Server, multi-scene multi-task content recommendation method, and storage medium

Abstract

The present disclosure relates to a server, a multi-scene multi-task content recommendation method and a storage medium. The server comprises a controller configured to: acquire a to-be-recommended data set from a media database according to identification information of a target display device; the to-be-recommended data set comprises a plurality of historical media play data of the target display device in different classified media recommendation scenes; input the to-be-recommended data set into a multi-task prediction model to acquire a plurality of prediction probabilities output by the multi-task prediction model, the plurality of prediction probabilities corresponding to each media play task related to each piece of media data in the to-be-recommended data set; the multi-task prediction model comprises a plurality of task networks, each task network being used for predicting a completion probability of one media play task; and acquire at least one piece of media data in at least one scene according to the plurality of prediction probabilities. The present method can improve the accuracy of content recommendation in multi-scene multi-task.

Description

Technical Field

[0001] This disclosure relates to the field of recommendation algorithm technology, and in particular to a server, a content recommendation method for multiple scenarios and multiple tasks, and a storage medium. Background Technology

[0002] In the era of the intelligent internet, due to the data explosion of content information, intelligent recommendation has become an important means for people to access internet content. With the popularization of smart TVs and the diversification of TV media content, TV recommendation has also entered the era of content recommendation. How to provide users with more accurate recommended content has become the core competitiveness of smart TVs.

[0003] In the field of television recommendation, the recommended scenarios encompass a wide range of content, including TV dramas, movies, variety shows, animation, children's content, and documentaries. These scenarios exhibit significant differences, and building a separate recommendation model for each scenario would increase the resource consumption of the recommendation system. If data from all scenarios is directly input into a single model for training, the significant variations in user behavior across different scenarios will lead to a decline in the overall performance of the model, resulting in lower prediction accuracy. Therefore, improving the accuracy of content recommendation across multiple scenarios and tasks is a pressing issue that needs to be addressed. Summary of the Invention

[0004] To address, or at least partially address, the aforementioned technical problems, this disclosure provides a server, a multi-scenario, multi-task content recommendation method, and a storage medium, which can improve the accuracy of content recommendation under multi-scenario, multi-task conditions.

[0005] In a first aspect, this disclosure provides a server, including: The controller is configured as follows: The set of data to be recommended is obtained from the media asset database based on the identification information of the target display device; the set of data to be recommended includes multiple historical media asset playback data of the target display device under different categories of media asset recommendation scenarios. The dataset to be recommended is input into a multi-task prediction model to obtain multiple prediction probabilities output by the multi-task prediction model. The multiple prediction probabilities correspond to each media asset playback task related to each media asset data in the dataset to be recommended. The multi-task prediction model includes multiple task networks, and each task network is used to predict the completion probability of a media asset playback task. Based on the multiple predicted probabilities, at least one media asset data in at least one scenario is obtained, and the at least one media asset data in at least one scenario is sent to the target display device so that the target display device displays the at least one media asset data in at least one scenario.

[0006] As an optional implementation of this disclosure, the controller is further configured to: The step of obtaining at least one piece of media asset data for at least one scenario based on the multiple predicted probabilities includes: The predicted probabilities of each media asset playback task related to each media asset in the data set to be recommended are weighted and summed to obtain the recommendation probability of each media asset. The recommendation probabilities of each media asset are sorted, and based on the sorting results of the recommendation probabilities of each media asset, at least one media asset in at least one scenario is obtained.

[0007] As an optional implementation of this disclosure, the controller is further configured to: The multi-task prediction model was trained using the following method: Obtain training samples, the training samples including: User data; the user data includes user attribute information and media asset playback behavior information in various scenarios; Media asset data; Contextual data corresponding to media asset playback behavior information in various scenarios; The probability of tags corresponding to each media asset playback task related to each media asset data; The initial multi-task prediction model is trained based on the training samples to obtain the multi-task prediction model.

[0008] As an optional implementation of this disclosure, the controller is further configured to: The initial multi-task prediction model includes: an input layer, an embedding layer, a feature layer, a model layer, and an output layer; The step of training the initial multi-task prediction model based on the training samples to obtain the multi-task prediction model includes: The user data, media asset data, and context data corresponding to the media asset playback behavior information in various scenarios are input through the input layer. The embedding layer performs data transformation on the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in various scenarios to obtain user features, media asset features, and context features corresponding to the media asset playback behavior information in various scenarios. The feature layer obtains the overall preference features of the user for the media asset data in multiple scenarios based on the user features, the media asset features, and the context features corresponding to the media asset playback behavior information in various scenarios. The model layer obtains the unique and shared features corresponding to each media asset playback task based on the media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios. The output layer obtains the predicted probability of each media asset playback task based on the unique and shared features corresponding to each media asset playback task.

[0009] As an optional implementation of this disclosure, the controller is specifically configured as follows: The step of performing data transformation on the user data, media asset data, and context data corresponding to the media asset playback behavior information in various scenarios through the embedding layer to obtain user features, media asset features, and context features corresponding to the media asset playback behavior information in various scenarios includes: The user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario are input into the embedding layer to obtain the feature vector representation of the user data, the feature vector representation of the media asset data, and the feature vector representation of the context data corresponding to the media asset playback behavior information in each scenario, respectively. The feature vector representations of the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario are obtained based on the self-attention mechanism, respectively corresponding to user features, media asset features, and context features.

[0010] As an optional implementation of this disclosure, the controller is specifically configured as follows: The step of obtaining the user's overall preference features for the media asset data in multiple scenarios through the feature layer based on the user features, the media asset features, and the context features corresponding to the media asset playback behavior information in various scenarios includes: Based on the user characteristics and the context features corresponding to the media asset playback behavior information in each scenario, calculate the user's preference probability for each scenario; Obtain the user's feature vectors in various scenarios; Based on the user's preference probability for each scenario and the user's feature vector in each scenario, the overall preference features of the user for the media asset data in multiple scenarios are obtained.

[0011] As an optional implementation of this disclosure, the controller is specifically configured as follows: The process involves obtaining unique and shared features for each media asset playback task through the model layer based on the media asset features, the contextual features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios. This includes: The media asset features and the user's overall preference features for the media asset data in multiple scenarios are input into at least one dedicated expert network to obtain the unique features of each media asset playback task. The media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios are input into at least one shared expert network to obtain the shared features of each media asset playback task.

[0012] Secondly, a multi-scenario, multi-task content recommendation method is provided, the method comprising: The set of data to be recommended is obtained from the media asset database based on the identification information of the target display device; the set of data to be recommended includes multiple historical media asset playback data of the target display device under different categories of media asset recommendation scenarios. The dataset to be recommended is input into a multi-task prediction model to obtain multiple prediction probabilities output by the multi-task prediction model. The multiple prediction probabilities correspond to each media asset playback task related to each media asset data in the dataset to be recommended. The multi-task prediction model includes multiple task networks, and each task network is used to predict the completion probability of a media asset playback task. Based on the multiple predicted probabilities, at least one media asset data in at least one scenario is obtained, and the at least one media asset data in at least one scenario is sent to the target display device so that the target display device displays the at least one media asset data in at least one scenario.

[0013] As an optional implementation of this disclosure, obtaining at least one piece of media asset data for at least one scenario based on the plurality of predicted probabilities includes: The predicted probabilities of each media asset playback task related to each media asset in the data set to be recommended are weighted and summed to obtain the recommendation probability of each media asset. The recommendation probabilities of each media asset are sorted, and based on the sorting results of the recommendation probabilities of each media asset, at least one media asset in at least one scenario is obtained.

[0014] As an optional implementation of this disclosure, the multi-task prediction model is trained using the following method: Obtain training samples, the training samples including: User data; the user data includes user attribute information and media asset playback behavior information in various scenarios; Media asset data; Contextual data corresponding to media asset playback behavior information in various scenarios; The probability of tags corresponding to each media asset playback task related to each media asset data; The initial multi-task prediction model is trained based on the training samples to obtain the multi-task prediction model.

[0015] As an optional implementation of this disclosure, the initial multi-task prediction model includes: an input layer, an embedding layer, a feature layer, a model layer, and an output layer; The step of training the initial multi-task prediction model based on the training samples to obtain the multi-task prediction model includes: The user data, media asset data, and context data corresponding to the media asset playback behavior information in various scenarios are input through the input layer. The embedding layer performs data transformation on the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in various scenarios to obtain user features, media asset features, and context features corresponding to the media asset playback behavior information in various scenarios. The feature layer obtains the overall preference features of the user for the media asset data in multiple scenarios based on the user features, the media asset features, and the context features corresponding to the media asset playback behavior information in various scenarios. The model layer obtains the unique and shared features corresponding to each media asset playback task based on the media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios. The output layer obtains the predicted probability of each media asset playback task based on the unique and shared features corresponding to each media asset playback task.

[0016] As an optional implementation of this disclosure, the step of performing data transformation on the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in various scenarios through the embedding layer to obtain user features, media asset features, and context features corresponding to the media asset playback behavior information in various scenarios includes: The user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario are input into the embedding layer to obtain the feature vector representation of the user data, the feature vector representation of the media asset data, and the feature vector representation of the context data corresponding to the media asset playback behavior information in each scenario, respectively. The feature vector representations of the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario are obtained based on the self-attention mechanism, respectively corresponding to user features, media asset features, and context features.

[0017] As an optional implementation of this disclosure, the step of obtaining the user's overall preference features for the media asset data in multiple scenarios through the feature layer based on the user features, the media asset features, and the context features corresponding to the media asset playback behavior information in various scenarios includes: Based on the user characteristics and the context features corresponding to the media asset playback behavior information in each scenario, calculate the user's preference probability for each scenario; Obtain the user's feature vectors in various scenarios; Based on the user's preference probability for each scenario and the user's feature vector in each scenario, the overall preference features of the user for the media asset data in multiple scenarios are obtained.

[0018] As an optional implementation of this disclosure, the step of obtaining unique and shared features corresponding to each media asset playback task through the model layer based on the media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios includes: The media asset features and the user's overall preference features for the media asset data in multiple scenarios are input into at least one dedicated expert network to obtain the unique features of each media asset playback task. The media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios are input into at least one shared expert network to obtain the shared features of each media asset playback task.

[0019] Thirdly, a computer-readable storage medium is provided, comprising: storing a computer program on the computer-readable storage medium, wherein when the computer program is executed by a processor, it implements the multi-scenario, multi-task content recommendation method as shown in the second aspect.

[0020] The technical solution provided in this disclosure has the following advantages compared with the prior art: A set of data to be recommended is obtained from a media asset database based on the identification information of the target display device. The set of data to be recommended includes multiple historical media asset playback data of the target display device under different media asset recommendation scenarios. The set of data to be recommended is input into a multi-task prediction model to obtain multiple prediction probabilities output by the multi-task prediction model. Each prediction probability corresponds to a media asset playback task related to each media asset data in the set of data to be recommended. The multi-task prediction model includes multiple task networks, each task network being used to predict the completion probability of a media asset playback task. Based on the multiple prediction probabilities, at least one media asset data in at least one scenario is obtained, and the at least one media asset data in the at least one scenario is sent to the target display device so that the target display device displays the at least one media asset data in the at least one scenario. Since the multi-task prediction model includes multiple task networks, and each task network is used to predict the completion probability of different media asset playback tasks, the multi-task prediction model can predict the completion probability of each media asset playback task related to each media asset data in different media asset recommendation scenarios. Based on multiple predicted probabilities, it can obtain at least one media asset data in at least one scenario, thereby avoiding the problem of the overall performance of the model being degraded due to inputting data from all scenarios into the same network model for training, and further improving the accuracy of content recommendation in multiple scenarios and multiple tasks. Attached Figure Description

[0021] 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.

[0022] 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.

[0023] Figure 1A This is a schematic diagram of a media asset recommendation interface for a smart TV in the prior art; Figure 1B This is a schematic diagram illustrating an application scenario of multi-scenario, multi-task content recommendation in an embodiment of this disclosure; Figure 2 This is a hardware configuration block diagram of a server according to one or more embodiments of the present disclosure; Figure 3 This is one of the flowcharts illustrating a multi-scenario, multi-task content recommendation method provided in this embodiment of the disclosure; Figure 4AThis is one of the schematic diagrams of a media asset recommendation interface for a smart TV provided in this embodiment of the disclosure; Figure 4B This is a second flowchart illustrating a multi-scenario, multi-task content recommendation method provided in this embodiment of the disclosure. Figure 4C A second schematic diagram of a media asset recommendation interface for a smart TV provided in an embodiment of this disclosure; Figure 4D A third schematic diagram of a media asset recommendation interface for a smart TV provided in this embodiment of the present disclosure; Figure 5A This is one of the flowcharts illustrating a multi-task prediction model training method provided in this embodiment of the disclosure; Figure 5B This is a schematic diagram of the structure of a multi-task prediction model provided in an embodiment of the present disclosure; Figure 5C This is a second flowchart illustrating a multi-task prediction model training method provided in an embodiment of this disclosure. Detailed Implementation

[0024] 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.

[0025] 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.

[0026] The terms "first," "second," "third," etc., used in this disclosure, in the specification, claims, and accompanying drawings are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0027] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0028] With the widespread use of smart terminals, people are spending more and more time consuming on them, and their demands for the services they provide are also increasing. As an important part of family life, smart TVs are a vital source of family entertainment and information. Currently, there is a wide variety of video content available, so providing efficient and accurate video recommendation services to families can greatly enhance people's life experience.

[0029] In the field of television recommendation, the recommended scenarios encompass a wide range of content, including TV dramas, movies, variety shows, animation, children's content, and documentaries. These scenarios exhibit significant differences, and building a separate recommendation model for each scenario would increase the resource consumption of the recommendation system. If data from all scenarios is directly input into a single model for training, the significant variations in user behavior across different scenarios will lead to a decline in the overall performance of the model, resulting in lower prediction accuracy. Therefore, improving the accuracy of content recommendation across multiple scenarios and tasks is a pressing issue that needs to be addressed.

[0030] To address this issue, existing recommendation algorithms primarily rely on historical viewing records or user-favorited media assets for similarity-based recommendations, neglecting the differences between various recommendation scenarios. A uniform model struggles to capture the complex correlations across multiple scenarios, impacting the accuracy of the recommendation results. For example, refer to... Figure 1A As shown, Figure 1A This is a schematic diagram of a media asset recommendation interface for a smart TV. Based on historical viewing records or user-favorited media asset data, it recommends movies 1, 2, 3, and 4 that the user might be interested in. To address the inconsistency in data distribution across different scenarios, transfer learning can be used to align the data distribution. However, this method is not suitable for situations where multiple scenarios coexist, as it can lead to a trade-off between different scenarios. To simultaneously represent the differences across multiple scenarios, some methods incorporate auxiliary networks to model the commonalities and differences across multiple media asset recommendation scenarios, improving the efficiency of the multi-scenario ranking model. However, such methods become more complex with the addition of scenarios, increasing the computational complexity of training and inference, making them unsuitable for unified multi-scenario recommendations on the product's homepage.

[0031] Based on this, in this embodiment, firstly, a set of data to be recommended is obtained from the media asset database according to the identification information of the target display device. The set of data to be recommended includes multiple historical media asset playback data of the target display device under different media asset recommendation scenarios. Then, the set of data to be recommended is input into a multi-task prediction model to obtain multiple prediction probabilities output by the multi-task prediction model. The multiple prediction probabilities correspond to each media asset playback task related to each media asset data in the set of data to be recommended. The multi-task prediction model includes multiple task networks, each task network is used to predict the completion probability of a media asset playback task. Finally, based on the multiple prediction probabilities, at least one media asset data in at least one scenario is obtained, and the at least one media asset data in at least one scenario is sent to the target display device so that the target display device displays the at least one media asset data in the at least one scenario. Since the multi-task prediction model includes multiple task networks, and each task network is used to predict the completion probability of different media asset playback tasks, the multi-task prediction model can predict the completion probability of each media asset playback task related to each media asset data in different media asset recommendation scenarios. Based on multiple predicted probabilities, it can obtain at least one media asset data in at least one scenario, thereby avoiding the problem of the overall performance of the model being degraded due to inputting data from all scenarios into the same network model for training, and further improving the accuracy of content recommendation in multiple scenarios and multiple tasks.

[0032] For example, such as Figure 1B As shown, Figure 1B This is a schematic diagram illustrating an application scenario of multi-scenario, multi-task content recommendation provided in an embodiment of this disclosure. Figure 1B In this process, server 100 retrieves a set of recommended data from a media asset database based on the identification information of the target display device 200 (e.g., the display device currently used by the user). From this set, server 100 retrieves at least one media asset data item for at least one scenario. The set of recommended data contains multiple historical media asset playback data items for the target display device under different media asset recommendation scenarios. Server 100 then sends this at least one media asset data item for at least one scenario to the target display device 200, enabling the target display device 200 to display this at least one media asset data item for at least one scenario. For example, the display device 200 in this scenario could be a smart device with smart display capabilities, such as a smart TV, smart tablet, or smartphone.

[0033] In some embodiments, the display device 200 also communicates with the server 100. The display device 200 may be allowed to communicate via a local area network (LAN), wireless local area network (WLAN), and other networks. The server 100 may provide various content and interactive features to the display device 200. The server 100 may be a cluster or multiple clusters, and may include one or more types of servers.

[0034] The multi-scenario, multi-task content recommendation method provided in this disclosure can be implemented based on a server, or a functional module or functional entity within the server.

[0035] For example, Figure 2 This is a hardware configuration block diagram of a server according to one or more embodiments of this disclosure. Figure 2 As shown, the server includes at least one of the following: a tuner / demodulator 210, a communicator 220, a controller 230, a display interface 240, a memory, and a power supply. The controller 230 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 interface 240 may be a VGA (Video Graphics Array). Generally, server graphics card requirements are not high, so the graphics card interface of a typical server is a VGA interface, generally used for installing the server operating system or for daily debugging. The tuner / demodulator 210 receives broadcast television signals via wired or wireless means, and demodulates audio and video signals, such as EPG audio and video data signals, from multiple wireless or wired broadcast television signals. The communicator 220 is a component used to communicate with external devices according to various communication protocol types. For example, the communicator may include at least one of the following: a Wi-Fi module, a Bluetooth module, a wired Ethernet module, other network communication protocol chips or near-field communication protocol chips, and an infrared receiver. The server can establish the transmission and reception of control signals and data signals with the display device 200 or the local control device through the communicator 220. The controller 230 and the tuner 210 can be located in different separate devices, that is, the tuner 210 can also be in an external device of the main device where the controller 230 is located, such as an external set-top box.

[0036] In some embodiments, the controller 230 controls the operation of the server and responds to user operations through various software control programs stored in memory. The controller 230 controls the overall operation of the server. Users can input commands through a graphical user interface (GUI) displayed on the display interface 240, and the user input interface receives user input commands through the graphical user interface (GUI).

[0037] To illustrate this solution in more detail, the following will use examples to illustrate it. Figure 3 To explain, it is understandable that, although Figure 3 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 3At least some steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps. The method should be designed to implement the multi-scenario, multi-task content recommendation method provided in the embodiments of this disclosure.

[0038] like Figure 3 As shown, the method specifically includes the following steps: S31. Obtain the set of data to be recommended from the media asset database based on the identification information of the target display device.

[0039] The data set to be recommended includes multiple historical media asset playback data of the target display device under different media asset recommendation scenarios.

[0040] Specifically, different media asset recommendation scenarios include, but are not limited to: TV dramas, movies, variety shows, animation, children's programs, and documentaries. Media asset data includes, but is not limited to: multiple TV drama program data, multiple movie program data, multiple variety show program data, multiple animation program data, multiple children's program data, and multiple documentary program data. The target display device is the display device currently used by the user, which can be a smart TV, smart tablet, laptop, etc. The identification information of the target display device serves as a unique identifier for the target display device; for example, it can be the target display device's ID (identification). In this embodiment, taking a smart TV as an example, the identification information of the target display device can carry information such as the display device's IP (Internet Protocol) address, model, and screen size.

[0041] For example, the server first obtains the set of data to be recommended through the controller. Assume the set of data to be recommended contains 6 media assets, refer to... Figure 4A As shown, they are: Media Asset 1, Media Asset 2, Media Asset 3, Media Asset 4, Media Asset 5, and Media Asset 6; among them, Media Asset 1 and Media Asset 2 are from TV drama scenes, Media Asset 3 is from movie scenes, Media Asset 4 and Media Asset 5 are from variety show scenes, and Media Asset 6 is from documentary scenes.

[0042] S32. Input the dataset to be recommended into the multi-task prediction model and obtain multiple prediction probabilities output by the multi-task prediction model.

[0043] The multiple predicted probabilities correspond to various media asset playback tasks associated with each media asset in the dataset to be recommended. The multi-task prediction model includes multiple task networks, each used to predict the completion probability of a media asset playback task. Multiple tasks may include: a user clicking on a media asset, and whether the user effectively watches the media asset after clicking on it.

[0044] Specifically, the multi-task prediction model performs probability predictions for each media asset playback task related to each media asset data in the recommended dataset.

[0045] For example, suppose there are 6 media assets in the dataset to be recommended: Media Asset 1, Media Asset 2, Media Asset 3, Media Asset 4, Media Asset 5, and Media Asset 6. The multi-task prediction model predicts the following probabilities for each media asset playback task related to each media asset in the dataset: For example, the probability of a user clicking on Media Asset 1 is 0.8, and the probability of the user effectively watching it after clicking on Media Asset 1 is 0.7; the probability of a user clicking on Media Asset 2 is 0.6, and the probability of the user effectively watching it after clicking on Media Asset 2 is 0.8; the probability of a user clicking on Media Asset 3 is 0.5, and the probability of the user effectively watching it after clicking on Media Asset 3 is 0.5; the probability of a user clicking on Media Asset 4 is 0.5, and the probability of the user effectively watching it after clicking on Media Asset 4 is 0.6; the probability of a user clicking on Media Asset 5 is 0.9, and the probability of the user effectively watching it after clicking on Media Asset 5 is 0.7; the probability of a user clicking on Media Asset 6 is 0.3, and the probability of the user effectively watching it after clicking on Media Asset 6 is 0.5.

[0046] S33. Based on the multiple predicted probabilities, obtain at least one media asset data in at least one scenario, and send the at least one media asset data in at least one scenario to the target display device so that the target display device displays the at least one media asset data in at least one scenario.

[0047] In some embodiments, Figure 3 Based on, refer to Figure 4B As shown, step S33 (obtaining at least one media asset data point in at least one scenario based on the multiple predicted probabilities) can be implemented in the following way: S331. The predicted probabilities of each media asset playback task related to each media asset data in the data set to be recommended are weighted and summed to obtain the recommendation probability of each media asset data.

[0048] Specifically, the weight coefficients of click tasks and effective viewing tasks can be preset. Then, based on the weight coefficients of click tasks and effective viewing tasks, the predicted probabilities of each media asset playback task related to each media asset in the data set to be recommended are weighted and summed to obtain the recommendation probability of each media asset.

[0049] For example, in this embodiment of the disclosure, the weight coefficient of the click task can be set to 1, and the weight coefficient of the effective viewing task can also be set to 1. Assuming the probability of a user clicking on media asset 1 is 0.8, and the probability of the user effectively watching it after clicking on media asset 1 is 0.7; the probability of a user clicking on media asset 2 is 0.6, and the probability of the user effectively watching it after clicking on media asset 2 is 0.8; the probability of a user clicking on media asset 3 is 0.5, and the probability of the user effectively watching it after clicking on media asset 3 is 0.5; the probability of a user clicking on media asset 4 is 0.5, and the probability of the user effectively watching it after clicking on media asset 4 is 0.6; the probability of a user clicking on media asset 5 is 0.9, and the probability of the user effectively watching it after clicking on media asset 5 is 0.7; the probability of a user clicking on media asset 6 is 0.3, and the probability of the user effectively watching it after clicking on media asset 6 is 0.5; then the calculated recommendation probability for media asset 1 is 1.5, the recommendation probability for media asset 2 is 1.4, the recommendation probability for media asset 3 is 1, the recommendation probability for media asset 4 is 1.1, the recommendation probability for media asset 5 is 1.6, and the recommendation probability for media asset 6 is 0.8.

[0050] S332. Sort the recommendation probabilities of each media asset data, and based on the sorting results of the recommendation probabilities of each media asset data, obtain at least one media asset data in at least one scenario.

[0051] Specifically, based on the ranking results of the recommendation probabilities of each media asset data, at least one media asset data in at least one scenario is obtained.

[0052] For example, refer to Figure 4C As shown, sorting the media assets according to their recommendation probabilities, the recommendation order is: Media Asset 5 > Media Asset 1 > Media Asset 2 > Media Asset 4 > Media Asset 3 > Media Asset 6. Assuming we need to recommend 3 media assets to a user, refer to... Figure 4D As shown, they are, in order: Media Asset 5, Media Asset 1, Media Asset 2.

[0053] In this embodiment of the disclosure, a set of data to be recommended is obtained from a media asset database based on the identification information of the target display device. The set of data to be recommended includes multiple historical media asset playback data of the target display device under different media asset recommendation scenarios. The set of data to be recommended is input into a multi-task prediction model to obtain multiple prediction probabilities output by the multi-task prediction model. The multiple prediction probabilities correspond to each media asset playback task related to each media asset data in the set of data to be recommended. The multi-task prediction model includes multiple task networks, each task network is used to predict the completion probability of a media asset playback task. Based on the multiple prediction probabilities, at least one media asset data in at least one scenario is obtained, and the at least one media asset data in at least one scenario is sent to the target display device so that the target display device displays the at least one media asset data in the at least one scenario. Since the multi-task prediction model includes multiple task networks, and each task network is used to predict the completion probability of different media asset playback tasks, the multi-task prediction model can predict the completion probability of each media asset playback task related to each media asset data in different media asset recommendation scenarios. Based on multiple predicted probabilities, it can obtain at least one media asset data in at least one scenario, thereby avoiding the problem of the overall performance of the model being degraded due to inputting data from all scenarios into the same network model for training, and further improving the accuracy of content recommendation in multiple scenarios and multiple tasks.

[0054] Figure 5A This is a flowchart illustrating a multi-task prediction model training method provided in this embodiment. Optionally, this embodiment mainly describes the training process of the multi-task prediction model.

[0055] In some embodiments, refer to Figure 5A As shown, the multi-task prediction model is trained using the following method: S51. Obtain training samples.

[0056] The training samples include: (1) user data, which includes user attribute information and media asset playback behavior information in various scenarios. (2) media asset data. (3) context data corresponding to the user's media asset playback behavior information in various scenarios. (4) label probabilities corresponding to each media asset playback task related to each piece of media asset data.

[0057] Specifically, user data includes user attribute information and media asset playback behavior information in various media asset recommendation scenarios. For example, user attribute information includes home address, TV model, TV screen size, membership status, and user viewing time preferences.

[0058] For example, viewing time preferences are quite complex. In the context of media asset recommendations within a family entertainment environment, the entire family is considered as the user group, resulting in a large number of users and clear user segmentation among family members. For instance, a user might prefer watching variety shows during weekday lunch breaks, where shorter shows are suitable. In the evening, children returning home from school might watch anime or children's programs. Various media asset recommendation scenarios include, but are not limited to: movies, TV series, variety shows, animation, documentaries, and others. User media asset playback behavior information in each scenario includes information such as actors, directors, categories, and release dates associated with the media assets the user prefers to click or watch. Media asset data includes information such as directors, actors, categories, and release dates associated with each piece of media asset data. The contextual data corresponding to user media asset playback behavior information in each scenario includes the user's recent viewing history and historical viewing history for the same period.

[0059] S52. The initial multi-task prediction model is trained based on the training samples to obtain the multi-task prediction model.

[0060] Specifically, after obtaining the training samples, based on the initial multi-task prediction model and the training samples, the predicted probabilities of each media asset playback task related to each media asset data are generated. Based on the preset loss function, the initial multi-task prediction model is trained according to the training samples to obtain the multi-task prediction model. Since the media asset playback task is whether a user clicks on a certain media asset, and whether the user effectively watches the media asset after clicking on it, that is, the media asset playback task is a binary classification problem, the logarithmic loss function is used as the loss function.

[0061] During model training, the optimizer uses Adam (adaptive moment estimation), and the initial learning rate is 0.001, which decays at a rate of 0.9 every 10 complete training cycles.

[0062] In some embodiments, refer to Figure 5B As shown, the initial multi-task prediction model includes: an input layer, an embedding layer, a feature layer, a model layer, and an output layer.

[0063] Specifically, in the initial multi-task prediction model, the following mathematical symbols are defined: U represents the set of all users; V represents the set of all media assets; Represents a specific user; Represents a certain media asset; On behalf of users Features; Representative media assets Features; On behalf of users Media resources Contextual features. This represents a set of media asset playback tasks, such as... ,Right now For users Did you click on the media assets? , For users Click Media Assets Whether a viewing session constitutes a valid viewing session depends on the specific scenario. For example, for a TV series, watching for more than 10 minutes constitutes a valid viewing session; for a movie, watching for more than 5 minutes constitutes a valid viewing session. Additionally, valid viewing times can be set for variety shows, documentaries, anime, etc. It should be noted that the valid viewing time for different scenarios can be set according to the actual situation; no specific restrictions are imposed here.

[0064] In some embodiments, refer to Figure 5C As shown, S52 (training the initial multi-task prediction model based on the training samples to obtain the multi-task prediction model) can be implemented in the following way: S521. Input the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario through the input layer.

[0065] Specifically, the input layer is used to input training samples, which can be divided into three types of data: user data. Media data Context data corresponding to media asset playback behavior information in various scenarios Furthermore, user data is further segmented into more granular categories across various scenarios. This means that user data includes user attribute information and media asset playback behavior information within each scenario. ; in, Indicates user The attribute information, that is, Indicates the first One scenario, The value can be 4, 6, 7, or other reasonable values. Indicates user In the scene This refers to information about media playback behavior. Scenarios can include, but are not limited to: movies, TV series, variety shows, animation, documentaries, and others.

[0066] For example, Indicates user The attributes include home address, TV model, TV screen size, membership status, and preferred viewing time. Indicates user Media asset playback behavior information in Scenario 1. For example, when Scenario 1 is a movie, user 1's media asset playback behavior information in this scenario includes: the actors, directors, categories, and release dates associated with the media asset that was clicked to play.

[0067] S522. The user data, media asset data, and context data corresponding to the media asset playback behavior information in various scenarios are transformed through the embedding layer to obtain user features, media asset features, and context features corresponding to the media asset playback behavior information in various scenarios.

[0068] In some embodiments, the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario are input into the embedding layer to obtain the feature vector representation of the user data, the feature vector representation of the media asset data, and the feature vector representation of the context data corresponding to the media asset playback behavior information in each scenario, respectively.

[0069] The feature vector representations of the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario are obtained based on the self-attention mechanism, respectively corresponding to user features, media asset features, and context features.

[0070] Specifically, the embedding layer primarily transforms sparse data into dense data. Essentially, the embedding layer is a very large table whose function is to find the vector representation at the corresponding position based on the input. ; Since the input to the embedding layer consists of user data, media asset data, and contextual data corresponding to the user's media asset playback behavior information in various scenarios, the feature vector representation of the user data is determined according to the following formula:

[0071] ; The feature vector representation of media asset data is determined according to the following formula: ; The feature vector representation of the context data corresponding to the user's media asset playback behavior information in various scenarios is determined according to the following formula: ; S523. Based on the user features, the media asset features, and the context features corresponding to the media asset playback behavior information in each scenario, the feature layer obtains the overall preference features of the user for the media asset data in multiple scenarios.

[0072] In some embodiments, the above-mentioned acquisition of users' overall preference characteristics for the media asset data in multiple scenarios can be achieved in the following ways: (1) Calculate the user's preference probability for each scenario based on the user characteristics and the context features corresponding to the media asset playback behavior information in each scenario.

[0073] Specifically, the probability of user preference for each scenario is calculated using the following formula: ; in, The function represents the probability of a user's preference for each scenario. This can be expressed using a deep neural network (DNN), the function For activation function, Indicates user Attribute information, Indicates user Media resources Contextual features.

[0074] For example, activation functions can use The ReLU activation function is not used here to prevent some feature vectors from having excessively large magnitudes, which could lead to training instability. It also effectively avoids the problem of uneven distribution of latent feature vectors. Furthermore, softmax is not used because its activation function sums to 1, making it unsuitable for feature preference extraction in various scenarios. The number of elements can be determined based on the actual number of elements in the scenario. For example, it can be 4, 5, 6, or other reasonable values. There are no specific restrictions here.

[0075] (2) Obtain the feature vectors of users in various scenarios.

[0076] Specifically, feature extraction is performed on user characteristics in various scenarios to obtain all user characteristics in different scenarios, such as user preferences for actors, directors, and categories in different scenarios.

[0077] For example, the user's feature vector in various scenarios can be obtained using the following formula: ; in, Indicates the user in the scenario The feature representation below is a column vector, a function It can be represented by a deep neural network (DNN).

[0078] In this disclosure embodiment, four scenarios are taken as examples. In practical applications, More reasonable values ​​can be taken, and no specific restrictions are imposed here.

[0079] (3) Based on the user's preference probability for each scenario and the user's feature vector in each scenario, obtain the user's overall preference features for the media asset data in multiple scenarios.

[0080] Specifically, the overall user preference characteristics for media asset data across multiple scenarios are obtained using the following formula: ; in, Indicates user Overall preference characteristics for media asset data in multiple scenarios. This means concatenating the feature vectors from different scenarios.

[0081] Specifically, the feature representations from various scenarios are concatenated and then multiplied by the user's preference probability in each scenario to obtain the user's overall preference features across multiple scenarios.

[0082] S524. Based on the media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios, the model layer obtains the unique features and shared features corresponding to each media asset playback task.

[0083] Specifically, at the model layer, feature extraction is performed through multiple expert networks. The same features in multiple scenarios are extracted through a shared expert network, while the different features in multiple scenarios are extracted through a dedicated expert network.

[0084] In some embodiments, the acquisition of the unique and shared features corresponding to each media asset playback task can be achieved in the following ways: (1) Input the media asset features and the overall preference features of the user for the media asset data in multiple scenarios into at least one exclusive expert network to obtain the exclusive features of each media asset playback task.

[0085] For example, in this embodiment of the disclosure, two tasks are used. For example, This indicates that the user clicked on the task. This indicates a valid movie-watching task.

[0086] Specifically, the unique characteristics of each media asset playback task are obtained according to the following formula: ; in, Indicates task The exclusive expert network output feature map.

[0087] (2) Input the media asset features, the context features corresponding to the media asset playback behavior information in each scenario, and the overall preference features of the user for the media asset data in multiple scenarios into at least one shared expert network to obtain the shared features of each media asset playback task.

[0088] For example, in this embodiment of the disclosure, four shared expert networks are used as an example. Indicates the first A shared network of experts.

[0089] Specifically, the shared features of each media asset playback task are obtained according to the following formula: ; in, Indicates the first The output feature map of a shared expert network.

[0090] S525. In some embodiments, the output layer obtains the predicted probability corresponding to each media asset playback task based on the unique and shared features corresponding to each media asset playback task.

[0091] Specifically, the output layer primarily predicts the playback tasks for each media asset and outputs the probability value for that task. The task tower is also a neural network; the inputs for different tasks include shared experts and their unique task expert outputs. A gating network controls the network weights of each expert input to the task tower, i.e.: ; in, , This indicates that the final task output activation function is softmax. This represents the weight of each expert.

[0092] ; in, As a mission tower The input is processed through a neural network structure to obtain the final output, namely: ; in, Indicates media asset playback task The probability output.

[0093] The multi-task prediction model structure provided in this disclosure effectively solves the problem of increased model complexity. At the feature layer, it proposes a user-adaptive feature extraction architecture for multiple scenarios, extracting differences between users in different scenarios. At the model layer, it uses different category feature inputs for different expert network types, effectively capturing the differences between different tasks. Furthermore, as a general multi-scenario, multi-task architecture, this structure can be extended to different recommendation domains, such as product recommendation.

[0094] In this embodiment of the disclosure, a set of data to be recommended is obtained from a media asset database based on the identification information of the target display device. The set of data to be recommended includes multiple historical media asset playback data of the target display device under different media asset recommendation scenarios. The set of data to be recommended is input into a multi-task prediction model to obtain multiple prediction probabilities output by the multi-task prediction model. The multiple prediction probabilities correspond to each media asset playback task related to each media asset data in the set of data to be recommended. The multi-task prediction model includes multiple task networks, each task network is used to predict the completion probability of a media asset playback task. Based on the multiple prediction probabilities, at least one media asset data in at least one scenario is obtained, and the at least one media asset data in at least one scenario is sent to the target display device so that the target display device displays the at least one media asset data in the at least one scenario. Since the multi-task prediction model includes multiple task networks, and each task network is used to predict the completion probability of different media asset playback tasks, the multi-task prediction model can predict the completion probability of each media asset playback task related to each media asset data in different media asset recommendation scenarios. Based on multiple predicted probabilities, it can obtain at least one media asset data in at least one scenario, thereby avoiding the problem of the overall performance of the model being degraded due to inputting data from all scenarios into the same network model for training, and further improving the accuracy of content recommendation in multiple scenarios and multiple tasks.

[0095] This disclosure also provides a storage medium containing computer-executable instructions. When executed by a computer processor, the computer-executable instructions implement the various processes of the methods provided in any of the above embodiments and achieve the same technical effects. To avoid repetition, further details are omitted here.

[0096] The computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0097] For ease of explanation, the above description has been provided in conjunction with specific embodiments. However, the discussion in some embodiments above is not intended to be exhaustive or to limit the embodiments to the specific forms disclosed above. Various modifications and variations can be obtained based on the above teachings. The selection and description of the above embodiments are for the purpose of better explaining the principles and practical applications, thereby enabling those skilled in the art to better utilize the embodiments and various different variations of the embodiments suitable for specific application considerations.

Claims

1. A server, characterized in that, include: The controller is configured as follows: The set of data to be recommended is obtained from the media asset database based on the identification information of the target display device; the set of data to be recommended includes multiple historical media asset playback data of the target display device under different categories of media asset recommendation scenarios. The dataset to be recommended is input into a multi-task prediction model to obtain multiple prediction probabilities output by the multi-task prediction model. The multiple prediction probabilities correspond to each media asset playback task related to each media asset data in the dataset to be recommended. The multi-task prediction model includes multiple task networks, and each task network is used to predict the completion probability of a media asset playback task. Based on the multiple predicted probabilities, at least one media asset data in at least one scenario is obtained, and the at least one media asset data in at least one scenario is sent to the target display device so that the target display device displays the at least one media asset data in at least one scenario; The multi-task prediction model is trained from an initial multi-task prediction model, which includes an input layer, an embedding layer, a feature layer, a model layer, and an output layer. The training process is as follows: The input layer inputs user data, media asset data, and context data corresponding to media asset playback behavior information in various scenarios. The embedding layer performs data transformation on the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in various scenarios to obtain user features, media asset features, and context features corresponding to the media asset playback behavior information in various scenarios. The feature layer obtains the overall preference features of users for the media asset data in multiple scenarios based on user features, media asset features, and context features corresponding to media asset playback behavior information in various scenarios. The model layer obtains the unique and shared features corresponding to each media asset playback task based on the media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios. The output layer obtains the predicted probability of each media asset playback task based on the unique and shared features of each media asset playback task. The inputs for different tasks include shared experts and their unique task expert outputs. A gating network controls the network weights of each expert input to the task stack. ； in, , Indicates the first The output feature map of a shared expert network {1, 2, 3, 4} This indicates that the final task output activation function is softmax. This represents the weight of each expert; ； in, Indicates task Exclusive expert network output feature map, As a mission tower The input is processed through a neural network structure to obtain the final output, namely: ； in, Indicates media asset playback task The probability output.

2. The server according to claim 1, characterized in that, The controller is also configured to: The step of obtaining at least one piece of media asset data for at least one scenario based on the multiple predicted probabilities includes: The predicted probabilities of each media asset playback task related to each media asset in the data set to be recommended are weighted and summed to obtain the recommendation probability of each media asset. The recommendation probabilities of each media asset are sorted, and based on the sorting results of the recommendation probabilities of each media asset, at least one media asset in at least one scenario is obtained.

3. The server according to claim 1, characterized in that, The controller is also configured to: The multi-task prediction model was trained using the following method: Obtain training samples, the training samples including: User data; the user data includes user attribute information and media asset playback behavior information in various scenarios; Media asset data; Contextual data corresponding to media asset playback behavior information in various scenarios; The probability of tags corresponding to each media asset playback task related to each media asset data; Based on a preset loss function, the initial multi-task prediction model is trained using the training samples to obtain the multi-task prediction model.

4. The server according to claim 1, characterized in that, The controller is specifically configured as follows: The step of performing data transformation on the user data, media asset data, and context data corresponding to the media asset playback behavior information in various scenarios through the embedding layer to obtain user features, media asset features, and context features corresponding to the media asset playback behavior information in various scenarios includes: The user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario are input into the embedding layer to obtain the feature vector representation of the user data, the feature vector representation of the media asset data, and the feature vector representation of the context data corresponding to the media asset playback behavior information in each scenario, respectively. The embedding layer determines the vector representation according to the following formula: The feature vector representation of user data is determined using the following formula: ； In the above formula, = The Indicates user The attribute information, that is, Indicates the first One scenario, Indicates user In the scene Information on media asset playback behavior; The feature vector representation of media asset data is determined according to the following formula: ； In the above formula, This represents media asset data; The feature vector representation of the context data corresponding to the user's media asset playback behavior information in various scenarios is determined according to the following formula: ； In the above formula, This represents the context data corresponding to media asset playback behavior information in various scenarios; The feature vector representations of the user data, the media asset data, and the context data corresponding to the media asset playback behavior information in each scenario are obtained based on the self-attention mechanism, respectively corresponding to user features, media asset features, and context features.

5. The server according to claim 4, characterized in that, The controller is specifically configured as follows: The step of obtaining the user's overall preference features for the media asset data in multiple scenarios through the feature layer based on the user features, the media asset features, and the context features corresponding to the media asset playback behavior information in various scenarios includes: Based on the user characteristics and the context features corresponding to the media asset playback behavior information in each scenario, calculate the user's preference probability for each scenario; The probability of a user's preference for each scenario is calculated as follows: ； in, The function represents the probability of a user's preference for each scenario. Expressed using a deep neural network (DNN), the function For activation function, Indicates user Attribute information, Indicates user Media resources Contextual features; Obtain the user's feature vectors in various scenarios; The methods for obtaining user feature vectors in various scenarios are as follows: ； in, Indicates the user in the scenario The feature representation below is a column vector, a function Represented using a deep neural network (DNN); Based on the user's preference probability for each scenario and the user's feature vector in each scenario, the overall preference features of the user for the media asset data in multiple scenarios are obtained. The methods for obtaining users' overall preference characteristics for media asset data across multiple scenarios are as follows: ； in, Indicates user Overall preference characteristics for media asset data in multiple scenarios This means concatenating the feature vectors from different scenarios.

6. The server according to claim 5, characterized in that, The controller is specifically configured as follows: The process involves obtaining unique and shared features for each media asset playback task through the model layer based on the media asset features, the contextual features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios. This includes: The media asset features and the user's overall preference features for the media asset data in multiple scenarios are input into at least one dedicated expert network to obtain the unique features of each media asset playback task. The methods for obtaining the unique characteristics of each media asset playback task are as follows: ； in, Indicates task Exclusive expert network output feature map; The media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios are input into at least one shared expert network to obtain the shared features of each media asset playback task. The methods for obtaining the shared characteristics of each media asset playback task are as follows: ； in, Indicates the first The output feature map of a shared expert network.

7. A content recommendation method for multiple scenarios and multiple tasks, characterized in that, The method includes: The set of data to be recommended is obtained from the media asset database based on the identification information of the target display device; the set of data to be recommended includes multiple historical media asset playback data of the target display device under different categories of media asset recommendation scenarios. The dataset to be recommended is input into a multi-task prediction model to obtain multiple prediction probabilities output by the multi-task prediction model. The multiple prediction probabilities correspond to each media asset playback task related to each media asset data in the dataset to be recommended. The multi-task prediction model includes multiple task networks, and each task network is used to predict the completion probability of a media asset playback task. Based on the multiple predicted probabilities, at least one media asset data in at least one scenario is obtained, and the at least one media asset data in at least one scenario is sent to the target display device so that the target display device displays the at least one media asset data in at least one scenario; The multi-task prediction model is obtained by training an initial multi-task prediction model based on training samples. The initial multi-task prediction model includes an input layer, an embedding layer, a feature layer, a model layer, and an output layer. The training process is as follows: The input layer inputs user data, media asset data, and context data corresponding to media asset playback behavior information in various scenarios. The embedding layer performs data transformation on user data, media asset data, and context data corresponding to media asset playback behavior information in various scenarios to obtain user features, media asset features, and context features corresponding to media asset playback behavior information in various scenarios. The model layer obtains the unique and shared features corresponding to each media asset playback task based on the media asset features, the context features corresponding to the media asset playback behavior information in various scenarios, and the overall preference features of the user for the media asset data in multiple scenarios. The output layer obtains the predicted probability of each media asset playback task based on the unique and shared features of each media asset playback task. The inputs for different tasks include shared experts and their unique task expert outputs. A gating network controls the network weights of each expert input to the task stack. ； in, , Indicates the first The output feature map of a shared expert network {1, 2, 3, 4} This indicates that the final task output activation function is softmax. This represents the weight of each expert; ； in, Indicates task Exclusive expert network output feature map, As a mission tower The input is processed through a neural network structure to obtain the final output, namely: ； in, Indicates media asset playback task The probability output.

8. The method according to claim 7, characterized in that, The step of obtaining at least one piece of media asset data for at least one scenario based on the multiple predicted probabilities includes: The predicted probabilities of each media asset playback task related to each media asset in the data set to be recommended are weighted and summed to obtain the recommendation probability of each media asset. The recommendation probabilities of each media asset are sorted, and based on the sorting results of the recommendation probabilities of each media asset, at least one media asset in at least one scenario is obtained.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-scenario, multi-task content recommendation method as described in any one of claims 7-8.

Images