A recommendation method, apparatus and electronic device

Abstract

The present disclosure provides a recommendation method, device and electronic equipment, relates to the multimedia technical field, and solves the problem of how to improve the recommendation accuracy of multimedia files. The method comprises the following steps: receiving request information sent by an electronic device; determining actual user data according to an actual user identifier; inputting pre-configured media data, actual user data and actual context data into a recommendation model to obtain actual click rates and actual conversion rates corresponding to each actual multimedia file in the media data; determining recommendation data according to the actual click rates and the actual conversion rates; and sending the recommendation data to the electronic device.

Description

Technical Field

[0001] This disclosure relates to the field of multimedia technology, and more particularly to a recommended method, apparatus, and electronic device. Background Technology

[0002] Currently, in the field of multimedia technology, with the widespread application of smart terminals, people are spending more and more time on smart terminals, and their demands for the services provided by smart terminals are also increasing. To enhance the user experience, smart terminals typically pre-recommend multimedia files for users to choose from.

[0003] However, the multimedia files recommended by smart devices are often not what the user needs, resulting in a low accuracy rate for multimedia file recommendations. Summary of the Invention

[0004] To address the aforementioned technical problems, this disclosure provides a recommended method, apparatus, and electronic device.

[0005] The technical solution disclosed herein is as follows:

[0006] In a first aspect, this disclosure provides a recommendation method, comprising: receiving request information sent by an electronic device; wherein the request information includes at least an actual user identifier and actual context data, and the actual context data includes at least time information and location information; determining actual user data based on the actual user identifier; inputting pre-configured media asset data, actual user data, and actual context data into a recommendation model to obtain the actual click-through rate and actual conversion rate corresponding to each actual multimedia file in the media asset data; wherein the recommendation model is trained based on the historical membership degree of each behavioral motivation of the feature information corresponding to historical multimedia files, historical user data, and historical context data, and the feature vector corresponding to the latent space containing each behavioral motivation, and the behavioral motivation is used to indicate the reason for performing the user behavior corresponding to the behavioral motivation in the target environment; determining recommendation data based on the actual click-through rate and actual conversion rate; wherein the recommendation data includes at least one actual multimedia file; and sending the recommendation data to the electronic device.

[0007] Secondly, this disclosure provides a recommendation device, comprising: a receiving unit for receiving request information sent by an electronic device; wherein the request information includes at least an actual user identifier and actual context data, and the actual context data includes at least time information and location information; a processing unit for determining actual user data based on the actual user identifier received by the receiving unit; the processing unit is further configured to input pre-configured media asset data, actual user data, and actual context data received by the receiving unit into a recommendation model to obtain the actual click-through rate and actual conversion rate corresponding to each actual multimedia file in the media asset data; wherein the recommendation model is trained based on the historical membership degree of each behavioral motivation of the feature information corresponding to historical multimedia files, historical user data, and historical context data, and the feature vector corresponding to the latent space containing each behavioral motivation, and the behavioral motivation is used to indicate the reason for performing the user behavior corresponding to the behavioral motivation in the target environment; the processing unit is further configured to determine recommendation data based on the actual click-through rate and actual conversion rate; wherein the recommendation data includes at least one actual multimedia file; the processing unit is further configured to control a sending unit to send the recommendation data to an electronic device.

[0008] Thirdly, this disclosure provides an electronic device, including: a memory and a processor, wherein the memory is used to store a computer program; and the processor is used to cause the electronic device to perform any of the recommended methods provided in the first aspect above when executing the computer program.

[0009] Fourthly, this disclosure provides a computer-readable storage medium, comprising: a computer program stored on the computer-readable storage medium, which, when executed by a computing device, causes the computing device to implement any of the recommended methods provided in the first aspect above.

[0010] Fifthly, the present invention provides a computer program product that, when run on a computer, causes the computer to perform the recommended method as provided in any of the first aspects.

[0011] It should be noted that the aforementioned computer instructions may be stored, in whole or in part, on the first computer-readable storage medium. The first computer-readable storage medium may be packaged together with the processor of the recommended device, or it may be packaged separately from the processor of the recommended device; this disclosure does not impose any limitations on this.

[0012] The descriptions of the second, third, fourth, and fifth aspects in this disclosure can be referenced to the detailed description of the first aspect; and the beneficial effects of the descriptions of the second, third, fourth, and fifth aspects can be referenced to the analysis of the beneficial effects of the first aspect, which will not be repeated here.

[0013] In this disclosure, the names of the recommended devices do not limit the devices or functional modules themselves. In actual implementation, these devices or functional modules may appear under other names. As long as the functions of each device or functional module are similar to those of this disclosure, they fall within the scope of the claims of this disclosure and their equivalents.

[0014] These or other aspects of this disclosure will become more readily apparent in the following description.

[0015] The technical solution provided in this disclosure has the following advantages compared with the prior art:

[0016] Upon receiving a request from an electronic device, the actual user data is determined based on the received actual user identifier. Pre-configured media asset data, actual user data, and actual context data are input into the recommendation model. Since the recommendation model is trained based on the historical membership degree of each behavioral motivation and the latent space feature vector corresponding to each behavioral motivation, derived from the feature information corresponding to historical multimedia files, historical user data, and historical context data, the model can analyze the behavioral motivation of the action performed in the current actual context data. Based on this behavioral motivation, the actual click-through rate (CTR) and actual conversion rate for each actual multimedia file in the media asset data can be determined. Then, based on the actual CTR and actual conversion rate, recommendation data is determined, sent, and recommended to the electronic device. Because the actual CTR and actual conversion rate are determined after analyzing the behavioral motivation, the recommendation accuracy of multimedia files can be improved more accurately, ensuring a better user experience and solving the problem of how to improve the recommendation accuracy of multimedia files. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0018] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram illustrating a scenario for the recommended method provided in the embodiments of this application;

[0020] Figure 2 This is one of the structural schematic diagrams of the display device in the recommended method provided in the embodiments of this application;

[0021] Figure 3 This is the second schematic diagram of the structure of the display device in the recommended method provided in the embodiments of this application;

[0022] Figure 4 One of the flowcharts of the recommended method provided in the embodiments of this application;

[0023] Figure 5 A second schematic flowchart of the recommended method provided in the embodiments of this application;

[0024] Figure 6 A schematic diagram of the recommendation model in the recommendation method provided in the embodiments of this application;

[0025] Figure 7 The third flowchart illustrating the recommended method provided in the embodiments of this application;

[0026] Figure 8 The fourth flowchart illustrating the recommended method provided in the embodiments of this application;

[0027] Figure 9 This is a schematic diagram of the server structure provided in an embodiment of this application;

[0028] Figure 10 This is a schematic diagram of a chip system provided in an embodiment of this application. Detailed Implementation

[0029] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0030] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0032] In this embodiment of the disclosure, the deep neural network refers to a deep neural network, or DNN for short.

[0033] In this disclosure, Adam refers to Adaptive Moment Estimation.

[0034] The multi-task learning model in this embodiment refers to Multi-gate Mixture-of-Experts, abbreviated as MMOE.

[0035] In this embodiment of the disclosure, the latent space refers to the Latent Space.

[0036] In this embodiment, "rule" refers to the linear rectification function, also known as the modified linear unit, which is a commonly used activation function in artificial neural networks. It usually refers to nonlinear functions represented by the ramp function and its variants.

[0037] In this embodiment of the disclosure, Sigmoid refers to the Sigmoid function, which is a mathematical function with a beautiful S-shaped curve and has wide applications in logistic regression and artificial neural networks.

[0038] In this embodiment of the disclosure, Softmax refers to the Softmax function, which is often used as an activation function in the output layer of a neural network. It maps the values ​​of the output layer to the 0-1 interval through the activation function, and constructs a probability distribution for the neuron output.

[0039] In this disclosure, CNN refers to Convolutional Neural Networks (CNN).

[0040] In this embodiment of the disclosure, RNN refers to Recurrent Neural Network (RNN).

[0041] Figure 1 This is a schematic diagram illustrating an operational scenario between a display device and a control device according to one or more embodiments of this application, such as... Figure 1As shown, a user can operate the display device 200 via a mobile terminal 300 and a control device 100. The control device 100 can be a remote control, and communication between the remote control and the display device includes infrared protocol communication, Bluetooth protocol communication, wireless or other wired methods to control the display device 200. The user can input user commands through buttons on the remote control, voice input, control panel input, etc., to control the display device 200. In some embodiments, a mobile terminal, tablet computer, computer, laptop computer, and other smart devices can also be used to control the display device 200.

[0042] In some embodiments, the mobile terminal 300 can install software applications with the display device 200 to achieve connection and communication via network communication protocols, enabling one-to-one control operations and data communication. It can also transmit audio and video content displayed on the mobile terminal 300 to the display device 200 for synchronous display. The display device 200 can also communicate with the display device 200 via various communication methods. It can be allowed to communicate via a local area network (LAN), wireless local area network (WLAN), and other networks. The display device 200 can provide various content and interactive features. The display device 200 can be a liquid crystal display, an OLED display, or a projection display device. In addition to providing broadcast television reception functions, the display device 200 can also be equipped with a smart network television function that provides computer support.

[0043] In some embodiments, the electronic device provided in this application can be the server 400 described above. After the display device 200 is powered on, it sends a request message carrying the actual user identifier and actual context data to the server 400. Upon receiving the request message from the display device 200, the server 400 determines the actual user data based on the received actual user identifier. The server 400 inputs pre-configured media asset data, actual user data, and actual context data into the recommendation model. Since the recommendation model is trained based on feature vectors corresponding to the latent space containing each behavioral motivation, it can unify differences across different domains and scenarios. Simultaneously, based on the analysis of the actual user data and actual context data, behavioral motivations can be determined. Furthermore, based on the behavioral motivations and media asset data, the actual click-through rate and actual conversion rate corresponding to each actual multimedia file in the media asset data can be obtained. Then, the server 400 determines the recommendation data based on the actual click-through rate and actual conversion rate. Since the actual click-through rate and actual conversion rate are determined after fully analyzing behavioral motivations, the recommendation accuracy of multimedia files can be improved more accurately, ensuring a better user experience. Finally, the server 400 sends the recommendation data to the electronic device. In this way, the display device 200 can generate a corresponding recommendation page for the user to choose from based on the recommendation data sent by the display server 400.

[0044] Figure 2 A hardware configuration block diagram of a display device 200 according to an exemplary embodiment is shown. For example... Figure 2 The display device 200 shown includes at least one of the following: a tuner / demodulator 210, a communicator 220, a detector 230, an external device interface 240, a controller 250, a display 260, an audio output interface 270, a memory, a power supply, and a user interface 280. The controller includes a central processing unit, a video processor, an audio processor, a graphics processor, RAM, ROM, and a first to nth interface for input / output. The display 260 may be a touch-enabled display, such as a touch screen display. The tuner / demodulator 210 receives broadcast television signals via wired or wireless reception and demodulates audio and video signals, such as EPG data signals, from multiple wireless or wired broadcast television signals. The detector 230 is used to collect signals from the external environment or signals interacting with the external environment. The controller 250 and the tuner / demodulator 210 may be located in different separate devices; that is, the tuner / demodulator 210 may also be located in an external device of the main device containing the controller 250, such as an external set-top box.

[0045] In some embodiments, the controller 250 controls the operation of the display device and responds to user operations through various software control programs stored in memory. The controller 250 controls the overall operation of the display device 200.

[0046] In some examples, the display device 200 of one or more embodiments is a television 1, and the operating system of the television 1 is the Android system, for example... Figure 3 As shown, TV 1 can be logically divided into an application layer (referred to as "application layer") 21, an application framework layer (referred to as "framework layer") 22, an Android runtime and system library layer (referred to as "system runtime library layer") 23, and a kernel layer 24.

[0047] The application layer 21 includes one or more applications. These applications can be system applications or third-party applications. For example, application layer 21 may include a first application that provides multimedia file recommendation services. The framework layer 22 provides application programming interfaces (APIs) and programming frameworks for the applications in application layer 21. The system runtime library layer 23 supports the upper layer, namely the framework layer 22. When the framework layer 22 is used, the Android operating system runs the C / C++ libraries contained in the system runtime library layer 23 to implement the functions required by the framework layer 22. The kernel layer 24 acts as software middleware between the hardware layer and application layer 21, managing and controlling hardware and software resources.

[0048] In some examples, after the display device 200 is powered on, the first application sends a request message carrying the actual user identifier and actual context data to the server 400. Subsequently, after the receiving unit 401 of the server 400 receives the request message from the display device 200, the processing unit 402 of the server 400 determines the actual user data based on the actual user identifier received by the receiving unit 401. The processing unit 402 inputs pre-configured media asset data, actual user data, and actual context data into the recommendation model. Since the recommendation model is trained based on feature vectors corresponding to the latent space containing each behavioral motivation, it can unify differences across different domains and scenarios. Simultaneously, based on the analysis of the actual user data and actual context data, behavioral motivations can be determined. Furthermore, based on behavioral motivations and media asset data, the actual click-through rate and actual conversion rate corresponding to each actual multimedia file in the media asset data can be obtained. Then, the processing unit 402 determines the recommendation data based on the actual click-through rate and actual conversion rate. Since the actual click-through rate and actual conversion rate are determined after thorough analysis of behavioral motivations, the recommendation accuracy of multimedia files can be improved more accurately, ensuring a better user experience. Finally, the processing unit 402 controls the sending unit 403 to send recommendation data to the electronic device. In this way, the display device 200 can generate a corresponding recommendation page for the user to choose from based on the recommendation data sent by the display server 400.

[0049] Specifically, storage unit 404 is used to store the correspondence between actual user identifiers and actual user data, model data of recommendation models, etc.

[0050] Specifically, location information can be provided by systems that offer positioning services, such as the BeiDou Navigation Satellite System (BDS), the Global Positioning System (GPS), and GLONASS.

[0051] The actual user identifiers and actual user data involved in this application can be data authorized by the user or fully authorized by all parties.

[0052] In the following embodiments, the server 400 described above is used as the execution subject for the recommended method provided in the embodiments of this disclosure to illustrate the method of this application.

[0053] This application provides a recommended method, such as... Figure 4 As shown, the recommended method may include S11-S15.

[0054] S11. Receive request information sent by an electronic device. The request information includes at least the actual user identifier and actual context data, and the actual context data includes at least time information and location information.

[0055] In some examples, the recommendation method provided in this disclosure analyzes the behavioral motivations behind various behaviors. For instance, a user might prefer watching variety shows during weekday lunch breaks because the lunch break is short, and relatively short variety shows during this lunch break satisfy the user's needs. In the evening, the user might watch more anime or children's programs because their children need to watch movies after school. Thus, by analyzing the user's behavioral motivations, the accuracy of multimedia file recommendations can be improved.

[0056] In some examples, the actual user identifier can be the device code of the electronic device (such as the device serial number, physical address, etc.) or the user account (such as the mobile phone number, etc.).

[0057] S12. Determine the actual user data based on the actual user identifier.

[0058] In some examples, actual user data includes registration information such as user gender and age, as well as preference data, such as the percentage of multimedia files viewed in each category within a specified time period.

[0059] In some examples, the server 400 pre-stores the mapping relationship between each actual user identifier and the actual user data in its memory. Thus, when the server 400 needs to determine the actual user data, it can query the mapping relationship in memory based on the actual user identifier to determine the actual user data corresponding to that actual user identifier.

[0060] In some examples, a user data model is pre-trained in server 400. When it is necessary to determine the actual user data, the actual user identifier can be input into the user data model to determine the actual user data.

[0061] The training process for the user data model is as follows:

[0062] Acquire training sample data and training supervision data; wherein, both training sample data and training supervision data include training user identifiers and training user data corresponding to those training user identifiers.

[0063] The network parameters of the neural network model are repeatedly adjusted based on the training sample data until the neural network model converges, resulting in a pre-trained neural network model.

[0064] The pre-trained neural network model is trained under supervised training data until the network parameters of the pre-trained neural network model no longer change, thus obtaining the user data model.

[0065] S13. Input the pre-configured media asset data, actual user data, and actual context data into the recommendation model to obtain the actual click-through rate and actual conversion rate corresponding to each actual multimedia file in the media asset data. The recommendation model is trained based on the historical membership degree of each behavioral motivation derived from the feature information corresponding to historical multimedia files, historical user data, and historical context data, as well as the feature vector corresponding to the latent space containing each behavioral motivation. The behavioral motivation is used to indicate the reason for the user's behavior corresponding to the behavioral motivation in the target environment.

[0066] In some examples, the server 400 pre-stores media asset data so that when an electronic device requests to play the actual multimedia file in the media asset data, the actual multimedia file can be quickly sent to the electronic device, ensuring a good user experience.

[0067] S14. Determine the recommended data based on the actual click-through rate and actual conversion rate. The recommended data shall include at least one actual multimedia file.

[0068] In some examples, a recommendation score can be determined based on actual click-through rate and actual conversion rate. The recommendation score indicates the probability of accessing the actual multimedia file corresponding to that score. Based on the recommendation score, actual multimedia files with a recommendation score greater than or equal to a score threshold are selected. These multimedia files with a recommendation score greater than or equal to the score threshold are then used as recommendation data.

[0069] In some examples, the recommendation score is equal to the sum of the actual click-through rate and the actual conversion rate, or the recommendation score is equal to the average of the actual click-through rate and the actual conversion rate; this is not a limitation here.

[0070] In some examples, when determining recommended data based on the relationship between recommendation scores and score thresholds, there may be a large number of actual multimedia files whose recommendation scores are greater than or equal to the score threshold. To better generate recommended data, the recommendation method provided in this disclosure further filters these actual multimedia files. For example, it sorts the recommendation scores of these files in descending order of their scores, obtaining a sorting result. Based on the sorting result, it filters recommendation scores that are greater than or equal to the target ranking (e.g., target ranking is 3). Finally, it uses the actual multimedia files corresponding to recommendation scores greater than or equal to the target ranking as recommended data. This significantly reduces the number of actual multimedia files in the recommended data and improves the accuracy of recommending actual multimedia files.

[0071] Alternatively, based on actual user data, determine the percentage of multimedia files of different media asset types (e.g., movies, TV series, news, etc.) accessed by users within a preset time period, and identify media asset types with a percentage greater than or equal to a preset percentage. Then, among the actual multimedia files with a recommendation score greater than or equal to a score threshold, filter out the actual multimedia files corresponding to media asset types with a percentage greater than or equal to a preset percentage, and use these filtered media asset types as recommendation data.

[0072] S15. Send recommendation data to electronic devices.

[0073] As described above, the recommendation method provided in this embodiment unifies the differences between different media asset types by mapping behavioral motivations into a latent space. Secondly, by analyzing actual user data and actual context data, the behavioral motivation corresponding to the current behavior can be determined. Then, the semantic expression corresponding to this behavioral motivation is obtained. Finally, by combining the semantic expression with the media asset features of the actual multimedia files, actual multimedia files that conform to the semantic expression can be selected. Since the recommendation data is based on semantic expression, the accuracy of multimedia file recommendations can be improved, ensuring a better user experience.

[0074] In some feasible examples, the recommendation model includes at least a feature extraction sub-model, a motivation identification sub-model, a data prediction sub-model, and a dual-tower sub-model. The feature extraction sub-model extracts the media asset features of each actual multimedia file; the motivation identification sub-model determines historical membership; the data prediction sub-model determines the click result for each actual multimedia file based on the media asset features determined by the feature extraction sub-model, the historical membership determined by the motivation identification sub-model, and the feature vector; and the dual-tower sub-model determines the actual click-through rate and actual conversion rate for each actual multimedia file based on the click result determined by the data prediction sub-model. The click result includes either "not clicked" or "clicked".

[0075] In some feasible examples, combining Figure 4 ,like Figure 5 As shown, the recommended method provided in this embodiment of the disclosure further includes: S16-S22.

[0076] S16. Obtain training sample data and the labeling results of the training sample data. The training sample data includes feature vectors, at least one historical multimedia file, and at least one set of training data, which includes historical user data and historical context data. The labeling results include the historical actual click-through rate and historical actual conversion rate corresponding to each historical multimedia file.

[0077] S17. Input the historical multimedia files into the feature extraction sub-model to obtain the extraction results of the historical multimedia files by the feature extraction sub-model.

[0078] S18. Input the training data into the motivation recognition sub-model to obtain the recognition result of the motivation recognition sub-model on the training data.

[0079] S19. Input the extraction results, recognition results, and feature vectors into the data prediction sub-model to obtain the prediction results of the data prediction sub-model for the prediction results and feature vectors.

[0080] In some examples, when inputting the extraction results, recognition results, and feature vectors into the data prediction sub-model, the extraction results, recognition results, and feature vectors need to be concatenated, and then the concatenated features are input into the data prediction sub-model.

[0081] S20. Input the prediction results into the dual-tower sub-model to obtain the recall results of the dual-tower sub-model for the prediction results.

[0082] S21. Calculate the loss values ​​of the recall and labeling results based on the target loss function.

[0083] S22. When the loss value is greater than or equal to the loss threshold, adjust the first network parameters of the feature extraction sub-model, the second network parameters of the motivation recognition sub-model, the third network parameters of the data prediction sub-model, and the fourth network parameters of the dual-tower sub-model until the loss value is less than the loss threshold, thus obtaining the recommendation model.

[0084] In some examples, the network structure of the recommendation model is as follows: Figure 6 As shown in the diagram. Here, "item" represents media asset data, "user" represents actual user data, and "context" represents actual context data. The embedding layer is used to transform media asset data, actual user data, and actual context data into a data format recognizable by the recommendation model. Item embedding refers to the media asset features of the item, user embedding refers to the user features corresponding to the actual user data, and context embedding refers to the context features corresponding to the actual context data.

[0085] In some examples, the recommended method provided in this disclosure defines behavioral motivation as the root cause of a user's behavior in a specific environment, and maps the root causes of all user behaviors to a specific latent space, i.e., L∈R. M*N Where M represents the vector length of behavioral motivation in the latent space, N represents the length of the set of all behavioral motivations, and L represents the behavioral motivation.

[0086] In some examples, since user motivations differ depending on the actual context data, the recommended method provided in this disclosure inputs the user features corresponding to the actual user data and the context features corresponding to the actual context data into the motivation recognition sub-model. This allows the method to obtain the membership degree of the user's behavior for each motivation corresponding to both the actual user data and the actual context data.

[0087]

[0088] Here, the function g(*) can be expressed using a deep neural network (DNN), and the function σ(·) = sigmoid(·) is the activation function, m∈R. Nc represents the membership degree of each behavioral motivation in the latent space corresponding to the actual user data and the actual context data. i，j Indicates user u i Clicked on item v j Contextual features of time, Indicates user u i User characteristics.

[0089] It's worth noting that the ReLU activation function isn't used here to prevent some feature vectors in the latent space from having excessively large magnitudes, which could lead to training instability. It also effectively avoids the problem of uneven distribution of latent space vectors. The softmax activation function, with an output sum of 1, is unsuitable for scenarios involving multiple behavioral motivations.

[0090] In some examples, based on the membership degree of the user's behavior in each behavioral motivation and the feature vector in the latent space corresponding to both the actual user data and the actual context data, the semantic expression of the user's behavior can be obtained, i.e.:

[0091]

[0092] Where ⊙ represents matrix dot product, rep(m T M, axis = 1) ∈ R M*N This represents transposing vector m and copying it M times along the second dimension. Vector m represents the membership degree of each behavioral motivation of the user's behavior in the latent space corresponding to the actual user data and the actual context data. Indicates user u i The semantic expression of behavior.

[0093] It should be noted that due to the significant differences between different scenarios (e.g., movies and animations, although both can be viewed as multimedia files, they target different groups, leading to differences in actual user data), the representation of L will also differ in different scenarios. Therefore, the recommended method provided in this disclosure embodiment imposes the following restrictions on the latent space:

[0094] Constraint 1: In the same scenario, the intra-class distance of the behavioral motivation representation of the samples should be as small as possible.

[0095] Limitation 2: In different scenarios, the boundary distance of the inter-class spacing representing the behavioral motivation of the samples should be as large as possible.

[0096] We address these two limitations using central loss, namely:

[0097] l c =-l intra +αl inter .

[0098] Among them, l intra Indicates restriction 1, l inter This represents constraint 2, and α is used to adjust the weight between constraint 1 and constraint 2.

[0099]

[0100] in, The cluster center representing a scene is the arithmetic mean of the feature maps of all samples belonging to that cluster. The set S = {s1, s2, s3, s4, s5, s6} represents the set of all scenes. Here we have 6 major scenes: TV series, variety shows, documentaries, movies, animation, and children's shows.

[0101]

[0102] In this way, the difference in behavioral motivation expression in different scenarios can be increased by increasing the inter-class spacing, thereby improving the recommendation accuracy of multimedia files.

[0103] As can be seen, the recommendation method provided in this disclosure, by re-embedding user features, unifies the differences across multiple scenarios, thereby adapting to situations with significant scenario variations. Simultaneously, expressing scenario-specific limitations through a loss function demonstrates enhanced capabilities in extracting behavioral motivations.

[0104] In some examples, the input to the feature extraction sub-model is the material features of the item, mainly used to extract higher-order features of the item, and its expression is:

[0105]

[0106] in, The feature map represents the extracted features of media assets. The function representing the extraction of media asset features. Represents item v j Its characteristics.

[0107] For example, f(*) can be a network structure such as DNN, CNN or RNN.

[0108] For example, in the recommended method provided in the embodiments of this disclosure, f(*) adopts a DNN network structure.

[0109] In some examples, the semantic representation of user behavior and the extracted feature map are used as input to the data prediction sub-model, and the output is the user u. i Did you click on item v? j ,Right now:

[0110]

[0111] Where, pi,j This indicates that the data prediction sub-model predicts user u i Did you click on item v? j The probability of is output as the result.

[0112] For example, P i,j It is a floating-point value in the interval [0, 1], representing user u. i Clicked on item v j The probability size.

[0113] For example, the semantic representation of user behavior and the extracted feature map can be concatenated, and the concatenated features can be input into the data prediction sub-model.

[0114] For example, gating can be used to assign weights to different tasks based on the features input to expert n.

[0115] In some examples, the dual-tower sub-model's tasks are click-through rate (CTR) prediction and conversion rate (CVR) prediction, where CTR represents the number of users. i Did you click on item v? j The first task is represented by a Tower 1 model. The second task, CVR, represents whether a user's viewing behavior constitutes a valid view; it's a 0 / 1 binary classification problem, represented by a Tower 2 model. Both tasks use a log loss function as the loss function, i.e.:

[0116]

[0117] Where D represents the number of data samples, and Ω is the regularization term. The output probability representing CTR (i.e., the actual click-through rate). βl represents the output probability (i.e., actual conversion rate) of CVR. c Let represent the central loss, where β is a hyperparameter.

[0118] For example, the Ω regularization term uses L2 regularization.

[0119] It should be noted that during the training of the recommendation model, the optimizer used is Adam, and the initial learning rate is 0.001. Due to the huge amount of data, the maximum number of steps in the entire training is set to 10, and the learning rate decays by a factor of 0.9 after each step.

[0120] In some examples, when the recommendation model is serving online, it uses a cluster averaging approach to sum the actual click-through rate and the actual conversion rate to obtain the final score, i.e.:

[0121]

[0122] Here, a and b are both hyperparameters.

[0123] For example, a + b = 1, a > 0, b > 0.

[0124] In some examples, after receiving recommendation data from the server (400), the electronic device will display the multimedia files in the recommendation data in reverse order based on their scores on the front end.

[0125] In some feasible examples, the feature extraction sub-model, the motivation recognition sub-model, and the dual-tower sub-model all include deep neural networks, and the data prediction sub-model includes a multi-task learning model.

[0126] In some feasible examples, the target loss function includes the log loss function.

[0127] In some feasible examples, combining Figure 4 ,like Figure 7 As shown, the above S14 can be implemented by the following S140-S142.

[0128] S140. Determine the recommendation score based on the actual click-through rate and actual conversion rate. The recommendation score indicates the probability of accessing the actual multimedia file corresponding to the recommended score.

[0129] S141. Based on the recommendation score, filter out actual multimedia files whose recommendation score is greater than or equal to the score threshold.

[0130] S142. Use actual multimedia files with a recommendation score greater than or equal to the score threshold as recommendation data.

[0131] In some feasible examples, combining Figure 7 ,like Figure 8 As shown, the above S142 can be specifically implemented through the following S1420-S1422.

[0132] S1420. For actual multimedia files with a recommended score greater than or equal to the score threshold, sort the recommended scores greater than or equal to the score threshold in descending order to obtain the sorting results.

[0133] S1421. Based on the sorting results, filter for recommended scores that are greater than or equal to the target sorting score. Each recommended score corresponds to one sorting score.

[0134] S1422. Use the actual multimedia files corresponding to the ranking scores that are greater than or equal to the ranking threshold as the recommendation data.

[0135] For example, given user 1, assume the media asset data includes 5 actual multimedia files: actual multimedia file 1, actual multimedia file 2, actual multimedia file 3, actual multimedia file 4, and actual multimedia file 5. After calculations in steps S11-S14, the recommendation score for actual multimedia file 1 is determined to be 0.95, the recommendation score for actual multimedia file 2 is 0.6, the recommendation score for actual multimedia file 3 is 0.72, the recommendation score for actual multimedia file 4 is 0.83, and the recommendation score for actual multimedia file 5 is 0.3. When the score threshold is 0.5, it can be determined that the recommendation scores for actual multimedia file 1, actual multimedia file 2, actual multimedia file 3, and actual multimedia file 4 are all greater than the score threshold of 0.5. Then, the recommendation scores greater than or equal to the score threshold are sorted in descending order, resulting in the sorting results shown in Table 1.

[0136] Table 1

[0137]

[0138] The higher the recommendation score, the higher the corresponding target ranking, meaning users can see media assets more easily.

[0139] Thus, if only two media assets are displayed on the front end, then actual multimedia file 1 and actual multimedia file 4 can be presented to the user as recommended data.

[0140] In some examples, the recommendation method provided in this disclosure can not only train the recommendation model using actual user data and actual context data, but also solve the problem of significant differences in domain tasks. In actual projects, the accuracy of user click-through rate prediction has been improved by 2.4%; and performance stability has also been improved, with a 3.2% reduction in the variance of long-term click-through rate prediction. Furthermore, this method is a general solution and can be applied to personalized recommendations in other scenarios, such as product recommendations.

[0141] The foregoing mainly describes the solutions provided by the embodiments of this application from a methodological perspective. To achieve the above functions, it includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0142] This application embodiment can divide the recommended device into functional modules based on the above method example. For example, each function can be divided into its own functional modules, or two or more functions can be integrated into one processing unit. The integrated modules can be implemented in hardware or as software functional modules. It should be noted that the module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0143] like Figure 9 As shown in the diagram, an embodiment of this application provides a schematic diagram of the structure of a server 400. It includes a communicator 101 and a processor 102.

[0144] The communicator 101 is used to receive request information sent by an electronic device; wherein the request information includes at least an actual user identifier and actual context data, and the actual context data includes at least time information and location information; the processor 102 is used to determine actual user data based on the actual user identifier received by the communicator 101; the processor 102 is also used to input pre-configured media asset data, actual user data and actual context data received by the communicator 101 into a recommendation model to obtain the actual click-through rate and actual conversion rate corresponding to each actual multimedia file in the media asset data; wherein the recommendation model is trained based on the feature information corresponding to historical multimedia files, historical user data and historical context data, the historical membership degree of each behavioral motivation, and the feature vector corresponding to the latent space containing each behavioral motivation, and the behavioral motivation is used to indicate the reason for the user to perform the corresponding user behavior in the target environment; the processor 102 is also used to determine recommendation data based on the actual click-through rate and actual conversion rate; wherein the recommendation data includes at least one actual multimedia file; the processor 102 is also used to control the communicator 102 to send the recommendation data to the electronic device.

[0145] In some feasible examples, the recommendation model includes at least a feature extraction sub-model, a motivation identification sub-model, a data prediction sub-model, and a dual-tower sub-model. The feature extraction sub-model extracts the media asset features of each actual multimedia file; the motivation identification sub-model determines historical membership; the data prediction sub-model determines the click result for each actual multimedia file based on the media asset features determined by the feature extraction sub-model, the historical membership determined by the motivation identification sub-model, and the feature vector; and the dual-tower sub-model determines the actual click-through rate and actual conversion rate for each actual multimedia file based on the click result determined by the data prediction sub-model. The click result includes either "not clicked" or "clicked".

[0146] In some feasible examples, the training process of the recommendation model is as follows: A communicator 101 is further configured to acquire training sample data and the labeling results of the training sample data; wherein, the training sample data includes feature vectors, at least one historical multimedia file, and at least one set of training data, the training data including historical user data and historical context data, and the labeling results including the historical actual click-through rate and historical actual conversion rate corresponding to each historical multimedia file; a processor 102 is further configured to input the historical multimedia files acquired by the communicator 101 into the feature extraction sub-model to obtain the extraction results of the historical multimedia files by the feature extraction sub-model; a processor 102 is further configured to input the training data acquired by the communicator 101 into the motivation recognition sub-model to obtain the recognition results of the motivation recognition sub-model on the training data; Processor 102 is further configured to input the extraction result, the recognition result, and the feature vector obtained by communicator 101 into the data prediction sub-model to obtain the prediction result of the data prediction sub-model for the prediction result and the feature vector; processor 102 is further configured to input the prediction result into the dual-tower sub-model to obtain the recall result of the dual-tower sub-model for the prediction result; processor 102 is further configured to calculate the loss value of the recall result and the labeling result based on the target loss function; processor 102 is further configured to adjust the first network parameter of the feature extraction sub-model, the second network parameter of the motivation recognition sub-model, the third network parameter of the data prediction sub-model, and the fourth network parameter of the dual-tower sub-model when the loss value is greater than or equal to the loss threshold, until the loss value is less than the loss threshold, to obtain the recommendation model.

[0147] In some feasible examples, the feature extraction sub-model, the motivation recognition sub-model, and the dual-tower sub-model all include deep neural networks, and the data prediction sub-model includes a multi-task learning model.

[0148] In some feasible examples, the target loss function includes the log loss function.

[0149] In some implementable examples, processor 102 is specifically used to determine a recommendation score based on the actual click-through rate and the actual conversion rate; wherein the recommendation score is used to indicate the access probability of accessing the actual multimedia file corresponding to the recommendation score; processor 102 is specifically used to filter actual multimedia files whose recommendation scores are greater than or equal to a score threshold based on the recommendation score; processor 102 is specifically used to use actual multimedia files whose recommendation scores are greater than or equal to the score threshold as recommendation data.

[0150] In some implementable examples, processor 102 is specifically used to sort the actual multimedia files whose recommendation scores are greater than or equal to a score threshold in descending order, obtaining a sorting result; processor 102 is specifically used to filter recommendation scores greater than or equal to a target sorting score based on the sorting result; wherein, one recommendation score corresponds to one sorting score. Processor 102 is specifically used to use the actual multimedia files corresponding to the scores greater than or equal to the sorting threshold as recommendation data.

[0151] All relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and their functions will not be repeated here.

[0152] Of course, the server 400 provided in this application embodiment includes, but is not limited to, the modules described above. For example, the server 400 may also include a memory 103. The memory 103 may be used to store the program code of the server 400, and may also be used to store data generated by the server 400 during operation, such as data in write requests.

[0153] As an example, combined Figure 3 The acquisition unit 401 and the sending unit 403 in the server 400 perform the same functions as the communicator 101. The processing unit 402 performs the same functions as the processor 102. The storage unit 404 performs the same functions as the memory 103.

[0154] like Figure 10 As shown, this application embodiment also provides a chip system that can be applied to the server 400 in the foregoing embodiments. The chip system includes at least one processor 1501 and at least one interface circuit 1502. The processor 1501 can be the processor in the server 400. The processor 1501 and the interface circuit 1502 can be interconnected via a line. The processor 1501 can receive and execute computer instructions from the memory of the server 400 through the interface circuit 1502. When the computer instructions are executed by the processor 1501, the server 400 can perform the various steps performed by the server 400 in the foregoing embodiments. Of course, the chip system may also include other discrete devices, which are not specifically limited in this application embodiment.

[0155] This application also provides a computer-readable storage medium for storing computer instructions executed by the server 400 described above.

[0156] This application also provides a computer program product, including computer instructions executed by the server 400 described above.

[0157] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A recommendation method, characterized in that, include: Receive request information sent by an electronic device; wherein the request information includes at least an actual user identifier and actual context data, and the actual context data includes at least time information and location information; Based on the actual user identifier, determine the actual user data; The pre-configured media asset data, the actual user data, and the actual context data are input into the recommendation model to obtain the actual click-through rate and actual conversion rate corresponding to each actual multimedia file in the media asset data. The recommendation model is trained based on the feature information corresponding to historical multimedia files, historical user data, and historical context data, as well as the historical membership degree of each behavioral motivation and the feature vector corresponding to the latent space containing each behavioral motivation. The behavioral motivation is used to indicate the reason for the user to perform the corresponding user behavior in the target environment. Recommended data is determined based on the actual click-through rate and the actual conversion rate; wherein the recommended data includes at least one actual multimedia file; The recommendation data is sent to the electronic device; wherein, The recommendation model includes at least a feature extraction sub-model, a motivation identification sub-model, a data prediction sub-model, and a dual-tower sub-model. The feature extraction sub-model extracts media asset features for each actual multimedia file. The motivation identification sub-model determines the historical membership degree. The data prediction sub-model determines the click result for each actual multimedia file based on the media asset features determined by the feature extraction sub-model, the historical membership degree determined by the motivation identification sub-model, and the feature vector. The dual-tower sub-model determines the actual click-through rate and actual conversion rate for each actual multimedia file based on the click result determined by the data prediction sub-model. The click result includes either "not clicked" or "clicked".

2. The recommended method according to claim 1, characterized in that, The training process of the recommendation model is as follows: Obtain training sample data and the labeling results of the training sample data; wherein, the training sample data includes the feature vector, at least one historical multimedia file, and at least one set of training data, the training data includes historical user data and historical context data, and the labeling results include the historical actual click-through rate and historical actual conversion rate corresponding to each historical multimedia file; The historical multimedia file is input into the feature extraction sub-model to obtain the feature extraction result of the historical multimedia file by the feature extraction sub-model. The training data is input into the motivation recognition sub-model to obtain the recognition result of the motivation recognition sub-model on the training data; The extraction result, the recognition result, and the feature vector are concatenated to obtain the concatenated feature. The concatenated features are input into the data prediction sub-model to obtain the prediction result of the data prediction sub-model for the concatenated features; The prediction result is input into the dual-tower sub-model to obtain the recall result of the dual-tower sub-model for the prediction result; The loss values ​​of the recall results and the labeling results are calculated based on the target loss function; If the loss value is greater than or equal to the loss threshold, adjust the first network parameters of the feature extraction sub-model, the second network parameters of the motivation recognition sub-model, the third network parameters of the data prediction sub-model, and the fourth network parameters of the dual-tower sub-model until the loss value is less than the loss threshold, thus obtaining the recommendation model.

3. The recommended method according to claim 2, characterized in that, The feature extraction sub-model, the motivation recognition sub-model, and the dual-tower sub-model all include deep neural networks, and the data prediction sub-model includes a multi-task learning model.

4. The recommended method according to claim 2, characterized in that, The target loss function includes the logarithmic loss function.

5. The recommended method according to claim 1, characterized in that, The step of determining the recommendation data based on the actual click-through rate and the actual conversion rate includes: A recommendation score is determined based on the actual click-through rate and the actual conversion rate; wherein, the recommendation score is used to indicate the probability of accessing the actual multimedia file corresponding to the recommendation score; Based on the recommendation score, filter out actual multimedia files whose recommendation score is greater than or equal to the score threshold; The actual multimedia files whose recommendation scores are greater than or equal to the score threshold are used as recommendation data.

6. The recommended method according to claim 5, characterized in that, The step of using actual multimedia files with recommendation scores greater than or equal to the score threshold as recommendation data includes: For actual multimedia files whose recommendation scores are greater than or equal to the score threshold, sort the recommendation scores that are greater than or equal to the score threshold in descending order to obtain the sorting result; Based on the sorting results, the recommended scores that are greater than or equal to the target sorting score are filtered out; wherein, one recommended score corresponds to one sorting score; The actual multimedia files corresponding to the ranking scores that are greater than or equal to the ranking threshold are used as recommendation data.

7. A recommended device, characterized in that, include: A receiving unit is configured to receive request information sent by an electronic device; wherein the request information includes at least an actual user identifier and actual context data, and the actual context data includes at least time information and location information; The processing unit is used to determine the actual user data based on the actual user identifier received by the receiving unit; The processing unit is further configured to input pre-configured media asset data, the actual user data, and the actual context data received by the receiving unit into the recommendation model to obtain the actual click-through rate and actual conversion rate corresponding to each actual multimedia file in the media asset data; wherein, the recommendation model is trained based on the feature information corresponding to historical multimedia files, historical user data, and historical context data, as well as the historical membership degree of each behavioral motivation and the feature vector corresponding to the latent space containing each behavioral motivation, and the behavioral motivation is used to indicate the reason for performing the user behavior corresponding to the behavioral motivation in the target environment; The processing unit is further configured to determine recommended data based on the actual click-through rate and the actual conversion rate; wherein the recommended data includes at least one actual multimedia file; The processing unit is also configured to control the sending unit to send the recommendation data to the electronic device; The recommendation model includes at least a feature extraction sub-model, a motivation identification sub-model, a data prediction sub-model, and a dual-tower sub-model. The feature extraction sub-model extracts media asset features for each actual multimedia file. The motivation identification sub-model determines the historical membership degree. The data prediction sub-model determines the click result for each actual multimedia file based on the media asset features determined by the feature extraction sub-model, the historical membership degree determined by the motivation identification sub-model, and the feature vector. The dual-tower sub-model determines the actual click-through rate and actual conversion rate for each actual multimedia file based on the click result determined by the data prediction sub-model. The click result includes either "not clicked" or "clicked".

8. An electronic device, characterized in that, include: A memory and a processor, the memory being used to store a computer program; the processor being used to cause the electronic device to implement the recommended method according to any one of claims 1-6 when executing the computer program.

9. A computer-readable storage medium, characterized in that, include: The computer-readable storage medium stores a computer program that, when executed by a computing device, causes the computing device to implement the recommended method according to any one of claims 1-6.

Images