Method, device, electronic device and storage medium for recommending media information

Abstract

The application provides a media information recommendation method and device, electronic equipment and a storage medium. The method comprises the following steps: obtaining a first object feature of a target object, a content feature of at least two media information sequences to be recommended, and a combination feature; wherein the media information sequence comprises at least two types of media information, and the combination feature comprises a type preference feature and a position preference feature of the media information; determining a recommendation score of each media information sequence based on the first object feature, the content feature of each media information sequence, and the combination feature; selecting a target media information sequence from the at least two media information sequences based on the recommendation score of each media information sequence, and recommending the target media information sequence to a terminal corresponding to the target object. Through the application, the recommendation accuracy and efficiency of the media information can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a media information recommendation method and device, an electronic device, and a storage medium. BACKGROUND

[0002] Artificial intelligence (AI) is the use of digital computers or digital computer-controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which aims to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is the design principle and implementation method of various intelligent machines, so that the machine has the functions of perception, reasoning and decision-making.

[0003] A recommendation system is an important application branch of artificial intelligence. In the related art, when media information is recommended based on a recommendation system, the most optimal media information corresponding to each recommendation position is usually predicted based on user basic features and media information basic features to perform information recommendation according to the recommendation position. However, there are many types of media information, and the needs of users in different scenarios are different. The recommendation based on only user basic features and media information basic features cannot achieve the corresponding effect, and the display position of the media information can also affect the recommendation effect. SUMMARY

[0004] The embodiments of the present application provide a media information recommendation method, device, electronic device and storage medium, which can improve the recommendation accuracy and efficiency of media information.

[0005] The technical solutions of the embodiments of the present application are as follows:

[0006] The embodiments of the present application provide a media information recommendation method, which includes the following steps:

[0007] obtaining first object features of a target object, content features of at least two media information sequences to be recommended, and combination features;

[0008] The media information sequence includes at least two types of media information, and the combination features include type preference features and position preference features of the media information.

[0009] The type preference features are used to indicate the preference of the target object for the type of media information in different scenarios, and the position preference features are used to indicate the preference of the target object for the display position of the media information.

[0010] determine a recommendation score of each of the media information sequences based on the first object feature, the content feature of each of the media information sequences, and the combination feature;

[0011] select a target media information sequence from the at least two media information sequences based on the recommendation score of each of the media information sequences, and recommend the target media information sequence to a terminal corresponding to the target object.

[0012] Embodiments of the present application further provide a media information recommendation device, comprising:

[0013] an obtaining module, configured to obtain a first object feature of a target object, a content feature of at least two media information sequences to be recommended, and a combination feature;

[0014] wherein the media information sequences comprise at least two types of media information, and the combination feature comprises a type preference feature and a position preference feature of the media information;

[0015] wherein the type preference feature is used to indicate a preference of the target object for the type of the media information in different scenarios, and the position preference feature is used to indicate a preference of the target object for a display position of the media information;

[0016] a determining module, configured to determine a recommendation score of each of the media information sequences based on the first object feature, the content feature of each of the media information sequences, and the combination feature;

[0017] a recommending module, configured to select a target media information sequence from the at least two media information sequences based on the recommendation score of each of the media information sequences, and recommend the target media information sequence to a terminal corresponding to the target object.

[0018] In the above scheme, the obtaining module is further configured to obtain a first group feature of the target object, and a second group feature of at least two pre-set object groups, each of the object groups comprising at least two objects;

[0019] perform feature matching of the first group feature of the target object and the second group feature of each of the object groups respectively, to determine a matching degree of the first group feature and each of the second group features, and take an object group corresponding to a second group feature with the highest matching degree with the first group feature as a target object group;

[0020] obtain the content feature of the at least two media information sequences to be recommended, and obtain a group first object feature of the target object group and a group combination feature of the target object group corresponding to the at least two media information sequences;

[0021] The first object feature of the group is taken as a first object feature of the target object, and the group combination feature is taken as a combination feature of the target object corresponding to the at least two media information sequences.

[0022] In the scheme, the acquisition module is further configured to acquire user data of the target object, content data of the at least two media information sequences to be recommended, and combination data; and the combination data includes type preference data and location preference data of the media information.

[0023] The feature extraction layer of the first recommendation model is used to perform feature extraction on the user data, the content data of the at least two media information sequences to be recommended, and the combination data, respectively, to obtain a first object feature of the target object, content features of the at least two media information sequences to be recommended, and a combination feature.

[0024] The recommendation score of each media information sequence is determined based on the first object feature, the content features of each media information sequence, and the combination feature, and the determination includes:

[0025] The feature prediction layer of the first recommendation model is used to perform prediction in combination with the first object feature, the content features of each media information sequence, and the combination feature to obtain the recommendation score of each media information sequence.

[0026] In the scheme, the device further includes:

[0027] The first training module is configured to acquire a sample first object feature, a sample content feature, and a sample combination feature corresponding to a training sample, and the training sample is labeled with a corresponding sample label.

[0028] The training sample includes at least two types of media information samples, and the sample combination feature includes a sample type preference feature and a sample location preference feature of the media information sample.

[0029] The sample type preference feature is used to indicate the preference of a corresponding sample user for a type of corresponding media information sample in different scenarios, and the sample location preference feature is used to indicate the preference of the corresponding sample user for a display position of the corresponding media information sample.

[0030] The sample first object feature, the sample content feature, and the sample combination feature are input into the first recommendation model, and the first recommendation model is used to perform prediction of the recommendation score to obtain a corresponding prediction result.

[0031] The difference between the prediction result and the corresponding sample label is acquired.

[0032] Based on the difference, the model parameters of the first recommendation model are updated.

[0033] In the above solution, the determining module is further configured to perform the following processing on each of the media information sequences respectively to determine a recommendation score of each of the media information sequences:

[0034] The neural network model is used to perform classification prediction on the click data by combining the first object feature, the content feature of the media information sequence, and the combined feature, to obtain a click prediction result of at least two categories corresponding to the media information sequence.

[0035] A recommendation weight value corresponding to the click prediction result of each category is obtained.

[0036] Based on the click prediction result of the at least two categories and the recommendation weight value corresponding to the click prediction result of each category, a recommendation score of the media information sequence is determined.

[0037] In the above solution, the recommendation module is further configured to receive a media information sequence acquisition request sent by the terminal, the acquisition request being triggered in response to a media information acquisition instruction for a first-screen recommendation position.

[0038] The target media information sequence is sent to the terminal to display the target media information sequence through a first-screen recommendation position of the terminal.

[0039] In the above solution, the target media information sequence is displayed through the first-screen recommendation position of the terminal, and the recommendation module is further configured to obtain at least two candidate media information for each information recommendation position in a candidate recommendation position sequence to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence, wherein the candidate recommendation position sequence is a non-first-screen and includes at least two information recommendation positions.

[0040] The second object feature of the target object, the content feature of each candidate media information included in each of the media information candidate sequences, and the position feature are obtained.

[0041] Based on the second object feature of the target object, the content feature of each candidate media information included in each of the media information candidate sequences, and the position feature, a sequence recommendation score of each of the media information candidate sequences is determined.

[0042] Based on the sequence recommendation score of each of the media information candidate sequences, a target media information candidate sequence is selected from the at least two media information candidate sequences, and the target media information candidate sequence is recommended to a terminal corresponding to the target object, so that

[0043] The target media information candidate sequence is displayed through the candidate recommendation position sequence in the non-first-screen.

[0044] In the scheme, the recommendation module is further configured to perform click data prediction on each first media information to be recommended, to obtain predicted click data corresponding to each first media information, for an information recommendation position in a first position in the candidate recommendation position sequence.

[0045] Based on the predicted click data corresponding to each first media information, at least two first media information are selected as at least two candidate media information corresponding to the recommendation information position in the first position.

[0046] For each information recommendation position in a position other than the first position in the candidate recommendation position sequence, the following processing is performed to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence:

[0047] In combination with at least two candidate media information corresponding to the information recommendation position in the previous position, click data prediction is performed on each second media information different from the candidate media information, to obtain predicted click data corresponding to each second media information.

[0048] Based on the predicted click data corresponding to each second media information, at least two second media information are selected as at least two candidate media information corresponding to the recommendation information position in the current position.

[0049] In the scheme, the recommendation module is further configured to perform mapping processing on the second object feature of the target object, the content feature and the position feature of each candidate media information, respectively, through a feature mapping layer of a second recommendation model, to obtain corresponding mapping features.

[0050] The sum of the mapping features corresponding to the content features and the position features of each candidate media information is encoded by a feature encoding layer of the second recommendation model, to obtain an encoded feature corresponding to each candidate media information.

[0051] The mapping feature corresponding to the second object feature and the encoded feature corresponding to each candidate media information are spliced by a feature connection layer of the second recommendation model, to obtain a spliced feature.

[0052] Based on the spliced feature, prediction is performed by a feature prediction layer of the second recommendation model, to obtain a sequence recommendation score of each media information candidate sequence.

[0053] In the scheme, the device further comprises:

[0054] The second training module is configured to obtain a sample second object feature, a sample content feature and a sample position feature corresponding to a training sample, and the training sample is labeled with a corresponding sample label.

[0055] The training sample includes at least two media information, the sample content feature includes a content feature of each media information, and the sample position feature includes a position feature of each media information.

[0056] The sample second object feature, the sample content feature, and the sample position feature are input into the second recommendation model, a prediction of a recommendation score is performed by the second recommendation model, and a corresponding prediction result is obtained.

[0057] A difference between the prediction result and a corresponding sample label is obtained.

[0058] Based on the difference, a model parameter of the first recommendation model is updated.

[0059] In the scheme, the recommendation module is further configured to perform the following processing on each media information candidate sequence to determine a sequence recommendation score of each media information candidate sequence:

[0060] A classification prediction of click data is performed by a neural network model in combination with the second object feature, the content feature of each candidate media information, and the position feature to obtain a click prediction result of at least two categories corresponding to the media information sequence.

[0061] A recommendation weight value corresponding to the click prediction result of each category is obtained.

[0062] Based on the click prediction result of the at least two categories and the recommendation weight value corresponding to the click prediction result of each category, a sequence recommendation score of the media information candidate sequence is determined.

[0063] In the scheme, the recommendation module is further configured to recommend the target media information sequence carrying display style information to a terminal corresponding to the target object.

[0064] The display style information is used to indicate a display style corresponding to each type of media information in the target media information sequence.

[0065] The application also provides an electronic device, which includes:

[0066] A memory is configured to store executable instructions.

[0067] A processor is configured to execute the executable instructions stored in the memory to implement the media information recommendation method provided in the application.

[0068] The application also provides a computer readable storage medium storing executable instructions, which are executed by a processor to implement the media information recommendation method provided in the application.

[0069] The embodiments of the present application have the following beneficial effects:

[0070] In the embodiments of the present application, when media information is recommended to a target object, first, the first object feature of the target object, the content features of the at least two media information sequences to be recommended, and the combination feature are obtained, then the recommendation scores of the media information sequences are determined based on the first object feature, the content features of the media information sequences, and the combination feature, so that the target media information sequence is selected from the at least two media information sequences based on the recommendation scores and recommended to the terminal corresponding to the target object; here, the media information sequence includes at least two types of media information, and the combination feature includes: a type preference feature for indicating the preference of the target object for the type of media information in different scenarios, and a position preference feature for indicating the preference of the target object for the display position of the media information.

[0071] In this way, when media information is recommended based on the first object feature of the target object, the content features of the at least two media information sequences to be recommended, and the combination feature, the preference of the user for the type of media information in different scenarios is considered, so that the recommended information types are diversified and more consistent with the scenario in which the user is located, the demand of the user is met, the preference of the user for the display position of the media information is considered, the user is facilitated to click and view the information of personal preference, and thus the recommendation accuracy and efficiency of the media information are improved. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 FIG. 1 is a schematic diagram of the architecture of a media information recommendation system 100 provided by the embodiments of the present application;

[0073] Figure 2 FIG. 5 is a schematic diagram of the structure of an electronic device 500 for implementing a media information recommendation method provided by the embodiments of the present application;

[0074] Figure 3 FIG. 6 is a schematic diagram of the flow of a media information recommendation method provided by the embodiments of the present application;

[0075] Figure 4 FIG. 7 is a schematic diagram of the prediction of the recommendation score of a first recommendation model provided by the embodiments of the present application;

[0076] Figure 5 FIG. 8 is a schematic diagram of the display of each type of media information provided by the embodiments of the present application;

[0077] Figure 6 FIG. 9 is a schematic diagram of the prediction of the sequence recommendation score of a second recommendation model provided by the embodiments of the present application;

[0078] Figure 7 FIG. 10 is a schematic diagram of the triggering of a media information acquisition instruction provided by the embodiments of the present application;

[0079] Figure 8 is a flowchart of a method for recommending media information provided by an embodiment of the present application.

[0080] Figure 9 is a structural diagram of a card sequence recommendation model provided by an embodiment of the present application.

[0081] Figure 10 is a structural diagram of a media information recommendation device 555 provided by an embodiment of the present application. DETAILED DESCRIPTION

[0082] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings, and the described embodiments should not be regarded as limiting the present application, and all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0083] In the following description, "some embodiments" are described, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0084] In the following description, the terms "first\second\third" are only to distinguish similar objects, and do not represent a specific order of the objects, and it can be understood that "first\second\third" can be interchanged in a specific order or sequence as allowed, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0085] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0086] The related data collection and processing in the embodiments of the present application should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and within the scope of authorization of laws and regulations and the personal information subject, carry out subsequent data use and processing.

[0087] Before the embodiments of the present application are further described in detail, the terms and terms involved in the embodiments of the present application are explained, and the terms and terms involved in the embodiments of the present application are applicable to the following explanations.

[0088] 1) Client, an application program running in a terminal for providing various services, such as an instant messaging client, a video playing client.

[0089] 2) Response to, for indicating the operation performed on the dependent conditions or state, when the dependent conditions or state is satisfied, the operation performed by one or more operations can be real-time, or with a set delay; in the absence of special instructions, there is no limit to the execution order of the operations performed.

[0090] 3) Rerank: refers to the mixed arrangement or rearrangement, which is the stage of determining the final recommendation list and its display order according to the candidate set generated by the recall and sorting stage of the recommendation system.

[0091] 4) Heterogeneous card: refers to different types of item cards (i.e. media information) in the recommendation system, which have great differences in content, display form and revenue value.

[0092] 5) Card combination: combination of two or more heterogeneous cards.

[0093] 6) Card sequence: sequence composed of sorted heterogeneous cards.

[0094] 7) Sliding behavior: gesture behavior of users viewing more content when browsing information flow on mobile phones, mainly including upslide loading behavior and downslide refreshing behavior.

[0095] 8) First screen: card area that can be directly exposed on the mobile phone when downslide refreshing.

[0096] 9) Non-first screen: card area that cannot be directly exposed on the mobile phone when downslide refreshing and all card sequences loaded by upslide.

[0097] Based on the above explanations of the terms and terms involved in the embodiments of the present application, the media information recommendation system provided by the embodiments of the present application is described below. Referring to Figure 1 , Figure 1 The architecture diagram of the media information recommendation system 100 provided by the embodiments of the present application is shown in the figure. In order to support an exemplary application, the terminal (exemplarily shown as terminal 400-1) connects the server 200 through the network 300, which can be a wide area network or a local area network, or a combination of the two, and uses wireless or wired links to realize data transmission.

[0098] The terminal (such as terminal 400-1, which can be installed with a media information client) is used to send a media information acquisition request to the server 200 in response to a media information acquisition instruction triggered by a target object.

[0099] The server 200 is configured to receive a media information acquisition request, acquire a first object feature of a target object, content features of at least two media information sequences to be recommended, and a combination feature, determine a recommendation score of each media information sequence based on the first object feature, the content features of each media information sequence, and the combination feature, select a target media information sequence from the at least two media information sequences based on the recommendation scores of the media information sequences, and recommend the target media information sequence to a terminal corresponding to the target object.

[0100] The terminal (e.g., the terminal 400-1) is configured to receive the target media information sequence recommended by the server 200, and present the target media information sequence on a graphical interface 410 (e.g., a graphical interface 410-1).

[0101] Here, the media information sequence includes at least two types of media information, and the combination feature includes a type preference feature and a position preference feature of the media information. The type preference feature is used to indicate the preference of the target object for the type of media information in different scenarios, and the position preference feature is used to indicate the preference of the target object for the display position of the media information.

[0102] In actual applications, the server 200 can be a physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and basic cloud computing services such as big data and artificial intelligence platforms. The terminal (e.g., the terminal 400-1) can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart television, a smart watch, etc., but is not limited thereto. The terminal (e.g., the terminal 400-1) and the server 200 can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application.

[0103] Referring to Figure 2 , Figure 2 FIG. 5 is a structural schematic diagram of an electronic device 500 for implementing a media information recommendation method according to an embodiment of the present application. In actual applications, the electronic device 500 can be a server or a terminal shown in Figure 1 FIG. 5, and the electronic device 500 can be a server or a terminal shown in Figure 1The terminal is shown as an example to describe the electronic device implementing the media information recommendation method of the embodiments of the present application. The electronic device 500 provided by the embodiments of the present application includes at least one processor 510, a memory 550, at least one network interface 520 and a user interface 530. The various components in the electronic device 500 are coupled together through a bus system 540. It can be understood that the bus system 540 is used to realize the connection communication between the components. In addition to including a data bus, the bus system 540 also includes a power bus, a control bus and a status signal bus. However, in order to clearly illustrate the present application, all the buses in the figure are marked as the bus system 540. Figure 2

[0104] The processor 510 can be an integrated circuit chip having a processing capability of a signal, for example, a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor.

[0105] The user interface 530 includes one or more output devices 531 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 530 also includes one or more input devices 532 that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0106] The memory 550 can be removable, non-removable or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 550 optionally includes one or more storage devices physically located in proximity to the processor 510.

[0107] The memory 550 includes volatile memory or non-volatile memory, and can also include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), and the volatile memory can be random access memory (RAM). The memory 550 described in the embodiments of the present application is intended to include any suitable type of memory.

[0108] In some embodiments, the memory 550 is capable of storing data to support various operations, examples of which include programs, modules and data structures or subsets or supersets thereof, which are exemplarily described below.

[0109] ​The operating system 551 includes system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, and the like, for implementing various basic services and processing hardware-based tasks.

[0110] The network communication module 552 is configured to reach other computing devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 including Bluetooth, wireless fidelity (WiFi), and universal serial bus (USB), and the like.

[0111] The presentation module 553 is configured to enable presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 531 (e.g., a display screen, a speaker, and the like) associated with the user interface 530.

[0112] The input processing module 554 is configured to detect and interpret one or more user inputs or interactions from one or more input devices 532.

[0113] In some embodiments, the media information recommendation apparatus provided by the embodiments of the present application can be implemented in a software manner, Figure 2 A media information recommendation apparatus 555 stored in the memory 550 is shown, which can be software in the form of programs and plug-ins, etc., including the following software modules: an acquisition module 5551, a determination module 5552, and a recommendation module 5553, which are logical, and thus can be combined or further split according to the implemented functions, and the functions of the various modules will be described below.

[0114] In some embodiments, the media information recommendation apparatus provided by the embodiments of the present application can be implemented in a software manner,

[0115] Based on the above description of the media information recommendation system and the electronic device provided by the embodiments of the present application, the media information recommendation method provided by the embodiments of the present application is described below. In some embodiments, the media information recommendation method provided by the embodiments of the present application can be implemented by a server or a terminal alone, or by a server and a terminal cooperatively. The media information recommendation method provided by the embodiments of the present application is described below by taking the implementation of the server as an example. Referring to Figure 3 , Figure 3 FIG. 1 is a flowchart of the media information recommendation method provided by the embodiments of the present application. The media information recommendation method provided by the embodiments of the present application includes the following steps.

[0116] Step 101: The server acquires the first object feature of a target object, the content feature of at least two media information sequences to be recommended, and the combination feature.

[0117] The media information sequence includes at least two types of media information, and the combination feature includes a type preference feature and a position preference feature of the media information. The type preference feature is used to indicate the preference of the target object for the type of media information in different scenarios, and the position preference feature is used to indicate the preference of the target object for the display position of the media information.

[0118] Herein, in actual application, the server for media information recommendation is a background server. The terminal held by the target object can be provided with a client for browsing and watching media information, such as a news client, an instant messaging client configured with a media information browsing page, a browser client, etc. The media information can be picture information, text information, picture-text combined information, video information, etc.

[0119] When the target object browses the media information through the client, a media information acquisition instruction can be triggered, such as by executing a page pull-to-refresh operation to trigger the media information acquisition instruction. When the terminal receives the media information acquisition instruction triggered by the target object, the terminal sends a media information acquisition request to the background server of the client. The server receives the media information acquisition request sent by the terminal, and in response to the media information acquisition request, sends the media information to be recommended to the terminal.

[0120] In the embodiments of the present application, when the server performs media information recommendation, the media information is recommended in the form of a media information sequence to improve the recommendation efficiency of the media information, i.e., a media information sequence including at least two media information is recommended to the target object. Herein, the media information sequence includes at least two types of media information, and the types of media information can include text type, picture type, short video type, video type, etc., or can include combined types of the above types, such as picture-text type, etc.

[0121] Here, when the server performs media information recommendation for a target object, the server can first acquire a first object feature of the target object, content features of at least two media information sequences to be recommended, and a combination feature. The first object feature of the target object can be a user basic feature, such as a user name, age, gender, identification, and the like. The content features can be content features of media information contained in the media information sequences, such as text content, picture content, video content, and the like. The combination feature of the media information sequences includes a type preference feature and a position preference feature of the media information. Specifically, the type preference feature is used to indicate a preference of the target object for types of media information contained in the media information sequences (i.e., a preference for combinations of various types, such as a combination of <text-picture-video-text>, a combination of <video-text-picture>, and the like) in different scenarios (such as different time periods, different network states, different location information, and the like). The position preference feature is used to indicate a preference of the target object for a display position of the media information, i.e., a display position preferred to be clicked by the target object, such as a central position of a terminal screen.

[0122] In some embodiments, the server can acquire the first object feature of the target object, the content features of the at least two media information sequences to be recommended, and the combination feature in the following manner: acquiring a first group feature of the target object, and second group features of at least two pre-set object groups, each object group including at least two objects; performing feature matching between the first group feature of the target object and the second group features of the object groups respectively to determine matching degrees of the first group feature and the second group features, and taking an object group corresponding to a second group feature having the highest matching degree with the first group feature as a target object group; acquiring the content features of the at least two media information sequences to be recommended, and acquiring a group first object feature of the target object group and a group combination feature of the target object group corresponding to the at least two media information sequences; taking the group first object feature as the first object feature of the target object, and taking the group combination feature as the combination feature of the target object corresponding to the at least two media information sequences.

[0123] Here, in the acquisition of the first object feature of the target object, the content feature of the at least two media information sequences to be recommended, and the combination feature, the group first object feature and the group combination feature of the target object group to which the target object belongs can be acquired. In actual application, each object group can include at least two objects, which are obtained by dividing a plurality of objects according to a pre-set object group division standard. For example, the object group division standard can be age division (for example, 10-20 years old as an object group, 20-30 years old as an object group, etc.), gender division (for example, women as an object group, men as an object group), occupation division (students as an object group, program developers as an object group, product designers as an object group, etc.).

[0124] Specifically, first, the first group feature of the target object and the second group feature of the at least two object groups are acquired, where the first group feature and the second group feature can include age features, gender features, activity features in a target time period, etc. Second, the first group feature of the target object is matched with the second group feature of each object group to determine the matching degree of the first group feature and each second group feature, and the object group corresponding to the second group feature with the highest matching degree of the first group feature is taken as the target object group. Then, the content feature of the at least two media information sequences to be recommended is acquired, and the group first object feature of the target object group and the group combination feature of the target object group corresponding to the at least two media information sequences are acquired. Finally, the group first object feature is taken as the first object feature of the target object, and the group combination feature is taken as the combination feature of the target object corresponding to the at least two media information sequences.

[0125] In actual application, the acquisition method of the first object feature of the target object, the content feature of the at least two media information sequences to be recommended, and the combination feature can be used when the target object is a new user or when the first object feature of the target object individual and the combination feature of the target object individual corresponding to the at least two media information sequences fail to be acquired.

[0126] In addition, when the target object is a real-time user to be recommended, the first object feature of the target object and the combination feature of the target object individual corresponding to the at least two media information sequences can be directly acquired, or can be acquired by acquiring the group first object feature and the group combination feature of the target object group to which the target object belongs. When the target object is a user group to be recommended, the group first object feature and the group combination feature of the user group are directly acquired.

[0127] Step 102: Based on the first object feature, the content feature of each media information sequence, and the combination feature, the recommendation score of each media information sequence is determined.

[0128] Here, after obtaining the first object feature of the target object, the content features of the at least two media information sequences to be recommended, and the combination feature, the recommendation scores of the media information sequences can be determined based on the first object feature, the content features of the media information sequences, and the combination feature. The recommendation scores can be used to indicate the corresponding click data when the corresponding media information sequence is recommended to the terminal of the target object as the target media information sequence. In actual applications, the click data can be the click probability (or click possibility) of the target object for the target media information sequence, data representing whether the target media information sequence is clicked by the target object, data representing the high or low of the click probability (or click possibility) of the target object for the target media information sequence, and the like.

[0129] Here, in the embodiments of the present application, the calculation of the recommendation score can be implemented by a machine learning model (such as the first recommendation model). In some embodiments, the server can obtain the sample first object feature, the sample content feature, and the sample combination feature corresponding to the training sample, and the training sample is labeled with a corresponding sample label. The sample first object feature, the sample content feature, and the sample combination feature are input into the first recommendation model, and the prediction of the recommendation score is performed by the first recommendation model to obtain the corresponding prediction result. The difference between the prediction result and the corresponding sample label is obtained. Based on the difference, the model parameters of the first recommendation model are updated.

[0130] The training sample includes at least two types of media information samples, and the sample combination feature includes a sample type preference feature and a sample position preference feature of the media information sample. The sample type preference feature is used to indicate the preference of the corresponding sample user for the type of the corresponding media information sample in different scenarios, and the sample position preference feature is used to indicate the preference of the corresponding sample user for the display position of the corresponding media information sample.

[0131] Here, in the embodiments of the present application, the calculation of the recommendation score can be implemented by a machine learning model. In actual implementation, the first recommendation model is constructed by a machine learning network (such as a convolutional neural network CNN, a deep neural network DNN, or the like), and initial model parameters, an activation function (such as a Sigmoid function, a linear rectification function, or the like) of the model, and a loss function (such as a cross-entropy loss function, a logarithmic loss function, or the like) of the model are set for the first recommendation model. After the initial first recommendation model is constructed, the constructed first recommendation model is trained.

[0132] Specifically, first, a training sample for training the first recommendation model is obtained, the training sample being annotated with a corresponding sample label and including at least two types of media information samples. Further, a sample first object feature, a sample content feature, and a sample combination feature corresponding to the training sample are obtained, wherein the sample combination feature includes a sample type preference feature and a sample position preference feature of the media information sample; the sample type preference feature is used to indicate a preference of a corresponding sample user for a type of the corresponding media information sample in different scenarios, and the sample position preference feature is used to indicate a preference of the corresponding sample user for a display position of the corresponding media information sample.

[0133] Then, the sample first object feature, the sample content feature, and the sample combination feature of the training sample are input into the first recommendation model, a prediction of a recommendation score is performed by the first recommendation model in combination with the sample first object feature, the sample content feature, and the sample combination feature, and a corresponding prediction result is obtained; finally, a difference between the prediction result and the sample label of the corresponding training sample is obtained, and model parameters of the first recommendation model are updated based on the difference. Specifically, a value of a loss function of the first recommendation model can be determined based on the difference, when the value of the loss function exceeds a preset threshold, an error signal of the recommendation model is determined based on the loss function, the error signal is back propagated in the first recommendation model, and model parameters of each layer are updated in the process of propagation. The above steps are iterated continuously in the whole training process until the loss function converges, and the model parameters of the first recommendation model updated when the loss function converges are used as the model parameters of the first recommendation model finally trained to reduce the error of the model output.

[0134] Based on this, after the first recommendation model is trained, the server can obtain the first object feature of the target object, the content feature and the combination feature of the at least two media information sequences to be recommended in the following manner: obtaining user data of the target object, content data of the at least two media information sequences to be recommended, and combination data; wherein the combination data includes type preference data and position preference data of the media information; the feature extraction layer of the first recommendation model is used to perform feature extraction on the user data, the content data of the at least two media information sequences to be recommended, and the combination data respectively, and the first object feature of the target object, the content feature and the combination feature of the at least two media information sequences to be recommended are obtained.

[0135] Correspondingly, the server can determine the recommendation score of each media information sequence based on the first object feature, the content feature and the combination feature of each media information sequence in the following manner: the feature prediction layer of the first recommendation model is used to perform prediction in combination with the first object feature, the content feature and the combination feature of each media information sequence, and the recommendation score of each media information sequence is obtained.

[0136] In actual application, the first recommendation model can be a regression model or a classification model. When predicting by the feature prediction layer of the first recommendation model, combining the first object feature, the content features of the media information sequences, and the combined features to obtain the recommendation scores of the media information sequences, the regression model can be used to combine the first object feature, the content features of the media information sequences, and the combined features to obtain the recommendation scores of the media information sequences, which can be the click probability (or click possibility) of the target object to the target media information sequence, or data representing the click probability (or click possibility) of the target object to the target media information sequence.

[0137] The classification model can also be used to combine the first object feature, the content features of the media information sequences, and the combined features to obtain the recommendation scores of the media information sequences, which can be data representing whether the target media information sequence is clicked by the target object, such as 1 representing being clicked and 0 representing not being clicked.

[0138] Referring to Figure 4 , Figure 4 is a prediction diagram of the recommendation score of the first recommendation model provided in the embodiments of the present application. Here, the first recommendation model includes a feature extraction layer and a feature prediction layer. After obtaining the user data of the target object, the content data of the at least two media information sequences to be recommended, and the combined data, the feature extraction layer of the first recommendation model is used to extract features from the user data, the content data of the at least two media information sequences to be recommended, and the combined data, respectively, to obtain the first object feature of the target object, the content features of the at least two media information sequences to be recommended, and the combined features. Then, the feature prediction layer of the first recommendation model is used to combine the first object feature, the content features of the media information sequences, and the combined features to obtain the recommendation scores of the media information sequences.

[0139] In some embodiments, the server can also determine the recommendation scores of the media information sequences based on the first object feature, the content features of the media information sequences, and the combined features in the following manner: the following processing is performed for each media information sequence to determine the recommendation score of the media information sequence: a neural network model is used to combine the first object feature, the content features of the media information sequence, and the combined features to perform classification prediction of the click data, to obtain the click prediction results of at least two categories corresponding to the media information sequence; the recommendation weight values corresponding to the click prediction results of the categories are obtained; and the recommendation score of the media information sequence is determined based on the click prediction results of the at least two categories and the recommendation weight values corresponding to the click prediction results of the categories.

[0140] Here, the neural network model can also be obtained by the above-mentioned construction and training of the first recommendation model, which is not repeated in the embodiments of the present application. In actual application, the neural network model can be a classification prediction model. When calculating the recommendation score of each media information sequence, the following processing is performed for each media information candidate sequence: the first object feature, the content feature of the media information sequence, and the combined feature can be input into the neural network model, and the neural network model combines the first object feature, the content feature of the media information sequence, and the combined feature to perform classification prediction of the click data, to obtain the click prediction result of at least two categories corresponding to the media information sequence; the recommendation weight value corresponding to each category can be set in advance, and after obtaining the click prediction result of at least two categories corresponding to the media information sequence, the recommendation weight value corresponding to the click prediction result of each category is obtained; and then, based on the click prediction result of at least two categories and the recommendation weight value corresponding to the click prediction result of each category, the recommendation score of the media information sequence is determined.

[0141] Exemplarily, the categories of the classification prediction corresponding to the above-mentioned neural network model can include three categories of no click, click once, and click two or more times; the neural network model predicts the click prediction result of at least two categories corresponding to the media information sequence: P(click once), P(click >=2 times), and P(no click); and then the recommendation weight value corresponding to the click prediction result of each category is obtained, such as the recommendation weight value 1 corresponding to P(click once), the recommendation weight value 2 corresponding to P(click >=2 times), and the recommendation weight value 0 corresponding to P(no click). Thus, based on the click prediction result of at least two categories and the recommendation weight value corresponding to the click prediction result of each category, the recommendation score of the media information sequence is determined as score=0*P(no click)+1*P(click once)+2*P(click >=2 times).

[0142] Step 103: selecting a target media information sequence from the at least two media information sequences based on the recommendation score of each media information sequence, and recommending the target media information sequence to the terminal corresponding to the target object.

[0143] Here, after determining the recommendation score of each media information sequence, the target media information sequence is selected from the at least two media information sequences based on the recommendation score. Specifically, the media information sequence with the highest recommendation score can be selected as the target media information sequence, so as to recommend the target media information sequence to the terminal corresponding to the target object.

[0144] In some embodiments, the server can recommend the target media information sequence to the terminal corresponding to the target object by: recommending the target media information sequence carrying display style information to the terminal corresponding to the target object; wherein the display style information is used to indicate the display style corresponding to each type of media information in the target media information sequence.

[0145] In actual application, when recommending the target media information sequence to the terminal corresponding to the target object, the server can carry display style information in the target media information sequence, and the display style information is used to indicate the display style corresponding to each type of media information in the target media information sequence, such as displaying in the form of media information card, displaying in the form of small card for text type media information, displaying in the form of medium card for graphic text type media information, and displaying in the form of large card for video type media information. The height of the large card is higher than the height of the medium card, and the height of the medium card is higher than the height of the small card. Referring to Figure 5 , Figure 5 is a display schematic diagram of each type of media information provided by the embodiments of the present application. Here, the video type media information A is displayed in the form of a large card, and the graphic text type media information B is displayed in the form of a medium card, that is, the height of the display card corresponding to the media information A is higher than the height of the display card corresponding to the media information B.

[0146] In some embodiments, the server can recommend the target media information sequence to the terminal corresponding to the target object by: receiving the media information sequence acquisition request sent by the terminal, the acquisition request being triggered in response to the media information acquisition instruction for the first screen recommendation position; and sending the target media information sequence to the terminal to display the target media information sequence through the first screen recommendation position of the terminal.

[0147] In actual application, when browsing media information, the user can only consume the media information displayed in the first screen recommendation position, such as when the user performs the pull-to-refresh operation on the media information stream page to update the media information displayed in the first screen recommendation position. After the user only browses the media information in the first screen recommendation position, the user can directly perform the pull-to-refresh operation on the media information stream page again to update the media information displayed in the first screen recommendation position. Therefore, the media information recommendation in the first screen recommendation position is important. Considering that the display positions of different types of media information are different in size and the information recommendation position of the first screen is limited, in order to balance the exposure efficiency of each type of media information, a media information sequence containing at least two types of media information can be used for the recommendation of the first screen media information. Specifically, the recommendation score of each media information sequence to be recommended is determined by the above-mentioned embodiments, and the target media information sequence is selected for recommendation based on the recommendation score.

[0148] When the target object is browsing the media information through the client, the media information acquisition instruction for the first-screen recommendation position can be triggered, such as by performing a page pull-down refresh operation to trigger the media information acquisition instruction for the first-screen recommendation position. When the terminal receives the media information acquisition instruction for the first-screen recommendation position triggered by the target object, the terminal sends a media information sequence acquisition request to the background server of the client. When the server receives the media information sequence acquisition request sent by the terminal, the target media information sequence is selected from the at least two media information sequences to be recommended based on the recommendation scores of the media information sequences in the above embodiment, and the target media information sequence is sent to the terminal of the target object to display the target media information sequence through the first-screen recommendation position of the terminal.

[0149] In some embodiments, the recommended target media information sequence is displayed through the first-screen recommendation position of the terminal, and the server can recommend media information for information recommendation positions that are not in the first screen in the following manner: for each information recommendation position in a candidate recommendation position sequence, at least two candidate media information are acquired to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence, wherein the candidate recommendation position sequence is not in the first screen and includes at least two information recommendation positions; the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence are acquired; based on the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence, the sequence recommendation score of each media information candidate sequence is determined; based on the sequence recommendation scores of the media information candidate sequences, a target media information candidate sequence is selected from the at least two media information candidate sequences, and the target media information candidate sequence is recommended to the terminal corresponding to the target object to display the target media information candidate sequence through the candidate recommendation position sequence that is not in the first screen.

[0150] After the recommendation method of the media information of the first-screen recommendation position is described, the recommendation method of the media information of the information recommendation position that is not in the first screen is described. In actual applications, the information recommendation position that is not in the first screen is the information recommendation position displayed by performing a pull-up loading operation on the media information stream page; or the information recommendation position that is not in the first screen can also be the information recommendation position before the first-screen recommendation position refreshed by performing a pull-down refresh operation on the media information stream page. The plurality of information recommendation positions that are not in the first screen form a candidate recommendation position sequence.

[0151] For the candidate recommendation position sequence in the non-first screen and including at least two information recommendation positions, the recommendation of media information can be performed in the following manner: first, for each information recommendation position in the candidate recommendation position sequence, at least two candidate media information are obtained to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence. In some embodiments, the server can obtain at least two candidate media information for each information recommendation position in the candidate recommendation position sequence to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence in the following manner:

[0152] For the information recommendation position in the first position of the candidate recommendation position sequence, the click data prediction is performed on each first media information to be recommended to obtain the predicted click data corresponding to each first media information; based on the predicted click data corresponding to each first media information, at least two first media information are selected as at least two candidate media information corresponding to the recommendation information position in the first position.

[0153] For each information recommendation position in the non-first position of the candidate recommendation position sequence, the following processing is performed to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence: in combination with at least two candidate media information corresponding to the information recommendation position in the previous position of the current position, the click data prediction is performed on each second media information different from the candidate media information to obtain the predicted click data corresponding to each second media information; based on the predicted click data corresponding to each second media information, at least two second media information are selected as at least two candidate media information corresponding to the recommendation information position in the current position.

[0154] In actual application, for the information recommendation position in the first position of the candidate recommendation position sequence, the click data of each first media information is calculated, specifically, the predicted click data of each first media information can be predicted through a click rate prediction model (i.e. CTR model). Then, based on the predicted click data of each first media information, at least two first media information (such as at least two first media information with the predicted click data in descending order) are selected as at least two candidate media information corresponding to the recommendation information position in the first position.

[0155] For each information recommendation position other than the first position in the candidate recommendation position sequence, the following processing can be performed to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence: in combination with at least two candidate media information corresponding to the information recommendation position of the previous position of the current position, click data prediction is performed on each second media information different from the candidate media information to obtain predicted click data corresponding to each second media information. Here, in actual application, click data prediction is performed on each second media information different from the candidate media information in combination with each candidate media information corresponding to the information recommendation position of the previous position of the current position to obtain predicted click data corresponding to each second media information. Thus, based on the predicted click data corresponding to each second media information, at least two second media information (such as at least two second media information ranked in descending order of predicted click data) are selected as at least two candidate media information corresponding to the recommendation information position of the current position.

[0156] Secondly, the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence are obtained. Here, the second object feature of the target object can include user basic features (such as name, age, gender, etc.), and can also include type preference features of the user (i.e., the user's preference for the type of media information in different scenarios (such as different time periods, network states, positions, etc.)) and the like. The content feature of each candidate media information included in each media information candidate sequence can include media information text content, picture content, video content and the like. The position feature of each candidate media information included in each media information candidate sequence can be the position of each candidate media information in the media information candidate sequence, such as the first position in the media information candidate sequence, the second position in the media information candidate sequence and the like.

[0157] Then, based on the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence, the sequence recommendation score of each media information candidate sequence is determined. The sequence recommendation score can be used to indicate the corresponding click data when the corresponding media information candidate sequence is recommended to the terminal of the target object as the target media information candidate sequence. In actual application, the click data can be the click probability (or click possibility) of the target object for the target media information candidate sequence, data representing whether the target media information candidate sequence is clicked by the target object, data representing the high or low of the click probability (or click possibility) of the target object for the target media information candidate sequence, and the like.

[0158] Finally, based on the sequence recommendation scores of the candidate sequences of media information, a target candidate sequence of media information is selected from the at least two candidate sequences of media information, such as the candidate sequence of media information with the highest sequence recommendation score is selected as the target candidate sequence of media information. The target candidate sequence of media information is recommended to the terminal corresponding to the target object, so as to display the target candidate sequence of media information in the candidate recommendation position sequence that is not the first screen.

[0159] Here, in the embodiments of the present application, the calculation of the sequence recommendation score can be implemented by a machine learning model (such as the second recommendation model). In some embodiments, the server can obtain the second recommendation model by training as follows: obtaining sample second object features, sample content features, and sample position features corresponding to a training sample, the training sample being labeled with a corresponding sample label; inputting the sample second object features, the sample content features, and the sample position features into the second recommendation model, predicting the recommendation score by the second recommendation model to obtain a corresponding prediction result; obtaining the difference between the prediction result and the corresponding sample label; based on the difference, updating the model parameters of the second recommendation model. The training sample includes at least two media information, the sample content features include content features of each media information, and the sample position features include position features of each media information.

[0160] Here, in the embodiments of the present application, the calculation of the sequence recommendation score can be implemented by a machine learning model. In actual implementation, the second recommendation model is constructed by a machine learning network (such as a convolutional neural network CNN, a deep neural network DNN, etc.), and initial model parameters, an activation function (such as a Sigmoid function, a linear rectifier function, etc.) of the model, and a loss function (such as a cross-entropy loss function, a logarithmic loss function, etc.) of the model are set for the second recommendation model. After the initial second recommendation model is constructed, the constructed second recommendation model is trained.

[0161] Specifically, first, a training sample for training the second recommendation model is obtained, the training sample being labeled with a corresponding sample label. Then, sample second object features, sample content features, and sample position features corresponding to the training sample are obtained, wherein the training sample includes at least two media information, the sample content features include content features of each media information, and the sample position features include position features of each media information.

[0162] Then, the sample second object feature, the sample content feature and the sample location feature corresponding to the training sample are input into the second recommendation model, the second recommendation model is used to combine the sample second object feature, the sample content feature and the sample location feature to predict the sequence recommendation score, and a corresponding prediction result is obtained; finally, the difference between the prediction result and the sample label of the corresponding training sample is obtained, and the model parameters of the second recommendation model are updated based on the difference. Specifically, the value of the loss function of the second recommendation model can be determined based on the difference, when the value of the loss function exceeds the preset threshold, the error signal of the recommendation model is determined based on the loss function, the error signal is back propagated in the second recommendation model, and the model parameters of each layer are updated during the propagation. The above steps are iterated continuously in the whole training process until the loss function converges. The updated model parameters of the second recommendation model when the loss function converges are used as the model parameters of the finally trained second recommendation model to reduce the error of the model output.

[0163] Based on this, after the training of the second recommendation model is completed, the server can determine the sequence recommendation score of each media information candidate sequence based on the second object feature of the target object, the content feature and the location feature of each candidate media information included in each media information candidate sequence by the following method: the second object feature of the target object, the content feature and the location feature of each candidate media information are respectively mapped by the feature mapping layer of the second recommendation model to obtain corresponding mapping features; the sum of the mapping features corresponding to the content features and the mapping features corresponding to the location features of each candidate media information are encoded by the feature encoding layer of the second recommendation model to obtain the encoding features corresponding to each candidate media information; the mapping features corresponding to the second object feature and the encoding features corresponding to each candidate media information are spliced by the feature connection layer of the second recommendation model to obtain corresponding splicing features; and the splicing features are predicted by the feature prediction layer of the second recommendation model to obtain the sequence recommendation score of each media information candidate sequence.

[0164] Here, referring to Figure 6 , Figure 6 is a schematic diagram of the prediction of the sequence recommendation score of the second recommendation model provided by the embodiments of the present application. Here, the second recommendation model includes a feature mapping layer, a feature encoding layer and a feature prediction layer. In actual application, the second recommendation model can be constructed based on a Transformer network model.

[0165] Thus, after obtaining the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence, first, the second object feature, the content feature and the position feature of each candidate media information are respectively mapped by the feature mapping layer of the second recommendation model to obtain corresponding mapping features, so as to process discrete features into continuous features; second, the sum of the mapping features corresponding to the content features of each candidate media information and the mapping features corresponding to the position features is encoded by the feature encoding layer of the second recommendation model to obtain the encoding features corresponding to each candidate media information; then, the mapping features corresponding to the second object feature and the encoding features corresponding to each candidate media information are spliced by the feature connection layer of the second recommendation model to obtain corresponding spliced features; finally, the spliced features are predicted by the feature prediction layer of the second recommendation model to obtain the sequence recommendation scores of each media information candidate sequence.

[0166] Specifically, the second recommendation model can be a regression model or a classification model. When the sequence recommendation scores of each media information candidate sequence are obtained by predicting the spliced features based on the feature prediction layer of the second recommendation model, the regression prediction of the spliced features can be performed by the regression model to obtain the sequence recommendation scores of each media information candidate sequence. The sequence recommendation scores can be the click probability (or click possibility) of the target object for the target media information candidate sequence, or data representing the high or low click probability (or click possibility) of the target object for the target media information candidate sequence, etc.

[0167] The prediction of the spliced features based on the classification model can also be performed to obtain the sequence recommendation scores of each media information candidate sequence. The recommendation scores can be data representing whether the target media information candidate sequence is clicked by the target object, such as 1 representing being clicked and 0 representing not being clicked.

[0168] In some embodiments, the server can determine the sequence recommendation scores of each media information candidate sequence based on the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence by the following manner: the following processing is performed for each media information candidate sequence to determine the sequence recommendation scores of each media information candidate sequence: the classification prediction of the click data is performed by the neural network model in combination with the second object feature, the content feature and the position feature of each candidate media information to obtain the click prediction results of at least two categories corresponding to the media information sequence; the recommendation weight values corresponding to the click prediction results of each category are obtained; and the sequence recommendation scores of the media information candidate sequence are determined based on the click prediction results of the at least two categories and the recommendation weight values corresponding to the click prediction results of each category.

[0169] Here, the neural network model can also be obtained by the above-mentioned construction and training manner of the second recommendation model, which is not repeated in the embodiments of the present application. In actual application, the neural network model can be a classification prediction model, and when calculating the sequence recommendation score of each media information candidate sequence, the following processing is performed for each media information candidate sequence: the second object feature, the content feature of each candidate media information, and the position feature can be input into the neural network model, and the classification prediction of the click data is performed by the neural network model combined with the second object feature, the content feature of each candidate media information, and the position feature to obtain the click prediction result of at least two categories corresponding to the media information candidate sequence; the recommendation weight value corresponding to each category of click prediction result can be obtained after obtaining the click prediction result of at least two categories corresponding to the media information candidate sequence; and thus the sequence recommendation score of the media information candidate sequence is determined based on the click prediction result of at least two categories and the recommendation weight value corresponding to each category of click prediction result.

[0170] Exemplarily, the categories of the classification prediction corresponding to the above-mentioned neural network model can include three categories of no click, click once, click twice, and click three times or more; the click prediction result of at least two categories corresponding to the media information sequence is obtained by the neural network model: P(click once), P(click >=2 times), P(click >=3 times), P(no click); and then the recommendation weight value corresponding to each category of click prediction result is obtained, such as the recommendation weight value 1 corresponding to P(click once), the recommendation weight value 2 corresponding to P(click >=2 times), the recommendation weight value 3 corresponding to P(click >=3 times), and the recommendation weight value 0 corresponding to P(no click). Thus, the recommendation score of the media information sequence is determined based on the click prediction result of at least two categories and the recommendation weight value corresponding to each category of click prediction result: score=0*P(no click)+1*P(click once)+2*P(click >=2 times)+3*P(click >=3 times).

[0171] By applying the above-mentioned embodiments of the present application, in the embodiments of the present application, when the media information is recommended to the target object, the first object feature of the target object, the content feature of at least two media information sequences to be recommended, and the combination feature are obtained first, and then the recommendation score of each media information sequence is determined based on the first object feature, the content feature of each media information sequence, and the combination feature, so that the target media information sequence is selected from the at least two media information sequences based on the recommendation score and recommended to the terminal corresponding to the target object; here, the media information sequence includes at least two types of media information, and the combination feature includes: a type preference feature for indicating the preference of the target object for the type of media information in different scenarios, and a position preference feature for indicating the preference of the target object for the display position of the media information.

[0172] Thus, when the media information is recommended based on the first object feature of the target object, the content feature of the at least two media information sequences to be recommended, and the combined feature, the preference of the user for the media information type in different scenarios is considered, the recommended information type is diversified and more consistent with the scenario in which the user is located, the demand of the user is met, the preference of the user for the media information display position is considered, the user is facilitated to click and view the information of personal preference, and therefore the recommendation accuracy and the recommendation efficiency of the media information are improved.

[0173] An exemplary application of the embodiment of the present application in an actual application scenario will be described below. Here, the media information exists in the form of an information card to be recommended to the terminal of the user, and therefore, in the embodiment of the present application, the media information card is the media information.

[0174] As the last stage of the recommendation system, the mixing arrangement is used to determine the recommended list and the display order thereof to be finally displayed to the user, and its potential has gradually been valued in recent years. In the embodiment of the present application, the recommendation of the media information can be implemented based on the point-wise model mixing arrangement, the sequence model mixing arrangement, and the reordering model mixing arrangement. Among them, the point-wise model mixing arrangement scheme is to use the first object feature, the media information feature, the context feature, and the like to construct a point-wise mixing arrangement model, and to predict the optimal media information according to the recommended position one by one until the length of the media information sequence reaches the maximum slot. The sequence model mixing arrangement scheme is to use the point-wise mixing arrangement model to predict according to the position or to predict all the media information uniformly, then to generate a plurality of media information sequence candidate sets according to the model prediction value and the strategy, and further to predict the optimal media information sequence result through the sequence model. The reordering model mixing arrangement scheme is to fine-tune the sorting result given by the sorting stage by modeling the mutual influence between the items, not only to optimize the loss function as much as possible, but also to explicitly model the mutual influence between the items from the feature space, and to reorder the sequence result once according to the prediction value of the model for each position card.

[0175] When the user uses the media information stream product on the terminal (such as a mobile phone terminal), because of the limitation of the screen size and the height difference of the heterogeneous media information cards, only a small number of media information cards can be displayed on each screen, and therefore the user mainly has two sliding behaviors to view more content, such as Figure 7 as shown in Figure 7is a trigger schematic diagram of a media information acquisition instruction provided by an embodiment of the present application. The left drawing is a recommended example of refreshing a heterogeneous card (a picture-text card, a short video card, and a small video card) on a first screen through a pull-down operation (i.e., the media information acquisition instruction is triggered through a pull-down refresh), and the right drawing is a recommended example of loading a non-first-screen card through a slide-up operation (i.e., the media information acquisition instruction is triggered through a slide-up operation). (1) Pull-down refresh: the user only consumes the media information card exposed on the first screen; (2) Slide-up loading: the user keeps sliding up to expose more content.

[0176] From Figure 7 It can be seen that the heights of different kinds of heterogeneous cards are quite different. For example, the height of a short video card is nearly twice that of a picture-text card, that is, the exposure efficiency of a picture-text card is twice that of a short video card, and about 40% of users only consume the media information card on the first screen, so the first-screen recommendation efficiency is extremely important. Therefore, while considering the first-screen recommendation efficiency, a sorting strategy must be designed for the exposure efficiency of different heterogeneous media information cards to balance the exposure efficiency of each type of media information card. Whether a user clicks a certain media information card is not only related to the media information card in front of the certain media information card, but also related to the media information card exposed behind the certain media information card under different sliding behaviors.

[0177] The applicant found in the implementation process that the above technical solution does not consider that the exposure efficiency of heterogeneous cards will affect the more important first-screen recommendation efficiency; and does not consider that whether a user clicks a certain media information card is not only related to the media information card in front of the certain media information card, but also related to the media information card exposed behind the certain media information card under different sliding behaviors, that is, does not consider that the different consumption tendency information of a user implied by different sliding behaviors of the user will affect the recommendation effect of media information.

[0178] Based on this, an embodiment of the present application further provides a media information recommendation method, specifically a media information card combination and media information card sequence recommendation method based on user sliding behavior and considering the exposure efficiency of heterogeneous cards, comprising:

[0179] First, the media information on the first screen is recommended in the form of a heterogeneous media information card combination (i.e., the media information sequence involved in the above embodiment). When the user performs a pull-down refresh, a card combination recommendation model (i.e., the first recommendation model involved in the above embodiment) is used to predict the recommendation scores (such as CTR) of various media information card combinations contained in the candidate set. In addition to considering the inherent property features of the user, the request context features, and the preference features of the user for the content of the media information card, the card combination recommendation model also inputs the preference features of the user for the media information card combination (i.e., the type preference features involved in the above embodiment) and the preference features of the user for the display position of the media information card (i.e., the position preference features involved in the above embodiment).

[0180] wherein the type preference feature characterizes the preference degree of the user to the combination of various heterogeneous media information cards in different time periods and network states; and the position preference feature characterizes the preference degree of the user to various heterogeneous media information cards in different positions, time periods and network states.

[0181] Second, for the media information other than the first screen, a heterogeneous media information card sequence (i.e., the media information candidate sequence involved in the above embodiment) is used for recommendation. The media information other than the first screen obtained by the user through the up-slip refresh and the down-pull is used to select the top N (N>=2) candidate media information cards in each position by using the media information CTR model combined with the bundle search strategy according to the display position of the media information (i.e., the information recommendation position involved in the above embodiment), so as to obtain a set of multiple media information card candidate sequences, and finally the optimal media information card candidate sequence is selected by using the card sequence recommendation model (i.e., the second recommendation model involved in the above embodiment). Here, when the media information CTR model predicts the candidate media information card in the current position, the features of the media information card displayed in the previous position are considered, so that the information that "whether the user clicks a certain media information card is related to the media information card before the media information card" is considered.

[0182] Referring to Figure 8 , Figure 8 is a flowchart of a media information recommendation method provided by the embodiment of the present application. The method comprises the following steps:

[0183] Step 201: Start.

[0184] Step 202: Determine whether a down-pull refresh instruction is received, if yes, execute step 203, and if no, execute step 204.

[0185] Here, the down-pull refresh instruction is used to obtain the media information corresponding to the first screen recommendation position.

[0186] Step 203: Obtain the media information card combination corresponding to the first screen recommendation position in the information recommendation sequence by using the card combination recommendation model.

[0187] Here, the information recommendation sequence is a plurality of media information corresponding to each down-pull refresh or up-slip load, such as 10 media information corresponding to each down-pull refresh or up-slip load.

[0188] Specifically, the recommendation score of each media information card combination is predicted by using the card combination recommendation model, so that the media information card combination with the highest recommendation score is selected as the media information card combination corresponding to the first screen recommendation position in the information recommendation sequence based on the recommendation score of each media information card combination.

[0189] Step 204: For the information recommendation position other than the first screen, a sequence set including at least two media information card candidate sequences is generated through a bundle search strategy.

[0190] Here, the information recommendation position other than the first screen includes: when the information recommendation sequence is refreshed, the information recommendation position other than the first screen recommendation position; and when the information recommendation sequence is loaded, each information recommendation position in the information recommendation sequence.

[0191] Specifically, the non-first-screen card is a card candidate predicted by the card ctr model according to the position, and finally a candidate sequence list is generated through bundle search of the fixed first-screen card and the non-first-screen card with two candidates in each position.

[0192] Step 205: Through the card sequence recommendation model, a media information card candidate sequence corresponding to the non-first-screen recommendation position in the recommendation sequence is predicted from the sequence set including at least two media information card candidate sequences.

[0193] Specifically, the sequence recommendation score of each media information card candidate sequence is predicted through the card sequence recommendation model, so that the media information card candidate sequence with the highest sequence recommendation score is selected as the media information card candidate sequence corresponding to the non-first-screen recommendation position in the information recommendation sequence based on the sequence recommendation score of each media information card candidate sequence.

[0194] Step 206: End.

[0195] Next, the card combination recommendation model and the card sequence recommendation model will be described in detail.

[0196] Here, the media information of the first screen is recommended in the form of a heterogeneous card combination, which is implemented by using the card combination recommendation model in the embodiments of the present application. The card combination recommendation model models a card combination composed of multiple media information cards in a position sequence as an item, and the definition of the features and labels is constructed for the card combination. The input features of the card combination model include user basic features, item basic features (including card height features), context features, card combination portrait features and card position portrait features designed for the card combination. The following will be described in detail.

[0197] (1) Card combination capacity. Because of the difference in user screen resolution and the height difference of heterogeneous cards, the number of cards that each user can directly expose on the first screen (card combination capacity) is not fixed, but based on statistics and consistency of the constructed sample, a fixed card combination capacity needs to be determined. Because the height of the graphic-text card is the smallest, the height of the graphic-text card is taken as the basis for calculation in this scheme. In the embodiments of the present application, the median of the card combination capacity can be obtained by taking 7-day product data statistics as 4, so the first 4 cards are taken as the first screen card, that is, the first 4 cards are bound into a card combination for modeling.

[0198] (2) label definition. Because the card combination is composed of 4 cards, the label definition and ordinary ctr estimation scene have some differences, wherein,

[0199] Exposure: The exposure of the card combination is considered valid when the first two cards in the card combination are completely exposed.

[0200] Click label: The label of the card combination model is a multi-classification label, which is divided into three categories: no click, click once, and click two or more times. The definition of click behavior is to click and consume a duration greater than or equal to 6s.

[0201] Therefore, when predicting the score online, each training sample gets 3 scores, and the final sorting score is calculated as P(click once) + 2*P(click >= 2 times).

[0202] (3) Type preference features and position preference features. The card combination portrait feature describes the user's preference for various heterogeneous card combinations of the business in different time periods and network states; the card position portrait feature describes the user's preference for various heterogeneous cards of the business in different positions, time periods, and network states.

[0203] In actual application, in order to describe the user's long-term, short-term, and real-time card preference, the card combination and position portrait are respectively calculated from three time windows of user behavior data, including: (1) Year portrait: statistics of one-year data from the current day; (2) Month portrait: statistics of one-month data from the current day; (3) Session portrait: statistics of the user's last 50 swipes.

[0204] In practical applications, the card combination portrait feature and the card position portrait feature can be consistent in the time period and the statistical dimension of network state. The time period is divided as follows: 1) divided into four segments per hour: early morning (2-5 am), morning (6-11 am), afternoon (12-18 pm), and night (19-1 am of the next day); 2) divided into weekdays and weekends: weekdays (Monday-Friday, excluding Friday night), and weekends (Friday night, Saturday, and Sunday), wherein the user consumption habit on Friday night is special, and thus is also divided into weekends; the network state is divided as follows: 1) WIFI; 2) 4G; 3) 3G; 4) 2G; and 5) unknown network.

[0205] That is, the basic data of the portrait is constructed by counting, for each user, the time period, network state, card or card combination corresponding to each consumption behavior in the time window. The difference between the card combination portrait feature and the card position portrait feature is that the card position portrait feature has an additional position statistical dimension, that is, the card position portrait needs to count the preference data of the user at the first to fourth positions, such as the exposure and click counts of the first position text card, the exposure and click counts of the second position text card, and the like.

[0206] In actual implementation, the statistical dimension of the user portrait can be more detailed, such as whether the WIFI in the network state is a commonly used WIFI, because the consumption habits of some users on the WIFI in the company and the WIFI at home are also different; the time period is divided into four segments per day because of data sparsity, and the division can also be more detailed.

[0207] The statistical features of the type preference feature and the position preference feature can include:

[0208] 1) Exposure count. Type preference feature: exposure count of the card combination, wherein the first two cards are considered to be exposed if they are completely exposed. Position preference feature: exposure count of each card at each position. Among them, the card must be exposed by 50% of the area to be considered exposed.

[0209] 2) Click count. Type preference feature: count of clicks once, twice, and more, that is, there are two click count features. Position preference feature: click count of each card at each position. Among them, the card consumption duration must be greater than or equal to 6s to be considered an effective click.

[0210] 3) Click rate: including click rate of the card combination and click rate of each card at each position.

[0211] Considering that the exposure count of the card or combination of some users is small, the click rate obtained by directly using the click count / exposure count is not reliable, and thus the Wilson Score is used for calculation, and the formula is as follows:

[0212] r = pos / n

[0213]

[0214] wherein pos: click number; n: exposure number; z: parameter, generally takes the value 2, that is, 95% confidence,

[0215] 4) Click preference: namely, the preference of a certain type (such as a video type) card and the preference of a certain card combination. Because there are some users whose click rates of all cards are high or low, the Wilson Score is used, but the click preference of a card combination: pos is the click number of a certain type of card (such as the click number of a video card); n is the total click number of all cards;

[0216] Click preference of a card type: pos is the click number of a certain card combination (such as the click number of a card combination of <text, video, text>); n is the click number of all card combinations.

[0217] 5) Consumption time, including the consumption time of a certain type (such as a video type) card and the consumption time of a certain card combination. The total consumption time of a user on each type of card or card combination is specifically counted.

[0218] In actual application, when the type preference feature or the location preference feature of a certain user fails to be acquired, the type preference feature or the location preference feature corresponding to the user portrait of the user is used as the type preference feature or the location preference feature of the user. Specifically, the statistical dimensions of the user portrait and the statistical dimensions of the user portrait are consistent, but instead of counting the consumption behavior of a single user, all users are divided into different crowds according to different dimensions, and the consumption behavior of the users in each crowd is counted. The division dimensions can be as follows:

[0219] 1) Age: 0-6 years old, 7-12 years old, 13-15 years old, 16-18 years old, 19-22 years old, 23-25 years old, 26-30 years old, 31-40 years old, 41-50 years old, and more than 50 years old.

[0220] 2) Gender: male, female, and unknown.

[0221] 3) Activity: according to the click number of a user in the last 30 days, the user is divided into: no exposure, low activity, medium activity, and high activity.

[0222] Among them, for the recommendation of non-first-screen media information, a heterogeneous card sequence is used for recommendation, in the embodiment of the application, combined with bundle search, and a card sequence recommendation model is used to implement. The card sequence recommendation model is to model the entire card sequence of the up-to-bottom brushing as an item, and the model used is a transformer deep learning model, which will be described in detail below.

[0223] (1)Beam search. In this application, when generating the candidate sequence of the media information card, under the premise of ensuring the recommendation effect, in order to reduce the time consumption of model prediction, the beam search with beam size=10 is used to generate according to the prediction result of the card ctr model. In actual application, since the card ctr model solves the problem of sequence prediction by predicting each position, but the time consumption and performance are a problem, some debias models can be considered to uniformly predict all items, that is, the model can be run only once.

[0224] Beam search is a heuristic graph search algorithm. In the case of a large solution space of a graph, in order to reduce the space and time occupied by the search, some nodes with poor quality are pruned and some nodes with high quality are retained at each step of depth expansion. The following takes beam size=2 as an example to explain the specific process: assuming that the candidate cards sorted by the card ctr model prediction value of the first position are [a, b, c, d], the two cards a and b with the maximum score are selected, and the current card sequence is a or b, which will be used as the input feature of the card ctr model when predicting the next card.

[0225] When the candidate cards sorted by the card ctr model prediction value of the second position of the current sequence a are [d, e], and the candidate cards sorted by the card ctr model prediction value of the second position of the current sequence b are [f, g], then the two optimal sequences ad and ae are selected from the four sequences ad, ae, bf and bg according to the prediction score.

[0226] (2) Features and labels. The features of the card sequence recommendation model mainly include user basic features, item basic features, context features and user card portrait features. The user card portrait features depict the preference degree of the user to various heterogeneous cards of the business in different time periods and network states.

[0227] Click label: The model is a multi-classification model, and the prediction result is divided into four categories: no click, click once, click twice, click three times and more.

[0228] The calculation method of the final sorting score of each card candidate sequence: score=P(click once)+2*P(click twice)+3*P(click>=3 times)

[0229] (3) Model structure. The card sequence recommendation model adopts a transformer model, which has the following advantages: 1) it can explicitly model the mutual influence between cards in the feature space, and the distance between two cards does not affect the calculation of their relevance; 2) the transformer is a parallel computation, which is high in performance efficiency.

[0230] As shown in Figure 9 , Figure 9 is a structural diagram of a card sequence recommendation model provided by an embodiment of the present application. Here, the card sequence recommendation model includes a feature mapping layer, a feature encoding layer, and a feature prediction layer. In actual applications, the card sequence recommendation model can be constructed based on a Transformer network model. Here, the other features refer to user basic features and context features. The embeddings of the item basic features, the position features, and the user card portrait features at each position (pos1, pos2, pos3, etc.) are added together and then input into a transformer encoder module (i.e., the feature encoding layer), the encoded features and other features are concatenated and then input into an MLP layer to obtain a prediction result, i.e., the sequence recommendation score of the input media information card candidate sequence. Here, the position feature is the sequence position of each media information card in the card sequence, such as the first position in the sequence, the second position in the sequence, etc.

[0231] By applying the above embodiments of the present application, first, the card combination recommendation model is constructed in the present application, and the type preference feature that characterizes the user's preference degree for various heterogeneous card combinations in different time periods and network states and the position preference feature that characterizes the user's preference degree for different cards in different display positions on the first screen are obtained, which improves the first screen recommendation efficiency and user experience and solves the problem of the exposure efficiency of heterogeneous card combinations affecting the user's first screen recommendation efficiency. Second, the model and the ranking strategy based on card combination and sequence are designed in the present application according to the different consumption tendencies of the user implied by the user's sliding behavior, which improves the recommendation efficiency of the entire brushing sequence and solves the problem of not considering the user's different sliding behaviors affecting the overall recommendation efficiency.

[0232] The media information recommendation device 555 provided by an embodiment of the present application will be described below. In some embodiments, the media information recommendation device can be implemented in the form of a software module. Referring to Figure 10 , Figure 10 is a structural diagram of a media information recommendation device 555 provided by an embodiment of the present application. The media information recommendation device 555 provided by an embodiment of the present application includes:

[0233] The acquisition module 5551 is configured to acquire a first object feature of a target object, a content feature of at least two media information sequences to be recommended, and a combination feature.

[0234] The media information sequence includes at least two types of media information, and the combination feature includes a type preference feature and a position preference feature of the media information.

[0235] The type preference feature is used to indicate a preference of the target object for the type of the media information in different scenes, and the position preference feature is used to indicate a preference of the target object for a display position of the media information.

[0236] The determining module 5552 is configured to determine a recommendation score of each of the media information sequences based on the first object feature, a content feature of each of the media information sequences, and a combination feature.

[0237] The recommendation module 5553 is configured to select a target media information sequence from the at least two media information sequences based on the recommendation score of each of the media information sequences, and recommend the target media information sequence to a terminal corresponding to the target object.

[0238] In some embodiments, the obtaining module 5551 is further configured to obtain a first group feature of the target object and a second group feature of at least two pre-set object groups, each of the object groups including at least two objects.

[0239] The first group feature of the target object is matched with the second group feature of each of the object groups respectively to determine a matching degree of the first group feature and each of the second group features, and an object group corresponding to a second group feature with the highest matching degree with the first group feature is taken as a target object group.

[0240] A content feature of the at least two media information sequences to be recommended is obtained, and a group first object feature of the target object group and a group combination feature of the target object group corresponding to the at least two media information sequences are obtained.

[0241] The group first object feature is taken as the first object feature of the target object, and the group combination feature is taken as the combination feature of the target object corresponding to the at least two media information sequences.

[0242] In some embodiments, the obtaining module 5551 is further configured to obtain user data of the target object, content data of at least two media information sequences to be recommended, and combination data; and the combination data includes type preference data and position preference data of the media information.

[0243] The feature extraction layer of the first recommendation model is used for performing feature extraction on the user data, content data of the at least two media information sequences to be recommended, and the combined data respectively, to obtain a first object feature of the target object, a content feature of the at least two media information sequences to be recommended, and a combined feature;

[0244] The determining, based on the first object feature, the content feature of each media information sequence, and the combined feature, of a recommendation score of each media information sequence includes:

[0245] The feature prediction layer of the first recommendation model is used for performing prediction in combination with the first object feature, the content feature of each media information sequence, and the combined feature, to obtain the recommendation score of each media information sequence.

[0246] In some embodiments, the apparatus further includes:

[0247] The first training module is configured to obtain a sample first object feature, a sample content feature, and a sample combined feature corresponding to a training sample, the training sample being labeled with a corresponding sample label;

[0248] The training sample includes at least two types of media information samples, and the sample combined feature includes a sample type preference feature and a sample position preference feature of the media information samples.

[0249] The sample type preference feature is used to indicate a preference of a corresponding sample user for a type of corresponding media information sample in different scenarios, and the sample position preference feature is used to indicate a preference of the corresponding sample user for a display position of the corresponding media information sample.

[0250] The sample first object feature, the sample content feature, and the sample combined feature are input into the first recommendation model, and a prediction of the recommendation score is performed by the first recommendation model to obtain a corresponding prediction result.

[0251] A difference between the prediction result and the corresponding sample label is obtained.

[0252] Based on the difference, a model parameter of the first recommendation model is updated.

[0253] In some embodiments, the determining module 5552 is further configured to perform the following processing on each media information sequence respectively to determine a recommendation score of each media information sequence:

[0254] The neural network model is used to perform classification prediction of the click data in combination with the first object feature, the content feature of the media information sequence, and the combined feature, to obtain a click prediction result of at least two categories corresponding to the media information sequence.

[0255] obtain a recommendation weight value corresponding to the click prediction result of each category;

[0256] determine a recommendation score of the media information sequence based on the click prediction results of the at least two categories and the recommendation weight value corresponding to the click prediction result of each category.

[0257] In some embodiments, the recommendation module 5553 is further configured to receive an acquisition request of a media information sequence sent by the terminal, the acquisition request being triggered in response to a media information acquisition instruction for a first-screen recommendation position;

[0258] send the target media information sequence to the terminal to display the target media information sequence on a first-screen recommendation position of the terminal.

[0259] In some embodiments, the target media information sequence is displayed on a first-screen recommendation position of the terminal, and the recommendation module 5553 is further configured to, for each information recommendation position in a candidate recommendation position sequence, obtain at least two candidate media information to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence, wherein the candidate recommendation position sequence is non-first-screen and includes at least two information recommendation positions.

[0260] obtain a second object feature of the target object, a content feature and a position feature of each candidate media information included in each media information candidate sequence;

[0261] determine a sequence recommendation score of each media information candidate sequence based on the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence;

[0262] select a target media information candidate sequence from the at least two media information candidate sequences based on the sequence recommendation score of each media information candidate sequence, and recommend the target media information candidate sequence to a terminal corresponding to the target object, so that the target media information candidate sequence is displayed on a first-screen recommendation position of the terminal.

[0263] display the target media information candidate sequence on a candidate recommendation position sequence in a non-first-screen.

[0264] In some embodiments, the recommendation module 5553 is further configured to, for an information recommendation position in a first position in the candidate recommendation position sequence, perform click data prediction on each first media information to be recommended to obtain predicted click data corresponding to each first media information;

[0265] select at least two first media information as at least two candidate media information corresponding to the information recommendation position in the first position based on the predicted click data corresponding to each first media information.

[0266] For each information recommendation position in the candidate recommendation position sequence other than the first position, the following processing is performed to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence:

[0267] In combination with the at least two candidate media information corresponding to the information recommendation position of the previous position of the current position, click data prediction is performed on each second media information different from the candidate media information to obtain predicted click data corresponding to each second media information;

[0268] Based on the predicted click data corresponding to each second media information, at least two second media information are selected as at least two candidate media information corresponding to the recommendation information position of the current position.

[0269] In some embodiments, the recommendation module 5553 is further configured to perform mapping processing on the second object feature of the target object, the content feature of each candidate media information, and the position feature by a feature mapping layer of a second recommendation model to obtain corresponding mapping features;

[0270] The recommendation module 5553 is further configured to perform encoding processing on the sum of the mapping features corresponding to the content features of each candidate media information and the mapping features corresponding to the position features by a feature encoding layer of the second recommendation model to obtain encoding features corresponding to each candidate media information;

[0271] The recommendation module 5553 is further configured to perform splicing processing on the mapping features corresponding to the second object feature and the encoding features corresponding to each candidate media information by a feature connection layer of the second recommendation model to obtain corresponding splicing features;

[0272] The recommendation module 5553 is further configured to perform prediction based on the splicing features by a feature prediction layer of the second recommendation model to obtain sequence recommendation scores of each media information candidate sequence.

[0273] In some embodiments, the apparatus further includes:

[0274] A second training module configured to obtain sample second object features, sample content features, and sample position features corresponding to a training sample, the training sample being labeled with a corresponding sample label;

[0275] The training sample includes at least two media information, the sample content features include content features of each media information, and the sample position features include position features of each media information;

[0276] The sample second object features, sample content features, and sample position features are input into the second recommendation model, and prediction of recommendation scores is performed by the second recommendation model to obtain a corresponding prediction result;

[0277] obtaining a difference between the prediction result and a corresponding sample label;

[0278] updating model parameters of the first recommendation model based on the difference.

[0279] In some embodiments, the recommendation module 5553 is further configured to perform the following processing on each of the media information candidate sequences respectively to determine a sequence recommendation score of each of the media information candidate sequences:

[0280] performing, by a neural network model, classification prediction on the click data in combination with the second object feature, the content feature of each of the candidate media information, and the position feature to obtain click prediction results of at least two categories corresponding to the media information sequence;

[0281] obtaining a recommendation weight value corresponding to the click prediction result of each of the categories;

[0282] determining a sequence recommendation score of the media information candidate sequence based on the click prediction results of the at least two categories and the recommendation weight value corresponding to the click prediction result of each of the categories.

[0283] In some embodiments, the recommendation module 5553 is further configured to recommend the target media information sequence carrying the display style information to a terminal corresponding to the target object.

[0284] The display style information is used to indicate a display style corresponding to each type of media information in the target media information sequence.

[0285] In the embodiments of the present application, when media information is recommended to a target object, first, a first object feature of the target object, a content feature of at least two media information sequences to be recommended, and a combination feature are obtained, and then a recommendation score of each of the media information sequences is determined based on the first object feature, the content feature of each of the media information sequences, and the combination feature, so that a target media information sequence is selected from the at least two media information sequences based on the recommendation score and recommended to a terminal corresponding to the target object. Here, the media information sequence includes at least two types of media information, and the combination feature includes a type preference feature used to indicate a preference of the target object for the types of media information in different scenarios, and a position preference feature used to indicate a preference of the target object for a display position of the media information.

[0286] Thus, when the media information is recommended based on the first object feature of the target object, the content feature of the at least two media information sequences to be recommended, and the combined feature, the preference of the user for the media information type in different scenarios is considered, the recommended information type is diversified and more suitable for the scenario in which the user is located, the demand of the user is met, the preference of the user for the media information display position is considered, the user is facilitated to click and view the information of personal preference, and therefore the recommendation accuracy and the recommendation efficiency of the media information are improved.

[0287] The embodiment of the present application further provides an electronic device, and the electronic device comprises:

[0288] a memory configured to store executable instructions;

[0289] a processor configured to execute the executable instructions stored in the memory, and implement the method for recommending media information provided by the embodiment of the present application.

[0290] The embodiment of the present application further provides a computer program product or a computer program, which comprises computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method for recommending media information provided by the embodiment of the present application.

[0291] The embodiment of the present application further provides a computer readable storage medium, which stores executable instructions. The executable instructions are executed by a processor, and the method for recommending media information provided by the embodiment of the present application is implemented.

[0292] In some embodiments, the computer readable storage medium can be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disc, or CD-ROM; or various devices comprising one or any combination of the above memories.

[0293] In some embodiments, the executable instructions can be in the form of a program, software, software module, script or code, written in any form of programming language (including a compiled or interpreted language, or a declarative or procedural language), and can be deployed in any form, including being deployed as an independent program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0294] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., files that store one or more modules, subroutines, or code sections).

[0295] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0296] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A method of recommending media information, characterized by, The method comprises: obtaining a first object feature of a target object, a content feature of at least two media information sequences to be recommended, and a combination feature; wherein the media information sequence comprises at least two types of media information, and the combination feature comprises a type preference feature and a position preference feature of the media information; wherein the type preference feature is used to indicate a preference degree of the target object for various heterogeneous media information card combinations in different time periods and network states, and the position preference feature is used to indicate a preference degree of the target object for various heterogeneous media information cards in different positions, time periods, and network states; performing the following processing by a card combination recommendation model: determining a recommendation score of each of the media information sequences based on the first object feature, the content feature of each of the media information sequences, and the combination feature; and selecting a target media information sequence from the at least two media information sequences based on the recommendation score of each of the media information sequences, wherein the target media information sequence adopts a combination mode of heterogeneous media information cards, and the heterogeneous media information cards are different types of media information cards that are different in content, display form, and revenue value in a recommendation system; in a case where a media information acquisition instruction for a first-screen recommendation position is received by a terminal corresponding to the target object, recommending the target media information sequence to the terminal to display the target media information sequence through a first-screen recommendation position of the terminal; in a case where a media information acquisition instruction for a non-first-screen recommendation position is received by the terminal corresponding to the target object, obtaining at least two media information candidate sequences corresponding to a candidate recommendation position sequence by a bundle search mode, determining a sequence recommendation score of each of the media information candidate sequences by a card sequence recommendation model, selecting a target media information candidate sequence from the at least two media information candidate sequences based on the sequence recommendation score, and recommending the target media information candidate sequence to the terminal corresponding to the target object to display the target media information candidate sequence through the non-first-screen recommendation position of the terminal.

2. The method of claim 1, wherein, The method comprises: obtaining a first object feature of a target object, a content feature of at least two media information sequences to be recommended, and a combination feature; wherein the media information sequence comprises at least two types of media information, and the combination feature comprises a type preference feature and a position preference feature of the media information; wherein the type preference feature is used to indicate a preference degree of the target object for various heterogeneous media information card combinations in different time periods and network states, and the position preference feature is used to indicate a preference degree of the target object for various heterogeneous media information cards in different positions, time periods, and network states; performing the following processing by a card combination recommendation model: determining a recommendation score of each of the media information sequences based on the first object feature, the content feature of each of the media information sequences, and the combination feature; and selecting a target media information sequence from the at least two media information sequences based on the recommendation score of each of the media information sequences, wherein the target media information sequence adopts a combination mode of heterogeneous media information cards, and the heterogeneous media information cards are different types of media information cards that are different in content, display form, and revenue value in a recommendation system; in a case where a media information acquisition instruction for a first-screen recommendation position is received by a terminal corresponding to the target object, recommending the target media information sequence to the terminal to display the target media information sequence through a first-screen recommendation position of the terminal; in a case where a media information acquisition instruction for a non-first-screen recommendation position is received by the terminal corresponding to the target object, obtaining at least two media information candidate sequences corresponding to a candidate recommendation position sequence by a bundle search mode, determining a sequence recommendation score of each of the media information candidate sequences by a card sequence recommendation model, selecting a target media information candidate sequence from the at least two media information candidate sequences based on the sequence recommendation score, and recommending the target media information candidate sequence to the terminal corresponding to the target object to display the target media information candidate sequence through the non-first-screen recommendation position of the terminal. The method comprises: obtaining a first object feature of a target object, a content feature of at least two media information sequences to be recommended, and a combination feature; wherein the media information sequence comprises at least two types of media information, and the combination feature comprises a type preference feature and a position preference feature of the media information; wherein the type preference feature is used to indicate a preference degree of the target object for various heterogeneous media information card combinations in different time periods and network states, and the position preference feature is used to indicate a preference degree of the target object for various heterogeneous media information cards in different positions, time periods, and network states; performing the following processing by a card combination recommendation model: determining a recommendation score of each of the media information sequences based on the first object feature, the content feature of each of the media information sequences, and the combination feature; and selecting a target media information sequence from the at least two media information sequences based on the recommendation score of each of the media information sequences, wherein the target media information sequence adopts a combination mode of heterogeneous media information cards, and the heterogeneous media information cards are different types of media information cards that are different in content, display form, and revenue value in a recommendation system; in a case where a media information acquisition instruction for a first-screen recommendation position is received by a terminal corresponding to the target object, recommending the target media information sequence to the terminal to display the target media information sequence through a first-screen recommendation position of the terminal; in a case where a media information acquisition instruction for a non-first-screen recommendation position is received by the terminal corresponding to the target object, obtaining at least two media information candidate sequences corresponding to a candidate recommendation position sequence by a bundle search mode, determining a sequence recommendation score of each of the media information candidate sequences by a card sequence recommendation model, selecting a target media information candidate sequence from the at least two media information candidate sequences based on the sequence recommendation score, and recommending the target media information candidate sequence to the terminal corresponding to the target object to display the target media information candidate sequence through the non-first-screen recommendation position of the terminal.

3. The method of claim 1, wherein, The first object feature of the target object, the content feature of the at least two media information sequences to be recommended, and the combination feature are obtained, including: Obtaining user data of the target object, content data of the at least two media information sequences to be recommended, and combination data; wherein the combination data includes type preference data and location preference data of the media information; The user data, the content data of the at least two media information sequences to be recommended, and the combination data are respectively subjected to feature extraction through a feature extraction layer of the card combination recommendation model, to obtain the first object feature of the target object, the content feature of the at least two media information sequences to be recommended, and the combination feature; The recommendation score of each of the media information sequences is determined based on the first object feature, the content feature of each of the media information sequences, and the combination feature, including: The recommendation score of each of the media information sequences is obtained through the feature prediction layer of the card combination recommendation model, combined with the first object feature, the content feature of each of the media information sequences, and the combination feature.

4. The method of claim 3, wherein, The method further includes: Obtaining sample first object features, sample content features, and sample combination features corresponding to training samples, wherein the training samples are labeled with corresponding sample labels; The training samples include at least two types of media information samples, and the sample combination features include sample type preference features and sample location preference features of the media information samples; The sample type preference features are used to indicate the preference of a corresponding sample user for the type of a corresponding media information sample in different scenarios, and the sample location preference features are used to indicate the preference of the corresponding sample user for the display position of the corresponding media information sample; The sample first object features, the sample content features, and the sample combination features are input into the card combination recommendation model, and the prediction of the recommendation score is performed through the card combination recommendation model to obtain a corresponding prediction result; Obtaining the difference between the prediction result and the corresponding sample label; Based on the difference, the model parameters of the card combination recommendation model are updated.

5. The method of claim 1, wherein, The recommendation score of each of the media information sequences is determined based on the first object feature, the content feature of each of the media information sequences, and the combination feature, including: The following processing is performed for each of the media information sequences to determine the recommendation score of each of the media information sequences: Through a neural network model, the first object feature, the content feature of the media information sequence, and the combination feature are combined to perform classification prediction of click data, to obtain click prediction results of at least two categories corresponding to the media information sequence; Obtaining recommendation weight values corresponding to the click prediction results of each of the categories; Based on the click prediction results of the at least two categories and the recommendation weight values corresponding to the click prediction results of each of the categories, the recommendation score of the media information sequence is determined.

6. The method of claim 1, wherein, The at least two media information candidate sequences corresponding to the candidate recommendation position sequence are obtained in a way of bundle search, including: For each information recommendation position in the candidate recommendation position sequence, at least two candidate media information are obtained to obtain at least two media information candidate sequences corresponding to the candidate recommendation position sequence, wherein the candidate recommendation position sequence is in a non-first screen and includes at least two information recommendation positions; The determining, by the card sequence recommendation model, of the sequence recommendation scores of the media information candidate sequences includes: The following processing is performed by the card sequence recommendation model: The second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence are obtained; Based on the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence, the sequence recommendation scores of the media information candidate sequences are determined; Based on the sequence recommendation scores of the media information candidate sequences, a target media information candidate sequence is selected from the at least two media information candidate sequences.

7. The method of claim 6, wherein, The obtaining of the at least two candidate media information for each information recommendation position in the candidate recommendation position sequence to obtain the at least two media information candidate sequences corresponding to the candidate recommendation position sequence includes: For the information recommendation position in the first position in the candidate recommendation position sequence, the click data prediction is performed on each first media information to be recommended to obtain the predicted click data corresponding to each first media information; Based on the predicted click data corresponding to each first media information, at least two first media information are selected as at least two candidate media information corresponding to the information recommendation position in the first position; For each information recommendation position in the candidate recommendation position sequence, the following processing is performed to obtain the at least two media information candidate sequences corresponding to the candidate recommendation position sequence: In combination with the at least two candidate media information corresponding to the information recommendation position in the previous position of the current position, the click data prediction is performed on each second media information different from the candidate media information to obtain the predicted click data corresponding to each second media information; Based on the predicted click data corresponding to each second media information, at least two second media information are selected as at least two candidate media information corresponding to the information recommendation position in the current position.

8. The method of claim 6, wherein, The determining of the sequence recommendation scores of the media information candidate sequences based on the second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence includes: The second object feature of the target object, the content feature and the position feature of each candidate media information included in each media information candidate sequence are obtained; The mapping features corresponding to the second object feature of the target object, the content feature and the position feature of each candidate media information are obtained by performing mapping processing on the second object feature of the target object, the content feature and the position feature of each candidate media information through the feature mapping layer of the card sequence recommendation model; The encoding features corresponding to each candidate media information are obtained by performing encoding processing on the sum of the mapping features corresponding to the content feature and the position feature of each candidate media information through the feature encoding layer of the card sequence recommendation model; The mapping features corresponding to the second object features and the encoding features corresponding to each of the candidate media information are spliced through a feature connection layer of the card sequence recommendation model to obtain corresponding spliced features; The spliced features are predicted through a feature prediction layer of the card sequence recommendation model to obtain sequence recommendation scores of each of the media information candidate sequences.

9. The method of claim 8, wherein, The method further includes: obtaining sample second object features, sample content features and sample position features corresponding to a training sample, the training sample being labeled with a corresponding sample label; wherein the training sample includes at least two media information, the sample content features include content features of each of the media information, and the sample position features include position features of each of the media information; inputting the sample second object features, sample content features and sample position features into the card sequence recommendation model to predict recommendation scores through the card sequence recommendation model to obtain a corresponding prediction result; obtaining a difference between the prediction result and the corresponding sample label; updating model parameters of the card combination recommendation model based on the difference.

10. The method of claim 6, wherein, The determination of the sequence recommendation scores of each of the media information candidate sequences based on the second object features of the target object, content features and position features of each of the candidate media information included in each of the media information candidate sequences includes: the following processing is performed for each of the media information candidate sequences to determine the sequence recommendation scores of each of the media information candidate sequences: classification prediction of click data is performed through a neural network model in combination with the second object features, content features and position features of each of the candidate media information to obtain click prediction results of at least two categories corresponding to the media information sequence; obtaining recommendation weight values corresponding to the click prediction results of each of the categories; determining the sequence recommendation scores of the media information candidate sequences based on the click prediction results of the at least two categories and the recommendation weight values corresponding to the click prediction results of each of the categories.

11. The method of claim 1, wherein, The recommendation of the target media information sequence to the terminal includes: recommending the target media information sequence carrying display style information to a terminal corresponding to the target object; wherein the display style information is used to indicate display styles corresponding to each type of media information in the target media information sequence.

12. A device for recommending media information, characterized by The apparatus includes: an obtaining module configured to obtain first object features of a target object, content features of at least two media information sequences to be recommended and combination features; wherein the media information sequence includes at least two types of media information, and the combination features include type preference features and position preference features of the media information; wherein the type preference features are used to indicate preference degrees of the target object for combinations of various heterogeneous media information cards in different time periods and network states, and the position preference features are used to indicate preference degrees of the target object for various heterogeneous media information cards in different positions, time periods and network states. The determining module is configured to determine, by the card combination recommendation model, a recommendation score of each of the media information sequences based on the first object feature, a content feature of each of the media information sequences, and the combination feature; The recommendation module is configured to select a target media information sequence from the at least two media information sequences based on the recommendation score of each of the media information sequences, wherein the target media information sequence is in a form of a heterogeneous media information card combination, and the heterogeneous media information card is a different type of media information card that is different in content, display form, and revenue value in a recommendation system; In a case where the terminal corresponding to the target object receives a media information acquisition instruction for a first-screen recommendation position, the target media information sequence is recommended to the terminal to display the target media information sequence on the first-screen recommendation position of the terminal. In a case where the terminal corresponding to the target object receives a media information acquisition instruction for a non-first-screen recommendation position, at least two media information candidate sequences corresponding to a candidate recommendation position sequence are acquired by a bundle search, a sequence recommendation score of each of the media information candidate sequences is determined by a card sequence recommendation model, a target media information candidate sequence is selected from the at least two media information candidate sequences based on the sequence recommendation score, and the target media information candidate sequence is recommended to the terminal corresponding to the target object to display the target media information candidate sequence on the non-first-screen recommendation position of the terminal.

13. An electronic device, comprising: The electronic device includes: a memory configured to store executable instructions; a processor configured to execute the executable instructions stored in the memory to implement the method for recommending media information according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The executable instructions are stored in the memory and are executed to implement the method for recommending media information according to any one of claims 1 to 11.

15. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the method for recommending media information according to any one of claims 1 to 11.

Images