Content recommendation method and device based on content recommendation model, equipment and medium

Abstract

The application discloses a content recommendation method and device based on a content recommendation model, equipment and a medium, and belongs to the field of artificial intelligence. The content recommendation model comprises an object branch and at least two content branches. The at least two content branches are matched with at least two views of at least two candidate recommendation contents, and at least part of parameters of the at least two content branches are shared. The method comprises: obtaining an object representation corresponding to a first object identifier; obtaining a content representation of candidate recommendation content for the first object identifier; inputting the object representation and the content representation into the content recommendation model to obtain an object representation and a single-view content representation; determining a target recommendation content from the candidate recommendation content based on the object representation and the single-view content representation, and performing content recommendation to a first object represented by the first object identifier. The above scheme can improve the feature extraction capability of the content recommendation model, thereby improving the effect of content recommendation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a content recommendation method, apparatus, device, and medium based on a content recommendation model. Background Technology

[0002] Personalized recommendation systems have been widely used in e-commerce services, online games, content recommendation, video recommendation and other fields. How to search for content that users are interested in in a timely and effective manner is an important challenge for personalized recommendation systems.

[0003] In related technologies, content recommendation models employ a dual-tower model. This model trains the features of various types of recommended content using a single content tower, and then learns the relationship between user characteristics and the features of these different content types to improve the overall quality of content recommendations. The various types of recommended content include short videos, long videos, news, and novels.

[0004] However, the existing technologies cannot effectively learn information about each type of recommended content, which can easily lead to low accuracy in content recommendations. Summary of the Invention

[0005] This application provides a content recommendation method, apparatus, device, and medium based on a content recommendation model. The technical solution is as follows:

[0006] According to one aspect of this application, a content recommendation method based on a content recommendation model is provided, the content recommendation model including an object branch and at least two content branches, the at least two content branches being matched with at least two views of at least two candidate recommended content, and at least a portion of the parameters of the at least two content branches being shared.

[0007] The method includes:

[0008] Obtain the object representation corresponding to the first object identifier; and obtain the content representation of the candidate recommendation content for the first object identifier;

[0009] The object representation is input into the object branch of the content recommendation model for characterization to obtain the object representation; and the content representation is input into the content branch of the content recommendation model that matches the view of the candidate recommended content for characterization to obtain the single-view content representation.

[0010] Based on the object representation and the single-view content representation, target recommended content that meets the filtering conditions is determined from the candidate recommended content;

[0011] The target recommended content is then recommended to the first object represented by the first object identifier.

[0012] According to another aspect of this application, a method for training a content recommendation model is provided, the content recommendation model including an object branch and at least two content branches, the at least two content branches being matched with at least two views of at least two types of recommended content, and at least a portion of the parameters of the at least two content branches being shared.

[0013] The method includes:

[0014] Obtain the object representation corresponding to the sample object identifier, and obtain at least two content representations of at least two types of recommended content for the sample object identifier, wherein the at least two content representations carry tag information;

[0015] The object representation is input into the object branch of the content recommendation model for characterization to obtain an object representation; and the at least two content representations are input into the content branch of the content recommendation model that matches the view of the at least two recommended contents for characterization to obtain at least two single-view content representations.

[0016] Based on the object representation and the at least two single-view content representations, the prediction information corresponding to the at least two single-view content representations is determined respectively;

[0017] The training loss is determined based on the tag information and the prediction information, and the model parameters of the content recommendation model are trained based on the training loss.

[0018] According to another aspect of this application, a content recommendation model is provided that includes an object branch and at least two content branches, wherein the at least two content branches are matched with at least two views of at least two candidate recommended content, and at least a portion of the parameters of the at least two content branches are shared.

[0019] The device includes:

[0020] The acquisition module is used to acquire the object representation corresponding to the first object identifier; and to acquire the content representation of the candidate recommendation content for the first object identifier.

[0021] The input module is used to input the object representation into the object branch of the content recommendation model for representation to obtain the object representation; and to input the content representation into the content branch of the content recommendation model that matches the view of the candidate recommended content for representation to obtain the single-view content representation.

[0022] The filtering module is used to determine target recommended content that meets the filtering conditions from the candidate recommended content based on the object representation and the single-view content representation.

[0023] The recommendation module is used to recommend the target recommended content to the first object represented by the first object identifier.

[0024] According to another aspect of this application, a training apparatus for a content recommendation model is provided, the content recommendation model including an object branch and at least two content branches, the at least two content branches being matched with at least two views of at least two types of recommended content, and at least a portion of the parameters of the at least two content branches being shared.

[0025] The device includes:

[0026] The acquisition module is used to acquire the object representation corresponding to the sample object identifier, and to acquire at least two content representations of at least two types of recommended content for the sample object identifier, wherein the at least two content representations carry tag information;

[0027] The representation module is used to input the object representation into the object branch of the content recommendation model for representation to obtain an object representation; and to input the at least two content representations into the content branch of the content recommendation model that matches the view of the at least two recommended contents for representation to obtain at least two single-view content representations.

[0028] The prediction module is used to determine prediction information corresponding to the at least two single-view content representations based on the object representation and the at least two single-view content representations respectively.

[0029] The training module is used to determine the training loss based on the label information and the prediction information, and to train the model parameters of the content recommendation model based on the training loss.

[0030] According to another aspect of this application, a computer device is provided, the computer device comprising: a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the content recommendation method based on the content recommendation model as described above, or to implement the training method of the content recommendation model as described above.

[0031] According to another aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, the computer program being loaded and executed by a processor to implement the content recommendation method based on the content recommendation model as described above, or to implement the training method of the content recommendation model as described above.

[0032] According to another aspect of this application, a computer program product is provided, the computer program product including computer instructions stored in a computer-readable storage medium, wherein a processor obtains the computer instructions from the computer-readable storage medium, causing the processor to load and execute to implement the content recommendation method based on the content recommendation model as described above, or to implement the training method of the content recommendation model as described above.

[0033] The beneficial effects of the technical solutions provided in this application include at least the following:

[0034] The content recommendation model in this embodiment includes an object branch and at least two content branches. During the training phase of the content recommendation model, at least two content branches are matched with at least two views of at least two types of recommended content, and at least a portion of the parameters of the at least two content branches are shared. Accordingly, each content branch can learn relevant information about the recommended content under all views, enabling transfer learning between recommended content from different views, so that the single-view content representation also carries relevant information about the recommended content from other views. The parameter-sharing part can also balance the learning process between recommended content from different views, while the parameter-independent part corresponds to the view, and can also establish differences between single-view content representations corresponding to recommended content from different views, improving the accuracy of single-view content representation. Through the same object branch, the object branch can learn relevant information about the recommended content under all views, obtaining a more accurate object representation. Adopting the above scheme can improve the feature extraction capability of the content recommendation model. During the usage phase of using this content recommendation model for content recommendation, it can alleviate the cold start problem caused by data sparsity, improve the recommendation effect for new objects or new recommended content, improve the accuracy of content recommendation, and also improve the content recommendation experience to a certain extent. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A block diagram of a computer system provided in an exemplary embodiment is shown;

[0037] Figure 2 A schematic diagram of the overall architecture of a content recommendation model provided in an exemplary embodiment is shown;

[0038] Figure 3A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment is shown;

[0039] Figure 4 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment is shown;

[0040] Figure 5 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment is shown;

[0041] Figure 6 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment is shown;

[0042] Figure 7 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment is shown;

[0043] Figure 8 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment is shown;

[0044] Figure 9 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment is shown;

[0045] Figure 10 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment is shown;

[0046] Figure 11 A flowchart illustrating a training method for a content recommendation model provided in an exemplary embodiment is shown;

[0047] Figure 12 A schematic diagram of the overall architecture of a content recommendation model provided in an exemplary embodiment is shown;

[0048] Figure 13 A structural block diagram of a content recommendation device based on a content recommendation model provided in an exemplary embodiment is shown.

[0049] Figure 14 A structural block diagram of a training apparatus for a content recommendation model provided in an exemplary embodiment is shown.

[0050] Figure 15 A structural block diagram of a computer device provided in an exemplary embodiment is shown. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0052] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0053] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0054] It should be understood that although the terms first, second, etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, a first parameter may also be referred to as a second parameter, and similarly, a second parameter may also be referred to as a first parameter. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0055] It should be noted that this application may display prompt interfaces, pop-ups, or output voice prompts before and during the collection of user data. These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their data is being collected. This ensures that the application only begins the steps for collecting user data after receiving confirmation from the user regarding the prompt interface or pop-up; otherwise (i.e., without user confirmation), the steps for collecting user data end, meaning no user data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of related user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the first object identifier, first object log, and first object information involved in this application are all obtained with the user's consent and authorization.

[0056] Figure 1A structural block diagram of a computer system provided in an exemplary embodiment of this application is shown. This computer system can be implemented as a system architecture for a content recommendation method and / or a training method for a content recommendation model. The computer system 100 includes a terminal 100 and a server 120.

[0057] Terminal 100 can be an electronic device such as a mobile phone, tablet computer, vehicle terminal (vehicle system), wearable device, PC (Personal Computer), or unmanned reservation terminal. A client application for the target application can be installed and run on terminal 100. This target application can be an application that stores a content recommendation model and performs content recommendation based on the model, an application that provides training functions for the content recommendation model, or other applications that provide content recommendation functions based on the model. This application does not limit the form of the target application, including but not limited to apps, mini-programs, etc., installed on terminal 100, and can also be in web page form.

[0058] Server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud servers, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Server 120 can be a backend server for the aforementioned target application, used to provide backend services to the clients of the target application.

[0059] Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network (WAN) or local area network (LAN) to achieve data computation, storage, processing, and sharing. Based on the cloud computing business model, cloud technology encompasses network technology, information technology, integration technology, management platform technology, and application technology. It can form resource pools, providing flexible and convenient on-demand access. Cloud computing technology will become a crucial support. Backend services of technical network systems require substantial computing and storage resources, such as video websites, image websites, and many portal websites. With the rapid development and application of the internet industry, every item may have its own identification mark in the future, requiring transmission to backend systems for logical processing. Data at different levels will be processed separately, and various industry data will require robust system support, which can only be achieved through cloud computing.

[0060] In some embodiments, the server described above can also be implemented as a node in a blockchain system. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and cryptographic algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer.

[0061] Terminal 100 and server 120 can communicate via a network, such as a wired or wireless network.

[0062] The content recommendation method and / or training method based on a content recommendation model provided in this application embodiment can be executed by a computer device, which refers to an electronic device with data computing, processing, and storage capabilities. Figure 1 Taking the implementation environment of the scheme shown as an example, the content recommendation method based on the content recommendation model can be executed by the terminal 100 (e.g., the client of the target application installed and running on the terminal 100 executes the content recommendation method based on the content recommendation model), or the server 120 executes the content recommendation method based on the content recommendation model, or the terminal 100 and the server 120 interact and cooperate to execute it. The training method of the content recommendation model can be executed by the terminal 100 (e.g., the client of the target application installed and running on the terminal 100 executes the training method of the content recommendation model), or the server 120 executes the training method of the content recommendation model, or the terminal 100 and the server 120 interact and cooperate to execute it. This application does not limit this.

[0063] In some embodiments, the content recommendation method based on the content recommendation model of this application can be applied to application scenarios such as content recommendation, music recommendation, product recommendation, and game item recommendation.

[0064] The embodiments of this application mainly involve artificial intelligence technology and machine learning technology.

[0065] Artificial intelligence (AI) is the theory, methods, technology, 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 that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0066] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0067] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learn-by-doing.

[0068] First, the content recommendation model of the embodiments of this application will be described.

[0069] The content recommendation model in this application includes an object branch and at least two content branches. The at least two content branches are matched with at least two views of at least two types of recommended content, and at least a portion of the parameters of the at least two content branches are shared.

[0070] For example, an object branch is a processing branch used to process information related to a sample object, and there is one object branch. Recommended content is data of interest to the sample object used during the training phase of the content recommendation model. A content branch is a processing branch used to process information related to the recommended content. The content branch corresponds to the view matching of the recommended content. In this embodiment, there are at least two content branches, and one content branch processes information related to one or more recommended contents under a certain view.

[0071] A view refers to a data type; different views represent different data types. Views can be pre-defined based on business experience, and the number of views is limited.

[0072] For example, the views can be pre-defined by the users of the content recommendation model based on their business experience. For instance, recommended content could be categorized into four types: long articles, short articles, long videos, and short videos, representing four different views. Alternatively, recommended content could be categorized into three types: news, applications, and movies, representing three different views. Depending on actual technical needs, more or fewer views can be created.

[0073] For example, one piece of recommended content corresponds to only one view, and one view may contain one or more pieces of recommended content.

[0074] For example, during the training phase of the content recommendation model, there are at least two content branches, each corresponding to a specific view of the recommended content, while there is only one object branch. Regardless of the view in which the object is represented, its related information is processed by the same object branch. During the usage phase of the content recommendation model, it is determined which specific view the candidate recommended content belongs to, thereby determining which content branch to use to process the relevant information of that candidate recommended content.

[0075] As an example, Figure 2 This illustration shows a schematic diagram of a content recommendation method based on a content recommendation model provided in an exemplary embodiment of this application.

[0076] Taking the recommended content as an example, including recommended content 1, recommended content 2, ..., recommended content N, the content corresponding to recommended content 1 is represented as 1, and the recommended content 1 corresponds to the view. Figure 1 Recommended content 2 corresponds to content 2, and recommended content 2 corresponds to video. Figure 2 The recommended content N corresponds to the content representation N, and the recommended content N corresponds to the view N.

[0077] The content recommendation model 200 includes one object branch and N content branches, with at least a portion of the parameters shared between the N content branches.

[0078] For example, Figure 2 The content recommendation model 200 shown includes: object branches corresponding to processing layers 203, 205, and 207; content branch 1 corresponding to processing layers 204, 206, and processing block 208-1; content branch 2 corresponding to processing layers 204, 206, and processing block 208-2; and content branch N corresponding to processing layers 204, 206, and processing block 208-N. The parameters of processing layers 204 and 206 are shared among the N content branches, while the parameters of processing blocks 208-1, 208-2, and 208-N are independent.

[0079] The object branch processes information related to object representation 201, resulting in object representation 209. The data flow of the object branch proceeds sequentially through processing layers 203, 205, and 207. Content branch 1 processes the view... Figure 1 The relevant information of content representation 1 (202-1) is used to obtain single-view content representation 210-1. The data flow of content branch 1 is sequentially to processing layer 204, processing layer 206, and processing block 208-1. Content branch 2 is used to process the view. Figure 2The content representation 2 (202-2) is processed to obtain the single-view content representation 210-2. The data flow of content branch 2 is sequentially to processing layer 204, processing layer 206, and processing block 208-2. Content branch N is used to process the content representation N (202-N) of view N to obtain the single-view content representation 210-N. The data flow of content branch N is sequentially to processing layer 204, processing layer 206, and processing block 208-N.

[0080] Then, training loss 1 is determined by calculating the similarity 211-1 between object representation 209 and single-view content representation 210-1, and the model parameters of object branch and content branch 1 are adjusted based on training loss 1. Training loss 2 is determined by calculating the similarity 211-2 between object representation 209 and single-view content representation 210-2, and the model parameters of object branch and content branch 2 are adjusted based on training loss 2. Training loss N is determined by calculating the similarity 211-N between object representation 209 and single-view content representation 210-N, and the model parameters of object branch and content branch N are adjusted based on training loss N, until training is completed, resulting in a trained content recommendation model. Content recommendation can then be performed based on this content recommendation model.

[0081] In some embodiments, depending on actual technical needs, more or fewer processing layers may be set, more or fewer parameters may be shared between N content branches, or more parameter sharing methods may be set for other processing layers, without limitation.

[0082] Next, we will introduce content recommendation methods based on content recommendation models. Figure 3 A flowchart illustrating an exemplary embodiment of this application provides a content recommendation method based on a content recommendation model, demonstrating how this method is applied... Figure 1 The terminal 100 (or a client installed on the terminal 100 that supports content recommendation based on a content recommendation model) or server 120 is shown as an example. The method includes:

[0083] Step 320: Obtain the object representation corresponding to the first object identifier; and obtain the content representation of the candidate recommended content for the first object identifier.

[0084] For example, the first object identifier is an identifier used to characterize the first object, which is the object that needs to be recommended content. In this embodiment, the first object can be a historical user who has already received content recommendations, or a new user who has not received content recommendations before.

[0085] Optionally, the first object identifier includes, but is not limited to, at least one of the first object's account, nickname, name, number, or code.

[0086] For example, the object representation corresponding to the first object identifier is used to characterize the relevant information of the first object. The object information of the first object is described from different types such as the attributes, behaviors, and associated information of the first object.

[0087] Optionally, the object information includes at least one of attribute description information, behavior description information, and association description information. Attribute description information includes, but is not limited to, attribute information such as name, age, gender, city, and city level; behavior description information includes, but is not limited to, behavior information such as historical click behavior, historical browsing behavior, and historical search behavior; and association description information includes, but is not limited to, association information such as satisfaction, preference, or attitude towards the recommended content.

[0088] Optionally, the representation of the object corresponding to the first object identifier may include, but is not limited to, at least one of feature vector, feature matrix, feature value or bit information.

[0089] For example, the candidate recommended content is a plurality of recommended content that can be recommended to the first object. Optionally, the candidate recommended content includes, but is not limited to, at least one of the following: videos, images, advertisements, news, etc.

[0090] In some embodiments, the candidate recommended content may also be music, goods, game items, etc.

[0091] Optionally, the candidate recommendation content in this embodiment includes multiple content. Each candidate recommendation content corresponds to a specific view. Multiple candidate recommendation contents may correspond to multiple different views or the same view.

[0092] Optionally, a content database is pre-set, storing multiple recommended content items for different views. This content database can be updated periodically, for example, hourly. Based on a first object identifier, candidate recommended content for that first object identifier is retrieved from the content database.

[0093] For example, content representation is used to characterize different types of information about candidate recommendation objects. For instance, if the candidate recommendation content is an image, the content representation can be determined based on different types of information such as people, animals, colors, and sizes of the image. Depending on the actual technical needs, each type of information can be further subdivided; for example, color can be further subdivided into chromatic and achromatic colors.

[0094] Optionally, the content representation may include, but is not limited to, at least one of feature vectors, feature matrices, feature values, or bit information.

[0095] In some embodiments, offline, a feature engineering module is used to process the object information of the first object to obtain an object representation; and to process the information of different types of candidate recommended objects to obtain a content representation.

[0096] Step 340: The object representation is represented in the object branch of the input content recommendation model to obtain the object representation; and the content representation is represented in the content branch of the input content recommendation model that matches the view of the candidate recommended content to obtain the single-view content representation.

[0097] For example, an object branch is a processing branch used to process information related to a first object, and there is one object branch. A content branch is a processing branch used to process information related to candidate recommended content. Content branches correspond to views, and one content branch processes information related to candidate recommended content for one view. There are at least two content branches.

[0098] For example, in the training phase, the content recommendation model of this application uses recommended content under at least two views for multi-view learning. Therefore, in the usage phase of the content recommendation model, the content representation is input into the content branch of the content recommendation model that matches the view of the candidate recommended content for representation, to obtain a single-view content representation. The single-view content representation also carries multi-view content information, which refers to the content feature information of the recommended content under at least two views.

[0099] For example, object representation refers to the representation of an object after it has been represented by a content recommendation model, while single-view content representation refers to the representation of a content after it has been represented by a content recommendation model.

[0100] Optionally, the object representation is represented in the object branch of the input content recommendation model to obtain the object representation; and the content representation is represented in the content branch of the input content recommendation model that matches the view of the candidate recommended content to obtain the single-view content representation, which carries multi-view content information.

[0101] Optionally, the steps of representing the object as the input object branch and representing the content as the input matching content branch can be executed simultaneously or sequentially, without restriction.

[0102] In some embodiments, the candidate recommended content includes multiple components. One candidate recommended content may be represented one at a time within the content branch of the content recommendation model that matches the view of that candidate recommended content, resulting in a single-view content representation. Alternatively, candidate recommended content for the same view may be grouped into a batch, and its content representation may be represented within the content branch of the content recommendation model that matches the view of that batch of candidate recommended content, resulting in a single-view content representation.

[0103] Step 360: Based on object representation and single-view content representation, determine the target recommended content that meets the filtering criteria from the candidate recommended content.

[0104] For example, the target recommendation content is one or more pieces of recommendation content determined from the candidate recommendation content and recommended to the first object. The filtering criteria are the conditions that the target recommendation content must meet, and can be set according to actual technical needs.

[0105] Optionally, based on object representation and single-view content representation, the similarity between the object representation and the single-view content representation is calculated, and target recommended content that meets the filtering criteria is determined from the candidate recommended content based on the similarity. The filtering criteria can be set to a similarity greater than or equal to a similarity threshold.

[0106] Optionally, based on object representation and single-view content representation, the similarity between the object representation and the single-view content representation is calculated. The content score of the candidate recommended content is determined based on the similarity. The target recommended content that meets the filtering criteria is then determined from the candidate recommended content based on the content score. The filtering criteria can be set to a content score greater than or equal to a score threshold, or the top n content scores, where n is a positive integer.

[0107] Step 380: Recommend the target content to the first object represented by the first object identifier.

[0108] Optionally, the target recommended content is retrieved from the content database through retrieval or recall, and then recommended to the first object represented by the first object identifier.

[0109] The target recommended content is recommended to the first object represented by the first object through at least one method, such as displaying a recommendation interface, pop-up window, or voice recommendation information.

[0110] The content recommendation model in this embodiment includes an object branch and at least two content branches. During the training phase of the content recommendation model, at least two content branches are matched with at least two views of at least two types of recommended content, and at least a portion of the parameters of the at least two content branches are shared. Accordingly, each content branch can learn relevant information about the recommended content under all views, enabling transfer learning between recommended content from different views, so that the single-view content representation also carries relevant information about the recommended content from other views. The parameter-sharing part can also balance the learning process between recommended content from different views, while the parameter-independent part corresponds to the view, and can also establish differences between single-view content representations corresponding to recommended content from different views, improving the accuracy of single-view content representation. Through the same object branch, the object branch can learn relevant information about the recommended content under all views, obtaining a more accurate object representation. Adopting the above scheme can improve the feature extraction capability of the content recommendation model. During the usage phase of using this content recommendation model for content recommendation, it can alleviate the cold start problem caused by data sparsity, improve the recommendation effect for new objects or new recommended content, improve the accuracy of content recommendation, and also improve the content recommendation experience to a certain extent.

[0111] In one exemplary embodiment of this application, at least two content branches include an embedding layer, a feature extraction layer, and a multi-view representation layer. The embedding layer and the feature extraction layer share parameters, and the multi-view representation layer includes at least two representation blocks that match at least two views of at least two candidate recommended content. The parameters of the at least two representation blocks are independent.

[0112] For example, at least two content branches include an embedding layer, a feature extraction layer, and a multi-view representation layer, with the embedding layer and the feature extraction layer sharing parameters. The multi-view representation layer includes at least two representation blocks that match at least two views of at least two candidate recommended content, with the parameters of the at least two representation blocks being independent.

[0113] Optionally, at least two content branches are also referred to as content towers. Depending on the number of views in the actual technology, more representation blocks can be set in the multi-view representation layer, that is, more content branches can be set in the content recommendation model.

[0114] Figure 4 The flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment of this application is shown. In step 340, the content representation is input into the content branch of the content recommendation model that matches the view of the candidate recommended content for representation, so as to obtain a single-view content representation, which can be implemented as steps 420 to 460:

[0115] Step 420: Input the content representation into the embedding layer to obtain the content embedding features.

[0116] For example, at least two content branches include an embedding layer.

[0117] Optionally, the content representation is input into the embedding layer to obtain the embedding features corresponding to each dimension of the content representation. The embedding features corresponding to each dimension of the content representation are then concatenated to obtain the content embedding features.

[0118] Step 440: Embed the content into the feature extraction layer to obtain content features.

[0119] For example, at least two content branches include a feature extraction layer.

[0120] Optionally, the content can be embedded into the feature extraction layer to obtain content features.

[0121] Step 460: Input the content features into the representation block in the multi-view representation layer that matches the view of the candidate recommended content for representation, and obtain the single-view content representation.

[0122] For example, at least two content branches include a multi-view representation layer, which includes multiple representation blocks that are respectively matched with multiple views of at least two candidate recommended content.

[0123] Optionally, the content features are input into the representation block in the multi-view representation layer that matches the view of the candidate recommended content for representation, to obtain a single-view content representation.

[0124] In one exemplary embodiment of this application, the object branch has the same number of layers as the at least two content branches. The object branch includes an embedding layer, a feature extraction layer, and a multi-view representation layer, where the multi-view representation layer includes only one representation block. Optionally, the object branch is also referred to as an object tower.

[0125] For example, in step 340, the object representation is represented in the object branch of the input content recommendation model to obtain the object representation, which can be implemented as follows:

[0126] The object representation is input into the embedding layer to obtain the object embedding features;

[0127] The object's embedded features are input into the feature extraction layer to obtain the object's features;

[0128] The object features are input into the multi-view representation layer for representation, resulting in an object representation.

[0129] In this embodiment, during the training phase of the content recommendation model, the learning process between recommended content from different views can be balanced through shared embedding and feature extraction layers. This allows at least two content branches to learn both the consistency and differences between recommended content from different views, enabling transfer learning between candidate recommended content from different views. Independent representation blocks in the multi-view representation layer can establish differences between single-view content representations corresponding to recommended content from different views, thereby improving the accuracy of single-view content representation during the usage phase. Meanwhile, the object branch can learn relevant information about objects and recommended content from all views, allowing for more effective information transfer between different views, thus improving the accuracy of object representation during the usage phase.

[0130] In one exemplary embodiment of this application, the feature extraction layer employs a multi-head self-attention neural network.

[0131] For example, the feature extraction layers of the object branch and at least two content branches of the content recommendation model each employ a multi-head self-attention (MHSA) neural network with residual connections. The multi-head self-attention neural network is a neural network that incorporates a multi-head self-attention mechanism.

[0132] Optionally, the parameters of the multi-head self-attention neural network in the object branch and at least two content branches are independent for each group.

[0133] Figure 5 A flowchart illustrating a content recommendation method based on a content recommendation model provided in an exemplary embodiment of this application is shown. Step 440 may further include steps 441 to 443:

[0134] Step 441: Based on the number of heads in the multi-head self-attention neural network, the content embedding features are split to obtain multiple content embedding sub-features with the same number of heads.

[0135] Optionally, based on the number of heads in the multi-head self-attention neural network, the content embedding features are split to obtain multiple content embedding sub-features with the same number of heads.

[0136] Step 442: Input multiple content embedding sub-features in parallel into a multi-head self-attention neural network for multi-head self-attention calculation, and obtain multiple content sub-features accordingly.

[0137] Optionally, multiple content embedding sub-features are input in parallel into a multi-head self-attention neural network for multi-head self-attention calculation, each yielding multiple content sub-features with attention.

[0138] Step 443: Concatenate multiple content sub-features to obtain the content features.

[0139] Optionally, multiple content sub-features can be concatenated to obtain the content feature.

[0140] In this embodiment, during the training phase of the content recommendation model, the multi-head self-attention neural network can effectively capture the potential feature interactions between multiple content embedding sub-features of the content embedding features, enabling the content branches to learn better and more accurate content features, thereby improving the accuracy of content features during the usage phase of the content recommendation model.

[0141] In one exemplary embodiment of this application, at least two representation blocks of the multi-view representation layer each employ at least one multilayer perceptron.

[0142] For example, the multi-view representation layer of the object branch of the content recommendation model uses at least one multilayer perceptron (MLP), and at least two representation blocks of the multi-view representation layers of at least two content branches each use at least one multilayer perceptron.

[0143] Figure 6 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment of this application is shown. Step 460 can be further implemented as step 461:

[0144] Step 461: Input the content features into at least one multilayer perceptron in the multi-view representation layer that matches the view of the candidate recommended content for mapping, and obtain a single-view content representation.

[0145] For example, the content features of the candidate recommended content are input into at least one multilayer perceptron in the multi-view representation layer that matches the view of the candidate recommended content for mapping, so as to obtain a single-view content representation.

[0146] In this embodiment, during the training phase of the content recommendation model, a parameter-independent multilayer perceptron is used to map content features. This establishes the differences between single-view content representations corresponding to recommended content from different views, enabling the content branch to learn better single-view content representations and thus improving the accuracy of single-view content representations during the usage phase. Furthermore, the feature extraction layer and the multi-view representation layer employ different types of neural networks to represent relevant features, which can further improve the accuracy of single-view content representation.

[0147] Figure 7 A flowchart of a content recommendation method based on a content recommendation model provided in an exemplary embodiment of this application is shown. Step 360 can be further implemented as steps 520 to 540:

[0148] Step 520: Calculate the similarity between object representation and single-view content representation based on object representation and single-view content representation.

[0149] For example, similarity is used to characterize the degree of interest that the first object corresponding to the object characterizes in the candidate recommended content corresponding to the single-view content characterizes.

[0150] Optionally, the similarity includes cosine similarity. A cosine similarity value closer to 1 indicates a greater similarity between the object representation and the single-view content representation, suggesting that the first object is more interested in the candidate recommended object. A cosine similarity value closer to 0 indicates a less similarity between the object representation and the single-view content representation, suggesting that the first object is less interested in the candidate recommended object.

[0151] Optionally, based on the object representation and the single-view content representation, the cosine similarity between the object representation and the single-view content representation is calculated.

[0152] Step 540: Determine the content score of the candidate recommended content based on the similarity, and determine the target recommended content that meets the screening criteria from the candidate recommended content based on the content score.

[0153] For example, the filtering criteria are the conditions that the target recommended content must meet, and can be set according to actual technical needs. In this embodiment, the filtering criteria can be set based on the content rating.

[0154] Optionally, a content score is determined based on the similarity of the candidate recommended content, and the target recommended content that meets the screening criteria is determined from the candidate recommended content based on the content score.

[0155] In this embodiment, content scoring is used to characterize the object's interest in candidate recommended content, thereby quantifying the object's interest level. This helps to determine the target recommended content that the object may be interested in and improves the recommendation effect.

[0156] Figure 8 A flowchart illustrating an exemplary embodiment of the content recommendation method based on a content recommendation model provided in this application is shown. Step 540 may include steps 541 to 542:

[0157] Step 541: Based on the content rating, sort the content ratings to obtain the sorting results.

[0158] Optionally, based on the multiple content ratings corresponding to multiple candidate recommended content, the multiple content ratings are sorted to obtain the sorting result.

[0159] Step 542: Based on the ranking results, the candidate recommended content corresponding to the top n content scores with the highest scores is determined as the target recommended content that meets the screening criteria, where n is an integer greater than 0.

[0160] Optionally, based on the ranking results, the candidate recommended content corresponding to the top n content scores with the highest scores is determined as the target recommended content that meets the screening criteria, where n is an integer greater than 0, that is, the target recommended content includes one or more.

[0161] In this embodiment, by sorting the content ratings, the candidate recommended content corresponding to the top n content ratings with the highest ratings is determined as the target recommended content. This can accurately identify the target recommended content that the target is interested in, thereby improving the recommendation effect.

[0162] Figure 9 The flowchart illustrates a content recommendation method based on a content recommendation model provided in an exemplary embodiment of this application. Step 320, obtaining the object representation corresponding to the first object identifier, can be implemented as steps 620 to 660:

[0163] Step 620: In response to the content recommendation request, obtain the first object identifier based on the content recommendation request.

[0164] For example, a content recommendation request is a request to recommend content to a first object, and the content recommendation request carries the identifier of the first object.

[0165] Optionally, the content recommendation request can be triggered by a first object, for example, by the first object clicking a recommendation button in the target application. The content recommendation request can also be event-driven, and the event includes at least one of the following: application startup, account login, or interface switching. For example, if the first object starts the application and logs in to its account, content recommendations can be automatically performed for that first object.

[0166] Optionally, in response to a content recommendation request, a first object identifier is obtained based on the content recommendation request.

[0167] Step 640: Obtain the object log corresponding to the first object identifier, and obtain object information from the object log; the object information includes at least one of attribute description class information, behavior description class information, and association description class information.

[0168] Optionally, the object log corresponding to the first object identifier is obtained, the data in the object log is cleaned, and the object information is obtained from the object log.

[0169] Optionally, the object information includes at least one of attribute description information, behavior description information, and association description information.

[0170] Step 660: Based on the object information, obtain the object representation corresponding to the first object identifier.

[0171] Optionally, based on object information, the object representation corresponding to the first object identifier is obtained through the feature engineering module.

[0172] In some embodiments, object information is related information under multiple views of a first object, and an object representation may correspond to one view or multiple views.

[0173] In this embodiment, determining the object representation based on object information can improve the accuracy of the object representation. Furthermore, processing the object representation offline first, and then inputting it online into the content recommendation model for further processing, can improve the data processing efficiency of the content recommendation model.

[0174] Figure 10 The flowchart illustrates a content recommendation method based on a content recommendation model provided in an exemplary embodiment of this application. Step 320, obtaining the content representation of candidate recommended content for a first object identifier, can be implemented as steps 720 to 740:

[0175] Step 720: Based on the first object identifier, obtain candidate recommended content for the first object identifier.

[0176] Optionally, based on the first object identifier, candidate recommended content for the first object representation can be obtained from a pre-set content database using a retrieval or recall method.

[0177] Step 740: Based on the content information corresponding to the candidate recommended content, obtain the content representation of the candidate recommended content for the first object identifier.

[0178] Optionally, based on the content information corresponding to the candidate recommended content, a specific view corresponding to the candidate recommended content is determined, and based on different types of information in the content information, a content representation of the candidate recommended content for the first object identifier is obtained.

[0179] In this embodiment, the content representation of candidate recommended content is obtained through the content information of the candidate recommended content, which enables a full and comprehensive representation of the candidate recommended content and improves the accuracy of the content representation. Furthermore, processing the content representation offline first, and then inputting the content representation online into the content recommendation model for processing, can improve the data processing efficiency of the content recommendation model.

[0180] In some embodiments, the method further includes: storing single-view content representations corresponding to candidate recommended content.

[0181] Optionally, the single-view content representation corresponding to the candidate recommendation content is stored. Once the single-view content representation of a candidate recommendation content has been stored, it can be directly retrieved next time without having to execute input step 340 again to obtain the single-view content representation.

[0182] In this embodiment, by storing the single-view content representation corresponding to the candidate recommendation content, it can be easily and directly called, which can effectively improve the data processing efficiency of content recommendation.

[0183] Figure 11 A flowchart illustrating a training method for a content recommendation model provided in an exemplary embodiment of this application is shown, in which the method is applied... Figure 1 The method, illustrated by the terminal 100 (or a client installed on the terminal 100 that supports the training process of the content recommendation model) or server 120, includes:

[0184] Step 820: Obtain the object representation corresponding to the sample object identifier, and obtain at least two content representations of at least two types of recommended content for the sample object identifier, wherein the at least two content representations carry tag information.

[0185] For example, a sample object identifier is an identifier used to characterize a sample object. Optionally, the sample object identifier includes, but is not limited to, at least one of the following: account name, nickname, name, number, code, etc.

[0186] For example, the object representation corresponding to the sample object identifier is used to characterize the object. For the same sample object, it can be described from different perspectives such as the sample object's attributes, behaviors, and associated information.

[0187] Optionally, the object information includes at least one of attribute description information, behavior description information, and association description information. Attribute description information includes, but is not limited to, attribute information such as name, age, gender, city, and city level; behavior description information includes, but is not limited to, behavior information such as historical click behavior, historical browsing behavior, and historical search behavior; and association description information includes, but is not limited to, association information such as satisfaction, preference, or attitude towards the recommended content.

[0188] For example, recommended content refers to multiple recommended content items that have been recommended to the sample object, and the recommended content is content that the sample object is interested in.

[0189] Optionally, the recommended content includes, but is not limited to, at least one of the following: videos, images, advertisements, and news. In this embodiment, the recommended content includes multiple elements; one recommended content corresponds to a specific view, and at least two candidate recommended contents correspond to at least two different views.

[0190] Optionally, the content representation of the recommended content may include, but is not limited to, at least one of feature vectors, feature matrices, feature values, or bit information.

[0191] For example, the content representation corresponding to the recommended content carries tag information, which is used to characterize the degree of interest of the sample object in the recommended content.

[0192] Optionally, the tag information can be the similarity between the object representation and the content representation. The similarity includes cosine similarity, which characterizes the degree of interest of the sample object in the recommended content. The closer the cosine similarity value is to 1, the more interested the sample object is in the recommended object.

[0193] Step 840: Represent the object in the object branch of the input content recommendation model to obtain an object representation; and represent the content in the content branch of the input content recommendation model that matches the view of at least two recommended contents to obtain at least two single-view content representations.

[0194] For example, the content recommendation model to be trained includes an object branch and at least two content branches, the at least two content branches being matched with different views of the recommended content, and at least a portion of the parameters of the at least two content branches being shared.

[0195] Optionally, the object representation is represented in the object branch of the input content recommendation model to obtain an object representation; and the at least two single-view content representations are represented in the content branch of the input content recommendation model that matches the views of at least two recommended contents.

[0196] Optionally, the steps of identifying the sample object corresponding to the input object branch and identifying at least one content as the input matching content branch can be performed simultaneously or sequentially, without limitation.

[0197] In some embodiments, the recommended content includes multiple components. It can be that one recommended content is represented at a time, with its corresponding content representation input into the content recommendation model and characterized within the content branch that matches the view of that recommended content, resulting in a single-view content representation. Alternatively, recommended content for the same view can be grouped into a batch, and its content representation input into the content recommendation model and characterized within the content branch that matches the view of that batch of recommended content, resulting in a single-view content representation.

[0198] Step 860: Based on the object representation and at least two single-view content representations, determine the prediction information corresponding to at least two single-view content representations respectively.

[0199] Optionally, based on the object representation and at least two single-view content representations, prediction information corresponding to at least two single-view content representations is determined. The prediction information is used to characterize the degree of interest the sample object, predicted by the content recommendation model, has in the recommended content corresponding to that single-view content representation.

[0200] Optionally, the predicted information may include the similarity between object representations and content representations, whereby the similarity includes cosine similarity, which characterizes the degree of interest of the sample object in the recommended content. The closer the cosine similarity value is to 1, the more interested the sample object is in the recommended object.

[0201] Step 880: Determine the training loss based on the label information and prediction information, and train the model parameters of the content recommendation model based on the training loss.

[0202] Optionally, the training loss is determined based on the label information and the prediction information, and the model parameters of the content recommendation model are trained using the backpropagation algorithm based on the training loss.

[0203] In some embodiments, an iteration of the content recommendation model processes an object representation corresponding to a sample object identifier and a content representation corresponding to a recommended content for the sample object identifier. A recommended content corresponds to a view. Thus, an iteration determines a prediction and adjusts the model parameters of an object branch and a content branch.

[0204] In some embodiments, the method further includes: storing single-view content representations corresponding to at least two types of recommended content, for convenient direct retrieval in subsequent searches.

[0205] In summary, the content recommendation model of this application can achieve transfer learning between recommended content from different views, resulting in more accurate object representations and single-view content representations, thus improving the feature extraction capability of the content recommendation model. During the training phase of the content recommendation model, prediction information is determined each time based on an object representation and a single-view content representation, and the model parameters of the corresponding object branch and content branch are adjusted, improving the parameter adjustment accuracy of the content recommendation model and thus enhancing the learning effect.

[0206] As an example, taking a specific content recommendation model as an example, the training method of the content recommendation model of this application embodiment and the content recommendation method based on the content recommendation model are described.

[0207] 1. Model Structure of Content Recommendation Model

[0208] For example, taking two views as an example, namely a first view and a second view, the content recommendation model includes an object branch and two content branches, namely a first content branch and a second content branch. The first content branch matches the first view of the first recommended content, and the second content branch matches the second view of the second recommended content. At least some parameters of the first content branch and the second content branch are shared.

[0209] Figure 12This illustration shows a schematic diagram of the overall architecture of a content recommendation model provided in an exemplary embodiment of this application. The content recommendation model is a Multi-View Approach with Hybrid Attentive Networks (MV-HAN), which includes an embedding layer 10, a feature extraction layer 20, and a multi-view representation layer 30.

[0210] The object branch of the content recommendation model includes an embedding layer 11, a feature extraction layer 21, and a multi-view representation layer 31. The data flow of the object branch is sequentially: embedding layer 11, feature extraction layer 21, and multi-view representation layer 31. The feature extraction layer 21 uses a multi-head self-attention (MHSA) neural network, and the multi-view representation layer uses a multilayer perceptron (MLP).

[0211] The first content branch of the content recommendation model includes an embedding layer 12, a feature extraction layer 22, and a representation block 32-1. The data flow of the first content branch is, in sequence, embedding layer 12, feature extraction layer 22, and representation block 32-1. The second content branch of the content recommendation model includes an embedding layer 12, a feature extraction layer 22, and a representation block 32-2. The data flow of the second content branch is, in sequence, embedding layer 12, feature extraction layer 22, and representation block 32-2. Representation blocks 32-1 and 32-2 belong to the multi-view representation layer 32.

[0212] The embedding layer 12 and feature extraction layer 22 of the first content branch and the second content branch share parameters. The feature extraction layer 22 adopts a multi-head self-attention (MHSA) neural network. The representation block 32-1 and the representation block 32-2 each adopt a multilayer perceptron (MLP).

[0213] 2. Training methods for content recommendation models

[0214] Step 1: Obtain the object representation corresponding to the sample object identifier, and obtain the first content representation corresponding to the first recommended content and the second content representation corresponding to the second recommended content for the sample object identifier. The first recommended content corresponds to the first view, and the second recommended content corresponds to the second view. The first content representation and the second content representation carry first tag information and second tag information, respectively.

[0215] For example, the object is represented as U = [u1, u2, ..., u Nu ], u i Let u be the i-th feature representation of the sample object, and Nu be the number of features. For example, in the object representation U, u1, u2, and u3 are the feature representations corresponding to the name, age, and gender of the sample object, respectively.

[0216] The first content is represented as Os = [os1, os2, ..., os Nsv ], os i Let Ot be the i-th feature representation of the first recommended content, and Nsv be the number of features. The second content is represented as Ot = [ot1, ot2, ..., ot]. Ntv ], ot i Let be the i-th feature representation of the second recommended content, and Ntv be the number of features. For example, if the first recommended content is an image, then os1, os2, and os3 in the first content representation Os are the features of animals, plants, and people contained in the image, respectively.

[0217] The first content representation and the second content representation carry first tag information and second tag information, respectively. The first tag information is used to characterize the sample object's interest in the first recommended content, and the second tag information is used to characterize the sample object's interest in the second recommended content.

[0218] Step 2: Represent the object representation in the object branch of the input content recommendation model to obtain the object representation; and represent the first content representation in the first content branch of the input content recommendation model that matches the first view of the first recommended content to obtain the first single-view content representation; and represent the second content representation in the second content branch of the input content recommendation model that matches the second view of the second recommended content to obtain the second single-view content representation.

[0219] For example, the object representation U is input into the embedding layer 11 to obtain object embedding features; the object embedding features are input into the feature extraction layer 21 to obtain object features; the object features are input into the multi-view representation layer 31 to obtain object representation; the formula is expressed as:

[0220] p i =Emb u (u i )

[0221]

[0222] Among them, Emb u p represents the embedding process of embedding layer 11. i Indicate u i The corresponding object embedding features, Indicate u i The corresponding object representation, MRL u Represents the mapping of the multi-view representation layer 31, FEL u This represents the mapping of feature extraction layer 21.

[0223] The first content representation Os is input into the embedding layer 12 to obtain the first content embedding feature; the first content embedding feature is input into the feature extraction layer 22 to obtain the first content feature; the first content feature is input into the representation block 32-1 of the multi-view representation layer 31 to obtain the first single-view content representation. Figure 12 Single-view content representation 1);

[0224] The second content representation Ot is input into the embedding layer 12 to obtain the second content embedding feature; the second content embedding feature is input into the feature extraction layer 22 to obtain the second content feature; the second content feature is input into the representation block 32-2 of the multi-view representation layer 31 to obtain the second single-view content representation. Figure 12 Single-view content representation 2);

[0225] The formula is expressed as:

[0226]

[0227]

[0228]

[0229]

[0230] Among them, Emb o This indicates the embedding process of embedding layer 12, FEL o This represents the mapping of feature extraction layer 22. Let j be the feature representation of the j-th feature of the first recommended content. express The corresponding first content embedding feature. This represents the mapping of representation block 32-1 in the multi-view representation layer 32. express The corresponding first single-view content representation. This represents the feature representation of the k-th feature of the second recommended content. express The corresponding second content embedding feature. This represents the mapping of representation block 32-2 in the multi-view representation layer 32. express The corresponding second single-view content representation.

[0231] The MRL consists of multiple independent multilayer perceptrons and multiple multi-head self-attention neural networks with residual connections, and can be formally expressed as:

[0232] MRL(x) = MLP n (…(MLP1(RMHSA m(…RMHSA2(RMHSA1(x))))))

[0233] Where n is the number of multilayer perceptrons, MLP(x) is the mapping function of the multilayer perceptron, m is the number of stacked networks, and RMHSA is a multi-head self-attention neural network with residual connections, which is formally expressed as:

[0234] RMHSA(x)=ReLU(x+MHSA(x))

[0235] Here, MHSA is the mapping function for a multi-head self-attention neural network. In a multi-head self-attention neural network, the input vector is split into h parts, called h heads. Each head passes through the self-attention network in parallel to obtain its own attentional representation. The parts are then concatenated as the output of MHSA. ReLU is a non-linear activation function, expressed as:

[0236] ReLU(x) = max(0,x)

[0237] Step 3: Based on the object representation and the first single-view content representation, determine the first prediction information corresponding to the first single-view content representation; based on the object representation and the second single-view content representation, determine the second prediction information corresponding to the second single-view content representation.

[0238] For example, based on the cosine function, the first similarity between the object representation and the first single-view content representation, and the second similarity between the object representation and the second single-view content representation are calculated, as expressed by the formula:

[0239]

[0240]

[0241] in, Indicate u i and The first similarity between them Indicate u i and The second similarity between them.

[0242] For example, first prediction information and second prediction information are determined based on first similarity and second similarity, respectively. The first prediction information and second prediction information can be predicted content scores, which represent the degree of interest of the sample object in the first recommended content and the second recommended content.

[0243] Step 4: Determine the training loss based on the label information and prediction information, and train the model parameters of the content recommendation model based on the training loss.

[0244] For example, the training loss is determined based on the first label information and the first prediction information, and the model parameters of the object branch and the first content branch are trained. The training loss is determined based on the second label information and the second prediction information, and the model parameters of the object branch and the second content branch are trained.

[0245] Once the content recommendation model is trained, it can be used online. Figure 12 The online service 40 is used for content recommendation.

[0246] 3. Content recommendation methods based on content recommendation models

[0247] Step 11: In response to the content recommendation request, obtain the first object identifier based on the content recommendation request; obtain the object log corresponding to the first object identifier, and retrieve object information from the object log; based on the object information, obtain the object representation corresponding to the first object identifier. From the pre-set content database, retrieve candidate recommended content for the first object identifier, and obtain the content representation of the candidate recommended content.

[0248] Step 12: Represent the object representation in the object branch of the input content recommendation model to obtain the object representation; and represent the content representation in the content branch of the input content recommendation model that matches the view of the candidate recommended content to obtain the single-view content representation.

[0249] Step 13: Calculate the similarity between the object representation and the single-view content representation; determine the content score of the candidate recommended content based on the similarity.

[0250] Step 14: Sort the content ratings to obtain the ranking results; based on the ranking results, select the candidate recommended content corresponding to the top n content ratings as the target recommended content, and recommend the target recommended content to the first object represented by the first object identifier; n is an integer greater than 0.

[0251] Furthermore, the method of this application embodiment was also tested online in an A / B experiment (control experiment). The experimental results show that, compared with the traditional Deep Structured Semantic Model (DSSM), the content recommendation model of this application embodiment can increase the click-through rate by 6.95% and the user reading time by 9.69%, which significantly improves the user experience.

[0252] Figure 13The diagram illustrates a structural block diagram of a content recommendation apparatus based on a content recommendation model provided in an exemplary embodiment of this application. The content recommendation model includes an object branch and at least two content branches, wherein the at least two content branches are matched with at least two views of at least two types of candidate recommended content, and at least a portion of the parameters of the at least two content branches are shared. The apparatus includes:

[0253] The acquisition module 910 is used to acquire the object representation corresponding to the first object identifier; and to acquire the content representation of the candidate recommended content for the first object identifier.

[0254] The input module 920 is used to input the object representation into the object branch of the content recommendation model for characterization to obtain an object representation; and to input the content representation into the content branch of the content recommendation model that matches the view of the candidate recommended content for characterization to obtain a single-view content representation.

[0255] The filtering module 930 is used to determine target recommended content that meets the filtering conditions from the candidate recommended content based on the object representation and the single-view content representation.

[0256] The recommendation module 940 is used to recommend the target recommended content to the first object represented by the first object identifier.

[0257] In one example, the at least two content branches include an embedding layer, a feature extraction layer, and a multi-view representation layer. The embedding layer and the feature extraction layer share parameters, and the multi-view representation layer includes at least two representation blocks that match at least two views of the at least two candidate recommended content, with the parameters of the at least two representation blocks being independent.

[0258] The input module 920 is used for:

[0259] The content representation is input into the embedding layer to obtain the content embedding feature;

[0260] The content embedding features are input into the feature extraction layer to obtain content features;

[0261] The content features are input into the representation block in the multi-view representation layer that matches the view of the candidate recommended content for representation, and the single-view content representation is obtained.

[0262] In one example, the feature extraction layer employs a multi-head self-attention neural network;

[0263] The input module 920 is used for:

[0264] Based on the number of heads in the multi-head self-attention neural network, the content embedding feature is split to obtain multiple content embedding sub-features with the same number of heads.

[0265] The multiple content-embedded sub-features are input in parallel into the multi-head self-attention neural network for multi-head self-attention calculation, thereby obtaining multiple content sub-features.

[0266] The multiple content sub-features are concatenated to obtain the content feature.

[0267] In one example, each of the at least two representation blocks in the multi-view representation layer employs at least one multilayer perceptron;

[0268] The input module 920 is used for:

[0269] The content features are input into at least one multilayer perceptron in the multi-view representation layer that matches the view of the candidate recommended content for mapping, to obtain the single-view content representation.

[0270] In one example, the filtering module 930 is used for:

[0271] Based on the object representation and the single-view content representation, calculate the similarity between the object representation and the single-view content representation;

[0272] Based on the similarity, a content score is determined for the candidate recommended content. Based on the content score, target recommended content that meets the screening criteria is determined from the candidate recommended content.

[0273] In one example, the recommendation module 940 is used for:

[0274] Based on the content ratings, the content ratings are sorted to obtain the sorting results;

[0275] Based on the ranking results, the candidate recommended content corresponding to the top n content scores with the highest scores is determined as the target recommended content that meets the screening criteria, where n is an integer greater than 0.

[0276] In one example, the acquisition module 910 is used for:

[0277] In response to a content recommendation request, the first object identifier is obtained based on the content recommendation request;

[0278] Obtain the object log corresponding to the first object identifier, and obtain object information from the object log; the object information includes at least one of attribute description class information, behavior description class information, and association description class information;

[0279] Based on the object information, the object representation corresponding to the first object identifier is obtained.

[0280] In one example, the acquisition module 910 is used for:

[0281] Based on the first object identifier, obtain candidate recommendation content for the first object identifier;

[0282] Based on the content information corresponding to the candidate recommended content, the content representation of the candidate recommended content for the first object identifier is obtained.

[0283] In one example, the device further includes a storage module; the storage module is used for:

[0284] Store the single-view content representation corresponding to the candidate recommended content.

[0285] Figure 14 The diagram illustrates a structural block diagram of a training apparatus for a content recommendation model provided in an exemplary embodiment of this application. The content recommendation model includes an object branch and at least two content branches, wherein the at least two content branches are matched with at least two views of at least two types of recommended content, and at least a portion of the parameters of the at least two content branches are shared. The apparatus includes:

[0286] The acquisition module is used to acquire the object representation corresponding to the sample object identifier, and to acquire at least two content representations of at least two types of recommended content for the sample object identifier, wherein the at least two content representations carry tag information.

[0287] The representation module is used to input the object representation into the object branch of the content recommendation model for representation to obtain an object representation; and to input the at least two content representations into the content branch of the content recommendation model that matches the view of the at least two recommended contents for representation to obtain at least two single-view content representations.

[0288] The prediction module is used to determine prediction information corresponding to the at least two single-view content representations based on the object representation and the at least two single-view content representations.

[0289] The training module is used to determine the training loss based on the label information and the prediction information, and to train the model parameters of the content recommendation model based on the training loss.

[0290] This application also provides a computer device, which includes: a processor and a memory, wherein the memory stores a computer program; the processor is used to execute the computer program in the memory to implement the content recommendation method and / or the training method of the content recommendation model based on the above-described method embodiments.

[0291] Alternatively, the computer device is a server. For example, Figure 15 This is a structural block diagram of a server 1000 provided in an exemplary embodiment of this application.

[0292] Typically, server 1000 includes a processor 1001 and memory 1002.

[0293] Processor 1001 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 1001 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). Processor 1001 may also include a main processor and a coprocessor. The main processor, also known as a central processing unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1001 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1001 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0294] The memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1002 is used to store at least one instruction, which is executed by the processor 1001 to implement the content recommendation model-based content recommendation method and / or the content recommendation model training method provided in the method embodiments of this application.

[0295] In some embodiments, the server 1000 may optionally include an input interface 1003 and an output interface 1004. The processor 1001, memory 1002, and input interfaces 1003 and 1004 can be connected via a bus or signal lines. Various peripheral devices can be connected to the input interfaces 1003 and 1004 via a bus, signal lines, or a circuit board. The input interfaces 1003 and 1004 can be used to connect at least one input / output (I / O) related peripheral device to the processor 1001 and memory 1002. In some embodiments, the processor 1001, memory 1002, and input interfaces 1003 and 1004 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1001, memory 1002, and input interfaces 1003 and 1004 can be implemented on separate chips or circuit boards, and this application does not limit this.

[0296] Those skilled in the art will understand that Figure 15 The structure shown does not constitute a limitation on the computer device 1000, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0297] This application also provides a chip, which includes programmable logic circuits and / or program instructions. When the chip is run on a computer device, it is used to implement the content recommendation method and / or the training method of the content recommendation model based on the content recommendation model provided in the above method embodiments.

[0298] This application also provides a computer-readable storage medium storing a computer program, which is loaded and executed by a processor to implement the content recommendation method and / or the training method of the content recommendation model based on the content recommendation model provided in the above method embodiments.

[0299] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, causing the processor to load and execute them to implement the content recommendation method and / or the content recommendation model training method based on the content recommendation model provided in the above-described method embodiments.

[0300] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0301] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0302] Those skilled in the art will recognize that the functions described in the embodiments of this application in one or more of the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another, and computer storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0303] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A content recommendation method based on a content recommendation model, characterized in that, The content recommendation model includes an object branch and at least two content branches. The at least two content branches are matched with at least two views of at least two candidate recommended contents. The at least two content branches include an embedding layer, a feature extraction layer, and a multi-view representation layer. The parameters of the embedding layer and the feature extraction layer are shared, while the parameters of the multi-view representation layer are independent. The method includes: Obtain the object representation corresponding to the first object identifier; and obtain the content representation of the candidate recommendation content for the first object identifier; The object representation is input into the object branch of the content recommendation model for representation, resulting in an object representation; and the content representation is input into the content branch of the content recommendation model that matches the view of the candidate recommended content, and represented through the embedding layer, the feature extraction layer, and the multi-view representation layer, resulting in a single-view content representation; wherein, during the training phase of the content recommendation model, each content branch learns relevant information about the recommended content under all views, realizing transfer learning between recommended content in different views; the single-view content representation carries multi-view content information, which refers to the content feature information of the recommended content under at least two views; Based on the object representation and the single-view content representation, target recommended content that meets the filtering conditions is determined from the candidate recommended content; The target recommended content is then recommended to the first object represented by the first object identifier.

2. The method according to claim 1, characterized in that, The multi-view representation layer includes at least two representation blocks that match at least two views of the at least two candidate recommended contents, and the parameters of the at least two representation blocks are independent. The step of inputting the content representation into the content branch of the content recommendation model that matches the view of the candidate recommended content, and representing it through the embedding layer, the feature extraction layer, and the multi-view representation layer to obtain a single-view content representation, includes: The content representation is input into the embedding layer to obtain the content embedding feature; The content embedding features are input into the feature extraction layer to obtain content features; The content features are input into the representation block in the multi-view representation layer that matches the view of the candidate recommended content for representation, and the single-view content representation is obtained.

3. The method according to claim 2, characterized in that, The feature extraction layer employs a multi-head self-attention neural network; The step of inputting the content embedding features into the feature extraction layer to obtain content features includes: Based on the number of heads in the multi-head self-attention neural network, the content embedding feature is split to obtain multiple content embedding sub-features with the same number of heads. The multiple content-embedded sub-features are input in parallel into the multi-head self-attention neural network for multi-head self-attention calculation, thereby obtaining multiple content sub-features. The multiple content sub-features are concatenated to obtain the content feature.

4. The method according to claim 2, characterized in that, Each of the at least two representation blocks in the multi-view representation layer employs at least one multilayer perceptron. The step of inputting the content features into the representation block in the multi-view representation layer that matches the view of the candidate recommended content for representation, to obtain the single-view content representation, includes: The content features are input into at least one multilayer perceptron in the multi-view representation layer that matches the view of the candidate recommended content for mapping, to obtain the single-view content representation.

5. The method according to any one of claims 1 to 4, characterized in that, The step of determining target recommended content that meets the filtering criteria from the candidate recommended content based on the object representation and the single-view content representation includes: Based on the object representation and the single-view content representation, calculate the similarity between the object representation and the single-view content representation; Based on the similarity, a content score is determined for the candidate recommended content. Based on the content score, target recommended content that meets the screening criteria is determined from the candidate recommended content.

6. The method according to claim 5, characterized in that, The step of determining target recommended content that meets the screening criteria from the candidate recommended content based on the content score includes: Based on the content ratings, the content ratings are sorted to obtain the sorting results; Based on the ranking results, the candidate recommended content corresponding to the top n content scores with the highest scores is determined as the target recommended content that meets the screening criteria, where n is an integer greater than 0.

7. The method according to any one of claims 1 to 6, characterized in that, The step of obtaining the object representation corresponding to the first object identifier includes: In response to a content recommendation request, the first object identifier is obtained based on the content recommendation request; Obtain the object log corresponding to the first object identifier, and obtain object information from the object log; the object information includes at least one of attribute description class information, behavior description class information, and association description class information; Based on the object information, the object representation corresponding to the first object identifier is obtained.

8. The method according to any one of claims 1 to 6, characterized in that, The step of obtaining the content representation of the candidate recommended content for the first object identifier includes: Based on the first object identifier, obtain candidate recommendation content for the first object identifier; Based on the content information corresponding to the candidate recommended content, the content representation of the candidate recommended content for the first object identifier is obtained.

9. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Store the single-view content representation corresponding to the candidate recommended content.

10. A training method for a content recommendation model, characterized in that, The content recommendation model includes an object branch and at least two content branches. The at least two content branches are matched with at least two views of at least two types of recommended content. The at least two content branches include an embedding layer, a feature extraction layer, and a multi-view representation layer. The parameters of the embedding layer and the feature extraction layer are shared, while the parameters of the multi-view representation layer are independent. The method includes: Obtain the object representation corresponding to the sample object identifier, and obtain at least two content representations of at least two types of recommended content for the sample object identifier, wherein the at least two content representations carry tag information; The object representation is input into the object branch of the content recommendation model for representation to obtain an object representation; and the at least two content representations are input into the content branch of the content recommendation model that matches the view of the at least two recommended contents, and represented by the embedding layer, the feature extraction layer and the multi-view representation layer to obtain at least two single-view content representations. Based on the object representation and the at least two single-view content representations, the prediction information corresponding to the at least two single-view content representations is determined respectively; The training loss is determined based on the label information and the prediction information, and the model parameters of the content recommendation model are trained based on the training loss. Each content branch learns relevant information of the recommended content under all views, realizing transfer learning between recommended content in different views. The single-view content representation carries multi-view content information, which refers to the content feature information of the recommended content under at least two views.

11. A content recommendation device based on a content recommendation model, characterized in that, The content recommendation model includes an object branch and at least two content branches. The at least two content branches are matched with at least two views of at least two candidate recommended contents. The at least two content branches include an embedding layer, a feature extraction layer, and a multi-view representation layer. The parameters of the embedding layer and the feature extraction layer are shared, while the parameters of the multi-view representation layer are independent. The device includes: The acquisition module is used to acquire the object representation corresponding to the first object identifier; and to acquire the content representation of the candidate recommendation content for the first object identifier. An input module is used to input the object representation into the object branch of the content recommendation model for representation, thereby obtaining an object representation; and to input the content representation into the content branch of the content recommendation model that matches the view of the candidate recommended content, and to represent it through the embedding layer, the feature extraction layer, and the multi-view representation layer, thereby obtaining a single-view content representation; wherein, during the training phase of the content recommendation model, each content branch learns relevant information about the recommended content under all views, realizing transfer learning between recommended content in different views; the single-view content representation carries multi-view content information, which refers to the content feature information of the recommended content under at least two views; The filtering module is used to determine target recommended content that meets the filtering conditions from the candidate recommended content based on the object representation and the single-view content representation. The recommendation module is used to recommend the target recommended content to the first object represented by the first object identifier.

12. A training device for a content recommendation model, characterized in that, The content recommendation model includes an object branch and at least two content branches. The at least two content branches are matched with at least two views of at least two types of recommended content. The at least two content branches include an embedding layer, a feature extraction layer, and a multi-view representation layer. The parameters of the embedding layer and the feature extraction layer are shared, while the parameters of the multi-view representation layer are independent. The device includes: The acquisition module is used to acquire the object representation corresponding to the sample object identifier, and to acquire at least two content representations of at least two types of recommended content for the sample object identifier, wherein the at least two content representations carry tag information; The representation module is used to input the object representation into the object branch of the content recommendation model for representation to obtain an object representation; and to input the at least two content representations into the content branch of the content recommendation model that matches the view of the at least two recommended contents, and to represent them through the embedding layer, the feature extraction layer and the multi-view representation layer to obtain at least two single-view content representations. The prediction module is used to determine prediction information corresponding to the at least two single-view content representations based on the object representation and the at least two single-view content representations respectively. The training module is used to determine the training loss based on the label information and the prediction information, and to train the model parameters of the content recommendation model based on the training loss; wherein, each content branch learns the relevant information of the recommended content under all views, realizing transfer learning between the recommended content in different views; the single-view content representation carries multi-view content information, and the multi-view content information refers to the content feature information of the recommended content under at least two views.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the content recommendation method based on the content recommendation model as described in any one of claims 1 to 9, or to implement the training method of the content recommendation model as described in claim 10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is loaded and executed by a processor to implement the content recommendation method based on the content recommendation model as described in any one of claims 1 to 9, or to implement the training method for the content recommendation model as described in claim 10.

15. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium. A processor retrieves the computer instructions from the computer-readable storage medium, causing the processor to load and execute them to implement the content recommendation method based on the content recommendation model as described in any one of claims 1 to 9, or to implement the training method of the content recommendation model as described in claim 10.

Images