co-occurrence frequency matrix.

[0056] Therefore, by setting neuron weights according to the co-occurrence frequency matrix, the correlation of functions can be introduced into the model, so that the model can learn the correlation between multiple functions and make user preference predictions based on this. This can more comprehensively predict users' potential function preferences, improve the accuracy of user preference predictions, and thus improve users' experience with the application.

[0057] based on Figure 2 In another preferred implementation of the illustrated embodiment, when obtaining the user's preference data, for each function of the application, the behavior data of multiple user accounts for the function can be obtained, and it is determined whether the behavior data of each user account for the function contains data corresponding to the preference behavior. If the behavior data of the user account for the function contains data corresponding to the preference behavior, the preference label of the user account for the function is determined to be a first label, and the first label is used to indicate that the user account prefers the function. If the behavior data of the user account for the function does not contain data corresponding to the preference behavior, the preference label of the user account for the function is determined to be a second label, and the second label is used to indicate that the user account does not prefer the function.

[0058] For example, a computer device may collect user usage data of an application to construct multiple sets of training sample data, such as collecting user behavior flow data to determine whether the user has a preferred behavior for each function of the application. Therefore, the computer device may obtain behavior data of multiple users for each function of the application, and determine whether each user's behavior data for the function contains data corresponding to the preferred behavior. If the user's behavior data for the function contains data corresponding to the preferred behavior, the user's preference label for the function is determined to be the first label; if the user's behavior data for the function does not contain data corresponding to the preferred behavior, the user's preference label for the function is determined to be the second label.

[0059] Among them, the preference behavior is used to characterize the behavior of the user being interested in the functions of the application. Specifically, a judgment condition can be preset. When the user's behavior meets this judgment condition, it is determined as a preference behavior; when it does not meet this judgment condition, it is determined as not being a preference behavior. For example, a music software provides a medal function. The medal function can generate matching medals for the user according to the user's music listening behavior and preferences. It can be preset that the preference behavior of this function is that the user clicks the button to wear the medal. When the behavior data includes the behavior that the user clicks the button to wear the medal, it can be determined that the user has a preference behavior for this medal function; or it can be preset that the preference behavior of this function is operations such as the user upgrading the medal or the user changing the skin of the medal. Then, when the behavior data includes the user upgrading the medal or includes the user changing the skin of the medal, it can be determined that the user has a preference behavior for this medal function. The preference behavior can be set according to the specific situation of the functions of the application, which is not limited here.

[0060] Therefore, in this way, the preference situations of each user for multiple functions of the application can be counted, and then the preference data of each user can be formed. For example, an example of the storage structure of a kind of preference data can be: user ID 丨 preference labels of function 1, preference labels of function 2,..., preference labels of function n.

[0061] Furthermore, in some optional ways, for each function of the application, the number of positive samples with the preference label of the first label for this function can be counted. Since there may be differences in the number of interested users among different functions. For example, there are many users interested in function 1, while there are few users interested in function 2. This difference will lead to too large an association difference between the functions learned during model training and the user characteristics, resulting in prediction deviation. In order to reduce the prediction deviation caused by the too large difference in the user magnitude of the preference behaviors among functions, the positive samples can be preprocessed according to the number of positive samples. If the number of positive samples is greater than the first preset number, undersampling processing is performed on the positive samples to reduce the number of positive samples of this function; if the number of positive samples is less than the second preset number, oversampling processing is performed on the positive samples to increase the number of positive samples of this function. Among them, the first preset number is greater than or equal to the second preset number. The positive samples obtained by undersampling processing or oversampling processing are used as the preference data of this function.

[0062] Therefore, by performing different sampling methods on the positive samples according to the number of positive samples, the data volume of the preference data among multiple functions can be balanced, the accuracy of model training and the prediction accuracy can be improved, and the poor model training effect and prediction deviation caused by the too large difference in the number of positive samples among multiple functions can be avoided.

[0063] Since each set of training sample data corresponds to a user account, the model can process a set of training sample data to output the preference prediction results of the user account corresponding to the set of training sample data, and the preference prediction results include the preference prediction results of the user account for each function of the application. Similarly, the model can process multiple sets of training sample data separately to output the preference prediction results of multiple users for multiple functions of the application. Figure 2 In another preferred implementation of the embodiment shown, when adjusting the model parameters of the initial prediction model according to the preference prediction results and the preference labels, for each function of the application, a loss function of the function can be constructed based on the preference prediction results of multiple user accounts for the function obtained by predicting multiple sets of training sample data and multiple preference labels about the function in multiple sets of training sample data. When the loss function corresponding to each function is constructed one by one, the model parameters of the initial prediction model can be further adjusted according to the loss values ​​of multiple loss functions of multiple functions of the application.

[0064] For example, the loss function Loss corresponding to each function of the application can be expressed as follows:

[0065]

[0066] Where N is the number of samples in the training sample data, i refers to the i-th sample, and p(y i ) refers to the prediction result of the user's preference for a function of the application predicted by the model, y i Refers to the preference label corresponding to this function.

[0067] This formula represents the loss function of a function, that is, the loss function of a function can be constructed according to this formula, and the loss function corresponding to each function of the application can be constructed based on this formula, and multiple loss functions corresponding to multiple functions can be obtained. The adjustment of model parameters is based on the loss values ​​of multiple loss functions corresponding to multiple functions of the application. With model training and parameter adjustment, the loss values ​​of multiple loss functions corresponding to multiple functions of the application can gradually decrease, indicating that the results predicted by the model are gradually approaching the real results and tend to be the same, and the model is constantly converging.

[0068] Therefore, by constructing loss functions corresponding to multiple functions and adjusting model parameters based on multiple loss functions, the model's training for preference prediction of each function can converge accurately, laying a foundation for the model to accurately predict user preferences for each function and improving the model's preference prediction ability.

[0069] After the target prediction model is trained, the target prediction model can be used to predict any user's preference for any function of the application. Therefore, the embodiment of the present application also proposes the application scenario and application process of the target prediction model. For details, please refer to Figure 4 In the embodiment of the present application, an embodiment of the method for predicting user preferences includes:

[0070] 401. Obtain a target prediction model, wherein the target prediction model is obtained by training a plurality of sets of training sample data and co-occurrence information between multiple functions in each function combination of the application program using a machine learning algorithm; wherein each set of the training sample data includes feature profile data and preference data of the user account, wherein the preference data includes a preference label of the user account for each function of the application program; the preference label is used to indicate the preference of the user account for the function of the application program; and the co-occurrence information is used to indicate that the user account prefers each function in the function combination;

[0071] The method of this embodiment can be applied to a computer device, which can be Figure 1 The target prediction model can be based on the business server 100 or each terminal device in the network framework shown. Figure 2 The embodiment shown and its various preferred implementations are obtained by training using the user preference prediction model training method.

[0072] The feature portrait data of the user in each set of training sample data and the explanation of the user's preference data have been described in detail in the previous text and will not be repeated here.

[0073] 402. Acquire feature portrait data of a target user account, and input the feature portrait data of the target user account into the target prediction model, so that the target prediction model outputs the preference of the target user account for each function of the application according to the feature portrait data of the target user account;

[0074] The target user is the user for whom preference prediction is required. Feature portrait data of the target user can be obtained and input into the target prediction model, so that the target prediction model extracts user portrait features such as the target user's behavioral characteristics and attribute characteristics from the target user's feature portrait data, and outputs the target user's preference for each function of the application based on the user portrait features such as the target user's behavioral characteristics and attribute characteristics and the co-occurrence information that characterizes the correlation between multiple functions of the application. The preference can be the degree of interest.

[0075] In the process of training multiple sets of training sample data by machine learning algorithms, the target prediction model continuously learns the correspondence between user portrait features such as user behavior characteristics and user attribute characteristics and preference labels. When the model training converges, it learns the accurate representation of the correspondence, and then can accurately predict the user's preference for application functions based on the learned knowledge and prediction ability.

[0076] Therefore, in a preferred embodiment, the target prediction model can output the target user account's preference score for each function of the application based on the target user account's feature portrait data based on the prediction ability obtained by model training, and sort the target user account's multiple preference scores for multiple functions of the application, and output the sorting result. This sorting result can be output to the downstream portrait platform for the business party to call, indicating the functions that the user prefers and the functions that the user does not prefer. For example, the data output example of the sorting result can be: <user id|TOP1 function name_corresponding preference score, TOP2 function name_corresponding preference score, ...>.

[0077] After determining the user's preferred functions, the preferred functions can be displayed on the interface, such as Figure 5 As shown, for example, when a user uses a music software client and performs an in-terminal search, such as searching for functions, music or other information in the terminal, the plurality of functions predicted above can be displayed for recommendation to the user, such as displaying a plurality of user-preferred functions such as "new sound effects", "simple mode" and "radar mode", so that the user can quickly access the desired function without searching, thereby improving the user experience of the client.

[0078] Therefore, in this embodiment, compared with using multiple binary classification models to predict user preferences, the target prediction model can predict the user's preferences for multiple functions of the application by reading the user's portrait data only once, saving computing resources. At the same time, combined with the co-occurrence information used to characterize the correlation between multiple functions for prediction, it is possible to explore the user's potential preferences for other functions, making preference predictions more accurate and comprehensive.

[0079] The computer device in the embodiment of the present application is described below. Figure 6 In the embodiment of the present application, one embodiment of the computer device includes:

[0080] The computer device 600 may include one or more central processing units (CPU) 601 and a memory 605 , wherein the memory 605 stores one or more application programs or data.

[0081] The memory 605 may be a volatile storage or a persistent storage. The program stored in the memory 605 may include one or more modules, each of which may include a series of instruction operations in the computer device. Furthermore, the central processing unit 601 may be configured to communicate with the memory 605 and execute a series of instruction operations in the memory 605 on the computer device 600.

[0082] The computer device 600 may also include one or more power supplies 602, one or more wired or wireless network interfaces 603, one or more input and output interfaces 604, and / or one or more operating systems, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0083] The CPU 601 can execute the aforementioned Figure 2 The operations performed by the computer device in the illustrated embodiment and its various preferred implementations are not described in detail here.

[0084] The present application also provides a computer storage medium, wherein one embodiment includes: the computer storage medium stores instructions, and when the instructions are executed on a computer, the computer executes the aforementioned Figure 2 The operations performed by the computer device in the illustrated embodiment and its various preferred implementations.

[0085] The present application also provides a computer program product, wherein one embodiment includes: when the computer program product is run on a computer device, the computer device executes the aforementioned Figure 2 The operations performed by the computer device in the illustrated embodiment and its various preferred implementations.

[0086] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0087] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0088] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0089] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0090] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), disk or optical disk and other media that can store program code.

Claims

1. A user preference prediction model training method, characterized in that: The method comprises: Acquire multiple sets of training sample data, each set of the training sample data includes feature portrait data of a user account and preference data of the user account, the preference data includes a preference label of the user account for a function of the application program; the preference label is used to indicate the preference of the user account for the function of the application program; Obtaining co-occurrence information between multiple functions in each function combination of the application, wherein the co-occurrence information is obtained based on preference data of each user account for multiple functions in the function combination; Obtaining an initial prediction model, and inputting the multiple sets of training sample data and the co-occurrence information into the initial prediction model, so that the initial prediction model outputs a user account's preference prediction result for each function of the application according to the feature portrait data and the co-occurrence information; The model parameters of the initial prediction model are adjusted according to the preference prediction results and the preference labels, and the model training is stopped until the convergence conditions of the model training are met to obtain a target prediction model; the target prediction model is used to predict the preference of any user account for any function of the application.

2. The method according to claim 1, characterized in that The feature profile data includes usage data of the user account's access to the application's functions, data on the user account's operation behavior on the application, and personal information data of the user account; The initial prediction model outputs a user account preference prediction result for each function of the application according to the feature portrait data and the co-occurrence information, including: The initial prediction model extracts the behavioral characteristics of the user account from the path usage data and the operation behavior data, extracts the attribute characteristics of the user account from the personal information data, and outputs the preference prediction result of the user account for each function of the application based on the behavioral characteristics, the attribute characteristics and the co-occurrence information.

3. The method according to claim 2, characterized in that The outputting the preference prediction result of the user account for each function of the application according to the behavior feature, the attribute feature and the co-occurrence information includes: The neuron weights in the initial prediction model for predicting the user account's preference for each function of the application are set based on the co-occurrence information, and the preference prediction results of the user account for each function of the application are output based on the set neuron weights and the behavioral characteristics and the attribute characteristics.

4. The method according to claim 1, characterized in that: The step of obtaining the preference data includes: For each function of the application, obtain behavior data of multiple user accounts for the function; Determine respectively whether the behavior data of each user account for the function contains data corresponding to the preferred behavior; If the behavior data of the user account for the function includes data corresponding to the preferred behavior, determining that the preference label of the user account for the function is a first label; the first label is used to indicate that the user account prefers the function; If the user account's behavior data for the function does not include data corresponding to the preferred behavior, it is determined that the user account's preference label for the function is a second label; the second label is used to indicate that the user account does not prefer the function.

5. The method according to claim 4, characterized in that The method further comprises: For each function of the application, counting the number of positive samples whose preference label for the function is the first label; If the number of the positive samples is greater than a first preset number, the positive samples are undersampled; if the number of the positive samples is less than a second preset number, the positive samples are oversampled; the first preset number is greater than or equal to the second preset number; The positive samples obtained by the undersampling process or the oversampling process are used as the preference data for the function.

6. The method according to any one of claims 1 to 5, characterized in that: The adjusting the model parameters of the initial prediction model according to the preference prediction result and the preference label includes: For each function of the application, construct a loss function for the function based on the preference prediction results of the multiple user accounts for the function obtained by predicting the multiple sets of training sample data and the multiple preference labels about the function in the multiple sets of training sample data; The model parameters of the initial prediction model are adjusted according to the loss values ​​of the multiple loss functions of the multiple functions of the application.

7. A method for predicting user preferences, characterized in that: The method comprises: Obtaining a target prediction model, wherein the target prediction model is obtained by training a plurality of sets of training sample data and co-occurrence information between multiple functions in each function combination of the application program using a machine learning algorithm; wherein each set of the training sample data includes feature portrait data and preference data of the user account, wherein the preference data includes a preference label of the user account for each function of the application program; the preference label is used to indicate the preference of the user account for the function of the application program; and the co-occurrence information is used to indicate that the user account prefers each function in the function combination; Acquire characteristic portrait data of the target user account, and input the characteristic portrait data of the target user account into the target prediction model, so that the target prediction model outputs the preference of the target user account for each function of the application according to the characteristic portrait data of the target user account.

8. The method according to claim 7, characterized in that The target prediction model outputs the preference of the target user account for each function of the application according to the feature portrait data of the target user account, including: The target prediction model outputs the target user account's preference score for each function of the application based on the feature portrait data of the target user account, sorts the target user account's multiple preference scores for multiple functions of the application, and outputs the sorting result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

10. A computer storage medium, characterized in that: The computer storage medium stores instructions, and when the instructions are executed on a computer, the computer is caused to perform the method according to any one of claims 1 to 8.

Images