Server and media recommendation method

Abstract

The application provides a server and a media asset recommendation method. The server is configured to: obtain a to-be-recommended data set corresponding to a target user, the to-be-recommended data set including media asset data of a plurality of target media assets; input the to-be-recommended data set into a multi-task recommendation model to obtain a prediction probability of a second-stage task of a target media asset output by the multi-task recommendation model, wherein the multi-task recommendation model is a network model obtained by performing transfer learning of a pre-trained model on a second-stage task, the pre-trained model is a network model obtained by training on a first-stage task, the first-stage task includes a task of predicting whether a user will watch and a task of predicting whether a user will pay, and the second-stage task includes a task of predicting whether a user will pay; and obtain a recommendation probability of the target media asset according to the prediction probability, and then determine a to-be-recommended media asset. The application improves the effect of media asset recommendation taking into account the media asset preferences of users and the member conversion rate.

Description

[0001] This application claims priority to the Chinese Patent Application No. 202311378458.6, filed on October 23, 2023, and titled "Server and Media Recommendation Method", the content of which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[0002] The present application relates to the technical field of media recommendation, and in particular to a server and a media recommendation method. BACKGROUND

[0003] To facilitate users to quickly find the media they want to watch, intelligent televisions usually display multiple media on a video recommendation page, such as a home page, for users to choose. The media displayed on the video recommendation page can include free-to-watch media and paid-to-watch media, such as media that members can only watch in full. Displaying paid-to-watch media is an important means to improve member conversion rates and thus ensure the sustainable development of the video platform. To balance the media preferences of users and member conversion rates, in related technologies, a video platform pre-trains a media recommendation model according to various historical behaviors of users, such as media on-demand operations and payment operations, and recommends media displayed on the video recommendation page to users based on the media recommendation model. However, in the above multi-task training process, different historical behaviors of users can cause repulsion effects between multi-tasks, resulting in poor performance of the media recommendation model and difficulty in meeting media recommendation requirements. SUMMARY

[0004] To improve the media recommendation experience, the present application provides a server and a media recommendation method.

[0005] In a first aspect, the present application provides a server, which is configured to:

[0006] obtain a to-be-recommended data set corresponding to a target user, the to-be-recommended data set including media data of multiple target media;

[0007] input the to-be-recommended data set into a multi-task recommendation model to obtain a prediction probability of a second-stage task of the target media output by the multi-task recommendation model, wherein the multi-task recommendation model is a network model obtained by performing transfer learning of a pre-trained model on the second-stage task, the pre-trained model is a network model obtained by training on a first-stage task, the first-stage task includes a task of predicting whether a user watches and a task of predicting whether a user pays, and the second-stage task includes the task of predicting whether a user pays;

[0008] obtain a recommendation probability of the target media according to the prediction probability;

[0009] According to the recommendation probability, a to-be-recommended media resource corresponding to the target user is determined from the target media resource.

[0010] In some embodiments, the pre-training model comprises an input layer, an embedding layer, a feature layer, an expert model layer and an output layer connected in sequence, wherein the embedding layer is configured to calculate a sample feature vector, and the feature layer is configured to calculate a scene preference feature vector representing media resource preference features of the user in each recommendation scene based on the sample feature vector; the transfer learning of the pre-training model for the second-stage task comprises:

[0011] constructing a target domain training sample, wherein the target domain training sample comprises media resource data of a second sample media resource, and the media resource data of the second sample media resource comprises a label probability corresponding to the second-stage task;

[0012] setting network parameters of the embedding layer and the feature layer as fixed parameters;

[0013] inputting the target domain training sample into the pre-training model with the fixed parameters to obtain a recommendation probability of the target domain training sample.

[0014] In some embodiments, the source domain training sample corresponding to the pre-training model comprises media resource data of a first sample media resource, and the media resource data of the first sample media resource comprises a label probability corresponding to the first-stage task, and the constructing of the target domain training sample comprises:

[0015] In the source domain training sample, the first sample media resource with a label probability corresponding to a task of predicting whether a user will pay is deleted, and the second sample media resource comprises the first sample media resource retained in the source domain training sample;

[0016] In the second sample media resource, the label probability corresponding to the first-stage task is deleted, and the label probability of the second sample media resource is set as the label probability corresponding to the second-stage task.

[0017] In some embodiments, the constructing of the target domain training sample further comprises:

[0018] obtaining a new sample media resource with a time attribute different from the first sample media resource, and the second sample media resource further comprises the new sample media resource.

[0019] In some embodiments, the expert model layer comprises a plurality of expert networks configured to output feature maps based on the sample feature vector and the scene preference feature vector, and the transfer learning of the pre-training model for the second-stage task further comprises:

[0020] Before inputting the target domain training sample into the pre-training model with the fixed parameters, weights of the expert network are set as a function of a loss function variation of the pre-training model, and network parameters of one of the expert networks are set to zero.

[0021] In some embodiments, the second-stage task transfer learning on the pre-training model further includes:

[0022] Before inputting the target domain training sample into the pre-training model with the fixed parameters, a feature map regularization loss is added to a loss function of the pre-training model.

[0023] In some embodiments, the expert model layer includes a gating network corresponding to the expert network, and the second-stage task transfer learning on the pre-training model further includes:

[0024] Before inputting the target domain training sample into the pre-training model with the fixed parameters, a gating network regularization loss is added to a loss function of the pre-training model.

[0025] In some embodiments, the task of predicting whether the user watches includes a task of predicting whether a on-demand operation is performed and a task of predicting whether a viewing duration exceeds a viewing duration threshold, the task of predicting whether the user pays in the first-stage task includes a task of predicting whether a payment page is entered, and the task of predicting whether the user pays in the second-stage task includes a task of predicting whether the payment page is entered, a task of predicting whether a display time of the payment page exceeds a payment duration threshold, and a task of predicting whether a payment operation is completed.

[0026] In some embodiments, the to-be-recommended data set further includes user data of the target user and context data of the target media asset, the target media asset includes a media asset determined based on a media asset recommendation of a user portrait of the target user, and the context data includes media asset data of a previous media asset before the target user watches the target media asset and media asset data of a next media asset after the target user watches the target media asset.

[0027] In a second aspect, the present application provides a media asset recommendation method, which includes:

[0028] Obtaining a to-be-recommended data set corresponding to a target user, the to-be-recommended data set including media asset data of a plurality of target media assets;

[0029] inputting the to-be-recommended data set into a multi-task recommendation model, to obtain a prediction probability of the target media asset corresponding to a second stage task output by the multi-task recommendation model, wherein the multi-task recommendation model is a network model obtained by performing transfer learning of the second stage task on a pre-trained model, the pre-trained model is a network model obtained by training through a first stage task, the first stage task includes a task of predicting whether a user performs viewing and a task of predicting whether the user performs payment, and the second stage task includes the task of predicting whether the user performs payment;

[0030] obtaining a recommendation probability of the target media asset according to the prediction probability;

[0031] determining a to-be-recommended media asset corresponding to the target user from the target media asset according to the recommendation probability.

[0032] The server and the media asset recommendation method provided in the application have the following beneficial effects:

[0033] The application performs media asset recommendation through a multi-task recommendation model. The multi-task model is a network model obtained by performing transfer learning of a second stage task on a pre-trained model obtained by training through a first stage task. The first stage task includes a task of predicting whether a user performs viewing and a task of predicting whether the user performs payment, so that the media asset recommended according to the pre-trained model has a relatively high probability of being viewed by the user and a relatively high probability of being paid for. The second stage task includes the task of predicting whether the user performs payment, so that the media asset recommended by the multi-task recommendation model obtained through the transfer learning has a relatively high probability of being viewed by the user and a further improved probability of being paid for, thereby improving the effect of the media asset recommendation taking into account the media asset preference of the user and the member conversion rate. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the application or the implementation manners in the related art, the following will briefly introduce the drawings needed to be used in the embodiment or related art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0035] Figure 1 Fig. 1 exemplarily shows a schematic diagram of an operation scenario between a display device and a control device according to some embodiments;

[0036] Figure 2 Fig. 2 exemplarily shows a hardware configuration block diagram of the control device 100 according to some embodiments;

[0037] Figure 3 Fig. 3 exemplarily shows a hardware configuration block diagram of the display device 200 according to some embodiments;

[0038] Figure 4 Fig. 1 illustrates a schematic diagram of a display device 200 according to some embodiments;

[0039] Figure 5 Fig. 2 illustrates a schematic diagram of a homepage of a video on demand program according to some embodiments;

[0040] Figure 6 Fig. 3 illustrates a schematic diagram of a content recommendation method according to some embodiments;

[0041] Figure 7 Fig. 4 illustrates a schematic diagram of a training method of a multi-task recommendation model according to some embodiments;

[0042] Figure 8 Fig. 5 illustrates a schematic diagram of a pre-training model according to some embodiments;

[0043] Figure 9 Fig. 6 illustrates a schematic diagram of a transfer learning method according to some embodiments;

[0044] Figure 10 Fig. 7 illustrates a schematic diagram of a model change of a first training stage and a second training stage according to some embodiments. DETAILED DESCRIPTION

[0045] For the purpose of clarity and a thorough understanding of the application, the application will be described in detail with reference to the drawings. It should be apparent that the following description is only for a partial embodiment of the application and is not meant in any way to limit the application.

[0046] It should be noted that the brief description of the terms in the present application is only for the convenience of understanding the following described embodiments, and is not intended to limit the embodiments of the present application. Unless otherwise specified, these terms should be understood according to their ordinary and general meanings.

[0047] The terms "first", "second", "third", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar or similar objects or entities, and do not necessarily mean to limit the specific order or sequence, unless otherwise noted. It should be understood that the terms used in this way can be interchanged under appropriate circumstances.

[0048] The terms "include" and "have" and any variations thereof are intended to cover but not exclusive inclusion, for example, a product or device including a series of components does not necessarily limit to all components clearly listed, but can include other components not clearly listed or inherent to these products or devices.

[0049] The display device provided by the embodiments of the present application can have various implementation forms, for example, can be a television, a smart television, a laser projection device, a monitor, an electronic bulletin board, an electronic table, etc. Figure 1 and Figure 2 is a specific implementation of the display device of the present application.

[0050] Figure 1 is a schematic diagram of an operation scenario between the display device and the control device according to an embodiment. As shown in Figure 1 , a user can operate the display device 200 through the smart device 300 or the control device 100.

[0051] In some embodiments, the control device 100 can be a remote controller, and the communication between the remote controller and the display device includes infrared protocol communication or Bluetooth protocol communication, and other short-distance communication modes, to control the display device 200 through wireless or wired mode. The user can input user instructions through the keys on the remote controller, voice input, control panel input, etc., to control the display device 200.

[0052] In some embodiments, the smart device 300 (such as a mobile terminal, a tablet computer, a computer, a notebook computer, etc.) can also be used to control the display device 200. For example, the display device 200 is controlled by using an application program running on the smart device.

[0053] In some embodiments, the display device can not receive instructions using the above-mentioned smart device or control device, but can receive user control through touch or gesture, etc.

[0054] In some embodiments, the display device 200 can also be controlled in a manner other than the control device 100 and the smart device 300, for example, can directly receive user voice instructions control through the voice instruction acquisition module configured inside the display device 200 device, or can receive user voice instruction control through the voice control device set outside the display device 200 device.

[0055] In some embodiments, the display device 200 also communicates data with the server 400. The display device 200 can be allowed to communicate and connect through a local area network (LAN), a wireless local area network (WLAN), and other networks. The server 400 can provide various content and interaction to the display device 200. The server 400 can be a cluster, or multiple clusters, and can include one or more types of servers.

[0056] Figure 2 An exemplary configuration block diagram of the control device 100 according to an exemplary embodiment is shown. As shown inFigure 2 As shown, the control device 100 includes a controller 110, a communication interface 130, a user input / output interface 140, a memory, a power supply. The control device 100 can receive the input operation instruction of the user, and convert the operation instruction into an instruction that can be recognized and responded by the display device 200, and play a role of an intermediary in the interaction between the user and the display device 200.

[0057] As shown, the display device 200 includes at least one of a tuner and 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. Figure 3

[0058] In some embodiments, the controller includes a processor, a video processor, an audio processor, a graphics processor, a RAM, a ROM, a first interface to an n-th interface for input / output.

[0059] The display 260 includes a display screen component for presenting a picture, and a driving component for driving the image display, a component for receiving the image signal originating from the controller output, and a component for displaying the video content, the image content, and the menu operation interface, and a user operation UI interface.

[0060] The display 260 can be a liquid crystal display, an OLED display, and a projection display, and can also be a projection device and a projection screen.

[0061] The communicator 220 is a component for communicating with external devices or servers according to various communication protocol types. For example, the communicator can include at least one of a Wifi module, a Bluetooth module, a wired Ethernet module, and other network communication protocol chips or near field communication protocol chips, and an infrared receiver. The display device 200 can establish the sending and receiving of control signals and data signals with the external control device 100 or the server 400 through the communicator 220.

[0062] The user interface can be used to receive the control signal of the control device 100 (such as an infrared remote controller, etc.).

[0063] The detector 230 is used to collect signals of the external environment or interaction with the outside. For example, the detector 230 includes a light receiver, a sensor for collecting ambient light intensity; or the detector 230 includes an image collector, such as a camera, which can be used to collect external environmental scenes, user attributes or user interaction gestures, or the detector 230 includes a sound collector, such as a microphone, etc., for receiving external sounds.

[0064] ​The external device interface 240 can include, but is not limited to, any one or more of a high-definition multimedia interface (HDMI), an analog or digital high-definition component input interface (component), a composite video input interface (CVBS), a USB input interface (USB), an RGB port, etc. It can also be a composite input / output interface formed by a plurality of the above interfaces.

[0065] The tuner and demodulator 210 receives broadcast television signals through wired or wireless reception, and demodulates audio / video signals and EPG data signals from a plurality of wireless or wired broadcast television signals.

[0066] In some embodiments, the controller 250 and the tuner and demodulator 210 can be located in different separate devices, i.e., the tuner and demodulator 210 can also be in an external device of the main device where the controller 250 is located, such as an external set-top box, etc.

[0067] The controller 250 controls the operation of the display device and responds to user operations by storing various software control programs in the memory. The controller 250 controls the overall operation of the display device 200. For example, in response to receiving a user command for selecting a UI object displayed on the display 260, the controller 250 can perform an operation related to the object selected by the user command.

[0068] In some embodiments, the controller includes at least one of a central processing unit (CPU), a video processor, an audio processor, a graphics processing unit (GPU), a random access memory (RAM), a read-only memory (ROM), a first to nth interface for input / output, a communication bus, etc.

[0069] The user can input a user command through a graphical user interface (GUI) displayed on the display 260, and the user input interface receives the user input command through the graphical user interface (GUI). Alternatively, the user can input a user command by inputting a specific sound or gesture, and the user input interface receives the user input command by recognizing the sound or gesture through a sensor.

[0070] A "user interface" is a medium interface for interaction and information exchange between an application program or operating system and a user, which realizes the conversion between the internal form of information and the form acceptable by the user. The commonly used form of a user interface is a graphic user interface (GUI), which refers to a user interface related to computer operation displayed in a graphical manner. It can be an icon, window, control, etc. interface element displayed in the display screen of an electronic device, wherein the control can include an icon, button, menu, tab, text box, dialog box, status bar, navigation bar, Widget, etc. visual interface element.

[0071] Referring to Figure 4 In some embodiments, the system is divided into four layers, from top to bottom, an Applications layer (referred to as "application layer" for short), an Application Framework layer (referred to as "framework layer" for short), an Android runtime and system library layer (referred to as "system runtime library layer" for short), and a kernel layer.

[0072] In some embodiments, at least one application program is running in the application layer, which can be a window (Window) program, a system setting program or a clock program, etc. provided by the operating system; or an application program developed by a third-party developer. In specific implementation, the application program package in the application layer is not limited to the above examples.

[0073] The framework layer provides an application programming interface (API) and a programming framework for the application program. The application framework layer includes some pre-defined functions. The application framework layer is equivalent to a processing center, which decides the action of the application program in the application layer. The application program can access the resources in the system and obtain the services of the system through the API interface in the execution.

[0074] As Figure 4As shown, the application framework layer in the embodiments of the present application includes managers (Managers), content providers (Content Provider), etc., wherein the managers include at least one of the following modules: an activity manager (ActivityManager) for interacting with all activities running in the system; a location manager (Location Manager) for providing system location services to system services or applications; a package manager (Package Manager) for retrieving various information related to application packages currently installed on the device; a notification manager (NotificationManager) for controlling the display and clearing of notification messages; and a window manager (Window Manager) for managing icons, windows, toolbars, wallpapers and desktop components on the user interface.

[0075] In some embodiments, the activity manager is used to manage the life cycle of each application and the general navigation back function, such as controlling the exit, opening, back, etc. of the application. The window manager is used to manage all window programs, such as obtaining the size of the display screen, determining whether there is a status bar, locking the screen, intercepting the screen, controlling the display window change (such as reducing the display window, shaking the display, twisting the display, etc.) and the like.

[0076] In some embodiments, the system runtime library layer provides support for the upper layer, i.e. the framework layer. When the framework layer is used, the Android operating system runs the C / C++ library contained in the system runtime library layer to realize the functions of the framework layer.

[0077] In some embodiments, the kernel layer is the layer between hardware and software. As Figure 4 As shown, the kernel layer includes at least one of the following drivers: an audio driver, a display driver, a Bluetooth driver, a camera driver, a WIFI driver, a USB driver, an HDMI driver, a sensor driver (such as a fingerprint sensor, a temperature sensor, a pressure sensor, etc.), and a power supply driver, etc.

[0078] In some embodiments, the hardware or software architecture can be based on the above-mentioned embodiments, and in some embodiments, it can be based on other similar hardware or software architectures, as long as it can realize the technical solutions of the present application.

[0079] In some embodiments, the applications in the application layer include a video on demand program, and the display device can be configured to enter the video on demand program after being powered on.

[0080] Referring to Figure 5The home page of the video on demand program is provided with a navigation bar 510 and a content display area 520 located below the navigation bar 510. The navigation bar 510 is used to display a plurality of title controls, such as a VIP control, an education control, a TV series control, a movie control and the like. When the user moves the focus of the display device 200 between the plurality of title controls, the recommended content corresponding to the title control that obtains the focus is respectively displayed in the content display area 520, and when the focus is located on one of the title controls, the page displayed by the display device 200 is the home page corresponding to the title control.

[0081] In some embodiments, the recommended content displayed in the content display area 520 includes a plurality of media recommendation controls 521, and each media recommendation control 521 corresponds to a media that a recommended user watches, such as Figure 5 As shown, the media 1, the media 2, the media 3, the media 4, the media 5 and the media 6 are different media.

[0082] In some embodiments, the media corresponding to the media recommendation control 521 is a media recommended by the server according to the user portrait corresponding to the current account, i.e., the user account currently logged in by the display device 200. The media can be a free-to-watch media or a paid-to-watch media.

[0083] To improve the probability that the recommended media is the favorite media of the user and improve the payment probability of the user, the embodiments of the present application provide a media recommendation method, which can filter part of the media recommended by the server 400 and recommend the filtered media to the user.

[0084] Referring to Figure 6 , a flowchart of the media recommendation method according to some embodiments is shown in Figure 6 , which can include the following steps:

[0085] Step S101: obtaining a to-be-recommended data set corresponding to a target user, the to-be-recommended data set including media data of a plurality of target media.

[0086] In some embodiments, the target user is a user to which media is to be recommended, such as a user who logs in a user account on a display device. When the user enters a media recommendation scene (hereinafter referred to as a scene) of the display device, the display device can send a media recommendation request to the server, and the media recommendation request can include an account identifier of the current account to obtain the media recommended by the server for the user. The scene can include Figure 5 , the scene corresponding to each title control of the navigation bar 510, for example, in a TV series scene corresponding to a TV series control, the user needs to be recommended a TV series media, and in a movie scene corresponding to a movie control, the user needs to be recommended a movie media.

[0087] In some embodiments, after receiving the media recommendation request, the server can obtain a to-be-recommended data set corresponding to the current account, which can include media data of a plurality of target media. The target media can be media determined by the server based on the user portrait of the target user for media recommendation, or the target media can also be media filtered from the media information library according to a preset rule, such as play heat, release time, etc.

[0088] Step S102: inputting the to-be-recommended data set into the multi-task recommendation model to obtain a prediction probability of the second stage task corresponding to the target media output by the multi-task recommendation model, wherein the multi-task recommendation model is a network model obtained by performing transfer learning of a pre-training model on the second stage task, the pre-training model is a network model obtained by training on a first stage task, the first stage task includes a task of predicting whether a user performs viewing and a task of predicting whether a user performs payment, and the second stage task includes the task of predicting whether a user performs payment.

[0089] In some embodiments, the server can calculate the prediction probability of the second stage task corresponding to the target media in the to-be-recommended data set through the pre-trained multi-task recommendation model, and determine whether to recommend the target media according to the prediction probability.

[0090] Among them, the multi-task recommendation model can be a network model obtained by performing transfer learning of a pre-training model trained by a first stage task on a second stage task.

[0091] In some embodiments, the first stage task includes a task of predicting whether a user performs viewing and a task of predicting whether a user performs payment, so that the pre-training model trained by the first stage task can predict the probability of a user performing viewing operation and the probability of performing payment operation on the source domain training sample corresponding to the pre-training model, such as the first sample media, and then generate the recommendation probability of the first sample media according to the probability of performing viewing operation and the probability of performing payment operation on the first sample media, for example, the probability of performing viewing operation and the probability of performing payment operation on the first sample media are weighted and summed to obtain the recommendation probability of the first sample media. The higher the recommendation probability is, the higher the probability of the user performing viewing operation and the probability of performing payment operation on the first sample media are.

[0092] It should be noted that the first sample media asset can be free media asset or paid media asset, and the probability of performing the payment operation on the first sample media asset refers to the probability of the user jumping from the detail page of the first sample media asset to the payment page, and after completing the payment operation, the user can watch the paid media asset, and if the payment operation is not completed, the free media asset can still be watched. Since the detail page of some first sample media assets can display media assets associated with the first sample media asset, and these media assets can be paid media assets, in order to watch these paid media assets, the user can perform a payment operation, and therefore, when the server recommends free media assets to the user, the user has a certain probability of performing a payment operation. The media assets associated with the first sample media asset refer to media assets having a mapping relationship with the first sample media asset, such as media assets of the same type, media assets of the same actor, and the like.

[0093] The first sample media asset can include media assets watched by the user in the past and media assets associated with the media assets. Since in these first sample media assets, the number of media assets on which the user has historically performed a payment operation can be relatively small compared to the media assets watched by the user in the past, this can result in a prediction probability obtained by the pre-training model, which represents the probability of the user performing a watching operation on the first sample media asset, having a relatively high accuracy, and the probability of the user performing a payment operation having a relatively low accuracy, which is difficult to meet the demand of taking into account the media preference of the user and the member conversion rate.

[0094] To solve the technical problem, the pre-training model can be subjected to transfer learning of a second stage task.

[0095] In some embodiments, the second stage task includes a task of predicting whether the user performs a payment, so that the multi-task recommendation model trained through the second stage task can improve the accuracy of predicting the probability of the user performing a payment operation on a target domain training sample, such as a second sample media asset, compared to the pre-training model. According to the probability of the user performing a payment operation on the second sample media asset, a prediction probability of the second sample media asset corresponding to the second stage task can be generated. The second sample media asset can include the first sample media asset, or the second sample media asset can also include media assets other than the first sample media asset to improve the transfer learning effect.

[0096] Since the multi-task recommendation model is obtained by transfer learning based on the pre-training model, when the prediction probability obtained by the multi-task recommendation model is relatively high, the recommendation probability obtained according to the prediction probability will exceed the recommendation probability threshold, which represents that the probability of the user performing a watching operation and a payment operation on the second sample media asset are both relatively high, thereby achieving the demand of taking into account the media preference of the user and the member conversion rate.

[0097] Step S103: obtaining a recommendation probability of the target media asset according to the prediction probability.

[0098] In some embodiments, the second stage task includes a plurality of tasks, and a comprehensive probability score is calculated as the recommendation probability of the target media asset based on the prediction probabilities corresponding to the plurality of tasks.

[0099] For example, the recommendation probability of the target media asset can be obtained by weighted sum of the plurality of prediction probabilities.

[0100] For example, the recommendation probability of the target media asset can be obtained by averaging the plurality of prediction probabilities.

[0101] Step S104: determining the target media asset to be recommended for the target user from the target media asset according to the recommendation probability.

[0102] In some embodiments, after the recommendation probability of the target media asset in the target media asset set to be recommended is calculated, the target media asset can be sorted according to the recommendation probability from large to small, and the target media asset with a high ranking is determined as the target media asset to be recommended for the target user.

[0103] As can be seen from the above embodiments, the multi-task model provided by the embodiments of the present application is a network model obtained by pre-training a model through a first stage task and then performing transfer learning on the model through a second stage task. The first stage task includes a task of predicting whether a user will watch and a task of predicting whether a user will pay, so that the media asset recommended according to the pre-trained model has a high probability of being watched and a high probability of being paid. The second stage task includes a task of predicting whether a user will pay, so that the multi-task recommendation model obtained after transfer learning further improves the probability of a user paying on the basis of having a high probability of being watched, thereby improving the effect of media asset recommendation taking into account user media asset preferences and member conversion rate.

[0104] To further illustrate the media asset recommendation method provided by the embodiments of the present application, the training process of the multi-task recommendation model is introduced below. Referring to Figure 7 The training process of the multi-task recommendation model can include the following steps:

[0105] Step S201: training a preset recommendation model through a first stage task by using source domain training samples.

[0106] In some embodiments, the training process of the multi-task recommendation model includes a pre-training phase and a transfer learning phase, wherein the pre-training phase is a first training phase, and the transfer learning phase is a second training phase. In the pre-training phase, the preset recommendation model is trained through a first stage task by using source domain training samples, and in the transfer learning phase, the preset recommendation model is trained through a second stage task by using target domain training samples.

[0107] In some embodiments, the set of all tasks of the first stage and the second stage is T, T = {t1, t2, …, t5}, |T| = 5, which means that the total tasks include 5 tasks. The 5 tasks are respectively: a task t1 of predicting whether to perform the on-demand operation, a task t2 of predicting whether the viewing duration exceeds the viewing duration threshold, a task t3 of predicting whether to enter the payment page, a task t4 of predicting whether the display time of the payment page exceeds the payment duration threshold, and a task t5 of predicting whether to complete the payment operation.

[0108] If the viewing duration exceeds the viewing duration threshold, it is determined that the user performs effective viewing, and if the viewing duration does not exceed the viewing duration threshold, it is determined that the user does not perform effective viewing. The effective viewing can represent that the user likes the media asset. In different scenarios, the viewing duration threshold can be different. For example, in a TV series scenario, the user can determine whether to like the media asset after a relatively long time of viewing because the plot progresses relatively slowly. Therefore, the viewing duration threshold is 10 minutes. In a movie scenario, the user can determine whether to like the media asset after a relatively short time of viewing because the plot progresses relatively quickly. Therefore, the viewing duration threshold can be 5 minutes.

[0109] The payment page refers to a checkout page. In the payment page, the user can finally complete the payment operation, or can finally not complete the payment operation due to weak payment intention or the like. If the display time of the payment page exceeds the payment duration threshold, it represents that the user has strong payment intention, and if the display time of the payment page does not exceed the payment duration threshold, it represents that the user has weak payment intention.

[0110] The first stage task is the source domain task T s , T s = {t1, t2, t3}, |T s | = 3, which means that the first stage task includes 3 tasks. The second stage task is the target domain task T τ , T τ = {t3, t4, t5}, |T τ | = 3, which means that the second stage task includes 3 tasks. T = T s ∪ T τ .

[0111] In some embodiments, the set of training samples can be represented by D, D = {(x1, y1), (x2, y2), … (x i , y i ), …}, which can be usually extracted from the log data of the system. Each training sample includes feature data x and label data y.

[0112] The feature data x includes user data x u , media asset data xv and the context data x of the first sample media asset uv , that is: x = [x u , x v , x uv ].

[0113] Wherein, the user data can include attribute data of the user and playing behavior data of the user in each scene. The attribute data can include user age, gender, region, TV model, TV screen size, whether a member, user's time preference for watching media assets, and the like. The playing behavior data can include statistical information of the user's time preference for watching media assets in each scene, such as s1 scene, s2 scene, …, sk scene, statistical information of scene preference, statistical information of actor preference, statistical information of content category preference, and the like. Exemplarily, s1 scene can be a TV series scene, and s2 can be a movie scene.

[0114] The media asset data includes attribute data of the media asset and statistical data of user behavior corresponding to the media asset. The attribute data can include director, actor, category, release time, language, and the like, and the statistical data can include the number of users who have watched the media asset, average playing time of the media asset, and the like.

[0115] The context data includes media asset data of the last media asset watched by the user before watching the media asset and media asset data of the next media asset watched after watching the media asset, and time information when the media asset is exposed to the user, geographic information of the user when the media asset is exposed to the user, and the like. For example, a first sample media asset is media asset B, the user jumps from a detail page of media asset A to a detail page of media asset B, and then jumps from the detail page of media asset B to a detail page of media asset C when watching media asset A in history, then the context data corresponding to media asset B includes media asset data of media asset A and media asset data of media asset C.

[0116] The label data y represents the behavior feedback of the user after the media asset is exposed to the user, y = [y t |t∈T]. For example, after media asset v is exposed to user u, the user u performs a on-demand operation on the media asset v, the watching time exceeds a watching time threshold, and enters a payment page, then in the pre-training stage, for the above-mentioned five tasks, the sample label is: y = {1, 1, 1, 0, 0}, wherein 1 indicates that the user performs the operation corresponding to the task, and 0 indicates that the user does not perform the operation corresponding to the task; if the media asset v is exposed to the user u, the user u does not perform the operation corresponding to the above-mentioned five tasks, then the sample label is: y = {0, 0, 0, 0, 0}.

[0117] The set of source domain training samples can be represented as D s , The number of samples is n s , and the label of the sample is: ys =[y t |t∈T s ] = [y 1 ,y 2 ,y 3 ].

[0118] The set of training samples for the target domain can be used in D τ express, The number of samples is n τ The sample label is: y τ =[y t |t∈T τ ] = [y 4 ,y 5 ,y 3 This ensures that the labels for the source domain task and the target domain task are aligned on task t3. The source domain training samples and the target domain training samples maintain consistency in the feature data x.

[0119] In some embodiments, the preset recommendation model may include an MMOE (Multi-gate Mixture-of-Experts) model, see [link to relevant documentation]. Figure 8 The preset recommendation model may include an input layer, an embedding layer, a feature layer, an expert model layer, and an output layer connected in sequence.

[0120] In some embodiments, the input layer is used to input source domain training samples D. s :

[0121]

[0122] x = [x u ,x v ,x uv ]

[0123] In some embodiments, the embedding layer is used to transform sparse data into dense data, and the input of the embedding layer is the source domain training samples input to the input layer.

[0124] The embedding layer may include multiple vector units, which perform calculations based on the data input from the input layer to obtain the source domain training samples D. s The corresponding sample feature vector Emb:

[0125] Emb = lookup(x, E).

[0126] The embedding layer is essentially a lookup table E, which is used to find the sample feature vector of the source domain training sample in the lookup table E based on the input.

[0127] In some embodiments, the feature layer is used to calculate a scene preference feature vector, representing the user's media asset preference features in each recommendation scenario, based on the sample feature vector. It also provides input data to the expert model layer. The feature layer can calculate the user's scene preference feature vector across multiple scenarios based on the sample feature vector output by the embedding layer. The calculation formula is as follows:

[0128] Fea = relu(g(Emb))

[0129] Here, Fea represents the user's scene preference feature vector, g(·) denotes a deep neural network (DNN), and ReLU is the activation function. The feature layer performs preliminary processing on the sample feature vectors.

[0130] In some embodiments, the expert model layer is used to extract feature diversity and output a feature map based on the extracted features. The expert model layer may include multiple expert networks for extracting feature diversity from the aforementioned sample feature vectors and scene preference feature vectors, such as six expert networks. Each expert network can output a feature map, and the formula for the feature map is shown below:

[0131] FM j =relu(g(Fea)),j∈{1,2,3,4,5,6}

[0132] Among them, FM j The feature map representing the output of the j-th expert network is represented by the function g(·), which can be expressed using a deep neural network (DNN), where ReLU is the activation function.

[0133] In some embodiments, the output of the expert model layer is FM, which is obtained by concatenating the outputs of six expert networks:

[0134] FM={FM1,FM2,FM3,FM4,FM5,FM6}

[0135] In some embodiments, the task tower layer is used to predict tasks and output features corresponding to those tasks. The task tower layer may include multiple task towers. During the pre-training phase, the output layer is configured with three task towers, corresponding to three source domain tasks T. s These are t1, t2, and t3. Among them, task t1, which predicts whether a video-on-demand action will be taken, and task t2, which predicts whether the viewing time will exceed the viewing time threshold, are tasks for predicting whether the user will watch the video. Task t3, which predicts whether the user will enter the payment page, is a task for predicting whether the user will make a payment.

[0136] Each task tower is a neural network. The task tower layers use gating networks to control the weights of each expert network input to the task tower, i.e.:

[0137] gate t = softmax(g(Fea)), t e T s

[0138] wherein softmax(·) represents that the final task output activation function is softmax.

[0139] gate t = {gate tj |j = {1, 2, 3, 4, 5, 6}, represents the weight size of each expert network, and the size is a vector with a length of 6, which is consistent with the number of expert networks, and each task corresponds to a gate output of an expert network t , gate t The weighted sum of the feature values of the expert network can obtain the features output by each task tower:

[0140] FT t = FM*gate t In some embodiments, the output layer includes a neural network structure for outputting the prediction probability p t corresponding to each task tower according to the features output by the task tower.

[0141] p t = sigmoid(g(FT t ))

[0142] wherein FT t , as the output of the task tower t, after the neural network structure of the output layer, the prediction probability p t corresponding to each task tower task t is obtained.

[0143] Since the first stage tasks are all CTR (click through rate) estimation, which are all 0 / 1 binary classification tasks, a log loss function can be used as the loss function, that is:

[0144]

[0145] wherein |D s | = n s , Ω is a regular term, which can be an L1 regular term, a is a hyperparameter, representing the label data of task t.

[0146] Step S202: training the pre-trained model through the target domain training sample for the second stage task.

[0147] In some embodiments, after obtaining the pre-training model through the first training stage, the pre-training model can be trained by the target domain training samples for the second stage task to obtain the multi-model recommendation model.

[0148] Referring to Figure 9 FIG. 1 shows a flowchart of a migration learning method according to some embodiments, which can include the following steps: Figure 9

[0149] Step S301: Constructing target domain training samples.

[0150] In some embodiments, the target domain training samples can include user data, media data of the second sample media, and context data of the second sample media.

[0151] The user data can be the same as the user data in the source domain training samples.

[0152] The second sample media can include samples in the first sample media for which the label probability of the prediction of entering the paid page is 0. Since the purpose of the migration learning stage is to improve the member conversion rate, the samples with the label probability of 0 indicate that the user will not perform a paid operation on the media corresponding to the samples, which is not conducive to improving the member conversion rate. Therefore, deleting these samples can improve the migration learning effect. After deleting these samples from the first sample media, the number of remaining samples can be small. To improve the migration learning effect, the time span of sample selection can be expanded to increase the number of samples. For example, the second sample media can include media corresponding to the user in the last month, which can be referred to as new sample media. The first sample media does not include the media in the last month, i.e., the second sample media includes the first sample media and the new sample media. The time attribute of the new sample media is different from that of the first sample media. The time attribute can include the time when the media becomes the media corresponding to the user, such as the time when the user historically watched the media.

[0153] Compared with the source domain training samples, the target domain training samples delete media data of part of the first sample media, and the context data corresponding to the part of the first sample media also needs to be deleted. The target domain training samples add media data of part of the sample media, and the context data corresponding to the part of the sample media also needs to be added.

[0154] In some embodiments, in the migration learning stage, the label probability corresponding to the first stage task in the second sample media can be deleted, and the label probability of the second sample media can be set to the label probability corresponding to the second stage task. After the pre-training model is input with the target domain training samples, the task tower of the output layer is updated to the target domain task T τ ​t3, t4, t5. In some embodiments, both the transfer learning stage and the pre-training stage include the task t3 of predicting whether to enter a paid page, which is an intermediate task that takes into account both the media preference of a user and the member conversion rate, and can improve the transfer learning effect.

[0155] In some embodiments, after obtaining the target domain training samples, the pre-trained model can be trained by inputting the target domain training samples into the pre-trained model to achieve transfer learning for the second stage task.

[0156] Step S302: Set the network parameters of the embedding layer and the feature layer as fixed parameters.

[0157] In some embodiments, in the transfer learning stage, the pre-trained model can be initialized. Referring to Figure 10 Before inputting the target domain training samples into the pre-trained model, the network parameters of the embedding layer and the feature layer can be set as fixed parameters, so that the network parameters of the embedding layer and the feature layer are frozen, which can ensure that the network parameters of the embedding layer and the feature layer do not participate in the gradient descent process during the entire transfer learning stage, so as to achieve the purpose of stabilizing the feature extraction performance of the pre-trained model.

[0158] Step S303: Set the network parameters of one of the expert networks to zero, and set the weight of the zeroed expert network based on the loss function.

[0159] In some embodiments, before inputting the target domain training samples into the pre-trained model, in order to fully utilize the expert network that plays an important role in the prediction result of the pre-trained model for the second stage task, and retrain the expert network that has no effect or poor effect on the prediction result of the second stage task, the expert network can be automatically adjusted by adjusting the weight of the expert network, wherein the weight of the expert network is the weight of the expert network in each task tower, and the greater the weight, the greater the effect of the output of the expert network on the prediction effect of the task tower.

[0160] By setting the network parameters of one of the expert networks to zero and setting the weight of the expert network as a function of the change in the loss function of the pre-trained model, the weight of the expert network can be automatically adjusted according to the update of the loss function, i.e., the weight of the expert network is updated by the loss function. After setting the network parameters of the expert network to zero, if the loss value output by the loss function increases at a high rate, it indicates that the expert network has a greater effect on the second stage task, and its weight needs to be increased, and if the loss value output by the loss function increases at a low rate, it indicates that the expert network has a smaller effect on the second stage task, and its weight needs to be decreased.

[0161] The weight of the expert network can be updated according to the following calculation formula:

[0162]

[0163] where z represents the pre-trained model in the pre-training stage, θ * represents the model parameters of the pre-trained model in the pre-training stage, i.e., the neuron parameters, θ *\j represents the model parameters of the pre-trained model when the neuron parameters of the jth expert network are set to zero and the model parameters of other expert networks remain consistent with the original model in the pre-training stage. z(x, θ * ) = [p 1 , p 2 , p 3 ], z(x, θ * ) represents the output of the pre-trained model in the pre-training stage on the training samples of the target domain, which is a vector with a dimension of 3, because the tasks of the target domain and the source domain are all three. y τ = [y 4 , y 5 , y 3 ], in y τ , the task t3 is placed at the last position, which can ensure that the knowledge in the pre-training stage can be accurately migrated to the tasks of the target domain.

[0164] The loss function L(·) remains consistent with the loss function in the pre-training stage. L(z(x, θ *\j ), y τ ) represents the loss function value of the supervised attention mechanism, which adopts the method of setting the neuron parameters of the jth expert network to zero and retraining the pre-trained model on the training samples of the target domain. Without loss of generality, only one epoch of model training can obtain the loss function value of L(z(x, θ *\j ), y τ ) on the validation set of the target domain. Similarly, keeping the model parameters of the pre-trained model in the pre-training stage consistent, one epoch of model training on the training samples of the target domain can obtain the loss function value of L(z(x, θ * ), y τ ) on the validation set of the target domain.

[0165] represents the weight of the jth expert network, and the calculation logic of the expert network selects the softmax function to normalize the weights of all expert networks to 1, which can also ensure that the weight of each expert network is not less than zero. Expert network

[0166] Based on this supervised attention mechanism, more attention can be provided to the expert network with a larger performance improvement in the second stage task, that is, such an expert network adapts better to the second stage task.

[0167] Step S304: Add the feature map regularization loss and the gating network regularization loss in the loss function.

[0168] In some embodiments, to improve the transfer learning effect, before training the pre-trained model through the target domain training samples, a feature map regularization loss can be added in the loss function of the pre-trained model to optimize the network parameters of the expert networks, so that the network parameters of the expert networks that have a large performance improvement on the second stage task are kept consistent with the original model as much as possible, and on the contrary, the network parameters of the expert networks that do not work or work poorly on the second stage task are changed greatly from the original model. The calculation formula of the feature map regularization loss is as follows:

[0169]

[0170] Wherein, Ω' is the feature map regularization loss, θ represents the model parameters of the pre-trained model in the transfer learning stage, and N represents the number of expert networks. In the embodiments of the present application, the value of N can be 6. j (z, θ, x) is the output value of the jth expert network in the transfer learning stage, FM j (z, θ * , x) is the output value of the jth expert network in the pre-training stage. Adding the feature map regularization loss is to ensure that as much knowledge learned in the source domain as possible can be retained in the transfer process of the pre-trained model.

[0171] In some embodiments, to improve the transfer learning effect, before training the pre-trained model through the target domain training samples, a gating network loss can be added in the loss function of the pre-trained model to optimize the network parameters of the expert networks and solve the problem of expert collapse, i.e., the problem that some expert networks basically do not work on improving the performance of the model. The calculation formula of the gating network loss is as follows:

[0172]

[0173] Wherein, Ω" is the feature map regularization loss.

[0174] In some embodiments, the problem of expert collapse can also be solved by increasing the loss rate of the gating network.

[0175] Based on the above embodiments, the loss function of the pre-trained model can be updated as follows:

[0176]

[0177] Wherein, b and c are hyperparameters.

[0178] Step S305: Model training through target domain training samples.

[0179] In some embodiments, after updating the loss function, the updated pre-training model can be trained by the target domain training samples to obtain the multi-task recommendation model. During the model training process, the optimizer adopts Adam (adaptive moment estimation), and the initial learning rate is 0.001, which is decayed by a factor of 0.9 every 10 complete training.

[0180] After the above steps, the pre-training model is migrated to obtain the multi-task recommendation model, and the user data containing the target user, the media asset data of the target media asset, and the context data of the target media asset are input into the multi-task recommendation model. The multi-task recommendation model can obtain the prediction probability of each second stage task through the output layer corresponding to the second stage task. After the plurality of prediction probabilities are weighted and controlled, the recommendation probability of the target media asset can be output. The higher the recommendation probability is, the higher the probability of the target user watching the media asset and performing a payment operation is, thereby achieving the purpose of taking into account the media asset preference of the user and the member conversion rate.

[0181] The weighting control can include the following calculation formula:

[0182]

[0183] In the above formula, w t represents the weight size of each second stage task, which changes with different business scenarios and different business stages, and the value thereof can be adjusted according to business experience. According to the above formula, the comprehensive probability score (score) of each target media asset can be obtained, which represents the recommendation probability of the target media asset. The top k media assets can be extracted according to the comprehensive probability score from high to low, and recommended to the user, wherein the value of k can be determined according to actual needs. As can be seen from the above embodiments, the embodiments of the present application propose a multi-task migration scheme to solve the scheme of deeper user behavior modeling in multiple scenarios. By freezing the embedding layer and the feature layer, the stability of feature generation can be ensured. By the expert adaptive strategy of the supervised attention mechanism, the effective expert network on the second stage task is retained, and the experts that are ineffective or have poor effects on the second stage task can also be fully trained, thereby increasing the utilization efficiency of the expert network. Through feature map adaptation, the network parameters of the optimized expert network are optimized, thereby improving the prediction effect on the second stage task. The gating network is self-adaptive, which effectively avoids the expert collapse problem. By retaining a same task in the pre-training stage and the migration learning stage, and the same task is the intermediate link of the user watching the media asset and performing payment, the purpose of taking into account the media asset preference of the user and the member conversion rate can be achieved, and the migration learning effect can be improved.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

[0185] The foregoing description has been set forth in conjunction with a particular embodiment and implementations, for purposes of convenience and illustration. However, the foregoing discussion is not intended to be exhaustive or to limit the embodiments to the precise forms disclosed. Various modifications and changes are possible in light of the teachings above. The selection and arrangement of various elements described or claimed in the foregoing embodiments are presented to provide the best description possible of the principles and the practical applications of the embodiments. The embodiments have been chosen and described in order to explain the principles of the embodiments and the practical applications they provide, and to allow others skilled in the art to understand for themselves the various implementations with various modifications as are suited to the particular use contemplated.

Claims

1. A server, characterized by The server is configured to: obtain a to-be-recommended data set corresponding to a target user, the to-be-recommended data set including media data of a plurality of target media; input the to-be-recommended data set into a multi-task recommendation model to obtain a prediction probability of a second stage task of the target media output by the multi-task recommendation model, wherein the multi-task recommendation model is a network model obtained by performing transfer learning of a pre-training model on the second stage task, the pre-training model is a network model obtained by training on a first stage task, the first stage task includes a task of predicting whether a user watches and a task of predicting whether a user pays, and the second stage task includes the task of predicting whether the user pays; obtain a recommendation probability of the target media according to the prediction probability; determine a to-be-recommended media corresponding to the target user from the target media according to the recommendation probability.

2. The server of claim 1, wherein, The pre-training model includes an input layer, an embedding layer, a feature layer, an expert model layer, and an output layer connected in sequence, wherein the embedding layer is used to calculate a sample feature vector, and the feature layer is used to calculate a scene preference feature vector representing media preference features of a user in each recommendation scene based on the sample feature vector; the transfer learning of the pre-training model on the second stage task includes: constructing a target domain training sample, the target domain training sample including media data of a second sample media, the media data of the second sample media including label probabilities corresponding to the second stage task; setting network parameters of the embedding layer and the feature layer as fixed parameters; inputting the target domain training sample into the pre-training model with the fixed parameters to obtain a recommendation probability of the target domain training sample.

3. The server of claim 2, wherein, The source domain training sample corresponding to the pre-training model includes media data of a first sample media, the media data of the first sample media including label probabilities corresponding to the first stage task, and the constructing of the target domain training sample includes: deleting, in the source domain training sample, the first sample media whose label probability corresponding to the task of predicting whether a user pays is zero, and the second sample media including the first sample media retained in the source domain training sample; deleting the label probabilities corresponding to the first stage task in the second sample media and setting the label probabilities of the second sample media as the label probabilities corresponding to the second stage task.

4. The server of claim 3, wherein, The constructing of the target domain training sample further includes: obtaining a new sample media whose time attribute is different from that of the first sample media, and the second sample media further including the new sample media.

5. The server of claim 2, wherein, The expert model layer includes a plurality of expert networks for outputting feature maps based on the sample feature vector and the scene preference feature vector, and the transfer learning of the pre-training model on the second stage task further includes: setting weights of the expert networks as a function of a loss function change amount of the pre-training model before inputting the target domain training sample into the pre-training model with the fixed parameters, and setting network parameters of one of the expert networks to zero.

6. The server of claim 5, wherein, The transfer learning of the pre-training model on the second stage task further includes: Before inputting the target domain training sample into the pre-training model with the fixed parameters, the feature map regularization loss is added to the loss function of the pre-training model.

7. The server of claim 5, wherein, The expert model layer includes a gating network corresponding to the expert network, and the second stage task migration learning of the pre-training model further includes: Before inputting the target domain training sample into the pre-training model with the fixed parameters, the gating network regularization loss is added to the loss function of the pre-training model.

8. The server of claim 1, wherein, The task of predicting whether the user watches includes the task of predicting whether the on-demand operation is performed, and the task of predicting whether the viewing time exceeds the viewing time threshold, the task of predicting whether the user pays in the first stage task includes the task of predicting whether the payment page is entered, and the task of predicting whether the user pays in the second stage task includes the task of predicting whether the payment page is entered, the task of predicting whether the display time of the payment page exceeds the payment time threshold, and the task of predicting whether the payment operation is completed.

9. The server of claim 1, wherein, The target user's user data and the target media asset's context data are also included in the to-be-recommended data set, the target media asset includes the media asset determined by the media asset recommendation based on the target user's user portrait, and the context data includes the media asset data of the last media asset before the target user watches the target media asset and the media asset data of the next media asset after watching the target media asset.

10. A media recommendation method characterized by, It includes: Obtain the target user's corresponding to-be-recommended data set, which includes media asset data of a plurality of target media assets; Input the to-be-recommended data set into a multi-task recommendation model to obtain the prediction probability of the second stage task of the target media asset output by the multi-task recommendation model, wherein the multi-task recommendation model is a network model obtained by performing migration learning of the second stage task on a pre-training model, the pre-training model is a network model obtained by training through a first stage task, the first stage task includes a task of predicting whether a user watches and a task of predicting whether a user pays, and the second stage task includes the task of predicting whether the user pays; According to the prediction probability, obtain the recommendation probability of the target media asset; According to the recommendation probability, determine the to-be-recommended media asset corresponding to the target user from the target media asset.

Images