Content push evaluation method, device, equipment and storage medium

Abstract

The present application provides a content push evaluation method, apparatus, device, and storage medium, relating to the field of artificial intelligence. The method includes: determining a set of behavioral features and candidate content features of a pending account; performing at least two feature integration operations, wherein in each feature integration operation: based on a first similarity corresponding to each behavioral feature, the behavioral feature is integrated and processed to obtain a corresponding sub-evaluation behavioral feature; the first similarity of each behavioral feature is determined by the content similarity between the behavioral feature and the candidate content feature at each interest point preset for the candidate content set; determining the evaluation behavioral feature of the pending account based on each obtained sub-evaluation behavioral feature; and determining the push evaluation value of the candidate content for the pending account based on at least a second similarity between the evaluation behavioral feature and the candidate content feature. This method can improve the accuracy of the push evaluation value of the determined candidate content.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to a content push evaluation method, apparatus, device and storage medium. Background Art

[0002] With the rapid development of Internet technology, recommendation technologies based on big data and artificial intelligence have emerged, and many content recommendation systems have emerged, which can intelligently recommend content of interest to users. Content recommendation systems can usually predict users' preferences by analyzing their historical behavior information, and thus recommend information of interest to users to meet their personalized recommendation needs. However, in general, whether to push the above-mentioned content to be recommended to users is determined only based on the analysis results of the user's historical behavior information and the single-dimensional features of the content to be recommended. The richness of the analyzed features in the analysis results is low, and therefore the accuracy of determining whether to push the content to be recommended to users is low. Summary of the Invention

[0003] The embodiments of the present application provide a content push evaluation method, apparatus, device, and storage medium for improving the accuracy of content push.

[0004] In a first aspect, the present application provides a content push evaluation method, comprising:

[0005] Determining a set of behavioral features and candidate content features of the account to be processed, wherein each behavioral feature in the set of behavioral features is determined based on historical behavioral information of the account to be processed, and the candidate content features are determined based on candidate content in the candidate content set;

[0006] performing at least two feature integration operations to obtain at least two sub-evaluation behavior features, wherein each feature integration operation includes: integrating each behavior feature in the behavior feature set based on a first similarity corresponding to the behavior feature to obtain a corresponding sub-evaluation behavior feature; wherein the first similarity of each behavior feature is determined based on a content similarity between the behavior feature and the candidate content feature at each interest point preset for the candidate content set;

[0007] Determining the evaluation behavior characteristics of the pending account based on the at least two sub-evaluation behavior characteristics;

[0008] Based at least on the second similarity between the evaluation behavior feature and the candidate content feature, a push evaluation value of the candidate content for the account to be processed is determined.

[0009] In a second aspect of the present application, a content push evaluation device is provided, comprising:

[0010] a feature determination unit, configured to determine a set of behavioral features and candidate content features of the account to be processed, wherein each behavioral feature in the set of behavioral features is determined based on historical behavioral information of the account to be processed, and each candidate content feature is determined based on candidate content in the candidate content set;

[0011] a first processing unit configured to perform at least two feature integration operations to obtain at least two sub-evaluation behavior features, wherein each feature integration operation includes: integrating each behavior feature in the behavior feature set based on its corresponding first similarity to obtain a corresponding sub-evaluation behavior feature; wherein the first similarity of each behavior feature is determined by the content similarity between the behavior feature and the candidate content feature at each interest point preset for the candidate content set;

[0012] a second processing unit, configured to determine the evaluation behavior characteristics of the account to be processed based on the at least two sub-evaluation behavior characteristics;

[0013] An evaluation unit is configured to determine a push evaluation value of the candidate content for the account to be processed based at least on the second similarity between the evaluation behavior feature and the candidate content feature.

[0014] In one possible implementation, the characteristic behavior includes a behavior vector, and the candidate content feature includes a candidate content vector; the first processing unit is specifically configured to: during a feature integration operation performed by an attention sub-network, perform the following operations on each of the behavior features:

[0015] Determining at least one comprehensive vector based on a behavior vector of a characteristic behavior in each of the behavior characteristics and the candidate content vector; the comprehensive vector includes at least one of an outer product, an inner product, and a difference between the characteristic vector and the candidate content vector;

[0016] performing a splicing process on the one behavior vector, the candidate content vector, and the at least one comprehensive vector;

[0017] Based on a set of fully connected networks in the one attention sub-network, the vector obtained by the splicing process is processed to obtain a first similarity corresponding to the one behavioral feature.

[0018] In one possible implementation, among the at least two attention sub-networks, the network structures of different attention sub-networks are the same, and the model parameters of different attention sub-networks are different.

[0019] In a possible implementation, the first processing unit is further configured to obtain the at least two attention sub-networks in the following manner:

[0020] Initialize the basic attention sub-network at least twice to obtain at least two attention sub-networks to be trained; the model parameters of different attention sub-networks to be trained are different;

[0021] Based on the target loss function, the at least two attention sub-networks to be trained are trained, and

[0022] When it is determined that the training reaches a preset end condition, each of the at least two attention sub-networks to be trained is output as a corresponding attention sub-network.

[0023] In a possible implementation, the evaluation unit is specifically configured to:

[0024] Obtaining at least one content push reference feature; the content push reference feature includes a profile feature of the account to be processed, a profile feature of the candidate content, and context information associated with the account to be processed;

[0025] Based on the at least one content push reference feature, the evaluation behavior feature, and the candidate content feature, a push evaluation value of the candidate content for the account to be processed is determined.

[0026] In a third aspect of the present application, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect when executing the program.

[0027] In a fourth aspect of the present application, a computer program product is provided, the computer program product comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method provided in the first aspect.

[0028] In a fifth aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions. When the computer instructions are executed on a computer, the computer executes the method described in the first aspect.

[0029] Since the embodiments of the present application adopt the above technical solution, at least the following technical effects are achieved:

[0030] In an embodiment of the present application, at multiple interest points pre-set for a set of candidate content to be pushed, the content similarity of each behavioral feature and the candidate content feature is analyzed respectively, and the first similarity of each behavioral feature is determined based on the content similarity. The first similarity obtained can reflect the similarity between the behavioral feature and the candidate content in multiple dimensions (i.e., multiple interest points). Therefore, the evaluation behavioral feature obtained by the sub-evaluation behavioral feature obtained after integrating the various behavioral features based on the first similarities can enhance the analysis of the similarity between the behavioral feature and the candidate content feature in multiple dimensions. Therefore, based on the second similarity between the evaluation behavioral feature and the candidate content feature, the push evaluation value of the determined candidate content for the account to be processed is highly accurate, thereby improving the accuracy of determining whether to push the candidate content to the account to be processed based on the push evaluation value. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A schematic diagram of an application scenario provided in an embodiment of the present application;

[0032] Figure 2 A schematic diagram of a point of interest set provided in an embodiment of the present application;

[0033] Figure 3 A schematic diagram of another point of interest set provided in an embodiment of the present application;

[0034] Figure 4 A schematic diagram of an interface for triggering a content push process provided in an embodiment of the present application;

[0035] Figure 5 A flowchart of a content push evaluation method provided in an embodiment of the present application;

[0036] Figure 6 A flowchart of obtaining sub-assessment behavior characteristics provided in an embodiment of the present application;

[0037] Figure 7 A schematic diagram of obtaining a first similarity corresponding to each behavioral feature provided in an embodiment of the present application;

[0038] Figure 8 An example diagram of the structure and processing process of a push evaluation model provided in an embodiment of the present application;

[0039] Figure 9 An example diagram of the structure and processing process of an attention sub-network provided in an embodiment of the present application;

[0040] Figure 10 This is an example diagram of the processing process of a push evaluation model provided in an embodiment of the present application;

[0041] Figure 11A schematic diagram of the structure of a content push evaluation device provided in an embodiment of the present application;

[0042] Figure 12 A structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to better understand the technical solutions provided by the embodiments of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods; in order to facilitate those skilled in the art to better understand the technical solutions of the present application, some concepts involved in the present application are described below.

[0044] 1) Accounts and pending accounts

[0045] Generally, an account represents the identity of a user on the Internet; in the embodiment of the present application, the account to be processed refers to the account to which content recommendation is currently required.

[0046] 2) Candidate content set and candidate content

[0047] In the era of self-media, content can generally refer to audio, video, graphics, etc.; in the embodiment of the present application, candidate content can be, but is not limited to, content that can be pushed to an account by a content recommendation system, and can be, but is not limited to, multimedia resources including at least one of text, audio, video, articles, graphics, pictures, etc., or any combination thereof; for example, candidate content can be a video containing a pet, or audio of a song, or graphics introducing a product, etc.; a candidate content set is a set of candidate content.

[0048] 3) Behavioral characteristics

[0049] The behavioral features in the embodiments of the present application may be features determined based on the historical behavioral information of an account, such as behavioral features may be features extracted from the historical behavioral information of an account, etc.; wherein, the embodiments of the present application do not limit the type and specific form of the historical behavioral information of an account, and those skilled in the art may set it according to actual needs.

[0050] 4) Points of Interest

[0051] In the embodiment of the present application, the points of interest may include the content type of the candidate content, and may also include the objects or elements contained in the candidate content; the details of the points of interest will be further explained below.

[0052] The embodiments of this application relate to artificial intelligence (AI), designed based on big data analysis and machine learning (ML) technologies within AI. AI is the theory, methods, techniques, and application systems for using 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 new intelligent machines that can respond in a manner similar to human intelligence.

[0053] Artificial intelligence (AI) is the study of the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making. AI technologies primarily encompass computer vision, natural language processing, and machine learning or deep learning. With the advancement of AI technology, research and application are expanding across a wide range of fields, including smart homes, smart customer service, virtual assistants, smart speakers, smart marketing, driverless and autonomous driving, robotics, and smart healthcare. As technology advances, AI will be applied in even more areas and play an increasingly important role.

[0054] Machine learning is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. 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, and inductive learning.

[0055] To make the objectives, technical solutions, and advantages of this application more clear, this application will be further described in detail below with reference to the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0056] The design concept of this application is described below.

[0057] In the process of content recommendation (or push), feature extraction is generally performed on the characteristics of a single dimension based on the user's historical behavior information and the content to be recommended, and then whether to push the above-mentioned content to be recommended to the user is determined based on the extraction result; however, in the above process, the dimension for extracting features of the user's historical behavior information and the current content to be recommended is very single, so it works better in recommendation scenarios with relatively single feature dimensions (also called content interest points) of the content contained in e-commerce and advertising recommendation objects. However, when applied to content recommendation scenarios with complex and diverse interest points such as entertainment and information, it is impossible to comprehensively and accurately extract the user's historical behavior information and the various interest points of the content to be recommended. Therefore, the accuracy of judging whether to push the content to be recommended to the user based on the result of feature extraction is also low. Therefore, how to improve the accuracy of content push is an issue that needs to be considered.

[0058] In view of this, the inventors have designed a content push evaluation method for improving the analysis dimension of the relationship between candidate content to be pushed and accounts to be processed, improving the accuracy of the push evaluation value of the determined candidate content for the accounts to be processed, and thus improving the accuracy of content recommendation. In this method, a set of behavioral features of the account to be processed can be determined based on the historical behavior information of the account to be processed, and candidate content features can be determined based on the candidate content; and based on information such as the content similarity between each behavioral feature and the candidate content features at each point of interest preset for the candidate content set, at least two feature integration operations are performed on each behavioral feature in the behavioral feature set to obtain at least two sub-evaluation behavioral features, and then based on each obtained sub-evaluation behavioral feature, an evaluation behavioral feature of the account to be processed is determined, and based on at least a second similarity between the evaluation behavioral feature and the candidate content feature, the push evaluation value of the candidate content for the account to be processed is determined.

[0059] In order to more clearly understand the design ideas of this application, the application scenarios in the embodiments of this application are introduced as examples below.

[0060] The content push evaluation method provided in this application can be applied to a computer device; wherein the computer device can be a terminal device or a server; that is, the content push evaluation method provided in this application can be applied to a terminal device, can also be applied to a server, can also be applied to a system including a terminal device and a server, and can be implemented through the interaction between the terminal device and the server.

[0061] See Figure 1 , an application scenario of a content push evaluation method is provided, the application scenario may include a terminal device 110 and a server 120; wherein the server 120 may be a server in a content recommendation system, and the terminal device 110 and the server 120 may communicate via a network, wherein:

[0062] The user can log in to the above-mentioned content recommendation system as a pending account through the terminal device 110 (such as but not limited to 110-1 or 110-2 shown in the figure); then the user can send a content recommendation request to the server 120 through the above-mentioned terminal device to request the server 120 to push candidate content in the candidate content set to it; wherein, after the terminal device 110 detects that the user logs in to the above-mentioned content recommendation system as a pending account, it can also directly send a content recommendation request to the server; after the terminal device 110 detects that the user opens the client of the above-mentioned content recommendation system on the terminal device 110, it can also directly send a content recommendation request to the server; the triggering method of the above-mentioned content recommendation request is only an exemplary introduction, and technicians in this field can set it according to actual needs.

[0063] After receiving the above-mentioned content recommendation request, server 120 (illustrated as 120-1, 120-2 or 120-3 in the figure) can determine the push evaluation value of each candidate content for the account to be processed based on the behavioral feature set of the account to be processed and the candidate content features of each candidate content in the candidate content set, and determine whether to push the candidate content in the candidate content set to the account to be processed based on the obtained push evaluation value; each of the above-mentioned behavioral features is determined based on the historical behavioral information of the above-mentioned account to be processed; wherein, the method of determining the push evaluation value of each candidate content for the account to be processed will be described in detail below.

[0064] The terminal device 110 in the embodiment of the present application can be a mobile terminal, a fixed terminal, a portable terminal or a vehicle-mounted terminal, such as a mobile phone, a station, a unit, a device, a multimedia computer, a multimedia tablet, an Internet node, a communicator, a desktop computer, a laptop computer, a notebook computer, a netbook computer, a tablet computer, a personal communication system (PCS) device, a personal navigation device, a personal digital assistant (PDA), an audio or video player, a digital camera or camcorder, a positioning device, a television receiver, a radio broadcast receiver, an e-book device, a gaming device or any combination thereof, including accessories and peripherals of these devices or any combination thereof.

[0065] The server 120 in the embodiment of the present application can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or multiple cloud servers (such as but not limited to server 120-1, server 120-2 or server 120-3 shown in the figure) that provide basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms in cloud computing technology; the functions of the above-mentioned server 120 can be implemented by one or more cloud servers, and can also be implemented by one or more cloud server clusters, etc.

[0066] Cloud computing is a computing model that distributes computing tasks across a resource pool consisting of a large number of computers, enabling various application systems to access computing power, storage space, and information services as needed. The network that provides these resources is called the "cloud." Cloud computing infrastructure providers establish a cloud computing resource pool (referred to as a cloud platform), commonly referred to as an Infrastructure as a Service (IaaS) platform. Within this resource pool, various virtual resources are deployed for external clients to choose from. The cloud computing resource pool primarily includes computing devices (virtualized machines, including operating systems), storage devices, and network devices. Based on logical functional divisions, the Platform as a Service (PaaS) layer can be deployed on top of the IaaS layer, and the Software as a Service (SaaS) layer can be deployed on top of the PaaS layer. SaaS can also be deployed directly on top of IaaS. PaaS is a platform for running software, such as databases and web containers. SaaS encompasses a wide range of business software, such as web portals and content recommendation systems.

[0067] In one possible application scenario, in an embodiment of the present application, cloud storage technology may be used to store the historical behavior information of the above-mentioned accounts to be processed, the behavioral feature set of the accounts to be processed, and information on at least one of the above-mentioned candidate content sets; wherein cloud storage (Cloud Storage) is a new concept extended and developed from the concept of cloud computing. A distributed cloud storage system (hereinafter referred to as a storage system) refers to a storage system that uses cluster applications, grid technology, and distributed storage file systems to bring together a large number of different types of storage devices (storage devices are also called storage nodes) in the network through application software or application interfaces to work together and provide external data storage and business access functions.

[0068] In one possible application scenario, to reduce communication latency, servers 120 can be deployed in various regions. Alternatively, for load balancing, different servers 120 can serve the regions corresponding to each terminal device 110. Blockchain technology is used to store historical behavior information of pending accounts, a set of behavioral characteristics of pending accounts, and information on at least one of the aforementioned candidate content sets, as well as to implement the content push evaluation method designed in the embodiments of this application. Multiple servers 120 share data through blockchain, and multiple servers 120 act as a data sharing system composed of multiple servers 120. For example, terminal device 110 is located at location a and communicates with server 120, while terminal device 110 is located at location b and communicates with other servers 120. Blockchain is a novel application model for computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Blockchain is essentially a decentralized database, a series of data blocks generated using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of the information (to prevent counterfeiting) and generate the next block. Blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer.

[0069] Each server 120 in the data sharing system has a node identifier corresponding to the server 120. Each server 120 in the data sharing system can store the node identifiers of other servers 120 in the data sharing system, so that the generated blocks can be broadcast to other servers 120 in the data sharing system according to the node identifiers of other servers 120.

[0070] Each server 120 may maintain a node identifier list as shown in the following table, storing the server 120 name and node identifier in the node identifier list. A node identifier may be an Internet Protocol (IP) address or any other information that can be used to identify the node. Table 1 uses the IP address as an example.

[0071] Table 1

[0072]

[0073] The following is a detailed introduction to the content push evaluation method in the embodiment of the present application. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the implementation methods of the present application are not limited in this respect.

[0074] based on Figure 1 In the application scenario, an example of a content push evaluation method involved in an embodiment of the present application is described below.

[0075] First, the points of interest and points of interest sets involved in the embodiments of the present application are described:

[0076] As an embodiment, the points of interest in the embodiment of the present application may include the content type of the candidate content; wherein, the specific method of dividing the content type is not limited, and those skilled in the art can set it according to actual needs; see Figure 2 In (a), when the candidate content is pictures and texts, the pictures and texts in the candidate content set can be divided into novels, news, finance, learning, job hunting, food, photography, language, and comics based on the semantics of the pictures and texts. Then, the novels, news, finance, learning, job hunting, food, photography, language, and comics categories can be used as the points of interest in the set of points of interest pre-set for the candidate content set; see Figure 2 In (b), when the candidate content is a video, the videos in the candidate content set can be divided into game category, news category, finance category, sports category, learning category, beauty category, job search category, food category, and comedy category based on the meaning expressed by the video, and then the game category, news category, finance category, sports category, learning category, beauty category, job search category, food category, comedy category, and job search category can be determined as the various interest points in the interest point set pre-set for the candidate content set.

[0077] As an embodiment, the interest point in the embodiment of the present application refers to an object (also referred to as an element) contained in the candidate content; see Figure 3 When the candidate content is a video, some or all of the objects contained in each video in the candidate content set can be selected and determined as the various points of interest in the pre-set point of interest set for the candidate content set; wherein the above-mentioned objects may include but are not limited to specific groups of people, animals, plants, smart devices, games, clothing, buildings, etc., wherein the specific groups of people may include but are not limited to doctors, students, teachers, athletes, etc.; animals may include but are not limited to pets (such as dogs, cats, rabbits, etc.), wild animals, poultry (such as chickens, ducks, geese, etc.); plants may include wheat, corn, flowers, fruit trees, etc.; smart devices may include mobile phones, PDAs, desktop notebooks, cameras, robots, etc.; games may include shooting games, competitive games, chess and card games, etc.; clothing may include short sleeves, T-shirts, jackets, sweaters, casual pants, jeans, long boots, sandals, etc.; buildings may include hospitals, railway stations, hotels, supermarkets, etc.; technicians in this field can set the above-mentioned objects according to actual needs.

[0078] The following describes the process of the push evaluation method for triggering content; in an embodiment of the present application, when the terminal device determines that the content push trigger conditions are met, it can send a content recommendation request to the server 120, and then the server 120 executes the content push process; wherein, during the content push process, the server can select part or all of the candidate content from the candidate content set, and execute the content push evaluation process for each selected candidate content; wherein the content push evaluation process will be described in detail below.

[0079] As an embodiment, the above-mentioned content push trigger conditions are not limited. Technical personnel in this field can set them according to actual needs. For example, it can be, but not limited to, satisfying one or any combination of the following trigger conditions 1 to 4 as satisfying the content push trigger conditions.

[0080] Trigger condition 1: The terminal device 110 determines that the user logs into the content recommendation system as the account to be processed.

[0081] Trigger condition 2: After detecting that a user has logged into the content recommendation system as the account being processed, terminal device 110 executes a request for recommendation of candidate content. The specific method for requesting recommendation is not limited and can be configured by those skilled in the art based on actual needs.

[0082] For example, see Figure 4 In (a), after the user logs in to the video application (i.e., the client of the above-mentioned content recommendation system) as the account to be processed, the terminal device 110 can display the video screen of the video currently recommended by the video application (i.e., the currently recommended candidate content) in area 401 on the page of the video application. The above-mentioned request recommendation operation can be a sliding operation on the above-mentioned video screen, wherein the sliding direction of the sliding operation is not limited. Those skilled in the art can set it according to actual needs, such as being able to set the sliding direction to any one or more of the sliding directions 1 to 6 shown in the figure, but not limited to the sliding direction.

[0083] See Figure 4 In (b), after the user logs in to the video application (i.e., the client of the above-mentioned content recommendation system) as the account to be processed, the terminal device 110 can display the video screen of the video currently recommended by the video application in area 402 on the page of the video application. The user can further trigger the above-mentioned request recommendation operation by clicking a video channel button (such as but not limited to 403-1 or 403-2 shown in the figure).

[0084] Trigger condition 3: The user opens the client of the content recommendation system on the terminal device 110 without logging into the content recommendation system as any account.

[0085] Trigger condition 4: The user browses the candidate content currently pushed by the above-mentioned content recommendation system through the terminal device 110 without logging into the above-mentioned content recommendation system as any account; for example, if the content pushed by the content recommendation system is a video, when the user browses the current video through the terminal device, it can be considered that the content push trigger condition is met.

[0086] It should be noted that the above trigger conditions 1 to 4 are only exemplary introductions, and those skilled in the art can set them according to actual needs.

[0087] As an example, please refer to Figure 5 , which is a flowchart of a content push evaluation method designed in an embodiment of the present application, is applied to perform a content push evaluation process for each selected candidate content, specifically including the following steps S501 to S504:

[0088] Step S501: Determine a behavioral feature set and candidate content features of the account to be processed, wherein the behavioral feature set includes at least one behavioral feature, each behavioral feature is determined based on historical behavioral information of the account to be processed, and the candidate content features are determined based on candidate content in the candidate content set.

[0089] As an embodiment, the above-mentioned historical behavior information may include information associated with the historical behavior of the account to be processed, such as historical behavior information may be but not limited to behavior description data associated with historical behavior, etc.; in the embodiment of the present application, information associated with a historical behavior can be regarded as a historical behavior information, and feature processing can be performed on the above-mentioned historical behavior information to obtain a behavior feature, such as feature extraction can be performed on the above-mentioned historical behavior information to obtain a behavior feature vector that represents the semantics contained in the historical behavior information, and the behavior feature vector can be regarded as a behavior feature, etc.

[0090] Among them, when the candidate content in the embodiment of the present application is a video, the above-mentioned behavioral features may be but are not limited to the vector corresponding to a certain video browsed by the account to be processed, that is, the above-mentioned behavioral features may be but are not limited to containing a positive feedback sequence, and the positive feedback sequence may be but is not limited to containing the vector corresponding to the identifier (such as id) of a certain video browsed, liked or commented on by the account to be processed; the candidate behavioral features may be a set of vectors corresponding to all videos that have been browsed, liked or commented on by the account to be processed in history; the candidate content set includes a set of candidate content calculated in batches. In this example, the candidate content set may include multiple candidate videos that can be pushed to the account to be processed; the candidate content is the candidate video corresponding to the candidate video for which the push evaluation value for the account to be processed currently needs to be predicted, and the candidate content feature is the vector corresponding to the candidate video.

[0091] As an embodiment, in the embodiment of the present application, part or all of the historical behavior information can be obtained from the historical behavior information associated with all historical behaviors triggered by the account to be processed within a preset time period, and then feature extraction is performed based on the obtained historical behavior information to obtain the behavior characteristics corresponding to each historical behavior information, and generate a behavior feature set of the account to be processed; wherein, the embodiment of the present application does not limit the above-mentioned preset time period, and technical personnel in this field can set it according to actual needs, such as being able to, but not limited to, setting the above-mentioned preset time period to 60 days, 30 days, 10 days, 7 days, 5 days or 1 day, etc.

[0092] Step S502, perform at least two feature integration operations to obtain at least two sub-evaluation behavior features, wherein each feature integration operation includes: for each behavior feature in the above-mentioned behavior feature set, based on the first similarity corresponding to the behavior feature, integrating the behavior feature to obtain a corresponding sub-evaluation behavior feature; wherein the first similarity of each behavior feature is determined based on the content similarity between the behavior feature and the above-mentioned candidate content feature at each interest point preset for the above-mentioned candidate content set.

[0093] As an embodiment, in step S502, a sub-feature integration operation can be performed separately for each behavioral feature in the above-mentioned behavioral feature set, wherein each sub-feature integration operation includes: determining the content similarity of one behavioral feature among the above-mentioned behavioral features with the above-mentioned candidate content feature at each of the above-mentioned interest points; and based on the obtained content similarities, processing the corresponding behavioral feature to obtain the above-mentioned corresponding sub-evaluation behavioral feature; wherein the specific process of the sub-feature integration operation will be further explained below.

[0094] As an embodiment, in order to improve the efficiency and accuracy of determining the at least two sub-evaluation behavior features, and thereby improve the efficiency and accuracy of determining the push evaluation value of the candidate content for the account to be processed, the embodiment of the present application may also utilize a push evaluation network obtained based on machine learning to process the above-mentioned individual behavior features and candidate content features to obtain the push evaluation value of the candidate content for the account to be processed; specifically, a multi-modal (i.e., multiple) attention sub-network is set in the above-mentioned push evaluation model, and then the individual behavior features in the above-mentioned behavior account set and the above-mentioned candidate content features can be input into the push evaluation network, and at least two feature integration operations are performed respectively through the at least two attention sub-networks included in the above-mentioned push evaluation network to obtain the above-mentioned at least two sub-evaluation behavior features; wherein, one attention sub-network performs one feature integration operation to obtain one sub-evaluation behavior feature; The sub-evaluation behavior features obtained by different attention sub-networks are different; in the process of obtaining the sub-evaluation behavior features, each attention sub-network can determine the first similarity of each behavior feature in the behavior feature set, and for each behavior feature, the first similarity obtained through different attention sub-networks is different. Therefore, in the process of obtaining at least two sub-evaluation behavior features through the above-mentioned push evaluation network, each behavior feature has a different first similarity generated in the process of being processed by different attention sub-networks. That is, it can be understood that in the process of processing each behavior feature through at least two sub-evaluation behavior features, a multi-way weight corresponding to each behavior feature is generated (that is, multiple first similarities are generated for each behavior feature), where the multi-way weight of a behavior feature can be used to describe the similarity between the behavior feature and the candidate content feature at multiple angles (that is, at multiple points of interest).

[0095] In the embodiment of the present application, the push evaluation network is not limited and can be configured by those skilled in the art according to actual needs. The push evaluation network can include, but is not limited to, models such as a convolutional neural network (CNN) and a recurrent neural network (RNN). Each attention sub-network can be composed of a group of fully connected networks, each of which includes a layer of fully connected networks or multiple fully connected networks. The specific structure and processing process of the push evaluation model will be further described below.

[0096] Step S503: determining the evaluation behavior characteristics of the account to be processed based on the at least two sub-evaluation behavior characteristics.

[0097] As an embodiment, in the embodiment of the present application, the above-mentioned at least two sub-evaluation behavior features can be directly spliced, and the splicing result can be determined as the evaluation behavior feature of the account to be processed; the average value of the above-mentioned at least two sub-evaluation behavior features can also be determined as the evaluation behavior feature of the account to be processed; if the sub-evaluation feature is represented in the form of a vector, the above-mentioned at least two sub-evaluation behavior features include a 128-dimensional sub-evaluation behavior vector 1, a 128-dimensional sub-evaluation behavior vector 2 and a 128-dimensional sub-evaluation behavior vector 3, then in the process of splicing sub-evaluation behavior vector 1 to sub-evaluation behavior vector 3, the lengths of sub-evaluation behavior vector 1 to sub-evaluation behavior vector 3 can be added to obtain a 384-dimensional vector, and the 384-dimensional vector can be determined as the evaluation behavior feature of the account to be processed; since the above-mentioned sub-evaluation behavior vector 1 to sub-evaluation behavior vector 3 are vectors of the same size, the same positions of the three vectors in the above-mentioned sub-evaluation behavior vector 1 to sub-evaluation behavior vector 3 can also be averaged one by one, and the vector composed of the obtained means can be determined as the evaluation behavior feature of the account to be processed.

[0098] Step S504 : determining a push evaluation value of the candidate content for the account to be processed based on at least the second similarity between the evaluation behavior feature and the candidate content feature.

[0099] As an embodiment, in step S504, the second similarity between the evaluation behavior feature and the candidate content feature may be directly determined as the push evaluation value of the candidate content for the account to be processed;

[0100] As an embodiment, in order to improve the accuracy of the determined push evaluation value, at least one content push reference feature can also be obtained in step S504; the above-mentioned content push reference feature includes the portrait feature of the above-mentioned account to be processed, the portrait feature of the above-mentioned candidate content, and the context information associated with the above-mentioned account to be processed; and based on the above-mentioned at least one content push reference feature, the above-mentioned evaluation behavior feature and the above-mentioned candidate content feature, the push evaluation value of the above-mentioned candidate content for the account to be processed is determined.

[0101] As an example, see Figure 6 , providing a flowchart for obtaining sub-assessment behavior characteristics, specifically including steps S601 and S602:

[0102] Step S601 : performing a sub-feature integration operation on each of the above-mentioned behavioral features to obtain a first similarity corresponding to each of the behavioral features.

[0103] As an embodiment, the sub-feature integration operation of one of the above-mentioned behavioral features may include, but is not limited to, the following steps S6011 and S6012:

[0104] Step S6011: determining the content similarity between the one behavior feature and the candidate content feature at each of the interest points.

[0105] It should be noted that in step S6011, the content similarity between a behavior feature and a candidate content feature at a point of interest can be understood as determining the similarity between the behavior feature and the candidate content feature for the point of interest; that is, in the embodiment of the present application, multiple points of interest are preset for the candidate content in the candidate content set, and each of the above-mentioned attention sub-networks can determine the content similarity between each behavior feature and the candidate content feature at one or more points of interest. For example, assuming that three points of interest, namely, point of interest A, point of interest B and point of interest C, are preset for the candidate content in the candidate content set, then three attention sub-networks can be set in the above-mentioned push evaluation model, and each attention sub-network randomly determines the content similarity between each behavior feature and the candidate content feature for one of the points of interest A, point of interest B and point of interest C. For example, a certain attention sub-network can determine the content similarity between each behavior feature and the candidate content feature at point of interest A, etc.

[0106] In step S6012, the sum of the obtained content similarities is determined as the first similarity corresponding to the above-mentioned behavior feature. As an embodiment, in steps S6011 and S6012, the attention distribution probability of the above-mentioned behavior feature on the above-mentioned candidate content feature at each point of interest can be determined as the corresponding content similarity based on the attention distribution mechanism; for ease of explanation, please refer to Figure 7 , here is a specific example: assuming that the interest point set pre-set for the candidate content set includes interest points 1 to interest point 3, and the above-mentioned one behavior feature is behavior feature A1 obtained based on the historical behavior information A of the account to be processed, then it can be seen from the figure that the content similarity between behavior feature A1 and candidate behavior feature B1 of candidate content B at interest point 1 is P1, the content similarity between behavior feature A1 and candidate behavior feature B1 of candidate content B at interest point 2 is P2, and the content similarity between behavior feature A1 and candidate behavior feature B1 of candidate content B at interest point 3 is P3, then the first similarity corresponding to behavior feature A1 is the sum of P1, P2 and P3.

[0107] Step S602: Based on the obtained first similarities, weighted summation is performed on the corresponding behavior features to obtain the above-mentioned corresponding sub-evaluation behavior features.

[0108] As an embodiment, the corresponding sub-assessment behavior characteristics can be obtained based on, but not limited to, the following formula (1):

[0109] Formula (1)

[0110] In the above formula (1), k is the identifier of the feature integration operation; i is the identifier of the behavior feature, is the behavioral feature identified as i; is the sub-evaluation behavior feature obtained by the feature integration operation identified as k; It is the first similarity corresponding to the behavioral feature identified as i in the feature integration operation identified as k.

[0111] As an embodiment, in order to improve the accuracy of each sub-evaluation behavior feature obtained and simplify the computational complexity of obtaining each sub-evaluation behavior feature, in the above-mentioned step S402, the embodiment of the present application can also normalize each of the above-mentioned first similarities based on the sum of the above-mentioned first similarities, and then perform weighted summation on the corresponding behavior features based on the normalized first similarities to obtain the above-mentioned corresponding sub-evaluation behavior features.

[0112] Specifically, in the embodiment of the present application, each first similarity can be normalized based on, but not limited to, the principle of the following formula (2), and each sub-evaluation feature is obtained based on the principle of formula (3).

[0113] Formula (2)

[0114] In the above formula (2), k is the identifier of the feature integration operation; i is the identifier of the behavior feature; is the first similarity corresponding to the behavioral feature identified as i in the feature integration operation identified as k; It is the result of normalizing the first similarity corresponding to the behavioral feature identified as i in the feature integration operation identified as k, that is, It is the normalized first similarity corresponding to the behavioral feature identified as i in the feature integration operation identified as k.

[0115] Formula (3)

[0116] In the above formula (1), k is the identifier of the feature integration operation; i is the identifier of the behavior feature, is the sub-evaluation behavior feature obtained by the feature integration operation identified as k; is the first similarity after normalization corresponding to the behavioral feature identified as i in the feature integration operation identified as k; It is the behavioral feature marked as i.

[0117] As an embodiment, after obtaining each sub-assessment behavior feature through steps S601 and S602, the assessment behavior feature of the account to be processed can be determined based on, but not limited to, the principle of the following formula (4):

[0118] Formula (4)

[0119] In the above formula (4), k is the identifier of the feature integration operation; is the sub-evaluation behavior feature obtained by the feature integration operation identified as k; Characterize each Perform splicing; It is the evaluation behavior characteristic of the account to be processed.

[0120] As an embodiment, the process of obtaining the push evaluation value of the candidate content for the account to be processed through the push evaluation network is further described below:

[0121] See Figure 8 , an embodiment of the present application provides a push evaluation network, which includes at least an attention network layer 802 (also known as an activation layer), and the attention network layer 802 includes at least two attention sub-networks, such as but not limited to including the attention sub-network 1 (Activation Net 1) to the attention sub-network k (Activation Net k) shown in the figure; each attention sub-network 1 is used to perform a feature integration operation to obtain a corresponding sub-evaluation behavior feature, and the sub-evaluation behavior features obtained by different attention sub-networks are different; the process of the above-mentioned feature integration operation can be found in the above description and will not be repeated here.

[0122] As an embodiment, in order to improve the efficiency of obtaining push evaluation values ​​of candidate content for pending accounts, the feature extraction network further includes an embedding layer 801 (Embedding Layer), a behavior feature concatenation layer 803, and a fully connected activation layer 804; wherein:

[0123] The above-mentioned embedding layer 801 can be used to extract features from various historical behavior information of the account to be processed, and obtain the behavior features corresponding to each historical behavior information (such as but not limited to including behavior features 1 to behavior features n as illustrated in the figure); the embedding layer 801 can also extract features from the candidate content in the candidate content set to obtain corresponding candidate content features, and can also process the content push reference information associated with the account to be processed to obtain at least one content push reference feature, wherein the content push reference feature includes the portrait features of the above-mentioned account to be processed, the portrait features of the above-mentioned candidate content, and the context information associated with the above-mentioned account to be processed; the above-mentioned context information can include but is not limited to information such as the device, time, and location where the account to be processed logs in.

[0124] The above-mentioned behavior feature splicing layer 803 is used to determine the evaluation behavior feature of the account to be processed based on the at least two sub-evaluation behavior features obtained by the attention network layer 802.

[0125] The fully connected activation layer 804 may be, but is not limited to, composed of one or any module among a PRelu module, a Dice module, a Softmax module, etc.; the fully connected activation layer 804 is used to determine the push evaluation value of the candidate content for the account to be processed based on at least the second similarity between the evaluation behavior feature and the candidate content feature.

[0126] Furthermore, in order to improve the accuracy of the determined push evaluation value, the fully connected activation layer 804 can be used to process at least one content push reference feature, the evaluation behavior feature, and the candidate content feature, and the result of the processing can be determined as the push evaluation value of the candidate content for the account to be processed.

[0127] The following further describes the attention network layer 802:

[0128] As an embodiment, in order to make different sub-evaluation behavior features in the at least two sub-evaluation behavior features obtained different, the embodiment of the present application can be based on the same network structure, but construct the above-mentioned at least two attention sub-networks with different model parameters, that is, in the above-mentioned at least two attention sub-networks, the network structures of different attention sub-networks are the same, and the model parameters of different attention sub-networks are different; in this way, after the same behavior features and the same candidate content are input into the above-mentioned at least two attention sub-networks, the sub-evaluation behavior features output by different attention sub-networks are different.

[0129] As an example, see Figure 9 , a schematic diagram of the network structure and processing process of an attention sub-network is provided, wherein each attention sub-network may include an attention unit 901 (Activation Unit) and a pooling layer 902 (SUM Pooling), wherein: the attention unit 901 is used to process each input behavior feature and candidate content feature in the feature integration operation to obtain the first similarity corresponding to each behavior feature; the pooling layer 902 is used to perform weighted pooling (SUM Pooling) on ​​the corresponding behavior features based on each first similarity obtained by the attention unit to obtain the corresponding sub-evaluation behavior feature; wherein the above-mentioned attention unit 901 can obtain the first similarity corresponding to each behavior feature based on the principle of the following formula (5).

[0130] Formula (5)

[0131] In the above formula (1), k is the identifier of the feature integration operation (that is, k is the identifier of the attention sub-network); is the behavioral feature identified as i; It is the behavioral feature identified as i in the feature integration operation identified as k The corresponding first similarity; is a candidate content feature; Represents the processing of a set of fully connected networks in the attention subnetwork identified as k.

[0132] As an embodiment, if the above-mentioned behavioral features are expressed in the form of vectors and the above-mentioned candidate content features are expressed in the form of vectors, then in the following content, the behavioral features will be referred to as behavioral vectors and the candidate content features will be referred to as candidate content vectors. Then, in the attention network layer 802, during a feature integration operation through an attention sub-network, the following operations can also be performed on each of the above-mentioned behavioral features: based on the behavioral vector of a characteristic behavior in each of the above-mentioned behavioral features and the above-mentioned candidate content vector, at least one comprehensive vector is determined; the above-mentioned one behavioral vector, the above-mentioned candidate content vector and the above-mentioned at least one comprehensive vector are spliced; based on a set of fully connected networks in the above-mentioned one attention sub-network, the vector obtained by the splicing process is processed to obtain a first similarity corresponding to the above-mentioned one behavioral feature; wherein the above-mentioned comprehensive vector includes at least one of the outer product, inner product and difference of the above-mentioned feature vector and the above-mentioned candidate content vector.

[0133] As an embodiment, in order to improve the accuracy of at least two attention sub-networks in the above-mentioned attention network layer 802, the embodiment of the present application uses a machine learning method for training to obtain the above-mentioned at least two attention sub-networks; specifically, the basic attention sub-network is initialized at least twice to obtain at least two attention sub-networks to be trained; and then based on the target loss function, the above-mentioned at least two attention sub-networks to be trained are trained, and when it is determined that the training is carried out to a preset end condition, each of the above-mentioned at least two attention sub-networks to be trained is output as a corresponding attention sub-network; wherein, in the above-mentioned at least two attention sub-networks to be trained, the model parameters of different attention sub-networks to be trained are different; in the embodiment of the present application, the at least two attention sub-networks obtained by the above-mentioned method have different model parameters of different attention sub-networks, so the sub-evaluation behavior characteristics obtained by performing feature integration operations on different attention sub-networks are also different.

[0134] In the embodiments of the present application, the specific structure of the above-mentioned basic attention sub-network is not limited. Those skilled in the art can set it according to actual needs. For example, at least one layer of fully connected network can be set as the above-mentioned basic attention sub-network, etc. There are no excessive restrictions on the above-mentioned target loss function and the preset end conditions. Those skilled in the art can set it according to actual needs.

[0135] As an embodiment, in the process of constructing the above-mentioned attention network layer 802, in order to improve the expression ability of the evaluation behavior characteristics obtained by the attention network layer 802, the embodiment of the present application can also determine the second number of attention sub-networks in the attention network layer 802 based on the first number of interest points contained in the above-mentioned interest point set; the specific method of determining the second number based on the first number is not limited in the embodiment of the present application, and technical personnel in this field can set it according to actual needs or business experience; if the interest point set contains 100 interest points, considering the expression ability of the obtained evaluation behavior characteristics and the efficiency of determining the evaluation behavior characteristics, 3 to 5 attention sub-networks can be set in the attention network layer 802; among them, different attention sub-networks may have strong capabilities for analyzing different interest points based on the principle of the attention mechanism, and one of the attention sub-networks may have strong capabilities for analyzing only one interest point, or may have strong capabilities for analyzing multiple interest points.

[0136] Based on the above method, the meaning of a sub-evaluation behavior feature expression in an embodiment of the present application may be: the similarity between the behavior feature set of the account to be processed and the candidate content feature on a certain point of interest is relatively high, that is, the account to be processed has a high degree of interest in the certain point of interest; the meaning of a sub-evaluation behavior feature expression may also be: the similarity between the behavior feature set of the account to be processed and the candidate content feature on at least two points of interest is relatively high, that is, the account to be processed has a high degree of interest in the at least two points of interest; for example, if the points of interest include games, beauties, cute pets, news, and finance, then the meaning of a sub-evaluation behavior feature expression may be that the account to be processed is more interested in videos of the game category, and the meaning of a sub-evaluation behavior feature expression may also be that the account to be processed is more interested in videos of the game category and the beauty category, etc.

[0137] To facilitate understanding of the method provided in the embodiments of the present application, a specific example is provided below. In this example, a video is used as candidate content in a candidate content set. The pre-set interest point set for the candidate content set includes 10 interest points, namely, games, beauties, pets, news, finance, sports, learning, makeup, job hunting, and food. Taking into account the expressive power of the determined evaluation behavior features and the efficiency of determining the evaluation behavior features, three attention sub-networks are set in the attention network layer 802. Here, account U is regarded as the account to be processed. The behavior feature set of account U includes behavior feature 1, behavior feature 2, and behavior feature 3. The candidate content is a video containing the content of "a beautiful woman walking a pet dog."

[0138] In this example, based on Figure 10 (Right now Figure 8The push evaluation network shown in the diagram is applicable to the case of this example, and determines the push evaluation value of the candidate content for account U. Specifically, the historical behavior information set of account U (such as the set including historical behavior information 1 to historical behavior information 3 in the figure), candidate content, profile data of account U, and profile data of candidate content can be input into the push evaluation network, and then:

[0139] Embedding layer 801 performs feature processing on the input historical behavior information 1 to historical behavior information 3, candidate content, account U's profile data, and candidate content's profile data, respectively, to obtain behavior features 1 to behavior features 3, candidate content features, account U's profile features, and candidate content's profile features;

[0140] The attention sub-networks 1 to 3 in the attention network layer 802 process the input behavior features 1 to 3 and the candidate content features respectively to obtain sub-evaluation behavior features 1 to 3; each of the attention sub-networks is based on the above Figure 9 The principle shown, as well as the processes of the above formulas (5), (2), and (3), obtain the corresponding sub-evaluation features;

[0141] The behavior feature concatenation layer 803 can process the sub-evaluation behavior features 1 to 3 obtained by the attention network layer 802 based on the principle of the above formula (4) to obtain the evaluation behavior features of account U;

[0142] The fully connected activation layer 804 is used to process the evaluation behavior characteristics of account U, the characteristics of candidate content, the portrait characteristics of account U and the portrait characteristics of candidate content to obtain the push evaluation value of the candidate content for account U.

[0143] In this example, multiple attention sub-networks are introduced to learn the features of candidate content in different directions, enabling the push evaluation network to capture points of interest from different angles (i.e., different points of interest), thereby improving the predictive ability of the push evaluation network. Experiments have shown that it performs well on offline recommendation datasets in the process of recommending streaming content, achieving significantly improved results.

[0144] Please refer to Figure 11 Based on the same inventive concept, an embodiment of the present application provides a content push evaluation device 1100, comprising:

[0145] A feature determination unit 1101 is configured to determine a set of behavior features and candidate content features of a pending account, wherein each behavior feature is determined based on historical behavior information of the pending account, and each candidate content feature is determined based on candidate content in the candidate content set;

[0146] The first processing unit 1102 is configured to perform at least two feature integration operations to obtain at least two sub-evaluation behavior features, wherein each feature integration operation includes: integrating each behavior feature in the behavior feature set based on a first similarity corresponding to the behavior feature to obtain a corresponding sub-evaluation behavior feature; wherein the first similarity of each behavior feature is determined based on a content similarity between the behavior feature and the candidate content feature at each interest point preset for the candidate content set;

[0147] The second processing unit 1103 is configured to determine the evaluation behavior characteristics of the account to be processed based on the at least two sub-evaluation behavior characteristics;

[0148] The evaluation unit 1104 is configured to determine a push evaluation value of the candidate content for the account to be processed based on at least the second similarity between the evaluation behavior feature and the candidate content feature.

[0149] As an embodiment, the first processing unit 1102 is specifically configured to:

[0150] For each of the above-mentioned behavioral features, a sub-feature integration operation is performed respectively, wherein each sub-feature integration operation includes: determining a content similarity between one of the above-mentioned behavioral features and the above-mentioned candidate content feature at each of the above-mentioned interest points; and determining the sum of the obtained content similarities as a first similarity corresponding to the above-mentioned one behavioral feature;

[0151] Based on the obtained first similarities, weighted summation is performed on the corresponding behavior features to obtain the above-mentioned corresponding sub-evaluation behavior features.

[0152] As an embodiment, the first processing unit 1102 is specifically configured to:

[0153] Based on the sum of all the obtained first similarities, normalizing each first similarity;

[0154] Based on the normalized first similarities, weighted summation is performed on the corresponding behavior features to obtain the corresponding sub-evaluation behavior features.

[0155] As an embodiment, the first processing unit 1102 is specifically configured to:

[0156] Inputting the above-mentioned behavioral features and the above-mentioned candidate content features into a push evaluation network, wherein the push evaluation network includes at least two attention sub-networks;

[0157] The feature integration operation is performed on the at least two attention sub-networks mentioned above respectively to obtain the at least two sub-evaluation behavior features mentioned above; wherein, one attention sub-network performs one feature integration operation to obtain one sub-evaluation behavior feature; different attention sub-networks obtain different sub-evaluation behavior features.

[0158] In one embodiment, the characteristic behavior includes a behavior vector, and the candidate content feature includes a candidate content vector. The first processing unit 1102 is specifically configured to perform the following operations on each of the behavioral features during a feature integration operation using an attention sub-network:

[0159] Determining at least one comprehensive vector based on a behavior vector of a characteristic behavior in each of the aforementioned behavior characteristics and the aforementioned candidate content vector; the comprehensive vector includes at least one of an outer product, an inner product, and a difference between the aforementioned characteristic vector and the aforementioned candidate content vector;

[0160] performing splicing processing on the one behavior vector, the candidate content vector, and the at least one comprehensive vector;

[0161] Based on a set of fully connected networks in the above attention sub-network, the vector obtained by splicing is processed to obtain the first similarity corresponding to the above behavioral feature.

[0162] As an embodiment, in the at least two attention sub-networks mentioned above, the network structures of different attention sub-networks are the same, and the model parameters of different attention sub-networks are different.

[0163] As an embodiment, the first processing unit 1102 is further configured to obtain the at least two attention sub-networks in the following manner:

[0164] Initialize the basic attention sub-network at least twice to obtain at least two attention sub-networks to be trained; the model parameters of different attention sub-networks to be trained are different;

[0165] Based on the target loss function, train the at least two attention sub-networks to be trained, and

[0166] When it is determined that the training reaches a preset end condition, each of the at least two attention sub-networks to be trained is output as a corresponding attention sub-network.

[0167] As an embodiment, the evaluation unit 1104 is specifically configured to:

[0168] Obtaining at least one content push reference feature; the content push reference feature includes a profile feature of the account to be processed, a profile feature of the candidate content, and context information associated with the account to be processed;

[0169] Based on the at least one content push reference feature, the evaluation behavior feature, and the candidate content feature, a push evaluation value of the candidate content for the account to be processed is determined.

[0170] As an example, Figure 11 The device in can be used to implement any of the methods discussed above.

[0171] Based on the same inventive concept as the above method embodiment, a computer device is also provided in the embodiment of the present application. The computer device can be used for data processing based on push content. In one embodiment, the computer device can be a server, such as Figure 1 In this embodiment, the structure of the computer device can be as follows: Figure 12 As shown, it includes a memory 1201 , a communication module 1203 and one or more processors 1202 .

[0172] Memory 1201 is used to store computer programs executed by processor 1202. Memory 1201 may primarily include a program storage area and a data storage area. The program storage area may store an operating system and programs required for running instant messaging functions, while the data storage area may store various instant messaging messages and operating instruction sets.

[0173] Memory 1201 can be volatile memory, such as random-access memory (RAM); non-volatile memory, such as read-only memory, flash memory, a hard disk drive (HDD), or a solid-state drive (SSD); or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 1201 can be a combination of the aforementioned memories.

[0174] The processor 1202 may include one or more central processing units (CPUs) or digital processing units, etc. The processor 1202 is configured to implement the above method when calling the computer program stored in the memory 1201 .

[0175] The communication module 1203 is used to communicate with terminal devices and other servers.

[0176] The specific connection medium between the memory 1201, the communication module 1203 and the processor 1202 is not limited in the embodiment of the present application. Figure 12 In the embodiment, the memory 1201 and the processor 1202 are connected via a bus 1204. The bus 1204 is connected to the processor 1202 via a bus 1204. Figure 12 The connections between the other components are shown in bold lines, which are only for illustration and are not intended to be limiting. The bus 1204 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, Figure 12 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0177] The memory 1201 stores a computer storage medium, which stores computer-executable instructions for implementing the content push evaluation method of the embodiment of the present application. The processor 1202 is configured to execute the content push evaluation method.

[0178] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0179] Alternatively, if the integrated units of the invention are implemented as software functional modules and sold or used as independent products, they can also be stored on a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product, stored on a storage medium, includes instructions for enabling a computer device (such as a personal computer, server, or network device) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage media include various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0180] Based on the same technical concept, an embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions. When the above-mentioned computer instructions are executed on a computer, the computer executes the method for starting an application for instant messaging as discussed above.

[0181] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0182] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A content push evaluation method, characterized in that: include: Determining a set of behavioral features and candidate content features of the account to be processed, wherein each behavioral feature in the set of behavioral features is determined based on historical behavioral information of the account to be processed, and the candidate content features are determined based on candidate content in the candidate content set; Perform at least two feature integration operations to obtain at least two sub-evaluation behavior features, wherein each feature integration operation includes: For each behavior feature in the behavior feature set, a sub-feature integration operation is performed, each sub-feature integration operation comprising: determining a content similarity between a behavior feature and the candidate content feature at each interest point preset for the candidate content set; and determining the sum of the obtained content similarities as a first similarity corresponding to the behavior feature; wherein the interest point includes at least one of the following: a content type of the candidate content, an object included in the candidate content; Based on the obtained first similarities, weighted summing is performed on the corresponding behavioral features to obtain corresponding sub-evaluation behavioral features; Determining the evaluation behavior characteristics of the pending account based on the at least two sub-evaluation behavior characteristics; Based at least on the second similarity between the evaluation behavior feature and the candidate content feature, a push evaluation value of the candidate content for the account to be processed is determined.

2. The method according to claim 1, wherein The step of performing weighted summation on the corresponding behavioral features based on the obtained first similarities to obtain corresponding sub-evaluation behavioral features includes: Based on the sum of all the obtained first similarities, normalizing each first similarity; Based on the normalized first similarities, corresponding behavioral features are weightedly summed to obtain the corresponding sub-evaluation behavioral features.

3. The method according to claim 1, wherein The performing of at least two feature integration operations to obtain at least two sub-evaluation behavior features includes: Inputting each of the behavioral features and the candidate content features into a push evaluation network, wherein the push evaluation network includes at least two attention sub-networks; A feature integration operation is performed on each of the at least two attention sub-networks to obtain the at least two sub-evaluation behavior features; wherein, one attention sub-network performs one feature integration operation to obtain one sub-evaluation behavior feature; and different attention sub-networks obtain different sub-evaluation behavior features.

4. The method according to claim 3, wherein The behavior feature includes a behavior vector, and the candidate content feature includes a candidate content vector; In the process of performing a feature integration operation through an attention sub-network, the following operations are performed for each of the behavioral features: determining at least one comprehensive vector based on a behavior vector of one of the behavior features and the candidate content vector; wherein the comprehensive vector includes at least one of an outer product, an inner product, and a difference between the behavior vector and the candidate content vector; performing a splicing process on the behavior vector, the candidate content vector, and the at least one comprehensive vector; Based on a set of fully connected networks in the one attention sub-network, the vector obtained by the splicing process is processed to obtain a first similarity corresponding to the one behavioral feature.

5. The method according to claim 3, wherein In the at least two attention sub-networks, the network structures of different attention sub-networks are the same, and the model parameters of different attention sub-networks are different.

6. The method according to claim 3, wherein The at least two attention sub-networks are obtained as follows: Initialize the basic attention sub-network at least twice to obtain at least two attention sub-networks to be trained; the model parameters of different attention sub-networks to be trained are different; Based on the target loss function, the at least two attention sub-networks to be trained are trained, and When it is determined that the training reaches a preset end condition, each of the at least two attention sub-networks to be trained is output as a corresponding attention sub-network.

7. The method according to any one of claims 1 to 6, wherein: The step of determining a push evaluation value of the candidate content for the account to be processed based on at least the second similarity between the evaluation behavior feature and the candidate content feature includes: Obtaining at least one content push reference feature; the content push reference feature includes a profile feature of the account to be processed, a profile feature of the candidate content, and context information associated with the account to be processed; Based on the at least one content push reference feature, the evaluation behavior feature, and the candidate content feature, a push evaluation value of the candidate content for the account to be processed is determined.

8. A content push evaluation device, characterized in that: include: a feature determination unit, configured to determine a set of behavioral features and candidate content features of the account to be processed, wherein each behavioral feature in the set of behavioral features is determined based on historical behavioral information of the account to be processed, and each candidate content feature is determined based on candidate content in the candidate content set; The first processing unit is configured to perform at least two feature integration operations to obtain at least two sub-evaluation behavior features, wherein each feature integration operation includes: For each behavior feature in the behavior feature set, a sub-feature integration operation is performed, each sub-feature integration operation comprising: determining a content similarity between a behavior feature and the candidate content feature at each interest point preset for the candidate content set; and determining the sum of the obtained content similarities as a first similarity corresponding to the behavior feature; wherein the interest point includes at least one of the following: a content type of the candidate content, an object included in the candidate content; Based on the obtained first similarities, weighted summing is performed on the corresponding behavioral features to obtain corresponding sub-evaluation behavioral features; a second processing unit, configured to determine the evaluation behavior characteristics of the account to be processed based on the at least two sub-evaluation behavior characteristics; An evaluation unit is configured to determine a push evaluation value of the candidate content for the account to be processed based at least on the second similarity between the evaluation behavior feature and the candidate content feature.

9. The device according to claim 8, wherein The first processing unit is specifically configured to: Based on the sum of all the obtained first similarities, normalizing each first similarity; Based on the normalized first similarities, corresponding behavioral features are weightedly summed to obtain the corresponding sub-evaluation behavioral features.

10. The device according to claim 8, wherein The first processing unit is specifically configured to: Inputting each of the behavioral features and the candidate content features into a push evaluation network, wherein the push evaluation network includes at least two attention sub-networks; A feature integration operation is performed on each of the at least two attention sub-networks to obtain the at least two sub-evaluation behavior features; wherein, one attention sub-network performs one feature integration operation to obtain one sub-evaluation behavior feature; and different attention sub-networks obtain different sub-evaluation behavior features.

11. A computer program product, characterized in that The computer program product includes computer instructions, which are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the method according to any one of claims 1 to 7.

12. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

13. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and when the computer instructions are executed on a computer, the computer is caused to perform the method according to any one of claims 1 to 7.

Images