Content recommendation method and device and storage medium
By using three neural networks to process domain indicators and features in a prediction model, combining private normalization and batch normalization, efficient prediction of multi-domain content recommendation is achieved, solving the time-consuming and labor-consuming multi-model Problem, reduces costs.
Patent Information
- Application Number
- CN202310594137.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-24
- Publication Date
- 2024-11-26
AI Technical Summary
In the prior art, multi-field content recommendation models are time-consuming and labor-intensive, and training models for each domain alone will lead to high labor and resource costs, making it difficult to reduce costs without reducing the prediction effect.
Using a prediction model, three neural networks jointly predict the domain indicator, the recommended domain specific features and the shared features of multiple domains, using a multi-layer perceptron and S-type growth curve function to process the features, combined with private normalization and batch normalization process to generate target prediction parameters for content recommendation.
On the basis of not reducing the prediction effect, the cost of manpower and resource is reduced, the problem of time-consuming and labor-consuming for multiple models is solved, and cost reduction and efficiency improvement is achieved.
Smart Images

Figure CN119025742A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of big data technology, and in particular to a content recommendation method, device and storage medium. Background Art
[0002] In the related art, the classic recall-rank method can be used to recommend content to users. When the recall-rank model is used to determine the target content to be recommended, one model can be used to predict different fields, or a separate model can be trained for each field to predict.
[0003] Among them, using one model for prediction in different fields may lead to insufficient prediction accuracy. On the other hand, training a separate model for each field for prediction will greatly increase the prediction cost. Summary of the invention
[0004] In order to overcome the problems existing in the related art, the present disclosure provides a content recommendation method, device and storage medium.
[0005] According to a first aspect of an embodiment of the present disclosure, a content recommendation method is provided, including:
[0006] Acquire a recommended domain, and determine a domain indicator based on the recommended domain, wherein the domain includes an area where content is displayed;
[0007] Acquire a user identifier and historical data of the user's operations on the content; determine specific features and shared features based on the historical data, wherein the specific features include specific features corresponding to the recommended field, and the shared features include shared features corresponding to multiple fields;
[0008] inputting the domain indicator into a first neural network, and obtaining a first prediction parameter using the first neural network;
[0009] Inputting the domain indicator and the specific feature into a second neural network, inputting the shared feature into a third neural network, and obtaining a second prediction parameter using the second neural network and the third neural network;
[0010] Summing the first prediction parameter and the second prediction parameter to obtain a target prediction parameter, and determining a prediction result based on the target prediction parameter;
[0011] Based on the prediction result, content is recommended for the user in the recommendation field.
[0012] In one embodiment, the specific features include at least one of the following:
[0013] Basic user characteristics;
[0014] User portrait characteristics;
[0015] User-specific behavior sequences;
[0016] Content characteristics;
[0017] The shared features include user sharing behavior sequences.
[0018] In one implementation, the first neural network includes a multi-layer perceptron MLP and an S-shaped growth curve function.
[0019] In one implementation, the using the second neural network and the third neural network to obtain the second prediction parameter includes:
[0020] Determining, based on the domain indicator, a specific domain corresponding to a specific feature in the historical data of the user;
[0021] Determining, using a second neural network, a specific scale and a specific deviation of the specific domain, and a specific moving average and a specific variance of all sample data in the specific domain;
[0022] Using a second neural network, performing private normalization processing on the specific feature based on the specific scale, specific deviation, specific moving average and specific variance to obtain first network parameters;
[0023] Determine, using a third neural network, a shared specific scale and a specific deviation corresponding to the shared feature, and a shared moving average and a shared variance of all sample data in multiple fields corresponding to the shared feature;
[0024] Using a third neural network to perform batch normalization on the shared features based on the shared scale, shared deviation, shared moving average, and shared variance to obtain second network parameters;
[0025] The first network parameter and the second network parameter are multiplied and / or summed to obtain a target network parameter, and the second prediction parameter is obtained based on the target network parameter.
[0026] In one implementation, the first network parameter includes a first network weight and a first network deviation; the second network parameter includes a second network weight and a second network deviation;
[0027] The performing product and / or sum processing on the first network parameter and the second network parameter to obtain the target network parameter includes:
[0028] The first network weight and the second network weight are multiplied to obtain a target network weight, and the first network deviation and the second network deviation are summed to obtain a target network deviation.
[0029] In one implementation, determining the prediction result based on the target prediction parameter includes:
[0030] Constructing a loss function based on the target prediction parameters;
[0031] The prediction result is determined based on the first neural network, the second neural network, the third neural network and the loss function.
[0032] According to a second aspect of an embodiment of the present disclosure, there is provided a content recommendation device, including:
[0033] an acquisition unit, the acquisition unit being configured to acquire a recommended field, a user identifier, and historical data of the user's operation on the content;
[0034] a processing unit, wherein the processing unit is configured to determine a domain indicator based on the recommended domain, the domain including an area where content is displayed; and determine specific features and shared features based on the historical data, wherein the specific features include specific features corresponding to the recommended domain, and the shared features include shared features corresponding to multiple domains;
[0035] A prediction unit, the prediction unit is configured to use a first neural network to obtain a first prediction parameter according to the domain indicator; use a second neural network and a third neural network to obtain a second prediction parameter according to the domain indicator, the specific feature and the shared feature; sum the first prediction parameter and the second prediction parameter to obtain a target prediction parameter, and determine a prediction result based on the target prediction parameter;
[0036] A recommendation unit is configured to recommend content to the user in the recommendation field based on the prediction result.
[0037] In one embodiment, the specific features include at least one of the following:
[0038] Basic user characteristics;
[0039] User portrait characteristics;
[0040] User-specific behavior sequences;
[0041] Content characteristics;
[0042] The shared features include user sharing behavior sequences.
[0043] In one implementation, the first neural network includes a multi-layer perceptron MLP and an S-shaped growth curve function.
[0044] In one embodiment, the using the second neural network and the third neural network to obtain the second prediction parameter according to the domain indicator, the specific feature and the shared feature includes:
[0045] Determining, based on the domain indicator, a specific domain corresponding to a specific feature in the historical data of the user;
[0046] Determining, using a second neural network, a specific scale and a specific deviation of the specific domain, and a specific moving average and a specific variance of all sample data in the specific domain;
[0047] Using a second neural network, performing private normalization processing on the specific feature based on the specific scale, specific deviation, specific moving average and specific variance to obtain first network parameters;
[0048] Determine, using a third neural network, a shared specific scale and a specific deviation corresponding to the shared feature, and a shared moving average and a shared variance of all sample data in multiple fields corresponding to the shared feature;
[0049] Using a third neural network to perform batch normalization on the shared features based on the shared scale, shared deviation, shared moving average, and shared variance to obtain second network parameters;
[0050] The first network parameter and the second network parameter are multiplied and / or summed to obtain a target network parameter, and the second prediction parameter is obtained based on the target network parameter.
[0051] In one implementation, the first network parameter includes a first network weight and a first network deviation; the second network parameter includes a second network weight and a second network deviation;
[0052] The performing product and / or sum processing on the first network parameter and the second network parameter to obtain the target network parameter includes:
[0053] The first network weight and the second network weight are multiplied to obtain a target network weight, and the first network deviation and the second network deviation are summed to obtain a target network deviation.
[0054] In one implementation, determining the prediction result based on the target prediction parameter includes:
[0055] Constructing a loss function based on the target prediction parameters;
[0056] The prediction result is determined based on the first neural network, the second neural network, the third neural network and the loss function.
[0057] According to a third aspect of an embodiment of the present disclosure, there is provided a content recommendation device, including:
[0058] processor;
[0059] a memory for storing processor-executable instructions;
[0060] The processor is configured to: execute the method as described in the first aspect or any embodiment of the first aspect.
[0061] According to a fourth aspect of an embodiment of the present disclosure, a storage medium is provided, in which instructions are stored. When the instructions in the storage medium are executed by a processor of a device, the device is enabled to execute a method as described in the first aspect or any embodiment of the first aspect.
[0062] The technical solution provided by the embodiments of the present disclosure may include the following beneficial effects: by setting three neural networks in one prediction model, the domain indicator, specific features of the recommended domain, and shared features of multiple domains are jointly predicted to determine the prediction results for content recommendation for users, thereby solving the time-consuming and labor-intensive technical problems of multiple models, reducing labor costs and resource costs without reducing the prediction effect, and achieving cost reduction and efficiency improvement.
[0063] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.
[0065] Figure 1 It is a schematic diagram of a large-scale information retrieval system according to an exemplary embodiment.
[0066] Figure 2A It is a schematic diagram showing a large-scale information retrieval application scenario according to an exemplary embodiment.
[0067] Figure 2B It is a schematic diagram showing a large-scale information retrieval application scenario according to an exemplary embodiment.
[0068] Figure 2C It is a schematic diagram showing a large-scale information retrieval application scenario according to an exemplary embodiment.
[0069] Figure 3 It is a schematic diagram of a modeling method in a multi-domain scenario according to an exemplary embodiment.
[0070] Figure 4It is a schematic diagram of multi-domain single modeling according to an exemplary embodiment.
[0071] Figure 5 is a schematic diagram of a multi-task learning model according to an exemplary embodiment.
[0072] Figure 6 The present invention is a flowchart of a content recommendation method according to an exemplary embodiment.
[0073] Figure 7 The present invention is a flowchart of a method for obtaining a second prediction parameter by using a second neural network and a third neural network according to an exemplary embodiment.
[0074] Figure 8 The invention is a flowchart of a method for determining a prediction result according to an exemplary embodiment.
[0075] Fig. 9 is a schematic diagram of a prediction model according to an exemplary embodiment.
[0076] Fig.10 is a schematic diagram of a fully connected neural network according to an exemplary embodiment.
[0077] Fig.11 The figure is a structural block diagram of a content recommendation device according to an exemplary embodiment.
[0078] Fig.12 The invention is a block diagram of a device for content recommendation according to an exemplary embodiment.
[0079] Fig.13 The invention is a block diagram of a device for content recommendation according to an exemplary embodiment. DETAILED DESCRIPTION
[0080] Here, exemplary embodiments will be described in detail, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present disclosure.
[0081] With the rapid development of technology, the amount of data on the Internet is also increasing day by day. It is impossible to display all goods or content in front of users. In order to solve the problem of information overload, large-scale information retrieval systems have emerged and are currently widely used in search, recommendation, advertising and other industries. Its main feature is that there are a large number of goods or items, and users usually have or do not have a clear purpose and do not know what items are available on the website or product. Large-scale retrieval systems use various retrieval algorithms to generate a list of items that users may be interested in based on the user's historical behavior, interest preferences or demographic characteristics. At the same time, in the face of massive data, the traditional practice of simply judging whether a user likes something based on manual rules has achieved little effect.
[0082] like Figure 1 As shown, a large-scale information retrieval system usually includes three stages: recall, sorting, and re-arrangement, and then returns N content to the user, where N is a positive integer. In the recall stage, a large number of materials can be recalled from an information library, such as a video library. The materials in the video library can be in the tens of millions, and the materials processed in the recall stage can be in the thousands or tens of thousands. The sorting stage includes rough sorting and fine sorting. Rough sorting is responsible for the preliminary screening of relevant materials. The processing level of this stage can be in the thousands, while fine sorting is directly responsible for the relevance sorting of materials, which determines the order in which users see the materials. The processing level of this stage can be in the hundreds or thousands. In the sorting stage, user features, project features, and / or cross features can be used to achieve sorting. After the fine sorting is completed, the re-sorting operation is performed, and the sorting results after re-shooting can be displayed to the user at the front end, so the processing level during re-sorting is usually in tens.
[0083] It can be considered that accurate ranking models are the core competitiveness of many personalized Internet products. Therefore, the industry's systems usually do a lot of work in fine ranking, usually combining a large number of data sets and fitting parameters to train deep learning models, so that the model is closer to reality. For example, Google recently released the first trillion-level model SwitchTransformer, with 1.6 trillion parameters, which is 4 times faster than the largest language model T5-XXL developed by Google. For another example, the fine ranking models currently used by Kuaishou all have trillion parameters, with a total parameter volume of more than 1.9 trillion, which is larger in scale and has been put into practical use.
[0084] These ultra-large-scale models mean more accurate fitting results, but also higher costs and consumption. For example, the implementation cost of a terabyte-scale model alone requires 256 graphics processing units (GPUs) for 8 days of offline training. The generated model of more than terabytes costs at least tens of millions of yuan in terms of storage and online machine loading per month for data and machine costs.
[0085] In addition, Internet products usually include multiple scenarios and pages, which requires estimates for multiple scenarios, multiple pages, and even multiple businesses. Xiaomi's app store has multiple scenarios, such as Figure 2A As shown in Figure 1, the banner at the top of the app store homepage and the "Today's Made for You" column both require personalized search to provide recommended items and ads. Figure 2B As shown in the figure, the App Store game (tab) page also has a top banner and module recommendation bar. Figure 2C As shown in the figure, on the homepage of the game center of the app store, the top banner and the downward flow are also separate. In terms of model principles, the processes and models used by these systems are basically the same, which are the classic processes of recall sorting and click-through rate (CTR) estimation models. However, in actual applications, CTR models usually focus on the estimation of a single field. Each field has a separate location in the mobile application or web page. The main method is to collect user samples and examples that appear in this location for offline training, and then serve this single location. For example, the top banner of the Xiaomi App Store homepage and the "Today for you" on the homepage use two CTR estimation models for personalized retrieval.
[0086] Usually, these multi-domain pages and scenes may share some common user groups and content materials, and each domain will also have its own unique user groups and materials. This will lead to different behaviors of the same user in different domains, and even more different behaviors of different users in different domains.
[0087] There are two common approaches in the industry: multi-domain single modeling: using all data from different domains to train a single model to serve all domains; multi-domain multi-model: modeling these scenarios and pages separately, that is, each domain uses its own unique data to train a separate model and deploy it online separately. Figure 3 As shown, modeling methods (a) and (c) correspond to multi-domain single modeling, and modeling method (b) corresponds to multi-domain multi-model.
[0088] Among them, when using a single model for multiple fields, user samples from different fields are usually directly merged together, and then the CTR model is used to estimate the user's preference for the material. Commonly used CTR models include: the traditional multi-layer neural network model (Deep Neural Network, DNN), the deep interest network model (Deep Interest Network, DIN) that considers the user's medium- and short-term temporal interests, the evolutionary deep interest network model (Deep Interest Evolution Network, DIEN) that considers the user's medium- and short-term interest changes, and the search-based user interest model (Search-based User Interest Model, SIM) that considers the user's long-term interests. Figure 4 As shown, the home page banner user behavior, the home page downlink user behavior, the game page banner user behavior, and the game page downlink user behavior can be integrated and trained to obtain models such as DNN, DIN, DIEN or SIM.
[0089] There are usually two ways to use multiple models in multiple fields. One is to use M single models to model M fields, that is, to train M single models separately with user data from MN fields. Usually, the selection of single models is the same as the CTR model in multi-field single modeling, including DNN, DIN, DIEN and SIM. Another method is to find modeling tasks with similarities in multiple fields, and only use M' models to model M' similar modeling tasks in M fields. For example, in the banner field of the homepage of Xiaomi App Store, users will have a variety of user behaviors in this field, including "click" and "download", which requires the CTR model to estimate whether the user clicks and downloads. Traditional modeling uses two CTR models to model the click and download of the homepage banner respectively. In related technologies, a multi-task model can also be used to simultaneously estimate the "click" and "download" behaviors of the homepage banner. Multi-objective models learn correlations from behaviors in the same field by learning hidden parameter sharing, such as multi-task learning models.
[0090] Figure 5 Several common multi-task learning models are shown. Among them, the first model is Figure 5 The model marked (a) is a hard parameter sharing model. The bottom layer of the model is a shared hidden layer, and the upper layer uses a specific fully connected layer to learn specific tasks. Different tasks correspond to different towers. The second model is Figure 5The model marked (b) is a one-gate mixture of experts (OMOE) model. In the shared layer at the bottom of the model, there are multiple expert layers, each of which is equivalent to a multi-layer neural network, that is, a multi-layer network is used to replace the single multi-layer network sharing layer in the Hard Parameter Sharing model. The OMOE model includes a gating network, which can assign a weight parameter to each expert layer, that is, assign different attention weights to each expert layer, so as to use different combinations of expert layers for prediction of different tasks, thereby improving prediction flexibility. Similarly, different tasks correspond to different towers. The third model is Figure 5 The model marked (c) is a multi-gate mixture of experts (MMOE) model. Compared with the OMOE model, the MMOE model includes multiple gating networks, each of which corresponds to a task, that is, a tower. Different gating networks are used to obtain different combinations of expert layers for different tasks, which further improves the prediction accuracy.
[0091] However, since each field has its own unique user and scenario characteristics, using a single model for multiple fields to build a model that serves all fields is often not as effective as modeling each field separately. Because users and materials in different fields have certain intersections, information sharing in different fields can improve the effect of the CTR prediction model to a certain extent. However, at the same time, there are certain differences in user behavior in different fields, resulting in certain differences in data distribution. Simply mixing data from all fields to learn a shared model for CTR prediction in all fields often does not achieve the expected results.
[0092] There are also many problems with modeling each domain separately. On the one hand, some business domains generate much less data than other domains, so splitting the data for separate training will ignore the commonalities between scenarios and make the data for such scenarios more scarce, making the model more difficult to learn. On the other hand, maintaining multiple models at the same time will cause a lot of resource consumption and requirements, as well as more labor costs. When the number of business domains reaches hundreds, this part of the cost will surge.
[0093] In view of this, the embodiments of the present disclosure provide a content recommendation method, which uses a single model to recommend content in various fields. By assigning field indicators to the recalled and / or roughly sorted content, and distinguishing the specific features of the recalled and / or roughly sorted content corresponding to the specific field and the shared features common to all fields, the field indicators, specific features and shared features are processed by different task towers and then jointly predicted, thereby solving the time-consuming and labor-intensive technical problem of multiple models, reducing labor and resource costs without reducing the prediction effect, and achieving cost reduction and efficiency improvement.
[0094] Figure 6 is a flow chart of a content recommendation method according to an exemplary embodiment. Figure 6 As shown, the method includes the following steps.
[0095] In step S11 , a recommended domain is acquired, and a domain indicator is determined based on the recommended domain.
[0096] Among them, the field includes the area where the content is displayed.
[0097] In step S12, the user identification and historical data of the user's operations on the content are obtained.
[0098] In step S13, specific features and shared features are determined based on the historical data.
[0099] Among them, the specific features include specific features corresponding to the recommended field, and the shared features include shared features corresponding to multiple fields.
[0100] In step S14, the domain indicator is input into the first neural network, and the first prediction parameter is obtained using the first neural network.
[0101] In step S15, the domain indicator and the specific features are input into the second neural network, the shared features are input into the third neural network, and the second prediction parameters are obtained using the second neural network and the third neural network.
[0102] In step S16, the first prediction parameter is summed with the second prediction parameter to obtain a target prediction parameter, and a prediction result is determined based on the target prediction parameter.
[0103] In step S17, content is recommended for the user in the recommendation field based on the prediction result.
[0104] In the disclosed embodiments, the content recommendation method can be used to recommend content to users in the recommendation field. In one example, content can be recommended to users on the homepage of a terminal application, or on a subpage of a terminal application (e.g., a game page of a terminal application store). Furthermore, content can include applications, products, pages, and other content that can be displayed to users and that users can click, purchase, or download.
[0105] In the disclosed embodiment, a recommended domain may be obtained, and a domain indicator may be determined based on the recommended domain. The recommended domain may include an area for displaying content. In one example, the area for displaying content may be the homepage of the terminal application, a subpage of the terminal application, or other areas. The domain indicator is used to indicate the recommended domain. Different recommended domains correspond to different domain indicators.
[0106] In the disclosed embodiment, the user identification and the historical data of the user's operation on the content can be obtained. Content recommendation needs to be performed based on the user's personal information and the user's historical operation records. The recommended content obtained for different users may be different. For the same user, if he has different historical data of operating the content at different time periods, the recommended content obtained at different times may also be different.
[0107] In the disclosed embodiment, specific features and shared features can be determined based on the historical data obtained. Among them, specific features can include specific features corresponding to the recommended field, and shared features can include shared features corresponding to multiple fields. That is to say, the historical data of the user's operation on the content can include specific features of the user's operation on the content of the recommended field, and can also include shared features of the user's operation on the content of the recommended field and other fields. Taking the content as the application in the terminal application store and the recommended field as the game page of the application store as an example, the specific feature can be a specific behavior sequence of the user on the game page of the application store, such as the behavior of clicking, downloading, etc. on each application on the game page, and the content corresponding to the operation, that is, each application; the shared feature can be a shared behavior sequence that the user may perform on the game page and other pages (such as the homepage), and the content corresponding to the execution behavior, that is, the corresponding application, for example, a click operation may be performed on a certain application on both the homepage and the game page.
[0108] In the disclosed embodiment, the domain indicator may be input into the first neural network, and the first prediction parameter may be obtained using the first neural network. Furthermore, the domain indicator and the specific feature may be input into the second neural network, and the shared feature may be input into the third neural network, and the second neural network and the third neural network may be used to obtain the second prediction parameter. The first prediction parameter and the second prediction parameter are summed to obtain the target prediction parameter, and the prediction result is determined based on the target prediction parameter.
[0109] That is to say, the domain indicator, specific features and shared features can be jointly predicted based on the first neural network, the second neural network and the third neural network, that is, one model can be used to predict the recommendation results of different fields. Among them, the first neural network processes the domain indicator, the second neural network processes the domain indicator and the specific features of the recommended field, and the third neural network processes the shared features.
[0110] In the disclosed embodiment, content can be recommended to users in the recommended field based on the prediction results. In one example, the rough ranking results in different fields can be finely ranked based on the prediction model composed of the first neural network, the second neural network and the third neural network to obtain ranked prediction results, and content can be recommended to users based on the ranked prediction results.
[0111] By adopting the technical solution of the embodiment of the present disclosure, three neural networks are set in one prediction model to jointly predict the field indicator, the specific features of the recommended field, and the shared features of multiple fields to determine the prediction results for content recommendation for users. This solves the time-consuming and labor-intensive technical problems of multiple models, reduces labor and resource costs without reducing the prediction effect, and achieves cost reduction and efficiency improvement.
[0112] In the disclosed embodiment, the first neural network, the second neural network and the third neural network may constitute a prediction model, and the prediction model may be pre-trained. The prediction model may be trained in the following manner: obtain multiple sample data, wherein the multiple sample data include user identifiers of multiple users, and historical data of operations performed by each user in each field. Set a field identifier for each field, and extract specific features of each field and shared features of all fields in the historical data. Input the field identifier into the first neural network, the field indicator and specific features into the second neural network, and the shared features into the third neural network, and jointly train the first neural network, the second neural network and the third neural network until the model converges. The training of the model may be implemented based on a mini-batch method.
[0113] In the embodiment of the present disclosure, the specific feature may include at least one of a basic user feature, a user portrait feature, a user specific behavior sequence, and a content feature. The shared feature may include a user shared behavior sequence.
[0114] In the embodiment of the present disclosure, the first neural network may include a multi-layer perceptron MLP and an S-shaped growth curve (sigmoid) function. After the domain indicator is processed by the MLP, the first prediction parameter sa is calculated by the sigmoid function. Among them, sa can be calculated by the formula sa=sigmoid(MLP(s)), where s is the domain indicator.
[0115] In the embodiment of the present disclosure, the second neural network and the third neural network may include a star topology fully connected neural network (Star Topology Full Connected Network, Star Topology FCN).
[0116] During the training process of a neural network, the input distribution of the hidden layer network will change with the change of the parameters of the previous layer. This change may cause the distribution of the neural network to gradually shift to the upper and lower limits of the value range of the nonlinear function. For example, if the nonlinear function is sigmoid, the above changes may cause the gradient of back propagation to disappear, resulting in slower model training. Normalization can be used to solve this problem.
[0117] In one implementation, batch normalization (BN) can be performed on all features. BN achieves global normalization by accumulating the normalization parameters of all samples. The calculation formula of BN is as follows: Among them, z1' is the global output vector, β is the globally learnable scale, i.e., the shared scale, γ is the globally learnable bias, i.e., the shared bias, μ is the average value of the current batch, and σ 2 is the variance of the current batch, z is the input vector, and ε is a small value used to prevent the division by 0. Among them, μ can be replaced by the moving average E of all samples, σ 2 It can be replaced by the variance Var of all samples, that is, batch normalization assumes that all samples use shared statistics. At this time, the calculation formula of BN is:
[0118] However, in multi-domain application scenarios, samples are not global, and samples from different fields have their own normalized moments. Therefore, PN processing can be performed on samples from different fields separately. The calculation formula of PN is as follows: Among them, z2' is the output vector of a specific domain, γ p is a specific scale for a specific field, β p is a domain-specific bias. In each mini-batch of model training, the shared scale γ is combined with the domain-specific scale γ p The final scale is determined by element-wise multiplication, i.e., PN adaptively scales the representation according to the domain indicator. Similarly, the bias of PN is also obtained by adding the shared bias β and the specific bias β p Similarly, μ can be replaced by the moving average E of all samples in the field, and σ can be replaced by the variance Var of all samples in the field. 2 , the calculation formula of PN is:
[0119] In another implementation, a combination of Partitioned Normalization (PN) processing and BN processing may be used to normalize the input features.
[0120] Figure 7 is a flowchart of a method for obtaining a second prediction parameter using a second neural network and a third neural network according to an exemplary embodiment. Figure 7 As shown, the method includes the following steps.
[0121] In step S21, a specific domain corresponding to a specific feature in the user's historical data is determined based on the domain indicator.
[0122] In step S22, a second neural network is used to determine a specific scale and a specific deviation of a specific domain, as well as a specific moving average and a specific variance of all sample data in the specific domain.
[0123] In step S23, a second neural network is used to perform private normalization processing on the specific features based on a specific scale, a specific deviation, a specific moving average, and a specific variance to obtain first network parameters.
[0124] In step S24, a third neural network is used to determine the shared specific scale and specific deviation corresponding to the shared features, as well as the shared moving average and shared variance of all sample data in multiple fields corresponding to the shared features.
[0125] In step S25, a third neural network is used to perform private normalization processing on the shared features based on the shared scale, shared deviation, shared moving average and shared variance to obtain second network parameters.
[0126] In step S26, the first network parameter and the second network parameter are processed by product and / or summation to obtain the target network parameter, and the second prediction parameter is obtained based on the target network parameter.
[0127] In the embodiment of the present disclosure, the specific domain corresponding to the specific feature in the user's historical data can be determined based on the domain indicator, and then the specific scale β of the specific domain can be determined using the second neural network. p and specific deviation γ p , and a specific moving average E of all sample data in a specific field p and variance var p The second neural network determines the β p , γ p 、E p and Var p Perform PN processing to obtain first network parameters.
[0128] In the disclosed embodiment, a third neural network may be used to determine the shared specific scale β and specific deviation β corresponding to the shared feature, as well as the shared moving average E and shared variance Var of all sample data in multiple fields corresponding to the shared feature. The third neural network performs BN processing based on the determined β, γ, E and Var to obtain the second network parameters.
[0129] The second neural network and the third neural network can form a star topology fully convolutional neural network (StarTopology Fully Convolutional Networks, Star Topology FCN). The fully connected neural network of Star Topology FCN may include a shared central FCN network and specific FCN networks in multiple fields, wherein the central FCN network is responsible for learning information common to all fields, and the specific FCN network is responsible for learning information unique to the field. The central FCN network may include second network parameters, which include second network weights W and second network deviations b. The specific FCN network may include first network parameters, which include first network weights W p and the first network deviation b p For domain p, its target network parameters include target network weight W p ' and target network deviation b p ', can be expressed as follows: b' p =b p +b, where the symbol Represents element-wise multiplication.
[0130] By adopting the technical solution of the embodiment of the present disclosure, by performing PN processing on specific features, the differences between different fields can be better captured; by performing BN processing on shared features, more precise target prediction can be performed. At the same time, in actual industrial use, since most parameters are contributed by the embedding layer, although multiple specific FCN networks are added in the embodiment of the present disclosure, the increase in parameters is not obvious compared with the total number of parameters. Therefore, the processing method combining PN and BN adopted in the present disclosure can effectively improve the prediction effect without significantly consuming additional resources.
[0131] In the embodiment of the present disclosure, the second prediction parameter can be obtained based on the target network parameter. In one example, the formula out p =φ((W' p ) T in p +b' p ) Calculate the output of the Star Topology FCN fully connected layer, where outp is the Star Topology FCN fully connected layer output of domain p, in p is the input of domain p, φ is the activation function of FCN. The output of StarTopology FCN fully connected layer out p The one-dimensional output of is recorded as sm, which is the first prediction parameter.
[0132] In the embodiment of the present disclosure, the first prediction parameter sa and the second prediction parameter sm can be summed to obtain the target prediction parameter. The target prediction parameter can be predicted using the sigmoid function to obtain the prediction probability
[0133] Figure 8 is a flowchart of a method for determining a prediction result according to an exemplary embodiment. Figure 8 As shown, the method includes the following steps.
[0134] In step S31, a loss function is constructed based on the target prediction parameters.
[0135] In step S32, a prediction result is determined based on the first neural network, the second neural network, the third neural network and the loss function.
[0136] In the disclosed embodiment, a loss function is constructed based on the target prediction parameters. The calculation formula of the loss function is as follows: It is the cross entropy loss function. represents the predicted probability of the i-th instance in domain p, represents the true value of the i-th instance in domain p, M represents the number of fields, N p Represents the number of instances in domain P.
[0137] In the embodiment of the present disclosure, the cross entropy loss function of all fields is minimized by jointly training the first neural network, the second neural network and the third neural network, and then based on the first neural network, the second neural network and the third neural network that minimize the loss function, the prediction results of the content that the user may be interested in in a specific field are determined, and then content recommendations are made to the user based on the prediction results.
[0138] By adopting the technical solution of the embodiment of the present disclosure, a prediction model is determined by jointly training the first neural network, the second neural network and the third neural network, and content that the user may be interested in in a specific field is predicted based on the trained model. Content is recommended to the user based on the prediction results, thereby increasing the recommendation accuracy without significantly increasing the computational complexity and improving the user experience.
[0139] Fig. 9FIG. 1 is a schematic diagram of a prediction model according to an exemplary embodiment. Fig. 9 As shown, the model may include three neural networks, which may also be called three towers in some cases. Among them, the first neural network may be an auxiliary network, whose input is a domain indicator. After the domain indicator is input by the first neural network, it is processed by MLP and then calculated using the Sigmoid function to obtain the first prediction parameter sa. The input of the second neural network may include user features, content features, and user behavior sequences, and may also include domain indicators. Among them, user features may include basic user features, such as the user's age, gender, city / current location, user terminal model, etc., and user portrait features, such as the content category clicked or downloaded by the user in the first historical period, the content category clicked or downloaded by the user in the second historical period, etc. Content features may include content identification, total exposure of content, total clicks on content, and click-through rate of content, etc. User behavior sequences may include content identifications clicked by users in the third historical period, and content identifications downloaded by users in the fourth historical period, etc. The user behavior sequence input by the second neural network may be a user-specific behavior sequence, that is, a user's behavior sequence in a specific field. The input of the third neural network may include a user sharing behavior sequence. The same content can share the same features. For example, if the same content is exposed in different fields, the embedding expression of the same content in the model is shared, and the final input to the model will be a low-dimensional embedding matrix with a specific length. This process does not require additional pre-labeling, and user feedback data can be directly read, and training can be carried out in a mini-batch manner.
[0140] Furthermore, the first neural network can be used to capture the difference between different domains, and the domain indicator is directly used as an input feature to enhance the relevant information. The first neural network performs a forward pass calculation on the connected features to obtain a one-dimensional output sa.
[0141] The second neural network can be a privatized Star Topology FCN, which is used to perform PN processing on data in a specific field. The third neural network can be a shared FCN, which is used to process common data shared by the business, perform BN processing on these common data, and use a fully connected neural network to uniformly model the PN-processed data and the BN-processed data, thereby avoiding interference from noise and sample sparsity.
[0142] Fig.10 is a schematic diagram of a fully connected neural network according to an exemplary embodiment. Fig.10As shown, the fully connected neural network includes a shared central FCN (also called shared FCN) and specific FCNs corresponding to multiple fields. The shared FCN is responsible for learning information common to all fields, and the specific FCN is responsible for learning information unique to a single field. By combining the network parameters of the shared FCN with the network parameters of the specific FCN, for example, by element-wise multiplication of some parameters and addition of other parameters, the parameters of the fully connected neural network can be obtained, and then the output sm of the fully connected neural network can be calculated.
[0143] Continue to refer Fig. 9 , the output sm of the fully connected neural network is the output of the FCN composed of the second neural network and the third neural network. The sum of sa and sm can be calculated using the sigmoid function to obtain the predicted value of the input sample. In the model training stage, it is necessary to minimize the loss function of the model through multiple training to obtain the optimal parameters of the first neural network, the second neural network, and the third neural network. In the prediction stage, the prediction results can be obtained using the trained model.
[0144] Based on the same concept, the embodiment of the present disclosure also provides a content recommendation device.
[0145] It is understandable that in order to achieve the above functions, the content recommendation device provided by the embodiment of the present disclosure includes hardware structures and / or software modules corresponding to the execution of each function. In combination with the units and algorithm steps of each example disclosed in the embodiment of the present disclosure, the embodiment of the present disclosure can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the technical solution of the embodiment of the present disclosure.
[0146] Fig.11 is a block diagram of a content recommendation device according to an exemplary embodiment. Fig.11 The device 100 includes an acquisition unit 110, a processing unit 120, a prediction unit 130 and a recommendation unit 140.
[0147] The acquisition unit 110 is configured to acquire a recommended field, a user identifier, and historical data of user operations on content.
[0148] The processing unit 120 is configured to determine a domain indicator based on the recommended domain, where the domain includes an area where content is displayed; and determine specific features and shared features based on historical data.
[0149] Among them, the specific features include specific features corresponding to the recommended field, and the shared features include shared features corresponding to multiple fields.
[0150] The prediction unit 130 is configured to use a first neural network to obtain a first prediction parameter according to a domain indicator; use a second neural network and a third neural network to obtain a second prediction parameter according to the domain indicator, specific features and shared features; sum the first prediction parameter and the second prediction parameter to obtain a target prediction parameter, and determine a prediction result based on the target prediction parameter.
[0151] The recommendation unit 140 is configured to make content recommendations for the user in the recommendation field based on the prediction result.
[0152] In the embodiments of the present disclosure, the specific features include at least one of the following:
[0153] Basic user characteristics;
[0154] User portrait characteristics;
[0155] User-specific behavior sequences;
[0156] Content characteristics;
[0157] The shared features include user shared behavior sequences.
[0158] In the disclosed embodiment, the first neural network includes a multi-layer perceptron MLP and an S-shaped growth curve function.
[0159] In the embodiment of the present disclosure, the second neural network and the third neural network are used to obtain the second prediction parameter according to the domain indicator, the specific feature and the shared feature, including:
[0160] Determine, based on the domain indicator, a specific domain corresponding to a specific feature in the user's historical data;
[0161] using a second neural network to determine a domain-specific scale and a domain-specific bias, as well as a domain-specific moving average and a domain-specific variance for all sample data in the domain;
[0162] Using the second neural network to perform private normalization processing on the specific features based on a specific scale, a specific deviation, a specific moving average, and a specific variance to obtain the first network parameters;
[0163] Using a third neural network to determine a shared specific scale and a specific deviation corresponding to the shared features, as well as a shared moving average and a shared variance of all sample data in multiple fields corresponding to the shared features;
[0164] Using a third neural network to perform batch normalization on the shared features based on a shared scale, a shared bias, a shared moving average, and a shared variance, to obtain parameters of a second network;
[0165] The first network parameter and the second network parameter are processed by product and / or summation to obtain a target network parameter, and a second prediction parameter is obtained based on the target network parameter.
[0166] In the embodiment of the present disclosure, the first network parameter includes a first network weight and a first network deviation; the second network parameter includes a second network weight and a second network deviation;
[0167] Performing product and / or sum processing on the first network parameter and the second network parameter to obtain the target network parameter includes:
[0168] The target network weight is obtained by multiplying the first network weight and the second network weight, and the target network deviation is obtained by summing the first network deviation and the second network deviation.
[0169] In the embodiment of the present disclosure, determining the prediction result based on the target prediction parameter includes:
[0170] Construct a loss function based on the target prediction parameters;
[0171] Based on the first neural network, the second neural network, the third neural network and the loss function, a prediction result is determined.
[0172] By adopting the technical solution of the embodiment of the present disclosure, three neural networks are set in one prediction model to jointly predict the field indicator, the specific features of the recommended field, and the shared features of multiple fields to determine the prediction results for content recommendation for users. This solves the time-consuming and labor-intensive technical problems of multiple models, reduces labor and resource costs without reducing the prediction effect, and achieves cost reduction and efficiency improvement.
[0173] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.
[0174] Fig.12 2 is a block diagram of a device 200 for content recommendation according to an exemplary embodiment. For example, the device 200 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.
[0175] Reference Fig.12 , the device 200 may include one or more of the following components: a processing component 202 , a memory 204 , a power component 206 , a multimedia component 208 , an audio component 210 , an input / output (I / O) interface 212 , a sensor component 214 , and a communication component 216 .
[0176] The processing component 202 generally controls the overall operation of the device 200, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 202 may include one or more processors 220 to execute instructions to perform all or part of the steps of the above-described method. In addition, the processing component 202 may include one or more modules to facilitate interaction between the processing component 202 and other components. For example, the processing component 202 may include a multimedia module to facilitate interaction between the multimedia component 208 and the processing component 202.
[0177] The memory 204 is configured to store various types of data to support operations on the device 200. Examples of such data include instructions for any application or method operating on the device 200, contact data, phone book data, messages, pictures, videos, etc. The memory 204 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.
[0178] The power component 206 provides power to the various components of the device 200. The power component 206 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device 200.
[0179] The multimedia component 208 includes a screen that provides an output interface between the device 200 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundaries of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 208 includes a front camera and / or a rear camera. When the device 200 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.
[0180] The audio component 210 is configured to output and / or input audio signals. For example, the audio component 210 includes a microphone (MIC), and when the device 200 is in an operation mode, such as a call mode, a recording mode, and a speech recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in the memory 204 or sent via the communication component 216. In some embodiments, the audio component 210 also includes a speaker for outputting audio signals.
[0181] I / O interface 212 provides an interface between processing component 202 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include but are not limited to: a home button, a volume button, a start button, and a lock button.
[0182] The sensor assembly 214 includes one or more sensors for providing various aspects of the status assessment of the device 200. For example, the sensor assembly 214 can detect the open / closed state of the device 200, the relative positioning of components, such as the display and keypad of the device 200, the sensor assembly 214 can also detect the position change of the device 200 or a component of the device 200, the presence or absence of user contact with the device 200, the orientation or acceleration / deceleration of the device 200 and the temperature change of the device 200. The sensor assembly 214 can include a proximity sensor configured to detect the presence of a nearby object without any physical contact. The sensor assembly 214 can also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 214 can also include an accelerometer, a gyroscope sensor, a magnetic sensor, a pressure sensor or a temperature sensor.
[0183] The communication component 216 is configured to facilitate wired or wireless communication between the device 200 and other devices. The device 200 can access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component 216 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 216 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.
[0184] In an exemplary embodiment, the apparatus 200 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.
[0185] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 204 including instructions, and the instructions can be executed by the processor 220 of the device 200 to perform the above method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.
[0186] Fig.13 is a block diagram of a device 300 for content recommendation according to an exemplary embodiment. For example, the device 300 may be provided as a server. Fig.13 The apparatus 300 includes a processing component 322, which further includes one or more processors, and a memory resource represented by a memory 332 for storing instructions executable by the processing component 322, such as an application. The application stored in the memory 332 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 322 is configured to execute the instructions to perform the above-mentioned content recommendation method.
[0187] The device 300 may also include a power supply component 326 configured to perform power management of the device 300, a wired or wireless network interface 350 configured to connect the device 300 to a network, and an input / output (I / O) interface 358. The device 300 may operate based on an operating system stored in the memory 332, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or the like.
[0188] It is to be understood that in the present disclosure, "plurality" refers to two or more than two, and other quantifiers are similar. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship. The singular forms "a", "the" and "the" are also intended to include plural forms, unless the context clearly indicates other meanings.
[0189] It is further understood that the terms "first", "second", etc. are used to describe various information, but such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other, and do not indicate a specific order or degree of importance. In fact, the expressions "first", "second", etc. can be used interchangeably. For example, without departing from the scope of the present disclosure, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information.
[0190] It can be further understood that, unless otherwise specified, “connection” includes a direct connection without other components between the two, and also includes an indirect connection with other components between the two.
[0191] It is further understood that, although the operations are described in a specific order in the drawings in the embodiments of the present disclosure, it should not be understood as requiring the operations to be performed in the specific order shown or in a serial order, or requiring the execution of all the operations shown to obtain the desired results. In certain environments, multitasking and parallel processing may be advantageous.
[0192] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modifications, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary technical means in the art that are not disclosed in the present disclosure.
[0193] It should be understood that the present disclosure is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the scope of the appended claims.
Claims
1. A content recommendation method, characterized in that: The method comprises: Acquire a recommended domain, and determine a domain indicator based on the recommended domain, wherein the domain includes an area where content is displayed; Acquire a user identifier and historical data of the user's operations on the content; determine specific features and shared features based on the historical data, wherein the specific features include specific features corresponding to the recommended field, and the shared features include shared features corresponding to multiple fields; inputting the domain indicator into a first neural network, and obtaining a first prediction parameter using the first neural network; Inputting the domain indicator and the specific feature into a second neural network, inputting the shared feature into a third neural network, and obtaining a second prediction parameter using the second neural network and the third neural network; Summing the first prediction parameter and the second prediction parameter to obtain a target prediction parameter, and determining a prediction result based on the target prediction parameter; Based on the prediction result, content is recommended for the user in the recommendation field.
2. The method according to claim 1, characterized in that: The specific features include at least one of the following: Basic user characteristics; User portrait characteristics; User-specific behavior sequences; Content characteristics; The shared features include user sharing behavior sequences.
3. The method according to claim 1, characterized in that The first neural network includes a multi-layer perceptron MLP and an S-shaped growth curve function.
4. The method according to claim 1, characterized in that: The using the second neural network and the third neural network to obtain a second prediction parameter includes: Determining, based on the domain indicator, a specific domain corresponding to a specific feature in the historical data of the user; Determining, using a second neural network, a specific scale and a specific deviation of the specific domain, and a specific moving average and a specific variance of all sample data in the specific domain; Using a second neural network, performing private normalization processing on the specific feature based on the specific scale, specific deviation, specific moving average and specific variance to obtain first network parameters; Determine, using a third neural network, a shared specific scale and a specific deviation corresponding to the shared feature, and a shared moving average and a shared variance of all sample data in multiple fields corresponding to the shared feature; Using a third neural network to perform batch normalization on the shared features based on the shared scale, shared deviation, shared moving average, and shared variance to obtain second network parameters; The first network parameter and the second network parameter are multiplied and / or summed to obtain a target network parameter, and the second prediction parameter is obtained based on the target network parameter.
5. The method according to claim 4, characterized in that The first network parameters include a first network weight and a first network deviation; the second network parameters include a second network weight and a second network deviation; The performing product and / or sum processing on the first network parameter and the second network parameter to obtain the target network parameter includes: The first network weight and the second network weight are multiplied to obtain a target network weight, and the first network deviation and the second network deviation are summed to obtain a target network deviation.
6. The method according to claim 1, characterized in that The determining of the prediction result based on the target prediction parameter comprises: Constructing a loss function based on the target prediction parameters; The prediction result is determined based on the first neural network, the second neural network, the third neural network and the loss function.
7. A content recommendation device, characterized in that: include: an acquisition unit, the acquisition unit being configured to acquire a recommended field, a user identifier, and historical data of the user's operation on the content; a processing unit configured to determine a domain indicator based on the recommended domain, the domain comprising an area where content is displayed; Based on the historical data, determining specific features and shared features, wherein the specific features include specific features corresponding to the recommended field, and the shared features include shared features corresponding to multiple fields; A prediction unit, the prediction unit is configured to use a first neural network to obtain a first prediction parameter according to the domain indicator; use a second neural network and a third neural network to obtain a second prediction parameter according to the domain indicator, the specific feature and the shared feature; sum the first prediction parameter and the second prediction parameter to obtain a target prediction parameter, and determine a prediction result based on the target prediction parameter; A recommendation unit is configured to recommend content to the user in the recommendation field based on the prediction result.
8. The device according to claim 7, characterized in that The specific features include at least one of the following: Basic user characteristics; User portrait characteristics; User-specific behavior sequences; Content characteristics; The shared features include user sharing behavior sequences.
9. The device according to claim 7, characterized in that The first neural network includes a multi-layer perceptron MLP and an S-shaped growth curve function.
10. The device according to claim 7, characterized in that The method of using the second neural network and the third neural network to obtain a second prediction parameter according to the domain indicator, the specific feature and the shared feature includes: Determining, based on the domain indicator, a specific domain corresponding to a specific feature in the historical data of the user; Determining, using a second neural network, a specific scale and a specific deviation of the specific domain, and a specific moving average and a specific variance of all sample data in the specific domain; Using a second neural network, performing private normalization processing on the specific feature based on the specific scale, specific deviation, specific moving average and specific variance to obtain first network parameters; Determine, using a third neural network, a shared specific scale and a specific deviation corresponding to the shared feature, and a shared moving average and a shared variance of all sample data in multiple fields corresponding to the shared feature; Using a third neural network to perform batch normalization on the shared features based on the shared scale, shared deviation, shared moving average, and shared variance to obtain second network parameters; The first network parameter and the second network parameter are multiplied and / or summed to obtain a target network parameter, and the second prediction parameter is obtained based on the target network parameter.
11. The device according to claim 10, characterized in that The first network parameters include a first network weight and a first network deviation; the second network parameters include a second network weight and a second network deviation; The performing product and / or sum processing on the first network parameter and the second network parameter to obtain the target network parameter includes: The first network weight and the second network weight are multiplied to obtain a target network weight, and the first network deviation and the second network deviation are summed to obtain a target network deviation.
12. The device according to claim 7, characterized in that The determining of the prediction result based on the target prediction parameter comprises: Constructing a loss function based on the target prediction parameters; The prediction result is determined based on the first neural network, the second neural network, the third neural network and the loss function.
13. A content recommendation device, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to: Execute the method according to any one of claims 1 to 6.
14. A storage medium, characterized in that: The storage medium stores instructions, and when the instructions in the storage medium are executed by a processor of a device, the device is enabled to execute the method according to any one of claims 1 to 6.
Citation Information
Patent Citations
Method for recommending target object, computing device and computer storage medium
CN111523044A
Resource recommendation method and device, electronic equipment and storage medium
CN113886674A
Advanced techniques to improve content presentation experiences for businesses and users
US20190139092A1
System and method for user specific apparel attribute recommendation
US20210192597A1
