Content pushing method and apparatus, computer device and storage medium

Abstract

A content pushing method, which is executed by a computer device. The method comprises: acquiring the current information sequence of a user, the current information sequence comprising a plurality of information items, each information item corresponding to one piece of browsing content of the user, the plurality of information items comprising the information item corresponding to the current browsing content and the information item corresponding to at least one piece of historical browsing content, and the information items in the current information sequence being arranged in a browsing order of the corresponding browsing content (202); encoding the current information sequence into a feature sequence, the feature sequence comprising feature items in one-to-one correspondence with the information items in the current information sequence, and each feature item representing an encoding feature obtained by encoding the corresponding information item (204); acquiring the encoding feature of the current browsing content, the encoding feature of the current browsing content being obtained by encoding description information of the current browsing content (206); extracting from the feature sequence a feature related to the encoding feature of the current browsing content, so as to obtain a related feature of the current browsing content (208); acquiring description information of at least one piece of candidate content, and separately encoding the description information of the at least one piece of candidate content to obtain an encoding feature of each piece of candidate content (210); and, according to the encoding feature of each piece of candidate content and the related feature of the current browsing content, selecting candidate content from amongst the at least one piece of candidate content, and pushing the selected candidate content to the user (212).

Description

Content push method, device, computer equipment and storage medium

[0001] Related applications

[0002] This application claims priority to Chinese patent application number 2023111564187, filed on September 8, 2023, entitled “Content Push Method, Apparatus, Computer Equipment and Storage Medium,” the entire text of which is incorporated herein by reference. Technical Field

[0003] The present application relates to the field of computer technology, and in particular to a content push method, apparatus, computer equipment, storage medium, and computer program product. Background Art

[0004] With the development of artificial intelligence and computer technology, AI-based content push technology has emerged. This technology is used to push content on content platforms through AI. Content platforms include, but are not limited to, video platforms, audio platforms, and news platforms. For example, video platforms can push content related to videos on the platform to users.

[0005] In traditional technology, user characteristics and the content they have watched in the past are usually used as reference information, and artificial intelligence is used to analyze the reference information to determine the content to be pushed to the user.

[0006] However, the traditional method of pushing content does not necessarily meet the real-time needs of users, and the push accuracy is low, resulting in the need to consume more resources in order to achieve certain push goals.

[0007] Summary of the Invention

[0008] Based on this, a content push method, apparatus, computer device, computer-readable storage medium, and computer program product are provided.

[0009] In one aspect, the present application provides a content push method, which is executed by a computer device, and the method includes:

[0010] Obtaining a current information sequence of a user, the current information sequence comprising a plurality of information items, each information item corresponding to a content viewed by the user, the plurality of information items comprising an information item corresponding to the currently viewed content and an information item corresponding to at least one previously viewed content, the information items in the current information sequence being arranged in the order in which the corresponding viewed content was viewed;

[0011] Encoding the current information sequence into a feature sequence, wherein the feature sequence includes feature items corresponding one-to-one to information items in the current information sequence, and each feature item represents a coding feature obtained by encoding the corresponding information item;

[0012] Acquire the coding feature of the currently viewed content, where the coding feature of the currently viewed content is obtained by encoding the description information of the currently viewed content;

[0013] Extracting features related to the coding features of the currently viewed content from the feature sequence to obtain relevant features of the currently viewed content;

[0014] Obtaining description information of at least one candidate content, encoding the description information of the at least one candidate content respectively, and obtaining encoding features of each candidate content; and

[0015] Select candidate content from the at least one candidate content according to the encoding feature of each candidate content and the relevant features of the currently viewed content, and push the selected candidate content to the user.

[0016] On the other hand, the present application also provides a content push device, comprising:

[0017] a sequence acquisition module, configured to acquire a user's current information sequence, wherein the current information sequence includes a plurality of information items, each of which corresponds to a content viewed by the user, the plurality of information items including an information item corresponding to the currently viewed content and an information item corresponding to at least one previously viewed content, and the information items in the current information sequence are arranged in the order in which the corresponding viewed content was viewed;

[0018] a first encoding module, configured to encode the current information sequence into a feature sequence, wherein the feature sequence includes feature items corresponding one-to-one to information items in the current information sequence, each feature item representing an encoding feature obtained by encoding the corresponding information item;

[0019] A feature acquisition module, configured to acquire a coding feature of the currently viewed content, wherein the coding feature of the currently viewed content is obtained by encoding description information of the currently viewed content;

[0020] A feature extraction module, configured to extract features related to the coding features of the currently viewed content from the feature sequence, thereby obtaining relevant features of the currently viewed content;

[0021] a second encoding module, configured to obtain description information of at least one candidate content, and encode the description information of the at least one candidate content to obtain encoding features of each candidate content;

[0022] The push module is configured to select candidate content from the at least one candidate content according to the coding feature of each candidate content and the related features of the currently viewed content, and push the selected candidate content to the user.

[0023] On the other hand, the present application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in the above-mentioned content push method when executing the computer program.

[0024] On the other hand, the present application also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps in the above-mentioned content push method when executed by a processor.

[0025] On the other hand, the present application also provides a computer program product, including a computer program, which implements the steps in the above-mentioned content push method when executed by a processor.

[0026] The details of one or more embodiments of the present application are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the present application will become apparent from the description, drawings, and claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] FIG1 is a diagram illustrating an application environment of a content push method according to an embodiment;

[0029] FIG2 is a schematic diagram of a flow chart of a content push method according to an embodiment;

[0030] FIG3 is a schematic diagram of an interface in which a terminal displays currently viewed content and push information in one embodiment;

[0031] FIG4 is a structural diagram of a network for generating relevant features of currently viewed content in one embodiment;

[0032] FIG5 is a structural diagram of a feature extraction network in one embodiment;

[0033] FIG6 is a structural diagram of a correlation feature generation network in one embodiment;

[0034] FIG7 is a schematic diagram of a method for generating content fusion features according to an embodiment;

[0035] FIG8 is a structural diagram of a signature feature extraction network according to one embodiment;

[0036] FIG9 is a structural diagram of a recommendation degree prediction network in one embodiment;

[0037] FIG10 is a structural diagram of a recommendation degree prediction model according to one embodiment;

[0038] FIG11 is a schematic diagram of a principle for training a recommendation prediction model in one embodiment;

[0039] FIG12 is a schematic flow chart of a content push method in another embodiment;

[0040] FIG13 is a structural block diagram of a content push device according to an embodiment;

[0041] FIG14 is a diagram showing the internal structure of a computer device according to one embodiment;

[0042] FIG15 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0044] The content push method provided in the embodiments of the present application can be applied in the application environment shown in FIG1 . In this embodiment, a terminal 102 communicates with a server 104 via a network. A data storage system can store data that the server 104 needs to process. The data storage system can be set up separately, integrated with the server 104, or placed on a cloud or other network server.

[0045] Specifically, in response to a user triggering a request to browse any content on the media platform, terminal 102 sends a browsing request for the content to server 104. The content is the currently browsed content, referred to as the currently browsed content. In response to the browsing request, server 104 obtains the user's current information sequence, which includes information items corresponding to multiple browsed contents of the user. The information items include description information of the corresponding browsed contents. The multiple browsed contents include the currently browsed content and historically browsed contents. The current information sequence is encoded into a feature sequence, which includes feature items corresponding one-to-one to the information items in the current information sequence. The encoding features of the currently browsed content are obtained by encoding the description information of the currently browsed content. Features related to the encoding features of the currently browsed content are extracted from the feature sequence to obtain relevant features of the currently browsed content. The description information of at least one candidate content is encoded to obtain encoding features for each candidate content. Based on the encoding features of each candidate content and the relevant features of the currently browsed content, server 104 selects a candidate content from the at least one candidate content and pushes it to the user. When pushing the candidate content, the server may send push information of the candidate content to the user's terminal 102. The terminal 102 receives the push information sent by the server 104 and may display the push information in an interface provided by the media platform.

[0046] Terminal 102 may include, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices may include smart speakers, smart TVs, smart air conditioners, and smart car devices. Portable wearable devices may include smart watches, smart bracelets, and head-mounted devices. Server 104 may be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0047] The content push method provided in this application can be based on artificial intelligence technology. For example, a neural network model can be trained based on artificial intelligence technology, and the trained neural network model can be used to encode the current information sequence to obtain a feature sequence. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve optimal results. In other words, AI is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI is the study of the design principles and implementation methods of various intelligent machines, enabling them to have the capabilities of perception, reasoning, and decision-making. AI technology is an interdisciplinary discipline covering a wide range of fields, encompassing both hardware and software technologies. Basic AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, pre-trained model technology, operation / interaction systems, and mechatronics. Pre-trained models, also known as large models or basic models, can be widely applied to downstream tasks in various AI fields after fine-tuning. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0048] The solutions provided in the embodiments of this application involve technologies such as machine learning based on artificial intelligence, and are specifically described through the following embodiments:

[0049] In some embodiments, as shown in FIG2 , a content push method is provided. The method can be executed by a terminal or a server, or can be executed by both the terminal and the server. The method is described by taking the server 104 in FIG1 as an example, and includes the following steps 202 to 212. Among them:

[0050] Step 202: Obtain the user's current information sequence. The current information sequence includes multiple information items. Each information item corresponds to a reading content of the user. The multiple information items include an information item corresponding to the current reading content and an information item corresponding to at least one historical reading content. The information items in the current information sequence are arranged in the reading order of the corresponding reading content.

[0051] The server can obtain the user's current information sequence, which includes information items corresponding to the user's multiple viewing contents. The information items include description information of the corresponding viewing contents. The multiple viewing contents include current viewing contents and historical viewing contents.

[0052] Here, "viewed content" refers to content viewed by users, and "reading" means watching or browsing. Content includes, but is not limited to, one or more of video, audio, images, text, short films, newspapers, e-books, and the like. Video includes, but is not limited to, at least one of long or short videos. Users may be users on media platforms. "Viewed content" refers to content viewed by users on media platforms. Media platforms disseminate content online and include, but are not limited to, at least one of live streaming platforms, video platforms, news platforms, novel platforms, or gaming platforms. Video platforms may provide short videos or both short and long videos.

[0053] Current viewing content refers to the content that the user is viewing at the current time, for example, it can be the videos that the user is currently watching on the video platform. Historical viewing content refers to the content that the user has viewed in the past, for example, it can be the videos that the user has watched in the past on the video platform.

[0054] The current information sequence includes an information item corresponding to the currently viewed content and an information item corresponding to at least one historically viewed content. For example, the current information sequence includes information items corresponding to N historically viewed contents. In the current information sequence, the information items corresponding to the viewed contents are arranged in sequence according to the viewing order of the viewed contents, for example, in sequence from front to back according to the viewing order. For example, the earlier the viewing order, the earlier the position of the information item corresponding to the viewed content in the current information sequence. Since the currently viewed content is the last viewed, the information item corresponding to the currently viewed content is the last information item in the current information sequence. For example, the current information sequence can be expressed as: Seq u ={(item1,side_info1),(item2,side_info2)…(item N ,side_info N ),(item src ,side_info src )}. Among them, (item i ,side_info i ) is the information item of the i-th historical reading content, item i Represents the identifier (id, identity document) of the i-th historical reading content, side_info i Represents the description information of the i-th historical reading content. 1≤i≤N. (item src ,side_info src ) is the information item of the current reading content, item src Represents the identifier of the current reading content, side_info src Represents the description of the currently viewed content.

[0055] The descriptive information of the reading content is information used to describe the reading content. The descriptive information of the reading content may include information in at least one form of text, image or video. The descriptive information of the reading content may include at least one of the name of the reading content, the label of the reading content or the header image of the reading content. The label of the reading content may be used to indicate the type of the reading content, including but not limited to at least one of humor or suspense. The header image refers to the image displayed at the beginning of the reading content. The descriptive information of the reading content may also include part or all of the reading content. The descriptive information of the reading content may also include information integrated with key characters or key plots in the reading content.

[0056] Specifically, the server may obtain description information of the user's currently viewed content and the user's historical information sequence, wherein the historical information sequence includes information items corresponding to the user's historically viewed content, and the information items corresponding to the historically viewed content include description information of the historically viewed content. The server may generate information items corresponding to the currently viewed content based on the description information of the currently viewed content, and generate the user's current information sequence based on the information items corresponding to the currently viewed content and the historical information sequence.

[0057] Step 204: Encode the current information sequence into a feature sequence, where the feature sequence includes feature items that correspond one-to-one to the information items in the current information sequence.

[0058] Specifically, the server may encode each information item in the current information sequence respectively to obtain the encoding feature corresponding to each information item in each current information sequence. The feature item corresponding to the information item may be the encoding feature corresponding to the information item.

[0059] In some embodiments, the server may arrange the feature items corresponding to each information item according to the arrangement order of the information items in the current information sequence to obtain a feature sequence.

[0060] Step 206: Obtain the coding feature of the currently viewed content. The coding feature of the currently viewed content is obtained by encoding the description information of the currently viewed content.

[0061] Specifically, the server may encode the description information of the currently viewed content, and determine the features obtained by encoding as the encoding features of the currently viewed content.

[0062] In some embodiments, the description information of the currently viewed content includes multiple pieces of information. The server can encode the multiple pieces of information separately to obtain the encoding value of each piece of information, and combine the encoding value of each piece of information into the encoding feature of the currently viewed content. The encoding feature of the currently viewed content can be the encoding feature obtained by combination.

[0063] In some embodiments, due to the different importance of each piece of information in the description information of the currently viewed content, some information contains less information, while some information contains more information. The server can generate corresponding weight features for the coding features of the currently viewed content. The weight features include the weights corresponding to the coding values ​​of each piece of information in the coding features. The server can perform weighted calculations on the coding features of the currently viewed content and the corresponding weight features, and use the results of the weighted calculations as the coding features of the currently viewed content. Specifically, the server can multiply the coding values ​​in the coding features of the currently viewed content with the corresponding weights to obtain the weighted values ​​of each coding value, and arrange the weighted values ​​of each coding value according to the arrangement of the coding values ​​in the coding features to obtain the coding features of the currently viewed content. The coding features can be in the form of vectors or matrices. Taking the vector form as an example, for example, the coding feature of the currently viewed content is E0=[e1,e2,…,e f ],e1,e2,…,e f are the encoding values ​​of different items of information, and the weight feature is A0=[a1,a2,…,a f ], a j for e j The weight of the current reading content is V0 = [v1, v2, ..., v f ], V0 is the multiplication result of the corresponding positions in E0 and A0, v j =a j ×e j , that is, V0 is the result of the Hadamard product operation on E0 and A0. The weight feature can be used to weaken the encoding value of information containing less information or unimportant information in the encoding feature. For example, when the weight is less than 1, it can have a weakening effect, thereby minimizing unimportant data in the encoding feature of the currently viewed content, reducing the interference caused by unimportant data, and thus improving the accuracy of push notifications.

[0064] In some embodiments, since the information item corresponding to the currently viewed content includes descriptive information of the currently viewed content, the encoding feature of the currently viewed content can be obtained by encoding the information item corresponding to the currently viewed content. Since the feature sequence includes the feature item corresponding to the currently viewed content, and the feature item corresponding to the currently viewed content is obtained by encoding the information item corresponding to the currently viewed content, the server can use the feature item corresponding to the currently viewed content in the feature sequence as the encoding feature of the currently viewed content. Of course, the server can also encode the information item corresponding to the currently viewed content to obtain the item encoding feature of the currently viewed content instead of obtaining the encoding feature item from the feature sequence as the encoding feature of the currently viewed content. The encoding feature of the currently viewed content can be the item encoding feature of the currently viewed content. The information item can also include multiple pieces of information. The process of generating the item encoding feature can refer to the process of generating the encoding feature of the currently viewed content. Alternatively, the server can generate a weighted feature corresponding to the item encoding feature and perform a weighted calculation on the item encoding feature and the corresponding weighted feature to obtain the encoding feature of the currently viewed content. The process of performing weighted calculation on the item coding feature and the corresponding weight feature may refer to the above-mentioned process of performing weighted calculation on the coding feature and the corresponding weight feature.

[0065] Step 208: extract features related to the coding features of the currently viewed content from the feature sequence to obtain relevant features of the currently viewed content.

[0066] The relevant features of the currently viewed content refer to features extracted from the feature sequence and relevant to the coding features of the currently viewed content.

[0067] Specifically, for each feature item, the server can extract features related to the encoding features of the currently viewed content from the feature item to obtain related items corresponding to the feature item. The server can generate related features of the currently viewed content based on the related items corresponding to each feature item.

[0068] In some embodiments, the server may arrange the related items corresponding to each feature item according to the arrangement order of the feature items in the feature sequence, and use the arrangement result as the related feature of the currently viewed content.

[0069] In some embodiments, the server may calculate statistics on the related items corresponding to each feature item to obtain the relevant features of the currently viewed content. The statistics may be at least one of addition or multiplication. For example, the server may add the related items corresponding to each feature item to obtain the relevant features of the currently viewed content.

[0070] Step 210: Encode the description information of at least one candidate content to obtain encoding features of each candidate content.

[0071] The candidate content is pre-selected content, and there may be multiple candidate contents, where multiple refers to at least two. The content pushed to the user is selected from the candidate contents.

[0072] Specifically, the server may encode the description information of the candidate content, and determine the features obtained by encoding as the encoding features of the candidate content.

[0073] In some embodiments, the description information of the candidate content includes multiple pieces of information. The server can encode the multiple pieces of information separately to obtain the encoding value of each piece of information, and combine the encoding value of each piece of information into the encoding feature of the candidate content. The encoding feature of the candidate content can be the encoding feature obtained by combination.

[0074] In some embodiments, due to the different importance of each piece of information in the description information of the candidate content, some information contains less information, while some information contains more information. The server can generate a corresponding weight feature for the coding feature of the candidate content. The weight feature includes the weight corresponding to the coding value of each piece of information in the coding feature. The server can perform a weighted calculation on the coding feature of the candidate content and the corresponding weight feature, and use the result of the weighted calculation as the coding feature of the candidate content. Specifically, the server can multiply the coding value in the coding feature of the candidate content by the corresponding weight to obtain the weighted value of each coding value, and arrange the weighted values ​​of each coding value according to the arrangement of the coding value in the coding feature to obtain the coding feature of the candidate content.

[0075] In some embodiments, since the information item corresponding to the candidate content includes the description information of the candidate content, the coding feature of the candidate content can be obtained by encoding the information item corresponding to the candidate content. The server can encode the information item corresponding to the candidate content to obtain the item coding feature of the candidate content, and the coding feature of the candidate content can be the item coding feature of the candidate content. The information item can also include multiple items of information, and the process of generating the item coding feature can refer to the process of generating the coding feature of the candidate content. Alternatively, the server can generate a weight feature corresponding to the item coding feature, and perform weighted calculation on the item coding feature and the corresponding weight feature to obtain the coding feature of the candidate content. The process of weighted calculation on the item coding feature and the corresponding weight feature can refer to the above-mentioned process of weighted calculation on the coding feature and the corresponding weight feature.

[0076] Step 212 : Select candidate content from at least one candidate content based on the encoding feature of each candidate content and the relevant features of the currently viewed content, and push the candidate content to the user.

[0077] Specifically, after obtaining the relevant features of the currently viewed content, the server may determine the recommendation degree of the candidate content based on the encoding features of the candidate content and the relevant features of the currently viewed content, select a candidate content from at least one candidate content based on the recommendation degree, and push it to the user.

[0078] In some embodiments, the relevant features of the currently viewed content obtained in step 208 are referred to as relevant features of the first currently viewed content. For each candidate content, the server can extract features related to the encoding features of the candidate content from the feature sequence to obtain relevant features of the second currently viewed content corresponding to the candidate content. For each candidate content, the server can generate comprehensive features corresponding to the candidate content based on the relevant features of the first currently viewed content and the relevant features of the second currently viewed content corresponding to the candidate content. The server can predict the recommendation degree of the candidate content based on the comprehensive features corresponding to the candidate content. The method for generating relevant features of the second currently viewed content refers to the method for generating relevant features of the first currently viewed content.

[0079] In some embodiments, in response to a user triggering a request to view any content on a media platform, the terminal sends a viewing request for the content to the server. The content is the currently viewed content, and the viewing request may include an identifier for the currently viewed content and a user identifier. The server may store a set of historical information sequences, which includes historical information sequences for multiple users. Historical information sequences in the set of historical information sequences can be uniquely identified by the user identifier. In response to the viewing request, the server extracts the identifier for the currently viewed content and the user identifier from the viewing request, searches the set of historical information sequences for a corresponding historical information sequence based on the user identifier, determines the currently viewed content based on the identifier of the currently viewed content, obtains description information for the user's currently viewed content, generates information items corresponding to the currently viewed content based on the identifier and description information of the currently viewed content, adds the information items corresponding to the currently viewed content to the user's historical information sequences, obtains the user's current information sequence, and executes steps 204-212.

[0080] In some embodiments, when pushing candidate content, the server may send push information of the candidate content to the user's terminal. The push information may include at least one of the candidate content's name, title, popularity, push reason, etc., and may also include other information. The push information may include at least one form of information such as pictures, text, audio, or video. There is no restriction on the push information here. The terminal receives the push information sent by the server and may display the push information in the interface provided by the media platform. The push information may be displayed on the same page as the currently viewed content, or on a different page. Taking the content as a video as an example, as shown in Figure 3, (a) in Figure 3 shows the currently viewed video, and (b) in Figure 3 shows the push information of the pushed video under the title "Recommended for you". In Figure 3, (a) and (b) belong to the same page. Since the page is relatively long, (a) and (b) only show a portion of the page. (b) is triggered to display by the upward swiping operation triggered in (a).

[0081] In the above-described content push method, since the current information sequence includes information items corresponding to multiple viewed contents of the user, each information item includes descriptive information of the corresponding viewed contents, and the multiple viewed contents include the currently viewed content and historically viewed contents, and since the currently viewed content reflects the user's real-time interests, the current information sequence includes information that can reflect the user's real-time interests, thereby improving the accuracy of real-time push notifications. Furthermore, features related to the encoding features of the currently viewed content are extracted from the feature sequence to obtain relevant features of the currently viewed content. Since the relevant features of the currently viewed content are related to the currently viewed content, candidate content is selected from at least one candidate content based on the encoding features of each candidate content and the relevant features of the currently viewed content, and is pushed to the user. This strengthens the influence of the currently viewed content in content push notifications, making push notifications more compliant with real-time needs, improving the accuracy of real-time push notifications, and consuming fewer resources to achieve a certain push notification goal, thereby avoiding resource waste. For example, to achieve the push notification goal of having M people click on a pushed content, the present solution can achieve more accurate push notifications, which means that the total amount of content required to be pushed to achieve this goal is reduced, and the time taken to implement the push function is also reduced, effectively improving resource utilization.

[0082] In some embodiments, obtaining the user's current information sequence includes: obtaining description information of the user's currently viewed content; obtaining the user's historical information sequence, the historical information sequence including information items corresponding to the user's historically viewed content, the information items corresponding to the historically viewed content including description information of the historically viewed content; generating information items corresponding to the currently viewed content based on the description information of the currently viewed content; generating the user's current information sequence based on the information items corresponding to the currently viewed content and the historical information sequence.

[0083] Among them, the historical information sequence includes multiple information items, and multiple means at least two. The information items in the historical information sequence are information items corresponding to the user's historical reading content. In the historical information sequence, the information items are arranged in sequence according to the reading order of the historical reading content, for example, they are arranged in sequence from front to back according to the reading order. For example, the earlier the reading order, the earlier the position of the information item in the historical information sequence. For example, if the historical information sequence includes N information items corresponding to historical reading content, the historical information sequence can be expressed as: Seq = {(item1, side_info1), (item2, side_info2)… (item N ,side_info N )}. (item i ,side_info i ) represents the information item corresponding to the i-th historical reading content, item i Represents the identifier of the i-th historical reading content, side_info i Represents the description of the i-th historically viewed content. 1≤i≤N. Before the user views the currently viewed content, the user's historical information sequence already exists.

[0084] Specifically, the server may combine the identifier of the currently viewed content with the description information of the currently viewed content to generate an information item corresponding to the currently viewed content. For example, the information item corresponding to the currently viewed content may be represented as (item src ,side_info src ), where item src Represents the identifier of the current reading content, side_info src Represents the description of the currently viewed content.

[0085] In some embodiments, the server may add the information item corresponding to the currently viewed content to the historical information sequence to generate the current information sequence. Specifically, the server may add the information item corresponding to the currently viewed content to the historical information sequence in the order in which the information items were viewed to generate the current information sequence. For example, if the information items in the historical information sequence are arranged from front to back in the order in which the information items were viewed, and the currently viewed content is viewed last in the order in which the information items were viewed, the information item corresponding to the currently viewed content is added after the last information item in the historical information sequence to generate the current information sequence. In other words, the last information item in the current information sequence is the information item corresponding to the currently viewed content.

[0086] In this embodiment, when a content item is triggered for viewing, the corresponding information item cannot be updated in time to the historical information sequence. That is, when a content item is triggered for viewing, the historical information sequence obtained by the server does not include the information item corresponding to the content item. Therefore, directly using the historical information sequence for content push may not meet real-time requirements. However, a current information sequence is generated based on the information item corresponding to the currently viewed content and the historical information sequence. Since the currently viewed content is the content currently being viewed by the user, it reflects the user's real-time interests. Therefore, the current information sequence includes information that reflects the user's real-time interests. This overcomes the drawback of the historical information sequence's inability to update in time. Content push based on the current information sequence improves the accuracy of real-time push and further reduces resource waste.

[0087] In some embodiments, features related to the coding features of the currently viewed content are extracted from the feature sequence to obtain relevant features of the currently viewed content, including: combining each feature item in the feature sequence with the coding features of the currently viewed content to obtain combined features corresponding to each feature item; fusing each feature item with the corresponding combined features to obtain relevant items corresponding to each feature item; and generating relevant features of the currently viewed content based on the relevant items corresponding to each feature item.

[0088] The combination includes but is not limited to at least one of concatenation and fusion, and fusion includes at least one of addition, subtraction, and multiplication. The features extracted from the feature items and related to the coding features of the currently viewed content are called related items corresponding to the feature items.

[0089] Specifically, taking a feature item as an example, the process of generating a combined feature corresponding to the feature item is described: the server may add the feature item to the coded feature of the currently viewed content to obtain a first added feature, where addition refers to summing the data at the same position. Alternatively, the server may subtract the feature item from the coded feature of the currently viewed content to obtain a first subtracted feature, where subtraction refers to performing a difference calculation on the data at the same position. Alternatively, the server may multiply the feature item with the coded feature of the currently viewed content to obtain a first multiplied feature, where multiplication refers to performing a product operation on the data at the same position, i.e., multiplication refers to performing a Hadamard product operation. The server may concatenate at least two of the first added feature, the first subtracted feature, or the first multiplied feature, and use the result of the concatenation as the combined feature corresponding to the feature item. Alternatively, the server may concatenate at least two of the first added feature, the first subtracted feature, the first multiplied feature, the feature item, or the coded feature of the currently viewed content, and use the result of the concatenation as the combined feature corresponding to the feature item.

[0090] In some embodiments, the server may perform further feature extraction on the combined features corresponding to each feature item to obtain a correlation feature corresponding to the feature item. This further feature extraction may be implemented using a neural network, including but not limited to at least one of a fully connected neural network and a convolutional neural network. The correlation feature is used to characterize the relationship between the feature item and the encoding features of the currently viewed content.

[0091] In some embodiments, taking a feature item as an example, the process of generating relevant features of the currently viewed content corresponding to the feature item is described: the server can fuse the feature item with the corresponding associated features, such as a Hadamard product operation, and determine the result of the operation as the relevant item corresponding to the feature item, that is, determine it as a feature related to the coding feature of the currently viewed content.

[0092] In some embodiments, the server may perform statistics, such as summing, on the related items corresponding to each feature item to obtain the related features of the currently viewed content. Of course, the server may also use other methods to generate the related features of the currently viewed content, such as using the principle of an attention mechanism to generate the related features of the currently viewed content.

[0093] In this embodiment, through combination, feature extraction and fusion, features related to the coding features of the currently viewed content are automatically generated for each feature item, related items corresponding to each feature item are obtained, and related features of the currently viewed content are generated based on each related item, thereby improving the efficiency of generating related features of the currently viewed content.

[0094] In some embodiments, each feature item in the feature sequence is combined with the coding features of the currently viewed content to obtain the combined features corresponding to each feature item, including: performing a first dimensionality increase processing on each feature item in the feature sequence to obtain the first dimensionality increase feature corresponding to each feature item, and performing a second dimensionality increase processing on each feature item to obtain the second dimensionality increase feature corresponding to each feature item; performing a third dimensionality increase processing on the coding features of the currently viewed content to obtain the third dimensionality increase feature; combining the first dimensionality increase feature and the third dimensionality increase feature corresponding to each feature item to obtain the combined features corresponding to each feature item; fusing each feature item with the corresponding combined features to obtain the related items corresponding to each feature item, including: fusing the second dimensionality increase feature and the corresponding combined features corresponding to each feature item to obtain the features of each feature item related to the coding features of the currently viewed content, and obtaining the related items corresponding to each feature item.

[0095] Dimensionality increase processing is used to increase the dimensionality, which can be achieved through upsampling. The upsampling methods used in the first dimensionality increase processing, the second dimensionality increase processing, and the third dimensionality increase feature can be the same or different. For the same feature item, the corresponding first dimensionality increase feature and the second dimensionality increase feature can be the same or different.

[0096] Dimension refers to the dimension of a vector or matrix. The dimension of a vector is the number of values ​​included in the vector. For example, the dimension of a vector including 4 values ​​is 4. The dimension of a matrix is ​​determined by the number of rows and columns. Of course, dimensionality increase can also be achieved through a neural network, including but not limited to at least one of a convolutional neural network or a fully connected neural network. Combination includes but is not limited to at least one of splicing or fusion, and fusion includes but is not limited to at least one of addition, subtraction, or multiplication.

[0097] Specifically, taking a feature item as an example, the process of generating the combined feature corresponding to the feature item is explained: the server can fuse the first dimensionality-increasing feature and the third dimensionality-increasing feature corresponding to the feature item to obtain the dimensionality-increasing fusion feature corresponding to the feature item, and splice the dimensionality-increasing fusion feature corresponding to the feature item, the first dimensionality-increasing feature and the third dimensionality-increasing feature corresponding to the feature item to obtain the combined feature corresponding to the feature item.

[0098] In some embodiments, the server may perform further feature extraction on the combined features corresponding to each feature item to obtain a correlation feature corresponding to the feature item. This further feature extraction may be implemented using a neural network, including but not limited to at least one of a fully connected neural network and a convolutional neural network. The correlation feature is used to represent the relationship between the feature item and the encoded features of the currently viewed content.

[0099] In some embodiments, taking a feature item as an example, the process of generating relevant features of the currently viewed content corresponding to the feature item is described: the server can fuse the second dimensionality-increased feature corresponding to the feature item with the associated feature corresponding to the feature item, such as a Hadamard product operation, and determine the result of the operation as the relevant item corresponding to the feature item.

[0100] In some embodiments, the server may perform statistics, such as summing, on the related items corresponding to each feature item to obtain a first statistical feature, and the related features of the currently viewed content may be the first statistical feature. Alternatively, the server may perform dimensionality reduction processing on the first statistical feature to obtain the related features of the currently viewed content. Dimensionality reduction processing is a process that reduces the number of dimensions and can be implemented by downsampling, although the specific downsampling method is not limited.

[0101] In some embodiments, the server can generate relevant features of the currently viewed content through a relevant feature generation network of the currently viewed content. The relevant feature generation network of the currently viewed content is a neural network and is trained. The relevant feature generation network of the currently viewed content includes at least one multi-head network, and the multi-head network is used to perform dimensionality increase processing. For example, the relevant feature generation network of the currently viewed content includes a first multi-head network and a second multi-head network. The first multi-head network is used to perform dimensionality increase processing on the encoding features of the candidate content and the encoding features of the currently viewed content. The second multi-head network is used to perform dimensionality increase processing on the feature items in the feature sequence. The multi-head network can include multiple sub-network units, each network unit has its own parameters, and each sub-network unit is used to transform the features input to the multi-head network into new features. Therefore, when one feature is input into the multi-head network, the multi-head network inputs multiple new features, thereby achieving the purpose of dimensionality increase. For example, if the multi-head network includes m sub-network units, m new features are output, and the m new features constitute the dimensionality increase feature.

[0102] As shown in Figure 4, the structure of the relevant feature generation network is shown. The third dimensionality increase process can be implemented by the first multi-head network. q2 represents the third dimensionality increase feature obtained by performing the third dimensionality increase process on the features of the currently viewed content. The first dimensionality increase process and the second dimensionality increase process can be implemented by the second multi-head network. n represents the number of feature items included in the feature sequence, and k1 to k n Respectively represent the first dimension-increasing features corresponding to the 1st to nth feature items. n Represent the second dimensionality-increasing features corresponding to the 1st to nth feature items respectively. If the second multi-head network includes m sub-network units, the first dimensionality-increasing features and the second dimensionality-increasing features both include m new features, for example, k n It can be expressed as k n =(k n 1 ,k n 2 …k n m ). k n 1 ~k n m are m new features generated by the second multi-head network.

[0103] In some embodiments, the server inputs the feature sequence and the encoded features of the currently viewed content into the relevant feature generation network of the currently viewed content, generates the combined features corresponding to each feature item through the relevant feature generation network of the currently viewed content, and fuses the second dimensionality-increased features and the corresponding combined features corresponding to each feature item through the relevant feature generation network of the currently viewed content, such as Hadamard product operation, to obtain the relevant items corresponding to each feature item, and generates the relevant features of the currently viewed content based on the relevant items corresponding to each feature item.

[0104] In this embodiment, by combining the features after the dimensionality increase process, the combination and fusion are made more in-depth, and the generated relevant features of the currently viewed content are made more accurate.

[0105] In some embodiments, the first dimensionality increase feature and the third dimensionality increase feature corresponding to each feature item are respectively combined to obtain the combined feature corresponding to each feature item, including: fusing the first dimensionality increase feature and the third dimensionality increase feature corresponding to each feature item to obtain the dimensionality increase fusion feature corresponding to each feature item; for each feature item, splicing the dimensionality increase fusion feature corresponding to the feature item, the first dimensionality increase feature and the third dimensionality increase feature corresponding to the feature item to obtain the combined feature corresponding to the feature item.

[0106] Specifically, fusion includes at least one of addition, subtraction, or multiplication. Taking a feature item as an example, the process of generating a combined feature corresponding to the feature item is described as follows: the server may add the first and third dimensionality-increasing features corresponding to the feature item to obtain a second added feature, where addition refers to summing the data at the same position. Alternatively, the server may subtract the first and third dimensionality-increasing features corresponding to the feature item to obtain a second subtracted feature, where subtraction refers to performing a difference calculation on the data at the same position. Alternatively, the server may multiply the first and third dimensionality-increasing features corresponding to the feature item to obtain a second multiplied feature, where multiplication refers to performing a product operation on the data at the same position, i.e., multiplication refers to performing a Hadamard product operation. The server may concatenate at least two of the second added feature, the second subtracted feature, or the second multiplied feature, and use the concatenated result as the combined feature corresponding to the feature item. Alternatively, the server may concatenate at least two of the second added feature, the second subtracted feature, the second multiplied feature, and the first and third dimensionality-increasing features corresponding to the feature item, and use the concatenated result as the combined feature corresponding to the feature item. For example, the server can concatenate the second additive feature, the second subtractive feature, the second multiplicative feature, the first dimensionality-increasing feature corresponding to the feature item, and the third dimensionality-increasing feature, and use the concatenated result as the combined feature corresponding to the feature item. The second additive feature, the second subtractive feature, and the second multiplicative feature are each a dimensionality-increasing fusion feature.

[0107] In this embodiment, the combined features are generated by fusion and splicing, so that the features are deeply combined, thereby improving the useful information contained in the combined features.

[0108] In some embodiments, encoding the current information sequence into a feature sequence includes: encoding each information item in the current information sequence separately to obtain the encoding features of each information item; generating a corresponding weight feature for the encoding features of each information item; weighting the encoding features of each information item with the corresponding weight feature to obtain the feature items corresponding to each information item; and arranging the feature items corresponding to each information item to obtain a feature sequence.

[0109] Specifically, an information item may include multiple pieces of information. For each information item, the server may encode each piece of information in the information item to obtain a code value for each piece of information, and combine the code values ​​for each piece of information into a code feature for the information item. The server may then arrange the code features of each information item according to their order in the current information sequence to obtain a code feature sequence. The server may generate a corresponding weight feature for the code feature of each information item in the code feature sequence. The weight feature corresponding to the code feature of the information item includes the weight corresponding to each code value in the code feature of the information item. For each information item, the server may perform a weighted calculation on the code feature of the information item and the corresponding weight feature, and use the result of the weighted calculation as the feature item corresponding to the information item.

[0110] In some embodiments, the server may multiply the coding value in the coding feature of the information item by the corresponding weight to obtain the weighted value of each coding value, and arrange the weighted values ​​of each coding value according to the arrangement of the coding values ​​in the coding feature to obtain the feature item corresponding to the information item.

[0111] In some embodiments, the server can generate weight features and weight the coding features and weight features based on a neural network. For example, the server can generate weight features and weight the coding features and weight features through a feature extraction network. The feature extraction network is trained. As shown in Figure 5, a structural diagram of the feature extraction network is shown. The feature extraction network includes a weight feature generation network and a weighted operation unit. The coding feature E of the information item is input into the weight feature generation network to obtain the weight feature A corresponding to E. E and A are input into the weighted operation unit. The weighted operation unit multiplies E and A to obtain the feature item corresponding to the information item.

[0112] In some embodiments, the feature extraction network includes a multi-layer sub-network, where multi-layer refers to at least two layers, and each sub-network can be implemented using a fully connected neural network (Multi-Layer Perception, MLP), that is, the sub-network can be a fully connected layer. Taking the feature extraction network including a 2-layer sub-network as an example, the parameters of the first-layer sub-network include w1 and c1, and the parameters of the second-layer sub-network include w2 and c2, and c1 and c2 are both activation functions. w1 is a matrix with f rows and f / c3 columns, w1 is a matrix with f / c3 rows and f columns, and c3 is a reduction ratio parameter. During training, w1 and w2 are parameters that need to be learned. Therefore, the weight feature A can be expressed as: A=Fex(E)=c2(w2c1(w1E)), where Fex represents the feature extraction network.

[0113] In this embodiment, the weight feature can be used to weaken the coding value of the coding feature that contains less information or unimportant information, thereby minimizing unimportant data in the feature items and reducing the interference caused by unimportant data, thereby improving the accuracy of the push.

[0114] In some embodiments, candidate content is selected from at least one candidate content based on the encoding features of each candidate content and the relevant features of the currently viewed content and is pushed to the user, including: for each candidate content, generating comprehensive features corresponding to the candidate content based on the encoding features of the candidate content and the relevant features of the currently viewed content; predicting the recommendation level of each candidate content based on the comprehensive features corresponding to each candidate content; and selecting candidate content from at least one candidate content according to the recommendation level and pushing it to the user.

[0115] Specifically, the server can generate a recommendation score for each of the multiple candidate content items. Taking the generation of a recommendation score for a single candidate content item as an example, the server can combine the encoding features of the candidate content item with the relevant features of the currently viewed content to generate a comprehensive feature corresponding to the candidate content item. Based on the comprehensive features corresponding to the candidate content item, the server can predict the recommendation score of the candidate content item. The server can further extract features related to the currently viewed content item and combine the extracted features with the encoding features of the candidate content item to generate a comprehensive feature corresponding to the candidate content item.

[0116] In some embodiments, the server may select the candidate content with the highest recommendation degree from the multiple candidate content and push the selected candidate content to the user. Alternatively, the server may select a preset number of candidate content from the multiple candidate content in descending order of recommendation degree and push each selected candidate content to the user. The preset number can be set as needed, for example, 5, 10, or 12.

[0117] In this embodiment, since the comprehensive features are generated based on the encoding features of the candidate content and the relevant features of the currently viewed content, the recommendation degree predicted based on the comprehensive features is more in line with the current situation. Therefore, candidate content that meets real-time needs can be selected and pushed based on the recommendation degree, thereby improving the accuracy of the push.

[0118] In some embodiments, the relevant features of the currently viewed content are the relevant features of the first currently viewed content. Generating the comprehensive features corresponding to the candidate content based on the encoding features of the candidate content and the relevant features of the currently viewed content includes: extracting features related to the encoding features of the candidate content from the feature sequence to obtain the relevant features of the second currently viewed content corresponding to the candidate content; generating the comprehensive features corresponding to the candidate content based on the relevant features of the first currently viewed content and the relevant features of the second currently viewed content corresponding to the candidate content.

[0119] The relevant features of the currently viewed content obtained in step 208 (i.e., the step of extracting features related to the coding features of the currently viewed content from the feature sequence to obtain relevant features of the currently viewed content) are referred to as the first relevant features of the currently viewed content. The combined features in the step of "combining each feature item in the feature sequence with the coding features of the currently viewed content to obtain combined features corresponding to each feature item" are referred to as the first combined features. The associated features obtained in the step of "performing further feature extraction on the combined features to obtain associated features corresponding to the feature items" are referred to as the first associated features. The associated items obtained in the step of "fusing each feature item with the corresponding combined features to obtain associated items corresponding to each feature item" are referred to as the first associated items.

[0120] Specifically, the server can perform a fourth dimensionality increase processing on the coding features of the candidate content to obtain a fourth dimensionality increase feature. For each feature item, the server can combine the first dimensionality increase feature corresponding to the feature item with the fourth dimensionality increase feature to obtain a second combined feature corresponding to the feature item. The fourth dimensionality increase processing can be implemented through a first multi-head network. As shown in Figure 4, q1 represents the fourth dimensionality increase feature obtained by performing a fourth dimensionality increase processing on the coding features of the candidate content. For each feature item, the server can perform feature extraction on the second combined feature to obtain a second associated feature. The second associated feature is used to represent the relationship between the feature item and the coding features of the candidate content. The server can fuse the second associated feature with the second dimensionality increase feature to obtain a second related item corresponding to the feature item. The server can perform statistics, such as summation, on the second related item corresponding to each feature item to obtain a second statistical feature, and the related feature of the second currently viewed content can be the second statistical feature. Alternatively, the server can perform dimensionality reduction processing on the second statistical feature to obtain the related feature of the second currently viewed content.

[0121] In some embodiments, as shown in FIG4 , the relevant feature generation network of the currently viewed content also includes an associated feature generation network. As shown in FIG6 , a structural diagram of the associated feature generation network is shown. The associated feature generation network includes a feature combination layer and an associated feature extraction layer. The above-mentioned combined features (first combined features, second combined features) can be generated through the feature combination layer, and the above-mentioned associated features (first associated features, second associated features) can be generated through the associated feature extraction layer. In FIG6 , q can be any one of q1 and q2, and k can be k n 1 ~k n m The correlation feature extraction layer can be implemented using a fully connected neural network (Multi-Layer Perception, MLP) or a convolutional neural network, and the feature combination layer is used to combine the input q and k. The output of the correlation feature generation network can be expressed as f att-out =MLP((q+k),(qk),(q*k),q,k), where MLP stands for correlation feature extraction layer, (q+k),(qk),(q*k) are obtained through the feature combination layer, (q+k),(qk),(q*k),q,k are concatenated to obtain a combination feature (e.g., the first combination feature or the second combination feature), and the combination feature is input into the MLP, i.e., correlation feature extraction layer, to obtain the correlation feature f att-out , f att-out It is a correlation feature such as the first correlation feature or the second correlation feature.

[0122] In some embodiments, in Figure 4, if the data input into the relevant feature generation network of the currently viewed content is a feature sequence and the encoding feature of the currently viewed content, then the relevant feature generation network of the currently viewed content outputs the first relevant feature of the currently viewed content; if the data input into the relevant feature generation network of the currently viewed content is a feature sequence and the encoding feature of the candidate content, then the relevant feature generation network of the currently viewed content outputs the second relevant feature of the currently viewed content. If the data input into the relevant feature generation network of the currently read content is a feature sequence, the coding features of the currently read content, and the coding features of the candidate content, the server can also determine the weighted weights of the relevant features of the first currently read content through the relevant feature generation network of the currently read content to obtain the first weighted weight, determine the weighted weights of the relevant features of the second currently read content corresponding to the candidate content to obtain the second weighted weight, and perform weighted calculation on the relevant features of the first currently read content and the relevant features of the second currently read content using the first weighted weight and the second weighted weight to obtain the relevant features of the target currently read content corresponding to the candidate content. For example, the relevant features of the target currently read content = the first weighted weight * the relevant features of the first currently read content + the second weighted weight * the relevant features of the second currently read content. The comprehensive features corresponding to the candidate content can be the relevant features of the target currently read content corresponding to the candidate content. For example, the relevant features of the target currently read content can be expressed as:

[0123] Among them, g is the second weighted weight, (1-g) is the first weighted weight, E src is the encoding feature of the current reading content, E t is the encoding feature of the candidate content, e i represents the i-th feature item in the feature sequence, f(q2,k i ,v i ) and f att (h(E SRC ),h(e i )) represent the first related item corresponding to the i-th feature item, Represents the relevant features of the first currently read content, f(q1,k i ,v i ) and f att (h(E t ),h(e i )) represent the second related item corresponding to the i-th feature item, Represents the relevant features of the second currently read content, k i On behalf of e i Perform the first dimensionality increase process to obtain the first dimensionality increase feature, v i On behalf of e iPerform a second dimensionality increase process to obtain a second dimensionality increase feature.

[0124] In some embodiments, the server may concatenate the encoded features of the currently viewed content with the relevant features of the target currently viewed content corresponding to the candidate content to obtain a comprehensive feature. Alternatively, the server may further extract features from the relevant features of the target currently viewed content to obtain relevant extracted features, and then obtain a comprehensive feature based on the relevant extracted features. For example, the server may use the relevant extracted features as the comprehensive feature, or may concatenate the encoded features of the currently viewed content with the relevant extracted features to obtain a comprehensive feature corresponding to the candidate content.

[0125] In this embodiment, the relevant features of the second currently viewed content are relevant to the candidate content, so that the comprehensive features are affected by the candidate content, and thus the recommendation degree predicted based on the comprehensive features is affected by the candidate content, thereby improving the rationality of the recommendation degree.

[0126] In some embodiments, generating comprehensive features corresponding to the candidate content based on the encoding features of the candidate content and the relevant features of the currently viewed content includes: fusing the encoding features of the currently viewed content on the basis of the encoding features of the candidate content to obtain content fusion features corresponding to the candidate content; generating comprehensive features corresponding to the candidate content based on the content fusion features corresponding to the candidate content and the relevant features of the currently viewed content.

[0127] Specifically, the server may perform at least one of addition, subtraction, or multiplication of the coding feature of the candidate content and the coding feature of the currently viewed content, and use the result of the calculation as the content fusion feature corresponding to the candidate content.

[0128] In some embodiments, the server can fuse the encoding features of the currently viewed content with the encoding features of the candidate content through a neural network to obtain a content fusion feature corresponding to the candidate content. The server can multiply the encoding features of the currently viewed content with the parameter matrix of the neural network, and then perform a Hadamard product operation on the result of the multiplication with the encoding features of the candidate content to obtain a content fusion feature corresponding to the candidate content. The parameter matrix refers to a matrix composed of parameters in the neural network. For example, if the encoding features of the currently viewed content are E1, the encoding features of the candidate content are E2, and the parameter matrix is ​​W, then the content fusion feature P = Hadamard product operation (E1*W, E2), where the Hadamard product operation (E1*W, E2) represents the calculation of the Hadamard product of E1*W and E2, and P represents the content fusion feature. The process of generating a content fusion feature can also be called a feature intersection process. Taking E1 and E2 as vectors and W as a matrix as an example, Figure 7 shows the principle diagram of generating a content fusion feature. In Figure 7, E1 and E2 are both 4-dimensional vectors, W is a matrix with 4 rows and 4 columns, and the resulting P is a 4-dimensional vector.

[0129] In some embodiments, the server may combine the content fusion features corresponding to the candidate content with the relevant features of the first currently viewed content to generate a comprehensive feature corresponding to the candidate content. Alternatively, the server may further extract features from the relevant features of the first currently viewed content to obtain relevant extracted features corresponding to the relevant features of the first currently viewed content, and combine the relevant extracted features corresponding to the relevant features of the first currently viewed content with the content fusion features corresponding to the candidate content to generate a comprehensive feature corresponding to the candidate content.

[0130] In some embodiments, the server may obtain user description information, including at least one of user ID, user age, user location, user device, points of interest, or user preferred categories. The server may encode the user description information to obtain user features, and then fuse the user features with the encoded features of the candidate content to obtain user fused features corresponding to the candidate content. The process of obtaining user fused features can refer to the process of obtaining content fused features.

[0131] In some embodiments, the server may encode the user's description information to obtain an information encoding feature. The user's description information may include multiple pieces of information. The server may encode each piece of information separately to obtain an encoding value for each piece of information, and combine the encoding values ​​for each piece of information into an information encoding feature. The user's feature may be the information encoding feature.

[0132] In some embodiments, the server may encode the discrete data in the user's description information while keeping the continuous data unchanged to obtain the user's encoded features. Discrete data, such as the user ID, may be encoded features. For discrete data in the user's description information, especially high-dimensional sparse discrete data, the server may generate an embedding vector corresponding to the discrete data.

[0133] In some embodiments, the server may use the information encoding feature as the user's feature. Alternatively, the server may generate a corresponding weight feature for the information encoding feature, where the weight feature includes the weight corresponding to the encoding value of each piece of information in the information encoding feature. The server may then perform a weighted calculation, i.e., a Hadamard product, on the information encoding feature and the corresponding weight feature, and use the result of the calculation as the user's feature. For details on the weighted calculation process, please refer to the relevant content related to obtaining the encoding feature of the currently viewed content.

[0134] In some embodiments, the server may generate a comprehensive feature corresponding to the candidate content based on the user fusion feature corresponding to the candidate content, the content fusion feature corresponding to the candidate content, and the relevant features of the first currently viewed content. The server may concatenate the user fusion feature, the content fusion feature, and the relevant features of the first currently viewed content to generate the comprehensive feature.

[0135] In some embodiments, the server can generate comprehensive features corresponding to the candidate content based on the relevant features of the first currently viewed content, the relevant features of the second currently viewed content corresponding to the candidate content, and the content fusion features corresponding to the candidate content. The server can determine the weighted weights of the relevant features of the first currently viewed content to obtain a first weighted weight, determine the weighted weights of the relevant features of the second currently viewed content corresponding to the candidate content to obtain a second weighted weight, perform weighted calculation on the relevant features of the first currently viewed content and the relevant features of the second currently viewed content using the first weighted weight and the second weighted weight to obtain the relevant features of the target currently viewed content corresponding to the candidate content, and concatenate the relevant features of the target currently viewed content corresponding to the candidate content and the content fusion features corresponding to the candidate content to generate comprehensive features corresponding to the candidate content. Alternatively, the server can perform further feature extraction on the relevant features of the target currently viewed content to obtain target extracted features corresponding to the candidate content, and concatenate the target extracted features corresponding to the candidate content and the content fusion features corresponding to the candidate content to generate comprehensive features corresponding to the candidate content.

[0136] In some embodiments, the server may generate comprehensive features corresponding to the candidate content based on the relevant features of the first currently viewed content, the content fusion features corresponding to the candidate content, and the encoding features of the currently viewed content. The server may concatenate the relevant features of the first currently viewed content, the content fusion features corresponding to the candidate content, and the encoding features of the currently viewed content to generate comprehensive features corresponding to the candidate content. Alternatively, the server may perform further feature extraction on the relevant features of the first currently viewed content to obtain relevant extracted features corresponding to the relevant features of the first currently viewed content, and concatenate the relevant extracted features corresponding to the relevant features of the first currently viewed content, the content fusion features corresponding to the candidate content, and the encoding features of the currently viewed content to generate comprehensive features corresponding to the candidate content.

[0137] In some embodiments, the server may encode the user's context information to obtain encoding features of the context information. Context information refers to the environment in which the currently viewed content is located, including the scene in which the currently viewed content is located in the media platform and information about the device displaying the currently viewed content. The context information includes multiple pieces of information, and the server may encode the multiple pieces of information separately to obtain encoding values ​​for each piece of information, and combine the encoding values ​​for each piece of information into encoding features of the context information. The server may obtain features of the context information based on the encoding features of the context information, for example, using the encoding features of the context information as features of the context information. Alternatively, the server may generate corresponding weight features for the encoding features of the context information, wherein the weight features include weights corresponding to the encoding values ​​of each piece of information in the encoding features of the context information, and the server may perform a weighted calculation, i.e., a Hadamard product operation, on the encoding features of the context information and the corresponding weight features, and use the result of the operation as features of the user's context information. The specific weighted calculation process may refer to the relevant content of obtaining the encoding features of the currently viewed content.

[0138] In some embodiments, the server may fuse features of contextual information with the encoding features of the candidate content to generate contextual fusion features corresponding to the candidate content. For methods for generating contextual fusion features, refer to the method for generating content fusion features. The server may generate comprehensive features corresponding to the candidate content based on the user's contextual fusion features, content fusion features corresponding to the candidate content, and features related to the first currently viewed content. The server may concatenate the contextual fusion features, the content fusion features corresponding to the candidate content, and features related to the first currently viewed content to generate comprehensive features corresponding to the candidate content.

[0139] In some embodiments, the server may combine at least one of content fusion features, user fusion features, context fusion features, and related features of the second currently viewed content corresponding to the candidate content with related features of the first currently viewed content to generate comprehensive features corresponding to the candidate content.

[0140] In some embodiments, the server can splice at least one of the content fusion feature, the user fusion feature, the context fusion feature, and the relevant features of the second currently read content, the relevant features of the first currently read content, and the coding features of the currently read content to generate a comprehensive feature corresponding to the candidate content. The server can splice at least one of the content fusion feature, the user fusion feature, the context fusion feature, the relevant features of the target currently read content corresponding to the candidate content, and the coding features of the currently read content to generate a comprehensive feature corresponding to the candidate content. The content fusion feature, the user fusion feature, and the context fusion feature can be obtained through a feature cross network. The feature cross network is trained, and the principle of the feature cross network is shown in Figure 7. W in Figure 7 is a parameter of the feature cross network. For example, the server can splice the coding features of the candidate content and the coding features of the currently read content and input them into the feature cross network to generate a content fusion feature. The same method can be used to generate user fusion features and context fusion features.

[0141] In this embodiment, the coding features of the currently viewed content are integrated with the coding features of the candidate content, thereby strengthening the influence of the currently viewed content on the generation of comprehensive features, making the recommendation degree predicted based on the comprehensive features more in line with real-time needs and improving the accuracy of real-time push.

[0142] In some embodiments, predicting the recommendation degree of each candidate content based on the comprehensive features corresponding to each candidate content includes: for each candidate content, determining the identification feature of the currently read content to obtain a first identification feature, and determining the identification feature of the candidate content to obtain a second identification feature; splicing the first identification feature and the second identification feature to obtain a spliced ​​identification feature corresponding to the candidate content; and predicting the recommendation degree of the candidate content based on the comprehensive features corresponding to the candidate content and the spliced ​​identification feature corresponding to the candidate content.

[0143] Specifically, the server may determine the user's identification feature to obtain a third identification feature, and then concatenate the first, second, and third identification features to obtain a concatenated identification feature corresponding to the candidate content. The first, second, and third identification features may be obtained by the server through encoding or may be pre-generated.

[0144] In some embodiments, the recommendation degree is obtained based on a recommendation degree prediction network, which may include an identification encoding network for encoding the identification to obtain identification features. The first identification feature, the second identification feature, and the third identification feature may all be generated by the identification encoding network.

[0145] In some embodiments, the server may further extract features from the concatenated identification features corresponding to the candidate content to obtain identification extraction features corresponding to the candidate content. The server may predict the recommendation degree of the candidate content based on the comprehensive features corresponding to the candidate content and the identification extraction features corresponding to the candidate content.

[0146] In this embodiment, based on the identification features of the currently viewed content and the identification features of the candidate content, the spliced ​​identification features corresponding to the candidate content are generated, and then based on the comprehensive features corresponding to the candidate content and the spliced ​​identification features corresponding to the candidate content, the recommendation degree of the candidate content is predicted. Through the spliced ​​identification features, the influence of the currently viewed content and the candidate content on the predicted recommendation degree is enhanced, so that the predicted recommendation degree meets real-time needs and improves the accuracy of the recommendation degree.

[0147] In some embodiments, the recommendation degree is obtained through a recommendation degree prediction network, which includes at least one feature extraction layer, an indicator prediction network corresponding to multiple preset indicators, and an identification feature extraction network; based on the comprehensive features corresponding to the candidate content and the splicing identification features corresponding to the candidate content, predicting the recommendation degree of the candidate content includes: inputting the comprehensive features corresponding to the candidate content into the recommendation degree prediction network, and obtaining the indicator features corresponding to each preset indicator after processing by at least one feature extraction layer; inputting the splicing identification features corresponding to the candidate content into the identification feature extraction network to obtain the identification extraction features; for each preset indicator, inputting the indicator features and identification extraction features corresponding to the preset indicator into the indicator prediction network corresponding to the preset indicator to obtain the predicted value of the preset indicator; determining the recommendation degree of the candidate content based on the predicted value of each preset indicator.

[0148] Among them, the recommendation prediction network is used to predict the values ​​of multiple preset indicators of the content. The preset indicators include but are not limited to at least one of the click-through rate, the number of subsequent content readings, or the subsequent reading time. Among them, the number of subsequent content readings refers to the number of contents related to the candidate content that the user reads when the candidate content is pushed to the user. Taking video as an example, the number of subsequent content readings can be called the number of subsequent video consumption. The subsequent reading time refers to the length of time the user reads the contents related to the candidate content when the candidate content is pushed to the user. The recommendation prediction network includes at least one feature extraction layer. Multiple preset indicators refer to at least two preset indicators. The recommendation prediction network includes an indicator prediction network corresponding to each preset indicator, and the output result of the indicator prediction network corresponding to the preset indicator represents the predicted value of the preset indicator. The identification feature extraction network is used to further extract the splicing identification features corresponding to the candidate content.

[0149] Specifically, the process of generating a prediction value is explained by taking a preset indicator as an example: the indicator prediction network can have multiple layers of cascaded fully connected layers, and the server can input the indicator features corresponding to the preset indicator into the corresponding indicator prediction network. Whenever the output result of a fully connected layer is obtained, the output result is fused with the identification extraction feature, and then the fusion result is input into the fully connected layer of the next layer. After the output result of the last fully connected layer is obtained, the output result of the last fully connected layer is fused with the identification extraction feature to obtain the prediction value of the preset indicator. For example, the numerical value represented by the fusion result can be determined to obtain the prediction value of the preset indicator. Wherein, fusion includes but is not limited to at least one of multiplication or addition, for example, it can be a Hadamard product operation.

[0150] In some embodiments, the identification feature extraction network is implemented using a neural network. The identification feature extraction network may include multiple neural network layers. Each neural network layer may be implemented using a fully connected neural network or a convolutional neural network. The neural network layer may also include an activation function. The activation functions in each neural network layer may be the same or different. The activation function includes at least one of relu or sigmoid. The identification feature extraction network may also include a gain unit. The gain unit has a gain coefficient. The gain coefficient may be preset or modified as needed. Taking the identification feature extraction network including two neural network layers as an example, as shown in FIG8 , the structural diagram of the identification feature extraction network is shown. The identification feature extraction network in FIG8 can be expressed as: x=RelU(input*W1+b1), RP output =r*sigmoid(x*W2+b2),RP output The value range is [0, r], r is the gain coefficient, which can be preset to 2, for example, x is the output result of the first neural network layer, input represents the input data of the first neural network layer, i.e., the splicing identification feature, W1 and b1 are the parameters of the first neural network layer, RelU is the activation function of the first neural network layer, W2 and b2 are the parameters of the second neural network layer, sigmoid is the activation function of the second neural network layer, RP output To identify the output of the feature extraction network, input and RP output Both are vectors or matrices, and input and RP output The dimensions are the same.

[0151] In some embodiments, the server may collect statistics on the predicted values ​​of each preset indicator to obtain the recommendation degree of the candidate content. The recommendation degree of the candidate content is positively correlated with the predicted value of each preset indicator.

[0152] In some embodiments, the server may multiply the predicted values ​​of the preset indicators to obtain the recommendation level of the candidate content. For example, the server may multiply the predicted values ​​of the click-through rate, the number of subsequent video consumption, and the subsequent consumption time to obtain the recommendation level of the candidate content.

[0153] In some embodiments, the server can multiply the predicted values ​​of each preset indicator by an exponential factor to obtain the recommendation degree of the candidate content. For example, the server can use the formula ranking score =pclk t1 *pvv t2 *pdur t3 Calculate the recommendation degree of candidate content, ranking score represents the recommendation level of the candidate content, pclk represents the predicted click-through rate, pvv represents the predicted number of subsequent content views, and pdur represents the predicted subsequent viewing duration. t1, t2, and t3 are the indices of pclk, pvv, and pdur, respectively. t1, t2, and t3 can be set as needed.

[0154] In this embodiment, the recommendation degree prediction network predicts the predicted values ​​of multiple preset indicators, and thus determines the recommendation degree of the candidate content according to the predicted values ​​of the preset indicators, thereby automatically generating the recommendation degree and improving the efficiency of generating the recommendation degree.

[0155] In some embodiments, each indicator prediction network corresponds to at least one identification feature extraction network; inputting the spliced ​​identification features corresponding to the candidate content into the identification feature extraction network to obtain the identification extraction features includes: for each indicator prediction network, inputting the spliced ​​identification features corresponding to the candidate content into each identification feature extraction network corresponding to the indicator prediction network, and obtaining the identification extraction features respectively output by each identification feature extraction network corresponding to the indicator prediction network; inputting the indicator features and identification extraction features corresponding to the preset indicator into the indicator prediction network corresponding to the preset indicator, and obtaining the predicted value of the preset indicator includes: inputting the indicator features corresponding to the preset indicator and each identification extraction features corresponding to the preset indicator into the indicator prediction network corresponding to the preset indicator, and obtaining the predicted value of the preset indicator.

[0156] The identification feature extraction networks corresponding to each indicator prediction network can be the same or different. For example, a recommendation prediction network includes two indicator prediction networks for preset indicators, each of which corresponds to a feature extraction network. As shown in FIG9 , the structure of the recommendation prediction network is shown. In FIG9 , the recommendation prediction network includes m feature extraction layers, where m ≥ 1, and includes two indicator prediction networks, namely, a first indicator prediction network and a second indicator prediction network. The first indicator prediction network is the indicator prediction network corresponding to the first preset indicator, and the second indicator prediction network is the indicator prediction network corresponding to the second preset indicator. The first identification feature extraction network is the identification feature extraction network corresponding to the first indicator prediction network, and the second identification feature extraction network is the identification feature extraction network corresponding to the second indicator prediction network.

[0157] The identification extraction features corresponding to the preset indicators refer to the identification extraction features respectively output by the identification feature extraction networks corresponding to the indicator prediction network of the preset indicators.

[0158] Specifically, each indicator prediction network includes multiple fully connected layers, each of which utilizes a fully connected neural network. Each fully connected layer in the indicator prediction network can have a one-to-one correspondence with a signature feature extraction network. Different indicator prediction networks correspond to different signature feature extraction networks. For example, if the first indicator prediction network includes two fully connected layers, each of these fully connected layers corresponds to a different signature feature extraction network. The process of generating a predicted value for a preset indicator is illustrated using the example of a server. The indicator prediction network can include multiple cascaded fully connected layers. The server can input the indicator features corresponding to the preset indicator into the corresponding indicator prediction network. After obtaining the output of each fully connected layer, the server can fuse the output with the signature extraction features output by the signature extraction features corresponding to that fully connected layer. The fused result is then input into the next fully connected layer. After obtaining the output of the last fully connected layer, the server can fuse the output of the last fully connected layer with the signature extraction features output by the signature extraction features corresponding to the last fully connected layer to obtain the predicted value for the preset indicator. For example, the server can determine the numerical value represented by the fused result to obtain the predicted value for the preset indicator. The fusion includes but is not limited to at least one of multiplication and addition, for example, it can be a Hadamard product operation.

[0159] In some embodiments, the feature extraction layer in the recommendation prediction network includes feature extraction networks corresponding to each preset indicator. For example, in FIG9 , the first feature extraction network includes: a feature extraction network corresponding to the first preset indicator, i.e., a first indicator feature extraction network, and a feature extraction network corresponding to the second preset indicator, i.e., a second indicator feature extraction network. The feature extraction layer also includes a shared network and a fusion network corresponding to each feature extraction network. For example, in FIG9 , the fusion network corresponding to the first indicator feature extraction network, i.e., a first fusion network, and the fusion network corresponding to the second indicator feature extraction network, i.e., a second fusion network. The fusion network is used to fuse the output results of the corresponding feature extraction network with the output results of the shared network. The fusion includes but is not limited to at least one of weighting, multiplication, or addition.

[0160] In some embodiments, the recommendation prediction network can be a network that optimizes the PLE structure, for example, based on the PLE structure, an identification feature extraction network is added. The feature extraction layer can be implemented using an expert network layer (CGC, Customized Gate Control).

[0161] In this embodiment, since each indicator prediction network corresponds to at least one identification feature extraction network, the output result of the indicator prediction network can be influenced by different identification extraction features, thereby reducing errors and improving the accuracy of the prediction value.

[0162] In some embodiments, the content push method provided by the present application can be based on a recommendation prediction model. As shown in FIG10 , which shows a structural diagram of the recommendation prediction model, the server inputs the user's description information, context information, description information of the candidate content, description information of the currently viewed content, and the current information sequence into the recommendation prediction model for encoding, thereby obtaining information encoding features, encoding features of context information, encoding features of the candidate content, encoding features of the currently viewed content, and an encoding feature sequence. The server inputs the information encoding features, encoding features of context information, encoding features of the candidate content, encoding features of the currently viewed content, and an encoding feature sequence into a feature extraction network to generate user features, features of context information, encoding features of the candidate content, encoding features of the currently viewed content, and a feature sequence. The server inputs the feature sequence, encoding features of the candidate content, and encoding features of the currently viewed content into a related feature generation network for the currently viewed content to generate related features of the target currently viewed content. The server concatenates the user's features with the candidate content's features and inputs them into a feature cross network to generate user fusion features. It concatenates the context information features with the candidate content's encoding features and inputs them into a feature cross network to generate context fusion features. It concatenates the encoding features of the currently viewed content with the encoding features of the candidate content and inputs them into a feature cross network to generate content fusion features. The server inputs the relevant features of the target currently viewed content into a feedforward neural network to obtain target extraction features. The server inputs the target extraction features, encoding features of the currently viewed content, user fusion features, context fusion features, and content fusion features into a concatenation layer for concatenation to obtain comprehensive features corresponding to the candidate content. The comprehensive features of the candidate content are input into a recommendation prediction network to obtain predicted values ​​for multiple preset indicators of the candidate content and generate a recommendation degree for the candidate content based on each predicted value. The description of each component of the recommendation prediction model has been previously described and will not be repeated here. This description will focus solely on the overall process of obtaining recommendation values ​​based on the recommendation prediction model.

[0163] In some embodiments, as shown in FIG11 , a schematic diagram of a principle for training a recommendation prediction model is shown. The process of training a recommendation prediction model is as follows:

[0164] The server can identify a sample user from multiple users and obtain the content viewed by the sample user at a historical moment as the sample viewing content. It can also obtain the content pushed to the sample user in real time when the sample viewing content is triggered, and obtain the label value of the pushed content under multiple preset indicators. The label value refers to the actual value generated by the sample user for the pushed content, such as the actual click-through rate. The server can also obtain the historical information sequence corresponding to the sample user stored in the server at the historical moment, as well as the sample user's descriptive information, contextual information, etc., and form training data based on the obtained data. During the training process, the server can input the training data into the recommendation prediction model to obtain the predicted value corresponding to the pushed content. For the specific process of obtaining the predicted value of the pushed content, refer to the method for generating the predicted value of the candidate content in the usage phase.

[0165] The server can determine the label value of each pushed content under multiple preset indicators. For preset indicators with discrete values, such as click-through rate, the classification loss function and label value can be used to calculate the loss value corresponding to the preset indicator. For preset indicators with continuous values, the server can calculate the difference between the predicted value of the preset indicator and the label value of the preset indicator to obtain the loss value corresponding to the preset indicator, and the loss value corresponding to the preset indicator is positively correlated with the difference. The server can update the parameters of the recommendation prediction model based on the loss values ​​corresponding to each preset indicator. For example, the stochastic gradient descent method can be used to update the parameters and iteratively train until the recommendation prediction model converges, thereby completing the training of the recommendation training model.

[0166] In the case where the value of the preset indicator is a continuous value, the label value of the preset indicator can be generated by a generalized normalization method. For example, if the value of the subsequent reading time is continuous, the actual value of the subsequent reading time can be generalized normalized to obtain the label value of the subsequent reading time. For example, the calculation formula is: label value of subsequent reading time = log2((play_dur / quantile_value)+1). Among them, play_dur represents the actual value of the subsequent reading time, quantile_value is a preset value, and quantile_value can be, for example, the average or median of the reading time of content of the same type. The type of content can be divided as needed. For example, the type of video can be divided into TV series, children's, etc.

[0167] Since the subsequent reading time has a causal relationship with clicks, that is, users will consume time only after clicking. Therefore, when the predicted value of click rate and the predicted value of subsequent reading time are predicted, a new predicted value of subsequent reading time can be generated based on the predicted value of click rate and the predicted value of subsequent reading time. Based on the difference between the new predicted value and the label value of subsequent reading time, the loss value corresponding to the subsequent reading time is generated. For example, the formula can be used, loss 时长 =loss(pclk*P 时长 , label 时长 ) Calculate the loss value corresponding to the subsequent reading time. 时 长 refers to the loss value corresponding to the subsequent reading time, pclk is the predicted value of click rate, P 时长 is the predicted value of subsequent reading time, label 时长 The tag value for the subsequent reading duration.

[0168] In some embodiments, all or part of the network used in training the recommendation prediction model may be pre-trained, and then the entire recommendation prediction model may be fine-tuned to complete the training of the recommendation prediction model. Alternatively, the server may pre-train information encoding features, contextual information encoding features, content encoding features, or encoding feature sequences. When training the recommendation prediction model, the pre-generated encoding features are directly obtained for training.

[0169] In this embodiment, the recommendation value of the candidate content is accurately predicted by the recommendation degree prediction model, thereby improving the push accuracy.

[0170] In some embodiments, as shown in FIG12 , a content push method is provided. The method can be executed by a terminal or a server, or can be executed by both the terminal and the server. The method is described by taking the server in FIG1 as an example, and includes the following steps 1202 to 1218. Among them:

[0171] Step 1202, obtain the description information of the user's current reading content, obtain the user's historical information sequence, generate the information item corresponding to the current reading content based on the description information of the current reading content, and generate the user's current information sequence based on the information item corresponding to the current reading content and the historical information sequence.

[0172] The historical information sequence includes information items corresponding to the user's historically viewed content, and the information items corresponding to the historically viewed content include description information of the historically viewed content.

[0173] Step 1204: Encode each information item in the current information sequence to obtain a coding feature of each information item, generate a corresponding weight feature for the coding feature of each information item, weight the coding feature of each information item with the corresponding weight feature to obtain a feature item corresponding to each information item, and arrange the feature items corresponding to each information item to obtain a feature sequence.

[0174] The feature sequence includes feature items that correspond one-to-one to the information items in the current information sequence.

[0175] Step 1206 : Obtain the coding feature of the currently viewed content, and respectively encode the description information of at least one candidate content to obtain the coding feature of each candidate content.

[0176] The encoding feature of the currently viewed content is obtained by encoding the description information of the currently viewed content.

[0177] Step 1208 : extracting features related to the coding features of the currently viewed content from the feature sequence to obtain first features related to the currently viewed content.

[0178] The server may perform a first dimensionality increase processing on each feature item in the feature sequence to obtain a first dimensionality increase feature corresponding to each feature item, perform a second dimensionality increase processing on each feature item to obtain a second dimensionality increase feature corresponding to each feature item, perform a third dimensionality increase processing on the encoding feature of the currently viewed content to obtain a third dimensionality increase feature, and combine the first dimensionality increase feature and the third dimensionality increase feature corresponding to each feature item to obtain a combined feature corresponding to each feature item. The server may fuse the second dimensionality increase feature and the corresponding combined feature corresponding to each feature item to obtain a related item corresponding to each feature item, and generate a related feature of the currently viewed content based on the related item corresponding to each feature item.

[0179] Step 1210 : For each candidate content, extract features related to the encoding features of the candidate content from the feature sequence to obtain related features of the second currently viewed content corresponding to the candidate content.

[0180] Step 1212 : For each candidate content, weighted calculation is performed on the relevant features of the first currently viewed content and the relevant features of the second currently viewed content corresponding to the candidate content to obtain the relevant features of the target currently viewed content corresponding to the candidate content.

[0181] Among them, the server can also determine the weighted weights of the relevant features of the first currently read content to obtain the first weighted weight, determine the weighted weights of the relevant features of the second currently read content corresponding to the candidate content to obtain the second weighted weight, and perform weighted calculation on the relevant features of the first currently read content and the relevant features of the second currently read content using the first weighted weight and the second weighted weight to obtain the relevant features of the target currently read content corresponding to the candidate content.

[0182] Step 1214: For each candidate content, the encoding features of the candidate content are fused with the encoding features of the currently viewed content to obtain the content fusion features corresponding to the candidate content. Based on the relevant features of the target currently viewed content corresponding to the candidate content and the content fusion features corresponding to the candidate content, the comprehensive features corresponding to the candidate content are generated.

[0183] Step 1216: For each candidate content, determine the identification feature of the currently viewed content to obtain a first identification feature, and determine the identification feature of the candidate content to obtain a second identification feature. Concatenate the first identification feature and the second identification feature to obtain a concatenated identification feature corresponding to the candidate content. Based on the comprehensive features corresponding to the candidate content and the concatenated identification feature corresponding to the candidate content, predict the recommendation degree of the candidate content.

[0184] Step 1218 : Select candidate content from the candidate content based on the recommendation degree and push it to the user.

[0185] In this embodiment, since the current information sequence includes information items corresponding to multiple user-viewed contents, each information item includes descriptive information of the corresponding viewed contents, and the multiple viewed contents include the currently viewed content and historically viewed contents, and since the currently viewed content reflects the user's real-time interests, the current information sequence includes information that can reflect the user's real-time interests, thereby improving the accuracy of real-time push notifications. Furthermore, features related to the encoding features of the currently viewed content are extracted from the feature sequence to obtain relevant features of the currently viewed content. Since the relevant features of the currently viewed content are strongly related to the currently viewed content, candidate content is selected from at least one candidate content based on the encoding features of each candidate content and the relevant features of the currently viewed content, and is pushed to the user. This strengthens the influence of the currently viewed content in content push notifications, makes push notifications more in line with real-time needs, and improves the accuracy of real-time push notifications. The content push method proposed in this application effectively solves the problems of insufficient relevance and insufficient characterization of historical and real-time interests in recommendation scenarios, thereby improving the accuracy of real-time push notifications.

[0186] The content push method provided by the present application can be applied to the recommendation system of any media platform and is used to push content in the media platform. Taking the video platform as an example, when the server corresponding to the video platform determines that the user has triggered a play operation for any video in the video platform, it determines the video corresponding to the play operation and obtains the currently played video. The currently played video is the above-mentioned currently viewed content. The server obtains the description information of the currently played video and the user's historical information sequence. The historical information sequence includes information items corresponding to the user's historically played videos. The information corresponding to the historically played videos includes description information of the historically played videos. The server generates information items corresponding to the currently played video based on the description information of the currently played video, and generates the user's current information sequence based on the information items corresponding to the currently played video and the historical information sequence. The server encodes the current information sequence into a feature sequence. The feature sequence includes feature items that correspond one-to-one to the information items in the current information sequence. The server obtains the features of the currently played video, extracts features related to the features of the currently played video from the feature sequence, and obtains relevant features of the currently viewed content. The server encodes the description information of at least one candidate video separately to obtain features of each candidate video. For each candidate video, the server predicts its recommendation score based on the features of the candidate video and the relevant features of the currently viewed content. Based on the recommendation score, the server selects a candidate video from the candidate videos and pushes the selected candidate video to the user. The content push method provided in this application is applied to video push, better meeting real-time requirements and improving the accuracy of real-time video push.

[0187] Through experiments in long video scenarios, the advantages of the content push method proposed in this application are obtained, as shown in Table 1.

[0188] Table 1

[0189] VV (VisitView) refers to the number of subsequent video consumption, which can also be understood as the number of user visits. CTR refers to the click-through rate. AUC (Area Under Curve) is the area under the ROC (Receiver Operating Characteristic Curve) curve and the coordinate axis, and is a model evaluation metric.

[0190] It should be understood that, although the steps in the flowcharts of the above embodiments are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the flowcharts of the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily to be performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0191] Based on the same inventive concept, the present application also provides a content push device for implementing the aforementioned content push method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more content push device embodiments provided below can be found in the above-mentioned limitations of the content push method and will not be repeated here.

[0192] In some embodiments, as shown in FIG13 , a content push device is provided, including: a sequence acquisition module 1302 , a first encoding module 1304 , a feature acquisition module 1306 , a feature extraction module 1308 , a second encoding module 1310 , and a push module 1312 , wherein:

[0193] Sequence acquisition module 1302 is used to obtain the user's current information sequence. The current information sequence includes multiple information items, each of which corresponds to a reading content of the user. The multiple information items include an information item corresponding to the current reading content and an information item corresponding to at least one historical reading content. The information items in the current information sequence are arranged according to the reading order of the corresponding reading content.

[0194] The first encoding module 1304 is configured to encode the current information sequence into a feature sequence, where the feature sequence includes feature items corresponding one-to-one to the information items in the current information sequence, and each feature item represents an encoding feature obtained by encoding the corresponding information item.

[0195] The feature acquisition module 1306 is used to acquire the coding features of the currently viewed content. The coding features of the currently viewed content are obtained by encoding the description information of the currently viewed content.

[0196] The feature extraction module 1308 is used to extract features related to the coding features of the currently viewed content from the feature sequence to obtain relevant features of the currently viewed content.

[0197] The second encoding module 1310 is configured to obtain description information of at least one candidate content, encode the description information of at least one candidate content, and obtain encoding features of each candidate content.

[0198] The push module 1312 is configured to select candidate content from at least one candidate content based on the encoding features of each candidate content and the relevant features of the currently viewed content, and push the selected candidate content to the user.

[0199] The content push device provided by the present application has a current information sequence including information items corresponding to multiple viewing contents of the user, each of which includes descriptive information of the corresponding viewing contents. The multiple viewing contents include the current viewing content and the historical viewing contents. Since the current viewing content reflects the real-time interests of the user, the current information sequence includes information that can reflect the real-time interests of the user, thereby improving the accuracy of the real-time push. Features related to the coding features of the current viewing content are extracted from the feature sequence to obtain the relevant features of the current viewing content. Since the relevant features of the current viewing content are strongly related to the current viewing content, candidate content is selected from at least one candidate content based on the coding features of each candidate content and the relevant features of the current viewing content, and is pushed to the user, thereby enhancing the influence of the current viewing content in the content push, making the push more in line with real-time needs, and improving the accuracy of the real-time push.

[0200] In some embodiments, the sequence acquisition module 1302 is further used to: obtain descriptive information of the user's current reading content; obtain the user's historical information sequence, the historical information sequence including information items corresponding to the user's historical reading content, the information items corresponding to the historical reading content including descriptive information of the historical reading content; generate information items corresponding to the current reading content based on the descriptive information of the current reading content; generate the user's current information sequence based on the information items corresponding to the current reading content and the historical information sequence.

[0201] In some embodiments, the feature extraction module 1308 is further used to: combine each feature item in the feature sequence with the encoding features of the currently viewed content to obtain the combined features corresponding to each feature item; fuse each feature item with the corresponding combined features to obtain the related items corresponding to each feature item, and generate the related features of the currently viewed content based on the related items corresponding to each feature item.

[0202] In some embodiments, the feature extraction module 1308 is further used to: perform a first dimensionality increase processing on each feature item in the feature sequence to obtain a first dimensionality increase feature corresponding to each feature item, and perform a second dimensionality increase processing on each feature item to obtain a second dimensionality increase feature corresponding to each feature item; perform a third dimensionality increase processing on the coding feature of the currently viewed content to obtain a third dimensionality increase feature; combine the first dimensionality increase feature and the third dimensionality increase feature corresponding to each feature item to obtain a combined feature corresponding to each feature item; fuse the second dimensionality increase feature and the corresponding combined feature corresponding to each feature item to obtain a feature of each feature item related to the coding feature of the currently viewed content, and obtain a related item corresponding to each feature item; and generate related features of the currently viewed content based on the related item corresponding to each feature item.

[0203] In some embodiments, the feature extraction module 1308 is further used to: fuse the first dimensionality increase feature and the third dimensionality increase feature corresponding to each feature item respectively to obtain the dimensionality increase fusion feature corresponding to each feature item; for each feature item, splice the dimensionality increase fusion feature corresponding to the feature item, the first dimensionality increase feature and the third dimensionality increase feature corresponding to the feature item to obtain the combined feature corresponding to the feature item.

[0204] In some embodiments, the first encoding module 1304 is further used to: encode each information item in the current information sequence separately to obtain the encoding features of each information item; generate a corresponding weight feature for the encoding feature of each information item; weight the encoding feature of each information item with the corresponding weight feature to obtain the feature item corresponding to each information item; and obtain a feature sequence based on the arrangement of the feature items corresponding to each information item.

[0205] In some embodiments, the push module 1312 is further used to: generate comprehensive features corresponding to each candidate content based on the encoding features of the candidate content and the relevant features of the currently read content; predict the recommendation level of each candidate content based on the comprehensive features corresponding to each candidate content; select candidate content from at least one candidate content according to the recommendation level and push it to the user.

[0206] In some embodiments, the relevant features of the currently read content are the relevant features of the first currently read content. The push module 1312 is further used to: extract features related to the encoding features of the candidate content from the feature sequence to obtain the relevant features of the second currently read content corresponding to the candidate content; and generate comprehensive features corresponding to the candidate content based on the relevant features of the first currently read content and the relevant features of the second currently read content corresponding to the candidate content.

[0207] In some embodiments, the push module 1312 is further used to: fuse the encoding features of the candidate content with the encoding features of the currently viewed content to obtain content fusion features corresponding to the candidate content; and generate comprehensive features corresponding to the candidate content based on the content fusion features corresponding to the candidate content and relevant features of the currently viewed content.

[0208] In some embodiments, the push module 1312 is further used to: for each candidate content, determine the identification feature of the currently viewed content to obtain a first identification feature, and determine the identification feature of the candidate content to obtain a second identification feature; splice the first identification feature and the second identification feature to obtain a spliced ​​identification feature corresponding to the candidate content; and predict the recommendation degree of the candidate content based on the comprehensive features corresponding to the candidate content and the spliced ​​identification features corresponding to the candidate content.

[0209] In some embodiments, the recommendation degree is obtained through a recommendation degree prediction network, which includes at least one feature extraction layer, an indicator prediction network corresponding to multiple preset indicators, and an identification feature extraction network; the push module 1312 is also used to: input the comprehensive features corresponding to the candidate content into the recommendation degree prediction network, and obtain the indicator features corresponding to each preset indicator after processing by at least one feature extraction layer; input the splicing identification features corresponding to the candidate content into the identification feature extraction network to obtain the identification extraction features; for each preset indicator, input the indicator features and identification extraction features corresponding to the preset indicator into the indicator prediction network corresponding to the preset indicator to obtain the predicted value of the preset indicator; determine the recommendation degree of the candidate content based on the predicted value of each preset indicator.

[0210] In some embodiments, each indicator prediction network corresponds to at least one identification feature extraction network; the push module 1312 is also used to: for each indicator prediction network, input the splicing identification features corresponding to the candidate content into each identification feature extraction network corresponding to the indicator prediction network, and obtain the identification extraction features respectively output by each identification feature extraction network corresponding to the indicator prediction network; input the indicator features corresponding to the preset indicator and each identification extraction features corresponding to the preset indicator into the indicator prediction network corresponding to the preset indicator, and obtain the predicted value of the preset indicator.

[0211] Each module in the above-mentioned content push device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor of the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0212] In some embodiments, a computer device is provided, which may be a server, and its internal structure diagram may be shown in Figure 14. The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store data involved in the content push method proposed in this application. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a content push method is implemented.

[0213] In some embodiments, a computer device is provided, which may be a terminal, and its internal structure diagram may be shown in Figure 15. The computer device includes a processor, memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals via wired or wireless means, and the wireless means may be implemented via Wi-Fi, a mobile cellular network, NFC (near field communication), or other technologies. When executed by the processor, the computer program implements a content push method. The display unit of the computer device is used to form a visually visible image, and may be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse.

[0214] Those skilled in the art will understand that the structures shown in Figures 14 and 15 are merely block diagrams of partial structures related to the solution of the present application, and do not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figures, or combine certain components, or have a different component arrangement.

[0215] In some embodiments, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned content push method when executing the computer program.

[0216] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned content push method are implemented.

[0217] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above-mentioned content push method when executed by a processor.

[0218] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant legal provisions.

[0219] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.

[0220] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0221] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

[0222] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0223] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A content push method, executed by a computer device, comprising: Acquire a current information sequence of a user, wherein the current information sequence includes a plurality of information items, each of which corresponds to a content viewed by the user, the plurality of information items including an information item corresponding to the current content viewed and an information item corresponding to at least one content viewed in history, and the information items in the current information sequence are arranged in a viewing order of the corresponding contents viewed; Encoding the current information sequence into a feature sequence, wherein the feature sequence includes feature items corresponding one to one to the information items in the current information sequence, and each feature item represents a coding feature obtained by encoding the corresponding information item; Acquire the coding feature of the currently viewed content, where the coding feature of the currently viewed content is obtained by encoding the description information of the currently viewed content; Extracting features related to the encoding features of the currently viewed content from the feature sequence to obtain relevant features of the currently viewed content; Acquire description information of at least one candidate content, encode the description information of the at least one candidate content respectively, and obtain encoding features of each candidate content; and According to the encoding feature of each candidate content and the related features of the currently viewed content, candidate content is selected from the at least one candidate content, and the selected candidate content is pushed to the user.

2. According to the method of claim 1, the obtaining of the user's current information sequence comprises: Obtaining description information of the content currently being viewed by the user; Acquire the user's historical information sequence, wherein the historical information sequence includes information items corresponding to the user's historically viewed content, and the information items corresponding to the historically viewed content include description information of the historically viewed content; Generate information items corresponding to the currently viewed content according to the description information of the currently viewed content; The current information sequence of the user is generated according to the information item corresponding to the currently viewed content and the historical information sequence.

3. The method according to claim 1 or 2, wherein extracting features related to the coding features of the currently viewed content from the feature sequence to obtain the related features of the currently viewed content comprises: Combining each feature item in the feature sequence with the coding feature of the currently viewed content to obtain a combined feature corresponding to each feature item; Respectively fusing each of the feature items with the corresponding combined features to obtain related items corresponding to each of the feature items; Generate relevant features of the currently viewed content based on the relevant items corresponding to each of the feature items.

4. The method according to claim 3, wherein each feature item in the feature sequence is combined with the coding feature of the currently viewed content to obtain the combined feature corresponding to each feature item, comprising: Performing a first dimension increase processing on each feature item in the feature sequence to obtain a first dimension increase feature corresponding to each feature item, and performing a second dimension increase processing on each feature item to obtain a second dimension increase feature corresponding to each feature item; Performing a third dimension-increasing process on the encoding feature of the currently viewed content to obtain a third dimension-increasing feature; Respectively combining the first dimension-increasing feature and the third dimension-increasing feature corresponding to each of the feature items to obtain a combined feature corresponding to each of the feature items; The step of fusing each of the feature items with the corresponding combined features to obtain the related items corresponding to each of the feature items comprises: The second dimension-increased features and the corresponding combined features corresponding to each of the feature items are fused respectively to obtain the related items corresponding to each of the feature items.

5. According to the method of claim 4, the step of combining the first dimension-increasing feature and the third dimension-increasing feature corresponding to each of the feature items to obtain the combined feature corresponding to each of the feature items comprises: Respectively fusing the first dimension-increasing feature corresponding to each of the feature items with the third dimension-increasing feature to obtain the dimension-increasing fusion feature corresponding to each of the feature items; For each of the feature items, the dimension-increasing fusion feature corresponding to the feature item and the first dimension-increasing fusion feature corresponding to the feature item are combined. The first dimension feature and the third dimension-increasing feature are concatenated to obtain a combined feature corresponding to the feature item.

6. The method according to any one of claims 1 to 5, wherein encoding the current information sequence into a feature sequence comprises: Encode each information item in the current information sequence respectively to obtain a coding feature of each information item; Generating a corresponding weight feature for the encoding feature of each of the information items; Weighting the encoding feature of each of the information items with the corresponding weight feature to obtain feature items corresponding to each of the information items; The feature sequence is obtained based on the arrangement of feature items corresponding to each of the information items.

7. The method according to any one of claims 1 to 6, wherein selecting candidate content from the at least one candidate content and pushing the candidate content to the user based on the encoding feature of each candidate content and the related features of the currently viewed content comprises: For each of the candidate contents, generating a comprehensive feature corresponding to the candidate content according to the encoding feature of the candidate content and the relevant features of the currently viewed content; Predicting the recommendation degree of each of the candidate contents based on the comprehensive features corresponding to each of the candidate contents; A candidate content is selected from the at least one candidate content according to the recommendation degree and is pushed to the user.

8. The method according to claim 7, wherein the relevant features of the currently viewed content are relevant features of the first currently viewed content, and generating the comprehensive features corresponding to the candidate content according to the encoding features of the candidate content and the relevant features of the currently viewed content comprises: Extracting features related to the encoding features of the candidate content from the feature sequence to obtain relevant features of the second currently viewed content corresponding to the candidate content; Based on the relevant features of the first currently viewed content and the relevant features of the second currently viewed content corresponding to the candidate content, a comprehensive feature corresponding to the candidate content is generated.

9. The method according to claim 7 or 8, wherein generating the comprehensive features corresponding to the candidate content according to the encoding features of the candidate content and the relevant features of the currently viewed content comprises: Based on the coding features of the candidate content, the coding features of the currently viewed content are integrated to obtain content integration features corresponding to the candidate content; Based on the content fusion features corresponding to the candidate content and the relevant features of the currently viewed content, a comprehensive feature corresponding to the candidate content is generated.

10. The method according to any one of claims 7 to 9, wherein the predicting the recommendation degree of each candidate content based on the comprehensive features corresponding to each candidate content comprises: For each of the candidate contents, determining a feature of the identification of the currently viewed content to obtain a first identification feature, and determining a feature of the identification of the candidate contents to obtain a second identification feature; Concatenate the first identification feature and the second identification feature to obtain a concatenated identification feature corresponding to the candidate content; Based on the comprehensive features corresponding to the candidate content and the splicing identification features corresponding to the candidate content, the recommendation degree of the candidate content is predicted.

11. The method according to claim 10, wherein the recommendation degree is obtained through a recommendation degree prediction network, wherein the recommendation degree prediction network comprises at least one feature extraction layer, an indicator prediction network corresponding to a plurality of preset indicators, and an identification feature extraction network; The predicting the recommendation degree of the candidate content based on the comprehensive features corresponding to the candidate content and the splicing identification features corresponding to the candidate content includes: Inputting the comprehensive features corresponding to the candidate content into the recommendation degree prediction network, and obtaining the indicator features corresponding to each of the preset indicators after being processed by the at least one feature extraction layer; Inputting the splicing identification features corresponding to the candidate content into the identification feature extraction network to obtain identification extraction features; For each of the preset indicators, the indicator features corresponding to the preset indicator and the identification extraction features are input Go to the indicator prediction network corresponding to the preset indicator to obtain the predicted value of the preset indicator; The recommendation degree of the candidate content is determined based on the predicted value of each of the preset indicators.

12. The method according to claim 11, wherein each of the indicator prediction networks corresponds to at least one identification feature extraction network; The step of inputting the splicing identification features corresponding to the candidate content into the identification feature extraction network to obtain the identification extraction features comprises: For each of the indicator prediction networks, the concatenated identification features corresponding to the candidate content are input into the identification feature extraction networks corresponding to the indicator prediction network, so as to obtain the identification extraction features respectively output by the identification feature extraction networks corresponding to the indicator prediction network; The step of inputting the indicator feature corresponding to the preset indicator and the identification extraction feature into the indicator prediction network corresponding to the preset indicator to obtain the predicted value of the preset indicator includes: The indicator feature corresponding to the preset indicator and the features of each identifier corresponding to the preset indicator are extracted and input into the indicator prediction network corresponding to the preset indicator to obtain the predicted value of the preset indicator.

13. A content push device, comprising: A sequence acquisition module, used to acquire a current information sequence of a user, wherein the current information sequence includes a plurality of information items, each of which corresponds to a content viewed by the user, and the plurality of information items include an information item corresponding to the current content viewed and an information item corresponding to at least one content viewed in history, and the information items in the current information sequence are arranged in a viewing order of the corresponding content viewed; A first encoding module, configured to encode the current information sequence into a feature sequence, wherein the feature sequence includes feature items corresponding one to one to the information items in the current information sequence, and each feature item represents a coding feature obtained by encoding the corresponding information item; A feature acquisition module, used to acquire the coding feature of the currently viewed content, wherein the coding feature of the currently viewed content is obtained by encoding the description information of the currently viewed content; A feature extraction module, used to extract features related to the coding features of the currently viewed content from the feature sequence, and obtain relevant features of the currently viewed content; a second encoding module, configured to obtain description information of at least one candidate content, encode the description information of the at least one candidate content respectively, and obtain encoding features of each candidate content; and The push module is used to select candidate content from the at least one candidate content according to the encoding feature of each candidate content and the relevant features of the currently viewed content, and push the selected candidate content to the user.

14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method according to any one of claims 1 to 12 when executing the computer program.

15. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 12.

Images