Artificial intelligence-based media account recommendation method, apparatus, and electronic device

Abstract

The application provides an artificial intelligence-based media account recommendation method and device, an electronic device, and a computer readable storage medium. The method comprises: obtaining a plurality of content features of at least one interactive media account of a user account, and obtaining a plurality of content features corresponding to a plurality of candidate media accounts; performing clustering processing and residual processing on the plurality of content features of each interactive media account to obtain a content vector of the interactive media account; performing clustering processing and residual processing on the plurality of content features of each candidate media account to obtain a content vector of each candidate media account; determining a candidate media account to be recommended based on the content similarity between the content vector of each interactive media account and the content vector of the plurality of candidate media accounts; and performing a recommendation operation corresponding to the user account based on the candidate media account to be recommended. Through the application, the account recommendation accuracy can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to artificial intelligence technology and blockchain technology, and in particular to a media account recommendation method and device based on artificial intelligence, an electronic device and a computer readable storage medium. BACKGROUND

[0002] Artificial intelligence (AI) is the use of digital computers or digital computer-controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0003] Media account recommendation is an important application of artificial intelligence, and an important part of media account recommendation is media account recommendation. The media account recommendation system in the related art learns the attention behavior of a user to obtain the similarity relationship between accounts, but it is difficult to depict the similarity relationship between accounts for accounts without posterior information (also known as cold start accounts), resulting in that the accounts recommended in this way cannot effectively meet the interests of the user, causing a poor experience for the user. SUMMARY

[0004] The embodiments of the present application provide a media account recommendation method and device based on artificial intelligence, an electronic device and a computer readable storage medium, which can mine the content vector of the account to improve the accuracy of recommendation.

[0005] The technical solution of the embodiments of the present application is as follows:

[0006] The embodiments of the present application provide a media account recommendation method based on artificial intelligence, comprising:

[0007] Obtain a plurality of content features of at least one interactive media account of a user account, and obtain a plurality of content features corresponding to a plurality of candidate media accounts respectively;

[0008] Cluster and residual process the plurality of content features of each of the interactive media accounts to obtain a content vector of the interactive media account;

[0009] Cluster and residual process the plurality of content features of each of the candidate media accounts to obtain a content vector of each of the candidate media accounts;

[0010] Determine a candidate media account to be recommended based on the content similarity between the content vector of each of the interactive media accounts and the content vector of the plurality of candidate media accounts;

[0011] Perform a recommendation operation corresponding to the user account based on the candidate media account to be recommended.

[0012] The embodiment of the application provides a media account recommendation device based on artificial intelligence, which comprises:

[0013] A feature module is configured to acquire a plurality of content features of at least one interactive media account of a user account and acquire a plurality of content features corresponding to a plurality of candidate media accounts respectively;

[0014] A vector module is configured to perform clustering processing and residual processing on the plurality of content features of each of the interactive media accounts to obtain a content vector of the interactive media account.

[0015] The vector module is further configured to perform clustering processing and residual processing on the plurality of content features of each of the candidate media accounts respectively to obtain a content vector of each of the candidate media accounts.

[0016] A similarity module is configured to determine a candidate media account to be recommended based on a content similarity between the content vector of each of the interactive media accounts and the content vectors of the plurality of candidate media accounts.

[0017] A recommendation module is configured to perform a recommendation operation corresponding to the user account based on the candidate media account to be recommended.

[0018] In the above scheme, the feature module is further configured to perform the following processing for each of the interactive media accounts: extract corresponding content features from a plurality of information published by the interactive media account as the content features of the interactive media account; and perform the following processing for each of the candidate media accounts: extract corresponding content features from a plurality of information published by the candidate media account as the content features of the candidate media account; wherein the type of the information comprises at least one of the following: video information, text information and image information.

[0019] In the above scheme, the vector module is further configured to perform the following processing for each of the interactive media accounts: perform clustering processing on the plurality of content features of the interactive media account to obtain at least one first clustering center; and perform residual processing on the plurality of content features of the interactive media account based on the at least one first clustering center to obtain the content vector of the interactive media account; and perform the following processing for each of the candidate media accounts: perform clustering processing on the plurality of content features of the candidate media account to obtain at least one second clustering center; and perform residual processing on the plurality of content features of the candidate media account based on the at least one second clustering center to obtain the content vector of the candidate media account.

[0020] In the above scheme, the content vector of the interactive media account is determined by a media account learning model, the media account learning model includes a local clustering core network and a normalization network; the vector module is further configured to: for each first clustering center, determining, by the local clustering core network, a first residual distribution of a plurality of content features of the interactive media account corresponding to the first clustering center; for each first clustering center, performing normalization processing on the first residual distribution corresponding to the first clustering center by the normalization network to obtain a normalization result corresponding to the first clustering center; and performing overall normalization processing on the normalization result of each first clustering center by the normalization network to obtain the content vector of the interactive media account.

[0021] In the above scheme, the media account learning model further includes a squeeze activation network; after the overall normalization processing on the normalization result of each first clustering center by the normalization network to obtain the content vector of the interactive media account, the vector module is further configured to: performing channel-based average pooling processing on the content vector of the interactive media account by the squeeze activation network to obtain global content features of the content vector of the interactive media account corresponding to each channel; performing full connection processing on the global content features of the content vector of the interactive media account corresponding to each channel by the squeeze activation network to obtain activation values of the content vector of the interactive media account corresponding to each channel; and performing dot multiplication processing on the activation values of the content vector of the interactive media account corresponding to each channel and the original content features of each channel in the content vector of the interactive media account, and updating the content vector of the interactive media account based on the dot multiplication processing result.

[0022] In the above scheme, the vector module is further configured to: performing convolution processing on each content feature of the interactive media account by the local clustering core network to obtain a corresponding convolution result; performing maximum likelihood function processing on the convolution result of each content feature of the interactive media account by the local clustering core network to obtain a corresponding maximum likelihood processing result; determining, by the local clustering core network, a first residual between each content feature of the interactive media account and the first clustering center; and performing weighted summation processing on the first residual corresponding to each content feature of the interactive media account with the corresponding maximum likelihood processing result of each content feature of the interactive media account as a weight to obtain the first residual distribution of the plurality of content features of the interactive media account corresponding to the first clustering center.

[0023] In the scheme, the content vector of the interactive media account is determined by a media account learning model, the media account learning model includes a local clustering core network and a normalization network; the vector module is further configured to: for each second clustering center, perform the following processing: determining, by the local clustering core network, a second residual distribution of a plurality of content features of the candidate media account corresponding to the second clustering center; for each second clustering center, performing the following processing: performing, by the normalization network, normalization processing on the second residual distribution of the plurality of content features of the candidate media account corresponding to the second clustering center to obtain a corresponding normalization result; and performing overall normalization processing on the normalization results corresponding to the plurality of content features of the candidate media account by the normalization network to obtain the content vector of the candidate media account.

[0024] In the scheme, the vector module is further configured to: performing convolution processing on each content feature of the candidate media account by the local clustering core network to obtain a corresponding convolution result; performing maximum likelihood function processing on the convolution result of each content feature of the candidate media account by the local clustering core network to obtain a corresponding maximum likelihood processing result; determining a second residual between each content feature of the candidate media account and the second clustering center by the local clustering core network; and performing weighted summation processing on the second residual corresponding to each content feature of the candidate media account with the corresponding maximum likelihood processing result of each content feature as a weight to obtain a second residual distribution of the plurality of content features corresponding to the second clustering center.

[0025] In the scheme, the media account learning model further includes a squeeze activation network; and the vector module is further configured to: after obtaining the content vector of the candidate media account by performing overall normalization processing on the normalization result of each second clustering center by the normalization network, performing channel-based average pooling processing on the content vector of the candidate media account by the squeeze activation network to obtain global content features of the content vector of the candidate media account corresponding to each channel; performing fully connected processing on the global content features of the content vector of the candidate media account corresponding to each channel by the squeeze activation network to obtain activation values of the content vector of the candidate media account corresponding to each channel; and performing dot multiplication processing on the activation values of the content vector of the candidate media account corresponding to each channel and the original content features of each channel in the content vector of the candidate media account, and updating the content vector of the candidate media account based on the dot multiplication processing result.

[0026] In the above scheme, the similarity module is further configured to: determine an account similarity between the account vector of each of the interactive media accounts and the account vectors of the plurality of candidate media accounts; fuse the content similarity and the account similarity between each of the interactive media accounts and the plurality of candidate media accounts to obtain a similarity between each of the interactive media accounts and the plurality of candidate media accounts; and sort the similarities between each of the interactive media accounts and the plurality of candidate media accounts in descending order, and select at least one candidate media account with a high ranking in the descending order as a candidate media account to be recommended.

[0027] In the above scheme, the similarity module is further configured to: for each of the candidate media accounts, extract a plurality of account features from account information of the candidate media account, and compress the plurality of account features of the candidate media account to obtain an account vector of the candidate media account; for each of the interactive media accounts, extract a plurality of account features from account information of the interactive media account, and compress the plurality of account features of the interactive media account to obtain an account vector of the interactive media account; determine an account similarity between the account vector of each of the interactive media accounts and the account vectors of the plurality of candidate media accounts; and for each of the interactive media accounts, determine an account similarity between the account vector of the interactive media account and the account vector of each of the candidate media accounts, and determine a content similarity between the content vector of the interactive media account and the content vector of each of the candidate media accounts; and average the account similarities and the content similarities.

[0028] In the above scheme, the similarity module is further configured to: perform embedding processing on the plurality of account features of the candidate media account to obtain a plurality of account embedding features of the candidate media account; perform weighted summation processing on the plurality of account embedding features of the candidate media account based on weights of the plurality of account features of the candidate media account to obtain an account vector corresponding to the candidate media account; perform embedding processing on the plurality of account features of the interactive media account to obtain a plurality of account embedding features of the interactive media account; and perform weighted summation processing on the plurality of account embedding features of the interactive media account based on weights of the plurality of account features of the interactive media account to obtain an account vector corresponding to the interactive media account.

[0029] In the above scheme, the similarity module is further configured to: perform embedding processing on the plurality of account features and the plurality of content features of the candidate media account to obtain a plurality of account embedding features and a plurality of content embedding features of the candidate media account; perform weighted summation processing on the plurality of account embedding features and the plurality of content embedding features of the candidate media account based on weights of the plurality of account embedding features and the plurality of content embedding features to obtain an account vector corresponding to the candidate media account; perform embedding processing on the plurality of account features and the plurality of content features of the interactive media account to obtain a plurality of account embedding features and a plurality of content embedding features of the interactive media account; and perform weighted summation processing on the plurality of account embedding features and the plurality of content embedding features of the interactive media account based on weights of the plurality of account embedding features and the plurality of content embedding features to obtain an account vector corresponding to the interactive media account.

[0030] In the above scheme, the content vector of the interactive media account is determined by a media account learning model, and the media account learning model includes a local clustering core network and a normalization network. The device further includes a training module configured to: before obtaining the plurality of content features of the at least one interactive media account of the user account and obtaining the plurality of content features corresponding to the plurality of candidate media accounts respectively, train the media account learning model by: obtaining a plurality of media account samples, and constructing a plurality of first-type triple samples based on the number of associated users of the plurality of media account samples; performing disassembly processing on the plurality of first-type triple samples to obtain a plurality of media account samples to predict a sample content vector of each of the plurality of media account samples by the media account learning model; determining second-type triple samples that meet a training condition according to the sample content vector; and substituting the sample content vector corresponding to each of the media account samples in the second-type triple samples into a triple loss function to determine parameters of the media account learning model when the triple loss function reaches a minimum value.

[0031] In the above scheme, the training module is further configured to: take any one of the media account samples as a first media account sample, obtain a second media account sample having the same label as the first media account sample from the plurality of media account samples, determine a first content distance between the sample content vector of the second media account sample and the sample content vector of the first media account sample; obtain a third media account sample having a different label from the first media account sample from the plurality of media account samples, determine a second content distance between the sample content vector of the third media account sample and the sample content vector of the first media account sample; extract the second media account sample and the third media account sample satisfying the following training condition from the plurality of second media account samples and the plurality of third media account samples: the second content distance corresponding to the third media account sample is greater than the first content distance corresponding to the second media account sample; the difference between the second content distance corresponding to the third media account sample and the first content distance corresponding to the second media account sample is less than a first threshold; and group the first media account sample, the second media account sample and the third media account sample satisfying the training condition to form the second type of triple sample.

[0032] In the above scheme, the training module is further configured to: obtain an associated user of each remaining media account sample, wherein the remaining media account sample is a media account sample different from the first media account sample in the plurality of media account samples; determine a user intersection between the associated user of each remaining media account sample and the associated user of the first media account sample; determine a remaining media account sample having a number of elements in the user intersection exceeding a second threshold as a second media account sample having the same label as the first media account sample; and determine a remaining media account sample having a number of elements in the user intersection less than a third threshold as a third media account sample having a different label from the first media account sample.

[0033] Embodiments of the present application provide an electronic device, comprising:

[0034] a memory configured to store executable instructions;

[0035] a processor configured to execute the executable instructions stored in the memory to implement the media account recommendation method based on artificial intelligence provided by the embodiments of the present application.

[0036] Embodiments of the present application provide a computer readable storage medium storing executable instructions for being executed by a processor to implement the media account recommendation method based on artificial intelligence provided by the embodiments of the present application.

[0037] Embodiments of the present application have the following beneficial effects:

[0038] By acquiring the content features of the media accounts, and clustering and residual processing the content features, the feature distribution difference of the content features itself can be hidden, only the distribution difference of the content features and the cluster center is reserved, the feature distribution of the media accounts in the content dimension is learned with higher efficiency, the relationship between the accounts in the content dimension is established, since the interaction between the accounts and the users in the recommendation system is carried out to a large extent with content as the carrier, the features of the media accounts are more accurately described, especially the cold-start media accounts, and then the recommendation accuracy of the media accounts based on the similarity of the content vectors is improved. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figures 1A-1D is a schematic diagram of the principle of the vector learning method in the related art;

[0040] Figure 2A is a structural schematic diagram of the media account recommendation system based on artificial intelligence provided by the embodiment of the present application;

[0041] Figure 2B is an application schematic diagram of the media account recommendation method based on the blockchain provided by the embodiment of the present application;

[0042] Figure 3 is a structural schematic diagram of an electronic device provided by the embodiment of the present application;

[0043] Figure 4 is a structural schematic diagram of the media account learning model of the media account recommendation method based on artificial intelligence provided by the embodiment of the present application;

[0044] Figures 5A-5C is a flow schematic diagram of the media account recommendation method based on artificial intelligence provided by the embodiment of the present application;

[0045] Figure 6 is an application architecture schematic diagram of the blockchain network provided by the embodiment of the present application;

[0046] Figure 7 is a structural schematic diagram of the blockchain in the blockchain network 600 provided by the embodiment of the present application;

[0047] Figure 8 is a functional architecture schematic diagram of the blockchain network 600 provided by the embodiment of the present application;

[0048] Figure 9 is a schematic diagram of the principle of the media account recommendation based on artificial intelligence provided by the embodiment of the present application;

[0049] Figure 10 is a training sample schematic diagram provided by the embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings, and the described embodiments should not be regarded as limiting the present application, and all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0051] In the following description, "some embodiments" are referred to, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0052] In the following description, the terms "first\second\third" are only to distinguish similar objects, and do not represent a specific order of the objects, and it can be understood that "first\second\third" can be interchanged in a specific order or sequence as allowed, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0054] Before the embodiments of the present application are further described in detail, the terms and phrases involved in the embodiments of the present application are explained, and the terms and phrases involved in the embodiments of the present application are applicable to the following explanations.

[0055] 1) InceptionV3: InceptionV3 network is a very deep convolutional network developed by Google, which is pre-trained on the pre-training set ImageNet and can be used to extract frame features.

[0056] 2) One-to-three scene: video floating layer. In the information flow product, after a user clicks on a certain video on the main information flow, enters the video playing page, and can continuously slide down and push other videos on the page, the video floating layer refers to the video recommendation page.

[0057] 3) Triplet loss: Triplet loss is first used in face recognition tasks, and its goal is to make the faces of the same person as "close" as possible in the feature space, and as "far" as possible from other faces. A new vector representation of the face is trained through online triplet minimization, which is defined on the concept of triplet <baseline sample, positive sample, negative sample>, and the goal is to make the distance between the baseline sample and the positive sample smaller than the distance between the baseline sample and the negative sample, or the absolute value of the difference between the two is greater than a certain threshold.

[0058] 4) Media account, an account for publishing (e.g., publishing in a news feed) articles or videos, some accounts are official accounts for publishing news information, such as news public accounts and newspaper public accounts, and some are self-media accounts for publishing articles in specific fields, such as entertainment public accounts and artificial intelligence public accounts.

[0059] 5) Collaborative filtering, using the preferences of a group of people with similar interests and common experiences to recommend information of interest to the user.

[0060] 6) One-to-three scenario: video floating layer in a news feed product, in response to a user's click operation on a certain video on the main news feed, entering a video playback page, in which other videos can be continuously pushed by sliding down, and the video floating layer refers to the video playback page.

[0061] 7) Blockchain network, a collection of nodes that add new blocks to the blockchain through consensus.

[0062] 8) Smart contract, also known as chain code or application code, a program deployed in the nodes of the blockchain network, the nodes execute the smart contract called in the received transaction to update or query the key-value data of the state database.

[0063] 9) Consensus, a process in the blockchain network for reaching an agreement on transactions in blocks among multiple nodes involved, the agreed block is appended to the end of the blockchain and used to update the state database.

[0064] In related technologies, user-based collaborative filtering can be used to recommend accounts, in which user information and account information are not considered, and the similarity between different users is explored based on user preference information, and user similarity is used for personalized recommendation. In user-based collaborative filtering, the user is the center, and a user group similar to the user's interest is observed, and other accounts of interest to the user group are recommended to the user.

[0065] In related technologies, an account vector learning model can also be used to learn account vectors, based on a user's account sequence, a user behavior information graph is constructed, and vector feature information of each account is learned by introducing a deep walk algorithm. See Figure 1A , Figure 1Ais a schematic diagram of the principle of the vector learning method in the related art, first, the sequence of accounts followed by the user is obtained, the sequence of accounts followed by user 1 is A, B, C, D and E, the sequence of accounts followed by user 2 is A, C, D and F, and the sequence of accounts followed by user 3 is E, C, B, F and A, see Figure 1B , Figure 1B is a schematic diagram of the principle of the vector learning method in the related art, based on the sequence of accounts in Figure 1A , a graph walk is constructed, for example, according to the sequence of user 1, a walk graph from A to B, from B to C, from C to D and from D to E can be constructed, and the same construction mode is continued to construct a walk graph according to the sequence of accounts of user 2 and user 3, see Figure 1C , Figure 1C is a schematic diagram of the principle of the vector learning method in the related art, Figure 1C , is shown based on the deep walk sequence obtained in Figure 1B , starting from any one account node, exploring edges of any length, obtaining multiple random deep walk sequences, for example, the sequence from A to B, see Figure 1D , Figure 1D is a schematic diagram of the principle of the vector learning method in the related art, Figure 1D shows a word vector learning model, finally Figure 1C , the deep walk sequence generated in the input word vector learning model for account vector learning, the first layer in the word vector learning model is the input layer, the second layer is the hidden layer (output learned account vector), and the third layer is the output layer (used for training), P represents a positive result and N represents a negative result, and the word vector learning model is trained through the output result.

[0066] In the embodiments of the present application, it is found that when the account vector is obtained based on user collaborative filtering, it tends to learn the similarity between accounts based on the user's follow behavior, thereby resulting in the need for post-data of the account. For accounts without post-data, this method cannot effectively learn the vector representation of the account, and it has a bias towards popular accounts. When using a word vector learning model to learn the account vector, although the similarity between accounts can be learned, for accounts that have few or no interactions with users, it is still difficult to effectively learn the account vector of the account. The above shortcomings result in the account vector obtained based on the collaborative method and the word vector model cannot be used in the account cold start scenario, because the accounts in the account cold start scenario are new accounts, neither have sufficient post-data nor are popular accounts.

[0067] To solve the above problems, the embodiment of the present application provides a media account recommendation method and device based on artificial intelligence, an electronic device and a computer readable storage medium, which can establish account vectors of multiple accounts from the content dimension, thereby effectively improving the recommendation accuracy for all accounts. The following describes an exemplary application of the electronic device provided by the embodiment of the present application. The electronic device provided by the embodiment of the present application can be implemented as a server. The following describes an exemplary application when the device is implemented as a server.

[0068] Referring to Figure 2A , Figure 2A FIG. 1 is a structural schematic diagram of a media account recommendation system based on artificial intelligence provided by the embodiment of the present application. The media account recommendation system can be used to support various media account recommendation scenarios, such as an application scenario of recommending media accounts publishing articles and / or videos, and the like. In the media account recommendation system, the terminal 400 is connected to the server 200 through a network. The network can be a wide area network or a local area network, or a combination of the two.

[0069] In some embodiments, the functions of the media account recommendation system are implemented based on various modules in the server 200. In the process of using the client by the user, the terminal 400 collects the media accounts as training samples, trains the media account learning model based on the obtained training samples, integrates the trained media account learning model in the server, and in response to the terminal 400 receiving a user's play operation on a certain video, the terminal 400 sends a recommendation request instruction to the server 200. The recommendation request instruction carries the interactive media account followed by the user. The server 200 determines the account vector of the interactive media account through the media account learning model, obtains the candidate media accounts from the database 500, determines the account vector of the candidate media account through the media account learning model, determines the similarity between the two account vectors, and determines the candidate media accounts meeting the similarity condition as the candidate media accounts to be recommended to the terminal 400. The server 200 sends the information of the candidate media accounts to be recommended to the terminal 400, such as the follow-up link of the candidate media account or the link of the content published by the candidate media account, so that the terminal 400 adds the candidate media account to the follow-up list or directly presents the published content of the candidate media account.

[0070] In some other embodiments, the terminal can also determine the account vector of the interactive media account through the media account learning model. The server 200 determines the account vector of the candidate media account through the media account learning model, determines the similarity between the two account vectors, and adds the candidate media accounts meeting the similarity condition to the follow-up list or directly presents the published content of the candidate media account.

[0071] In some embodiments, the server 200 can be a standalone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and basic cloud computing services such as big data and artificial intelligence platforms. The terminal can be a smartphone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, and the like, but is not limited thereto. The terminal and the server can be connected directly or indirectly through wired or wireless communication, and the present application is not limited in this regard.

[0072] Referring to Figure 2B , Figure 2B is an application schematic diagram of the media account recommendation method based on a blockchain provided by the embodiments of the present application, and the following describes an exemplary application of the blockchain network based on the embodiments of the present application. Figure 2B The blockchain network 600 (exemplarily showing that the blockchain network 600 includes the node 610-1 and the node 610-2), the server 200, the database 500, and the terminal 400 are included in the system 1000, and the following describes each of them.

[0073] The server 200 (mapped as the node 610-2) and the terminal 400 (mapped as the node 610-1) can both join the blockchain network 600 to become nodes therein, Figure 2B The terminal 400 is exemplarily mapped as the node 610-1 of the blockchain network 600 in the system 1000, and each node (such as the node 610-1 and the node 610-2) has a consensus function and a function of maintaining a state database (such as a key-value database).

[0074] The state database of each node (such as the node 610-1) records the interactive media account of the terminal 400, so that the terminal 400 can query the interactive media account recorded in the state database.

[0075] In some embodiments, in response to a play operation on a certain video, a plurality of servers 200 (each server is mapped as a node in the blockchain network) determine a candidate media account to be recommended, for a certain candidate media account to be recommended, when the number of nodes passing the consensus exceeds a node number threshold, the consensus is passed, the server 200 (mapped as the node 610-2) sends the candidate media account passing the consensus to the terminal 400 (mapped as the node 610-1), presents a video play page corresponding to the video, presents the video content of the video and the media account to be recommended in the video play page, and in response to the user's interactive operation on the media account to be recommended, displays that the media account to be recommended has been marked as an interactive media account, and stores the interactive media account on the chain.

[0076] Next, the structure of the electronic device for implementing the method for recommending a media account based on artificial intelligence provided by the embodiments of the present application is described. As described above, the electronic device provided by the embodiments of the present application can be a server 200 in Figure 2A . Referring to Figure 3 , Figure 3 is a structural schematic diagram of the electronic device provided by the embodiments of the present application, Figure 3 The server 200 shown in FIG. 1 includes at least one processor 210, a memory 250, and at least one network interface 220. The various components in the server 200 are coupled together by a bus system 240. It can be understood that the bus system 240 is used to realize the connection and communication between the components. The bus system 240 includes a data bus, a power bus, a control bus, and a status signal bus in addition to the data bus. However, for the purpose of clear illustration, all the buses are marked as the bus system 240 in Figure 3 .

[0077] The processor 210 can be an integrated circuit chip having a signal processing capability, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0078] The memory 250 can be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drive, optical disk drive, etc. The memory 250 can optionally include one or more storage devices that are physically located away from the processor 210.

[0079] The memory 250 includes volatile memory or non-volatile memory, and can also include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), and the volatile memory can be random access memory (RAM). The memory 250 described in the embodiments of the present application is intended to include any suitable type of memory.

[0080] In some embodiments, the memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, which are exemplarily described below.

[0081] The operating system 251 includes system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, and the like, for implementing various basic services and processing hardware-based tasks; and a network communication module 252 for reaching other computing devices via one or more (wired or wireless) network interfaces 220, such as Bluetooth, Wireless Fidelity (WiFi), Universal Serial Bus (USB), and the like.

[0082] In some embodiments, the artificial intelligence-based media account recommendation device provided by the embodiments of the present application can be implemented in a software manner, Figure 3 An artificial intelligence-based media account recommendation device 255 stored in the memory 250 is shown, which can be software in the form of programs and plug-ins, etc., including the following software modules: a feature module 2551, a vector module 2552, a similarity module 2553, a recommendation module 2554, and a training module 2555. These modules are logical, and thus can be combined or further split according to the implemented functions. The functions of each module will be described below.

[0083] The artificial intelligence-based media account recommendation method provided by the embodiments of the present application will be described below in conjunction with exemplary applications and implementations of the server 200 provided by the embodiments of the present application. Referring to Figure 4 , Figure 4 is a structural diagram of an artificial intelligence-based media account learning model provided by the embodiments of the present application, which can be applied to a public account recommendation system. The media account learning model includes a local clustering core network, a normalization network, and a squeeze-and-excitation network. The local clustering core network includes a convolution layer, a maximum likelihood function layer, and a local clustering core layer. The normalization network includes a first normalization layer and a second normalization layer. The squeeze-and-excitation network includes a squeeze layer and an activation layer. The squeeze-and-excitation network improves the quality of the representation generated by the network by explicitly modeling the interdependence between the channels of its convolution features.

[0084] The process of account learning through the media account learning model is as follows: Multiple content features are input into a local clustering core layer to obtain multiple cluster centers. These features are then input into a convolutional layer, which performs convolution processing on each content feature of the media account to obtain the corresponding convolution result. A maximum likelihood function layer processes the convolution result of each content feature of the media account using a maximum likelihood function to obtain the corresponding maximum likelihood processing result. The local clustering core layer determines the residual between each content feature of each media account and a certain cluster center. Using the maximum likelihood processing result corresponding to each content feature of the media account as weights, the residuals corresponding to each content feature of the media account are weighted and summed to obtain the residual distribution of multiple content features of the media account corresponding to the cluster center. Finally, the first normalization layer of the normalization network is used to normalize the residuals of the corresponding content features. The residual distribution of cluster centers is normalized to obtain the normalized result of the corresponding cluster center. The normalization result of each cluster center is then normalized as a whole through the second normalization layer of the normalization network to obtain the content vector of the media account. The content vector of the interactive media account is then averaged based on the channel through the squeezing layer of the squeezing activation network to obtain the global content feature of each channel corresponding to the content vector of the media account. The global content feature of each channel corresponding to the content vector of the media account is then fully connected through the activation layer of the squeezing activation network to obtain the activation value of each channel corresponding to the content vector of the media account. The activation value of each channel corresponding to the content vector of the media account is then multiplied by the original content feature of each channel in the content vector of the media account, and the content vector of the media account is updated based on the result of the multiplication.

[0085] Below, by Figure 2A Taking the execution of the media account recommendation system provided in this application embodiment by server 200 as an example, this paper describes the artificial intelligence-based media account recommendation method provided in this application embodiment. The media account recommendation system includes a training phase and an application phase. First, the training of the model in the artificial intelligence-based media account recommendation method provided in this application embodiment will be described.

[0086] In some embodiments, the content vector of the interactive media account is determined by a media account learning model, the media account learning model comprising at least a local cluster core network and a normalization network; the media account learning model is trained by: obtaining a plurality of media account samples, and constructing a plurality of first-type triple samples based on the number of associated users of the plurality of media account samples; performing disassembly processing on the plurality of first-type triple samples to obtain a plurality of media account samples, so as to predict the sample content vector of each media account sample by the media account learning model; determining a second-type triple sample meeting a training condition according to the sample content vector; and substituting the sample content vector corresponding to each media account sample in the second-type triple sample into a triple loss function to determine the parameters of the media account learning model when the triple loss function reaches a minimum value.

[0087] As an example, by training in an online triple minimization manner, assuming that there are one hundred thousand media account samples stored in the database, if the triple samples are directly constructed based on the one hundred thousand media samples, the calculation amount is too large, and it is difficult to traverse all the triples. Therefore, a plurality of media account samples are obtained from the one hundred thousand media account samples to construct a plurality of first-type triple samples. For example, when 10 first-type triple samples are constructed, three media account samples are included in each triple sample. Specifically, 10 media samples are first randomly obtained as the reference samples of each first-type triple sample. For the reference sample of any first-type triple sample, one positive sample and one negative sample corresponding to the reference sample are obtained, wherein the positive sample has the same label as the reference sample, and the negative sample has a different label from the reference sample. Here, whether the labels are the same is determined by the number of associated users (for example, the number of fans). If the number of associated users of two media account samples is more than a second threshold, it is determined that the two media account samples have the same label. If the number of associated users of two media account samples is less than a third threshold, it is determined that the two media account samples have different labels. The third threshold is less than the second threshold. When obtaining a positive sample corresponding to the reference sample, one positive sample can be randomly obtained from the media account samples having the same label as the reference sample, or the media account sample having the highest number of associated users can be obtained as the positive sample. The negative sample is obtained in the same way. Thus, the positive sample and the negative sample corresponding to the reference sample are combined with the reference sample to form a first-type triple sample corresponding to the reference sample. The 10 reference samples are processed in the same way to construct 10 first-type triples.

[0088] According to the above examples, the 10 first-type triple samples are disassembled to obtain a plurality of media account samples, i.e., 30 media account samples, which are the benchmark samples, positive samples and negative samples in the 10 first-type triple samples. Then, the media account learning model (initialized model or model obtained after several times of training) is used to predict the sample content vector of each media account sample, i.e., 30 sample content vectors corresponding to the 30 media account samples.

[0089] In some embodiments, the second-type triple sample meeting the training condition is determined according to the sample content vector, which can be realized by the following technical scheme: taking any one of the media account samples as a first media account sample, obtaining a second media account sample with the same label as the first media account sample from the plurality of media account samples, determining a first content distance between the sample content vector of the second media account sample and the sample content vector of the first media account sample; obtaining a third media account sample with a different label from the first media account sample from the plurality of media account samples, determining a second content distance between the sample content vector of the third media account sample and the sample content vector of the first media account sample; extracting the second media account sample and the third media account sample meeting the following training condition from the plurality of second media account samples and the plurality of third media account samples: the second content distance corresponding to the third media account sample is greater than the first content distance corresponding to the second media account sample; the difference between the second content distance corresponding to the third media account sample and the first content distance corresponding to the second media account sample is less than a first threshold; and combining the first media account sample, the second media account sample and the third media account sample meeting the training condition to form the second-type triple sample.

[0090] Taking the above example, taking any one of the 30 media account samples as the first media account sample A (which plays a role in the triple equivalent to the reference sample), obtaining at least one second media account sample with the same label as the first media account sample from the 29 media account samples (which plays a role in the triple equivalent to the positive sample), for example, there are 10 second media account samples, determine the first content distance between the sample content vector of the 10 second media account samples and the sample content vector of the first media account sample A, then there are 10 first content distances, then obtain at least one third media account sample with a different label from the first media account sample A from the 29 media account samples (which plays a role in the triple equivalent to the negative sample), for example, there are 5 third media account samples, determine the second content distance between the sample content vector of the 5 third media account samples and the sample content vector of the first media account sample A, then there are 5 second content distances, from the 10 second media account samples and the 5 third media account samples, extract the second media account samples and the third media account samples that meet the following training conditions: the second content distance corresponding to the third media account sample is greater than the first content distance corresponding to the second media account sample; the difference between the second content distance corresponding to the third media account sample and the first content distance corresponding to the second media account sample is less than the first threshold value, for example, the first distance corresponding to the second media account sample B is 1, the second distance corresponding to the third media account sample C is 5, and the first threshold value is 4.5, then since the first distance is less than the second distance, and the difference between the second distance and the first distance is 4, which is less than the first threshold value 4.5, it is determined that the second media account sample B and the third media account sample C meet the training conditions, thereby forming a second type of triple sample with the first media account sample A, the second media account sample B that meets the training conditions, and the third media account sample C that meets the training conditions. The second type of triple sample formed meets the training conditions. The above process is an example of taking any one of the 30 media account samples as the first media account sample. The 30 media account samples are traversed as the first media account sample to obtain multiple second type of triple samples that meet the training conditions.

[0091] As an example, the condition met by the first type of triple sample is in the label dimension, and the condition met by the second type of triple sample is in the label dimension and the content vector dimension. Based on 10 first type of triples (decomposed from 30 media account samples), 9000 (30*30*30) triples can be obtained in theory, thereby effectively expanding the triples. Then, from the 9000 triples, the second type of triples that meet the training conditions are determined based on the label, thereby effectively obtaining the second type of triples for parameter updating with less calculation.

[0092] In some embodiments, the above-mentioned obtaining, from the plurality of media account samples, a second media account sample having the same label as the first media account sample and obtaining, from the plurality of media account samples, a third media account sample having a different label from the first media account sample can be implemented by the following technical solutions: obtaining an associated user of each remaining media account sample; wherein the remaining media account sample is a media account sample in the plurality of media account samples that is different from the first media account sample; determining a user intersection between the associated user of each remaining media account sample and the associated user of the first media account sample; determining, as the second media account sample having the same label as the first media account sample, the remaining media account sample in which the number of elements in the user intersection exceeds a second threshold value; and determining, as the third media account sample having a different label from the first media account sample, the remaining media account sample in which the number of elements in the user intersection is less than a third threshold value.

[0093] As an example, taking any one of the 30 media account samples as the first media account sample A (which plays a role equivalent to a reference sample in a triple), at least one second media account sample having the same label as the first media account sample is obtained from the 29 media account samples (which plays a role equivalent to a positive sample in a triple), and at least one second media account sample having the same label as the first media account sample is obtained from the 29 media account samples (which plays a role equivalent to a positive sample in a triple). The specific second media account sample has the same label as the first media account sample, and the third media account sample has a different label from the first media account sample. Here, the number of associated users (which can be the number of fans) is used to determine whether the labels are the same. If the number of common associated users of the two media account samples exceeds a second threshold value, it is determined that the two media account samples have the same label. If the number of common associated users of the two media account samples is less than a third threshold value, it is determined that the two media account samples have different labels. The third threshold value is less than the second threshold value. When obtaining one second media account sample corresponding to the first media account sample, one can be randomly obtained from the media account samples having the same label as the first media account sample as the second media account sample, or the media account sample with the highest number of common associated users from the media account samples having the same label as the first media account sample as the second media account sample. The same applies to the third media account sample.

[0094] Continuing with the example of the server 200 in Figure 2A performing the method of the media account recommendation system provided by the embodiments of the present application, the following describes the application of the model in the artificial intelligence-based media account recommendation method provided by the embodiments of the present application. Referring to Figure 5A , Figure 5Ais a flowchart of a media account recommendation method based on artificial intelligence provided by an embodiment of the present application, which will be described in combination with Figure 5A The steps 101-105 shown will be described.

[0095] In step 101, a plurality of content features of at least one interactive media account of a user account are obtained, and a plurality of content features corresponding to a plurality of candidate media accounts are obtained.

[0096] In some embodiments, the plurality of content features of the at least one interactive media account of the user account obtained in step 101 can be realized by the following technical solution: performing the following processing for each interactive media account: extracting the corresponding content features from the plurality of information published by the interactive media account as the content features of the interactive media account; the plurality of content features corresponding to the plurality of candidate media accounts obtained in step 101 can be realized by the following technical solution: performing the following processing for each candidate media account: extracting the corresponding content features from the plurality of information published by the candidate media account as the content features of the candidate media account; wherein the type of information includes at least one of the following: video information, text information, image information.

[0097] As an example, there is at least one interactive media account associated with the user account, the interactive media account is a media account followed by the user account, or the interactive media account is a media account that the user account has interacted with, for each interactive media account, the content features corresponding to the interactive media account are obtained, the content features are derived from the information published by the interactive media account, for example, the interactive media account publishes 10 information, the 10 information includes at least one of the following: video information, text information, image information. The corresponding content features can be extracted from each information, so that 10 content features of the corresponding interactive media account are obtained, the candidate media account is a media account that is not followed by the user account or has not interacted with the user account, the cold start account is included in the candidate media account, the cold start account belongs to a media account that does not exist or has only a small amount of posterior data, which can be defined by account creation time or number of interactive users, for example, the account creation time does not exceed a time threshold (the account creation time does not exceed 7 days), for example, the number of interactive users of the account does not exceed a person-time threshold (the number of interactive users does not exceed 10 person-times), and the content features of the candidate media account can be obtained in the same way as the content features of the interactive media account.

[0098] In step 102, the plurality of content features of each interactive media account are clustered and residual processed to obtain a content vector of the interactive media account.

[0099] In some embodiments, referring to Figure 5B , Figure 5Bis a flowchart of a media account recommendation method based on artificial intelligence provided by an embodiment of the present application. In step 102, the plurality of content features of each interactive media account are clustered and residual processed to obtain the content vector of the interactive media account. The content vector of the interactive media account can be obtained by executing the following steps for each interactive media account: Figure 5B Steps 1021-1022 shown are implemented.

[0100] In step 1021, the plurality of content features of the interactive media account are clustered to obtain at least one first cluster center.

[0101] As an example, the plurality of content features of a certain interactive media account are clustered by the local clustering core network in the media account learning model. The operator for clustering processing has been updated in the training process of the media account learning model, so that in the inference stage, the plurality of content features of the interactive media account are inputted, and the plurality of content features are clustered by the local clustering core network in the media account learning model to obtain at least one first cluster center. In step 1021, the above-mentioned processing needs to be performed for each interactive media account respectively, because the processing object of the media account learning model is a single media account.

[0102] In step 1022, the plurality of content features of the interactive media account are residual processed based on the at least one first cluster center to obtain the content vector of the interactive media account.

[0103] In some embodiments, the content vector of the interactive media account is determined by the media account learning model, and the media account learning model includes a local clustering core network and a normalization network. In step 1022, the plurality of content features of the interactive media account are residual processed based on the at least one first cluster center to obtain the content vector of the interactive media account, which can be implemented by the following technical solution: for each first cluster center, the following processing is performed: by the local clustering core network, a first residual distribution of the plurality of content features of the interactive media account corresponding to the first cluster center is determined; for each first cluster center, the following processing is performed: by the normalization network, the first residual distribution corresponding to the first cluster center is normalized to obtain a normalization result corresponding to the first cluster center; and by the normalization network, the normalization results of each first cluster center are overall normalized to obtain the content vector of the interactive media account.

[0104] In some embodiments, the first residual distribution of the plurality of content features of the interactive media account corresponding to the first clustering center can be determined by the following technical solutions: performing convolution processing on each content feature of the interactive media account through the local clustering core network to obtain a corresponding convolution result; performing maximum likelihood function processing on the convolution result of each content feature of the interactive media account through the local clustering core network to obtain a corresponding maximum likelihood processing result; determining the first residual between each content feature of the interactive media account and the first clustering center through the local clustering core network; and performing weighted summation processing on the first residual corresponding to each content feature of the interactive media account by taking the corresponding maximum likelihood processing result of each content feature of the interactive media account as a weight to obtain the first residual distribution of the plurality of content features of the interactive media account corresponding to the first clustering center.

[0105] As an example, the plurality of content features are input to the convolution layer of the local clustering core network, the convolution processing is performed on each content feature of the interactive media account through the convolution layer to obtain a corresponding convolution result, the maximum likelihood function processing is performed on the convolution result of each content feature of the interactive media account through the maximum likelihood function layer of the local clustering core network to obtain a corresponding maximum likelihood processing result, the maximum likelihood function is a soft max function, the residual between each content feature of the interactive media account and a clustering center A is determined through the local clustering core layer of the local clustering core network, assuming that 10 content features are obtained, there are 10 residuals corresponding to the 10 content features, the residual is the difference between each content feature and the clustering center A, the weighted summation processing is performed on the residual corresponding to each content feature of the interactive media account by taking the corresponding maximum likelihood processing result of each content feature of the interactive media account as a weight to obtain the residual distribution of the plurality of content features of the interactive media account corresponding to the clustering center, the same processing is performed for each clustering center through the above processing mode, thereby obtaining the residual distribution corresponding to each clustering center, the residual distribution corresponding to each clustering center (for example, clustering centers A-C) is normalized through the first normalization layer of the normalization network to obtain the normalization result corresponding to each clustering center, that is, the normalization result corresponding to the clustering center A, the normalization result corresponding to the clustering center B, and the normalization result corresponding to the clustering center C, by obtaining the residual distribution of each clustering, the features representing the global meaning of the plurality of clusters are learned, thereby expressing the distribution of the features within the clustering range of each cluster, this distribution erases the feature distribution difference of the content features itself, only the effective distribution difference between the content features and the clustering center is retained, the normalization result of each clustering center (for example, clustering centers A-C) is overall normalized through the second normalization layer of the normalization network to obtain the content vector of the interactive media account.

[0106] In some embodiments, the media account learning model further comprises a squeeze activation network; after the normalization result of each first clustering center is overall normalized by the normalization network to obtain the content vector of the interactive media account, the content vector of the interactive media account is processed by channel-based average pooling by the squeeze activation network to obtain global content features of the content vector of the interactive media account corresponding to each channel; the global content features of the content vector of the interactive media account corresponding to each channel are processed by full connection by the squeeze activation network to obtain activation values of the content vector of the interactive media account corresponding to each channel; the activation values of the content vector of the interactive media account corresponding to each channel and the original content features of each channel in the content vector of the interactive media account are processed by point multiplication to update the content vector of the interactive media account based on the point multiplication result.

[0107] As an example, the content vector output by the normalization network has learned the features of the content dimension of the interactive media account, in order to further optimize the expression effect of the content vector on the content dimension features of the interactive media account, it is necessary to perform feature enhancement processing on the content vector output by the normalization network, the content vector of the interactive media account is processed by channel-based average pooling by the squeeze layer of the squeeze activation network to obtain global content features of the content vector of the interactive media account corresponding to each channel, the global content features of the content vector of the interactive media account corresponding to each channel are processed by full connection by the activation layer of the squeeze activation network to obtain activation values of the content vector of the interactive media account corresponding to each channel, the activation values of the content vector of the interactive media account corresponding to each channel and the original content features of each channel in the content vector of the interactive media account are processed by point multiplication to update the content vector of the interactive media account based on the point multiplication result.

[0108] The squeeze activation network mainly includes a squeeze layer and an activation layer. In the convolution process, the feature relationship of the channel and the spatial relationship learned by the convolution kernel are mixed together. The squeeze activation network is used to separate the mixture, so as to directly learn the feature relationship of the channel. Since the convolution is only operated in a local space, it is difficult to obtain enough information to extract the relationship between the channels. The squeeze layer encodes the entire spatial feature of a channel into a global feature (performs channel-based average pooling processing). In principle, a more complex aggregation strategy can also be used. The squeeze layer obtains the global description feature. Next, another operation is needed to grasp the relationship between the channels. The activation layer learns the nonlinear relationship between the channels. The learned relationship is not mutually exclusive. Based on this, a threshold mechanism in the form of an S-shaped activation function is used. In order to reduce complexity and improve generalization ability, a structure including two fully connected layers is used. The first fully connected layer serves as a dimension reduction function. The dimension reduction coefficient is a hyperparameter. Then, a linear rectified activation function is used for activation processing. The last fully connected layer is used to restore the original dimension. Finally, the learned activation values (values are 0 to 1) of each channel are multiplied by the original content features input to the squeeze layer. The above process actually learns the weight coefficients of each channel, so that the media account learning model has better discrimination ability for the features of each channel.

[0109] In step 103, the plurality of content features of each candidate media account are respectively subjected to clustering processing and residual processing, to obtain a content vector of each candidate media account.

[0110] In some embodiments, the plurality of content features of each candidate media account are respectively subjected to clustering processing and residual processing in step 103 to obtain a content vector of each candidate media account, which can be implemented by the following technical solution: for each candidate media account, the following processing is performed: clustering processing is performed on the plurality of content features of the candidate media account to obtain at least one second cluster center; residual processing is performed on the plurality of content features of the candidate media account based on the at least one second cluster center to obtain a content vector of the candidate media account.

[0111] In some embodiments, the content vector of the interactive media account is determined by a media account learning model, the media account learning model comprising a local clustering core network and a normalization network; the clustering processing of the plurality of content features of the candidate media account to obtain at least one second clustering center can be implemented by the following technical solution: for each second clustering center, the following processing is performed: determining, by the local clustering core network, a second residual distribution of the plurality of content features of the candidate media account corresponding to the second clustering center; for each second clustering center, the following processing is performed: performing, by the normalization network, normalization processing on the second residual distribution of the plurality of content features of the candidate media account corresponding to the second clustering center to obtain a corresponding normalization result; performing overall normalization processing on the normalization result corresponding to the plurality of content features of the candidate media account by the normalization network to obtain the content vector of the candidate media account.

[0112] In some embodiments, the determination of the second residual distribution of the plurality of content features of the candidate media account corresponding to the second clustering center by the local clustering core network can be implemented by the following technical solution: performing convolution processing on each content feature of the candidate media account by the local clustering core network to obtain a corresponding convolution result; performing maximum likelihood function processing on the convolution result of each content feature of the candidate media account by the local clustering core network to obtain a corresponding maximum likelihood processing result; determining, by the local clustering core network, a second residual between each content feature of the candidate media account and the second clustering center; performing weighted summation processing on the second residual corresponding to each content feature of the candidate media account with the corresponding maximum likelihood processing result of each content feature of the candidate media account as a weight to obtain the second residual distribution of the plurality of content features corresponding to the second clustering center.

[0113] In some embodiments, the media account learning model further comprises a squeeze activation network; after the overall normalization processing of the normalization result of each second clustering center by the normalization network to obtain the content vector of the candidate media account, performing channel-based average pooling processing on the content vector of the candidate media account by the squeeze activation network to obtain global content features corresponding to each channel of the content vector of the candidate media account; performing full connection processing on the global content features corresponding to each channel of the content vector of the candidate media account by the squeeze activation network to obtain activation values corresponding to each channel of the content vector of the candidate media account; performing dot multiplication processing on the activation values corresponding to each channel of the content vector of the candidate media account and the original content features of each channel in the content vector of the candidate media account, and updating the content vector of the candidate media account based on the dot multiplication processing result.

[0114] As an example, the specific embodiments of the corresponding candidate media account embodiment can refer to the specific embodiments introduced in the corresponding interactive media account embodiment, and the difference is only that the processing object is changed from an interactive media account to a candidate media account.

[0115] In some embodiments, when processing the content features through the media account learning model, the content features can be processed by dimensionality reduction and grouping, so as to effectively reduce the model parameters, thereby effectively improving the inference efficiency.

[0116] In step 104, based on the content similarity between the content vector of each interactive media account and the content vector of the plurality of candidate media accounts, the candidate media account to be recommended is determined.

[0117] In some embodiments, referring to Figure 5C , Figure 5C is a flowchart of the media account recommendation method based on artificial intelligence provided by the embodiments of the present application. The determination of the candidate media account to be recommended in step 104 based on the content similarity between the content vector of each interactive media account and the content vector of the plurality of candidate media accounts can be implemented by executing Figure 5C steps 1041-1043 shown in the figure.

[0118] In step 1041, the account similarity between the account vector of each interactive media account and the account vector of the plurality of candidate media accounts is determined.

[0119] In some embodiments, the determination of the account similarity between the account vector of each interactive media account and the account vector of the plurality of candidate media accounts in step 1041 can be implemented by the following technical solutions: for each candidate media account, the following processing is performed: extracting a plurality of account features from the account information of the candidate media account, compressing the plurality of account features of the candidate media account to obtain the account vector of the candidate media account; for each interactive media account, the following processing is performed: extracting a plurality of account features from the account information of the interactive media account, compressing the plurality of account features of the interactive media account to obtain the account vector of the interactive media account; determining the account similarity between the account vector of each interactive media account and the account vector of the plurality of candidate media accounts.

[0120] As an example, when learning the interactive media accounts and the candidate media accounts, learning can be performed from the content dimension, and learning can also be performed from the dimension of the accounts themselves, i.e., learning the account vectors of the interactive media accounts and the account vectors of the candidate media accounts respectively, and then determining the account similarities between the account vector of each interactive media account and the account vectors of the plurality of candidate media accounts. For example, there are two interactive media accounts A and B, 10 account similarities between the account vector of the interactive media account A and the account vectors of the 10 candidate media accounts are determined, and 10 account similarities between the account vector of the interactive media account B and the account vectors of the 10 candidate media accounts are determined.

[0121] In some embodiments, the compression processing of the plurality of account features of the candidate media account to obtain the account vector of the candidate media account can be implemented by the following technical solution: performing embedding processing on the plurality of account features of the candidate media account to obtain weights of the plurality of account embedding features of the candidate media account; performing weighted summation processing on the plurality of account embedding features of the candidate media account based on the weights of the plurality of account features of the candidate media account to obtain the account vector corresponding to the candidate media account.

[0122] In some embodiments, the compression processing of the plurality of account features of the candidate media account to obtain the account vector of the candidate media account can be implemented by the following technical solution: performing embedding processing on the plurality of account features of the candidate media account to obtain weights of the plurality of account embedding features of the candidate media account; performing weighted summation processing on the plurality of account embedding features of the candidate media account based on the weights of the plurality of account features of the candidate media account to obtain the account vector corresponding to the candidate media account.

[0123] As an example, when the plurality of account features of the candidate media account are compressed, the plurality of account features of the candidate media account are taken as input, and for each candidate media account, the account features corresponding to the candidate media account are obtained, the account features are derived from the account information of the candidate media account itself, and the account information includes at least one of the following: level information, category information, and label information. The corresponding account features can be extracted from each information. Since the account features are extracted from the above information, the account features are sparse features. The way of obtaining the account features of the interactive media account can refer to the way of obtaining the account features of the candidate media account. After obtaining the plurality of account features of the candidate media account, the plurality of account features can be embedded by a neural network model to obtain account embedding features of the plurality of account features of the candidate media account (the account embedding features are dense representations of the sparse account features). Based on the weights of the plurality of account features of the candidate media account, the plurality of account embedding features of the candidate media account are weighted and summed to obtain the account vector corresponding to the candidate media account. The weights involved in the embedding process are prior data or obtained through training. The training method is similar to the training method of the word vector model.

[0124] As an example, when the plurality of account features of the candidate media account are compressed, the plurality of account features of the candidate media account are taken as input, and for each candidate media account, the account features corresponding to the candidate media account are obtained, the account features are derived from the account information of the candidate media account itself, and the account information includes at least one of the following: level information, category information, and label information. The corresponding account features can be extracted from each information. Since the account features are extracted from the above information, the account features are sparse features. The way of obtaining the account features of the interactive media account can refer to the way of obtaining the account features of the candidate media account. After obtaining the plurality of account features of the candidate media account, the plurality of account features can be embedded by a neural network model to obtain account embedding features of the plurality of account features of the candidate media account (the account embedding features are dense representations of the sparse account features). Based on the weights of the plurality of account features of the candidate media account, the plurality of account embedding features of the candidate media account are weighted and summed to obtain the account vector corresponding to the candidate media account. The weights involved in the embedding process are prior data or obtained through training. The training method is similar to the training method of the word vector model.

[0125] In some embodiments, the compression processing of the plurality of account features of the interactive media account to obtain the account vector of the interactive media account can be implemented by the following technical solutions: embedding the plurality of account features of the interactive media account to obtain a plurality of account embedding features of the interactive media account; and based on the weights of the plurality of account features of the interactive media account, the plurality of account embedding features of the interactive media account are weighted and summed to obtain the account vector corresponding to the interactive media account.

[0126] In some embodiments, the compression processing of the plurality of account features of the interactive media account to obtain the account vector of the interactive media account can be implemented by the following technical solutions: embedding the plurality of account features of the interactive media account to obtain a plurality of account embedding features of the interactive media account; and based on the weights of the plurality of account features of the interactive media account, the plurality of account embedding features of the interactive media account are weighted and summed to obtain the account vector corresponding to the interactive media account.

[0127] As an example, the specific embodiments of the embodiment of the corresponding interactive media account can refer to the specific embodiments introduced in the embodiment of the corresponding candidate media account, and the difference is only that the processing object is changed from the candidate media account to the interactive media account.

[0128] In step 1042, the content similarity and the account similarity between each interactive media account and the plurality of candidate media accounts are fused to obtain the similarity between each interactive media account and the plurality of candidate media accounts.

[0129] In some embodiments, the fusion of the content similarity and the account similarity between each interactive media account and the plurality of candidate media accounts in step 1042 can be realized by the following technical solution: for each interactive media account, the following processing is performed: determining the account similarity between the account vector of the interactive media account and the account vector of each candidate media account, and determining the content similarity between the content vector of the interactive media account and the content vector of each candidate media account; and performing average processing on the account similarity and the content similarity.

[0130] As an example, when fusing the account similarity and the content similarity, a method other than average processing can be used, for example, weighted sum processing is performed on the account similarity and the content similarity according to a pre-allocated weight.

[0131] As an example, when fusing the content similarity and the account similarity between each interactive media account and the plurality of candidate media accounts, for example, for the interactive media account A, the account similarity between the account vector of the interactive media account A and the account vector of each candidate media account is determined, and the content similarity between the content vector of the interactive media account A and the content vector of each candidate media account is determined; the account similarity and the content similarity are averaged to obtain the similarity between the interactive media account A and each candidate media account, and similar processing is performed for all interactive media accounts.

[0132] In step 1043, the similarities between each interactive media account and the plurality of candidate media accounts are sorted in descending order, and at least one candidate media account with a high ranking in the descending order sorting result is selected as the candidate media account to be recommended.

[0133] In some embodiments, the matching pairs of the interactive media account and the candidate media account are constructed, the matching pairs are constructed by any one of the plurality of interactive media accounts and any one of the plurality of candidate media accounts, the matching pairs are globally sorted in descending order according to the similarity between the interactive media account and the candidate media account in the matching pair, and the candidate media account of at least one matching pair with a high ranking in the descending order sorting result is selected as the candidate media account to be recommended.

[0134] In some embodiments, the following processing is performed for each interactive media account: the similarities between the interactive media account and the plurality of candidate media accounts are ranked in descending order, and in the descending order ranking result of each interactive media account, the top at least one candidate media account is selected as the to-be-recommended media account corresponding to the interactive media account.

[0135] As an example, when descending order ranking is performed, all similarities can be ranked in descending order, for example, the similarities between interactive media account A and 3 candidate media accounts a-c are 0.5, 0.6, and 0.2 respectively, and the similarities between interactive media account B and 3 candidate media accounts a-c are 0.7, 0.3, and 0.8 respectively, then 0.5, 0.6, 0.2, 0.3, 0.7, and 0.8 are ranked in descending order, and the top two candidate media accounts b and c are obtained, and the top two candidate media accounts a and c are selected as the to-be-recommended media accounts. When descending order ranking is performed, descending order ranking can be performed for different interactive media accounts respectively, for example, the similarities between interactive media account A and 3 candidate media accounts a-c are 0.5, 0.6, and 0.2 respectively, the top one candidate media account b is obtained, the similarities between interactive media account B and 3 candidate media accounts a-c are 0.7, 0.3, and 0.8 respectively, the top one candidate media account c is obtained, and then the candidate media accounts b and c are selected as the to-be-recommended media accounts.

[0136] In some embodiments, the content vector or the account vector can be used alone. If the account vector is used alone for recommendation, the account similarities between each interactive media account and the plurality of candidate media accounts are ranked in descending order, and the top at least one candidate media account in the descending order ranking result is selected as the to-be-recommended candidate media account. If the content vector is used alone for recommendation, the content similarities between each interactive media account and the plurality of candidate media accounts are ranked in descending order, and the top at least one candidate media account in the descending order ranking result is selected as the to-be-recommended candidate media account. The specific ranking manner and the manner of selecting the to-be-recommended media account can refer to the above-described implementation manner of ranking based on the fused similarity.

[0137] In some embodiments, the vector utilization mode is selected in different scenarios. When the number of information published by the candidate media account is less than a first information number threshold, it is determined to utilize the account vector to represent the candidate media account and the interactive media account, and the account similarity is used for sorting. When the number of information published by the candidate media account is not less than the first information number threshold and less than a second information number threshold, it is determined to utilize the account vector and the content vector to represent the candidate media account and the interactive media account, and the account similarity and the content similarity are fused, and the similarity obtained through the fusion is used for sorting. When the number of information published by the candidate media account is not less than the second information number threshold, it is determined to utilize the content vector to represent the candidate media account and the interactive media account, and the content similarity is used for sorting.

[0138] In step 105, the recommendation operation of the corresponding user account is performed based on the candidate media account to be recommended.

[0139] As an example, in a one-to-three scenario, the information (for example, a video) published by the candidate media account to be recommended is recommended to the user, or the attention recommendation is directly performed on the candidate media account to be recommended.

[0140] Here, the embodiments of the present application can also combine the blockchain technology. After the terminal acquires the candidate media account to be recommended and marks the new interactive media account, a transaction for storing the new interactive media account is generated, and the generated transaction is submitted to the node of the blockchain network, so that the node stores the new interactive media account to the blockchain network after consensus of the transaction. Before being stored to the blockchain network, the terminal can also perform hash processing on the new interactive media account to obtain the digest information corresponding to the new interactive media account; and the obtained digest information of the new interactive media account is stored to the blockchain network. Through the above-mentioned manner, the new interactive media account is prevented from being tampered, the security of the new interactive media account is improved, and the user's interactive media account is prevented from being tampered by malicious users or malicious programs.

[0141] Referring to Figure 6 , Figure 6 The application architecture schematic diagram of the blockchain network provided by the embodiments of the present application includes a business subject 800, a blockchain network 600 (exemplarily showing nodes 610-1 to 610-3), an authentication center 700, which are described below respectively.

[0142] The type of the blockchain network 600 is flexible, for example, any one of public chain, private chain or consortium chain. For example, in the public chain, any electronic device of a business subject, such as a user terminal and a server, can access the blockchain network 600 without authorization; for example, in the consortium chain, the computer device (such as a terminal / server) under the jurisdiction of the business subject can access the blockchain network 600 after being authorized, and then the computer device becomes a client node in the blockchain network 600.

[0143] In some embodiments, the business subject can be a terminal, and the client node 410 can only act as an observer of the blockchain network 600, that is, to provide a function of supporting the business subject to initiate a transaction (for example, for storing data on the chain or querying data on the chain), and the client node can not implement or selectively implement (for example, depending on the specific business needs of the business subject) the functions of the nodes 610-1 to 610-3 in the blockchain network 600, such as sorting function, consensus service and ledger function. Thus, the data and business processing logic of the business subject can be migrated to the blockchain network 600 to the greatest extent, and the data and business processing process can be realized by the blockchain network 600 to be trusted and traceable.

[0144] The nodes in the blockchain network 600 receive the transaction submitted by the client node 410 of the business subject 800, execute the transaction to update the ledger or query the ledger, and the various intermediate results or final results of executing the transaction can be returned to the client node of the business subject for display.

[0145] For example, the client node 410 can subscribe to the events of interest in the blockchain network 600, such as transactions occurring in a specific organization / channel in the blockchain network 600, and the corresponding transaction notifications are pushed by the nodes 610-1 to 610-3 to the client node 410, thereby triggering the corresponding business logic in the client node 410.

[0146] The following describes an exemplary application of the blockchain by taking an example of a business subject accessing the blockchain network to realize media account recommendation.

[0147] Referring to Figure 6The media account recommendation involves a business subject 800, which is registered with the authentication center 700 to obtain a digital certificate, the digital certificate including a public key of the business subject and a digital signature of the authentication center 700 on the public key and identity information of the business subject, which is attached to a transaction together with a digital signature of the business subject for the transaction and sent to the blockchain network, so that the blockchain network extracts the digital certificate and the signature from the transaction to verify the reliability of the message (i.e., whether it has been tampered with) and the identity of the business subject sending the message, and the blockchain network verifies the identity, for example, whether it has the authority to initiate the transaction. The client running on the computer device (such as a terminal or server) under the jurisdiction of the business subject can request access to the blockchain network 600 to become a client node.

[0148] The business subject 800 client node 410 is used to present the media account to be recommended, for example, in response to a play operation on a video, a video playback page corresponding to the video is presented, and the video content of the video and the media account to be recommended are presented in the video playback page. In response to a user's interactive operation on the media account to be recommended, the media account to be recommended is displayed as an interactive media account, and here the terminal sends the interactive media account to the blockchain network 600.

[0149] The operation of sending the interactive media account to the blockchain network 600 can be pre-configured in the business logic of the client node 410, and when the terminal obtains the interactive media account, the client node 410 automatically sends the interactive media account to the blockchain network 600. When sending, the client node 410 generates a transaction corresponding to the storage operation according to the interactive media account, specifies the smart contract to be called to implement the storage operation and the parameters to be passed to the smart contract in the transaction, and the transaction also carries the digital certificate of the client node 410 and the signed digital signature (for example, using the private key in the digital certificate of the client node 410 to encrypt the digest of the transaction), and broadcasts the transaction to the nodes 610-1 to 610-3 in the blockchain network 600.

[0150] When the nodes 610-1 to 610-3 in the blockchain network 600 receive the transaction, they verify the digital certificate and the digital signature carried by the transaction. After successful verification, according to the identity of the business subject 800 carried in the transaction, it is confirmed whether the business subject 800 has the transaction authority, and any one of the digital signature and the authority verification will result in a failed transaction. After successful verification, the node 610-1 to 610-3 signs its own digital signature (for example, using the private key of the node 610-1 to encrypt the digest of the transaction), and continues to broadcast in the blockchain network 600.

[0151] After the nodes 610-1 to 610-3 in the blockchain network 600 receive the transaction whose verification is successful, the transaction is filled into a new block and broadcasted. After the nodes 610-1 to 610-3 in the blockchain network 600 broadcast the new block, a consensus process is performed on the new block (the nodes can be used as consensus nodes), if the consensus is successful, the new block is appended to the tail of the blockchain stored by the nodes, and the state database is updated according to the result of the transaction: for the transaction of submitting an updated interactive media account, the interactive media account is added to the state database.

[0152] As an example of the blockchain, refer to Figure 7 , Figure 7 The structure of the blockchain in the blockchain network 600 provided by the embodiments of the present application is shown in Figure 7 The genesis block, block 2 and block 3 are shown, each block has a different height, and the header of each block includes the hash value of all transactions in the block, and also includes the hash value of all transactions in the previous block. After the record of the newly generated transaction is filled into the block and the consensus of the nodes in the blockchain network is passed, the record is appended to the tail of the blockchain to form a chain-like growth, and the chain-like structure based on the hash value between the blocks ensures the tamper resistance and anti-forgery of the transactions in the blocks.

[0153] The exemplary functional architecture of the blockchain network provided by the embodiments of the present application is described below, refer to Figure 8 , Figure 8 The functional architecture of the blockchain network 600 provided by the embodiments of the present application is shown in

[0154] The resource layer 605 encapsulates the computing resources, storage resources and communication resources of the nodes 610-1 to 610-3 in the blockchain network 600.

[0155] The data layer 604 encapsulates various data structures for implementing the ledger, including the blockchain implemented by files in the file system, the state database of the key-value type and the proof of existence (for example, the hash tree of the transactions in the block).

[0156] The network layer 603 encapsulates the functions of the point-to-point (P2P, Point to Point) network protocol, the data propagation mechanism and the data verification mechanism, the access authentication mechanism and the business subject identity management.

[0157] The P2P network protocol realizes the communication between the nodes 610-1 to 610-3 in the blockchain network 600, the data propagation mechanism ensures the propagation of transactions in the blockchain network 600, and the data verification mechanism is used to realize the reliability of the data transmission between the nodes 610-1 to 610-3 based on the cryptography method (for example, digital certificate, digital signature, public / private key pair); the access authentication mechanism is used to authenticate the identity of the business subject joining the blockchain network 600 according to the actual business scenario, and the business subject is given the permission to access the blockchain network 600 when the authentication is passed; the business subject identity management is used to store the identity of the business subject allowed to access the blockchain network 600 and the permission (for example, the type of transaction that can be initiated).

[0158] The consensus layer 602 encapsulates the mechanism (i.e., consensus mechanism) for the nodes 610-1 to 610-3 in the blockchain network 600 to reach an agreement on the block, the functions of transaction management and ledger management. The consensus mechanism includes POS, POW and DPOS consensus algorithms, and supports the pluggable consensus algorithm.

[0159] The transaction management is used to verify the digital signature carried in the transaction received by the nodes 610-1 to 610-3, verify the identity information of the business subject, and determine whether the business subject has the permission to conduct the transaction according to the identity information (read the relevant information from the business subject identity management); for the business subject authorized to access the blockchain network 600, the business subject has a digital certificate issued by the authentication center, and the business subject signs the submitted transaction by using the private key in the digital certificate, thereby declaring the legal identity of the business subject.

[0160] The ledger management is used to maintain the blockchain and the state database. For the block that reaches the consensus, it is appended to the tail of the blockchain; the transactions in the block that reaches the consensus are executed, when the transaction includes an update operation, the key-value pair in the state database is updated, when the transaction includes a query operation, the key-value pair in the state database is queried and the query result is returned to the client node of the business subject. The state database supports various dimension query operations, including: querying the block according to the block sequence number (for example, the hash value of the transaction); querying the block according to the block hash value; querying the block according to the transaction sequence number; querying the transaction according to the transaction sequence number; querying the account data of the business subject according to the account (sequence number) of the business subject; querying the blockchain in the channel according to the channel name.

[0161] The application layer 601 encapsulates various businesses that the blockchain network can implement, including transaction traceability, evidence storage and verification, etc.

[0162] The following describes an exemplary application of the embodiments of this application in a real-world account recommendation scenario, which can effectively improve the recommendation accuracy of cold-start accounts. For example, based on the user's list of followed accounts, the most similar candidate media accounts are determined by the content vectors of the followed accounts (interactive accounts) and the content vectors of candidate media accounts (e.g., cold-start accounts). Then, in a one-to-three scenario, the information (e.g., videos) published by the determined candidate media accounts is recommended to the user, or the user can be directly recommended to follow the determined candidate media accounts.

[0163] In some embodiments, see Figure 9 , Figure 9 This is a schematic diagram illustrating the principle of media account recommendation based on artificial intelligence provided in this application embodiment. This application embodiment provides a media account recommendation method based on artificial intelligence that uses a media account learning model to learn the content features of videos published by media accounts. The media account learning model is trained by online triple minimization, as described below.

[0164] First, the videos and corresponding content features published by the media account are acquired, such as video embedding features. The content features of each video are then used as input to the media account learning model, which outputs the features of that media account. A squeeze activation network is then used to enhance these features, resulting in a content vector for the media account. During training, the media account learning model is trained using online triplet minimization, as opposed to offline triplet minimization. Offline triplet minimization requires all training data to be input into the media account learning model at the beginning, and before each iteration... The media account learning model is trained once by selecting the intermediate difficulty triplet from all triplets and using the intermediate difficulty triplet as input. Since all triplets obtained from the training data must be traversed before each training iteration, the efficiency of offline mining of intermediate difficulty triplets is very low. Online triplet minimization, on the other hand, traverses the batch of training samples without traversing all triplets to find intermediate difficulty triplets, thereby improving training efficiency. The media account learning model is trained using triplet loss (the loss function of the media account learning model) based on the intermediate difficulty triplet, so that the content vector of the media account can be obtained through the trained media account learning model.

[0165] In some embodiments, the media account learning model is a supervised model, and when the method for learning the content vector using the media account learning model is used, a supervision signal is also needed, which can be constructed by the fan data of the media account. For example, if two media accounts have more common fans, it means that the two media accounts are more similar, and therefore, a triplet can be constructed based on the supervision signal, and then the intermediate difficulty triplet that meets the supervision signal is obtained by online triplet minimization.

[0166] In some embodiments, the intermediate difficulty triplet is defined based on the distance definition of the reference sample and the negative sample and the positive sample in the triplet, see Figure 10 , Figure 10 is a training sample diagram provided by an embodiment of the present application. According to the difficulty level of the triplet, the triplet can be divided into three categories: simple triplet, difficult triplet (corresponding to difficult negative sample), and intermediate difficulty triplet (corresponding to intermediate difficulty negative sample). The simple triplet is a triplet corresponding to a loss of zero, which is formally defined as the distance between the reference sample and the negative sample > the distance between the reference sample and the positive sample + a set threshold. The difficult triplet is formally defined as the distance between the negative sample and the reference sample < the distance between the reference sample and the positive sample. The intermediate difficulty triplet is formally defined as the distance between the negative sample and the reference sample > the distance between the reference sample and the positive sample, but not so much that the loss is zero, which is formally defined as the distance between the reference sample and the positive sample < the distance between the reference sample and the negative sample < the distance between the reference sample and the positive sample + a set threshold. The simple triplet (corresponding to the simple negative sample) is simple because it is easy to identify, so it is not necessary to construct too many of this type of triplet, otherwise it will reduce the training efficiency. If the difficult triplet is used, it may affect the training effect, so the intermediate difficulty triplet is selected by online triplet minimization.

[0167] During training, the content features of the 100 videos recently published by the media account are selected as the video features used to learn the content vector. In the training data, 10 media accounts are randomly selected each time as the input of the media account learning model. During prediction, an indefinite number of videos can be accepted as the input of the media account learning model. In terms of model prediction effect, the click rate index is used to measure it. Compared with the test results before training, the click rate index after training is improved by 4 percentage points.

[0168] In some embodiments, an Enhanced Graph Embedding with Side Information (EGES) model is used to learn the account vectors of media accounts. By obtaining the sequence of interactive media accounts followed by each user, the content features and other features (such as account level, account verticality, etc.) of the interactive media accounts are extracted and input into the EGES model to obtain the account vectors of the interactive media accounts. Based on the EGES model, the user's follow sequence is fitted, and the content features and other features (such as account level, account verticality, etc.) of candidate media accounts (cold start accounts) are obtained and input into the EGES model to obtain the account vectors of the candidate media accounts. Since features such as account level are input, the account vectors of the learned media accounts contain richer information.

[0169] The AI-based media account recommendation method provided in this application does not rely on posterior data. Instead, it starts with the information published by the account and applies a media account learning model that performs clustering and residual processing to the account cold start task. Based on existing content features, the media account learning model learns the vector data of the videos published by the account. The vector representation of the account is trained by minimizing online triples. The content vector output by the media account learning model obtained in this way learns the content information of the videos published by the account, and establishes the relationship between accounts from the content dimension. This allows the cold start problem of new accounts to be solved from the perspective of the content vector (similarity) of the accounts.

[0170] The following description continues to illustrate the exemplary structure of the AI-based media account recommendation device 255 provided in this application embodiment as a software module. In some embodiments, such as... Figure 3 As shown, the software modules in the AI-based media account recommendation device 255 stored in the memory 250 may include: a feature module 2551, used to acquire multiple content features of at least one interactive media account of a user account, and acquire multiple content features corresponding to multiple candidate media accounts respectively; a vector module 2552, used to perform clustering and residual processing on the multiple content features of each interactive media account to obtain the content vector of the interactive media account; the vector module 2552 is also used to perform clustering and residual processing on the multiple content features of each candidate media account respectively to obtain the content vector of each candidate media account; a similarity module 2553, used to determine the candidate media accounts to be recommended based on the content similarity between the content vector of each interactive media account and the content vectors of multiple candidate media accounts; and a recommendation module 2554, used to perform a recommendation operation on the corresponding user account based on the candidate media accounts to be recommended.

[0171] In some embodiments, the feature module 2551 is further configured to: perform the following processing for each interactive media account: extract corresponding content features from multiple messages published by the interactive media account as content features of the interactive media account; and perform the following processing for each candidate media account: extract corresponding content features from multiple messages published by the candidate media account as content features of the candidate media account; wherein the type of information includes at least one of the following: video information, text information, and image information.

[0172] In some embodiments, the vector module 2552 is further configured to: perform the following processing for each interactive media account: cluster multiple content features of the interactive media account to obtain at least one first cluster center; perform residual processing on multiple content features of the interactive media account based on at least one first cluster center to obtain a content vector of the interactive media account; and perform the following processing for each candidate media account: cluster multiple content features of the candidate media account to obtain at least one second cluster center; and perform residual processing on multiple content features of the candidate media account based on at least one second cluster center to obtain a content vector of the candidate media account.

[0173] In some embodiments, the content vector of the interactive media account is determined by a media account learning model, which includes a local clustering core network and a normalization network. The vector module 2552 is further configured to: perform the following processing for each first cluster center: determine the first residual distribution of multiple content features of the interactive media account corresponding to the first cluster center through the local clustering core network; perform the following processing for each first cluster center: normalize the first residual distribution of the corresponding first cluster center through the normalization network to obtain the normalization result of the corresponding first cluster center; and perform overall normalization processing on the normalization result of each first cluster center through the normalization network to obtain the content vector of the interactive media account.

[0174] In some embodiments, the media account learning model further includes a squeeze activation network; after normalizing the normalization results of each first cluster center through a normalization network to obtain the content vector of the interactive media account, the vector module 2552 is further configured to: perform channel-based average pooling processing on the content vector of the interactive media account through the squeeze activation network to obtain the global content features of each channel corresponding to the content vector of the interactive media account; perform fully connected processing on the global content features of each channel corresponding to the content vector of the interactive media account through the squeeze activation network to obtain the activation value of each channel corresponding to the content vector of the interactive media account; perform dot product processing on the activation value of each channel corresponding to the content vector of the interactive media account and the original content features of each channel in the content vector of the interactive media account, and update the content vector of the interactive media account based on the dot product processing result.

[0175] In some embodiments, the vector module 2552 is further configured to: perform convolution processing on each content feature of the interactive media account through a local clustering core network to obtain the corresponding convolution result; perform maximum likelihood function processing on the convolution result of each content feature of the interactive media account through a local clustering core network to obtain the corresponding maximum likelihood processing result; determine the first residual between each content feature of each interactive media account and the first cluster center through a local clustering core network; and perform weighted summation processing on the first residual corresponding to each content feature of the interactive media account using the corresponding maximum likelihood processing result as weight to obtain the first residual distribution of multiple content features of the interactive media account corresponding to the first cluster center.

[0176] In some embodiments, the content vector of the interactive media account is determined by a media account learning model, which includes a local clustering core network and a normalization network. The vector module 2552 is further configured to: perform the following processing for each second cluster center: determine the second residual distribution of multiple content features of the candidate media account corresponding to the second cluster center through the local clustering core network; perform the following processing for each second cluster center: normalize the second residual distribution of multiple content features of the candidate media account corresponding to the second cluster center through the normalization network to obtain the corresponding normalization result; and perform overall normalization processing on the normalization result corresponding to multiple content features of the candidate media account through the normalization network to obtain the content vector of the candidate media account.

[0177] In some embodiments, the vector module 2552 is further configured to: perform convolution processing on each content feature of the candidate media account through a local clustering core network to obtain the corresponding convolution result; perform maximum likelihood function processing on the convolution result of each content feature of the candidate media account through a local clustering core network to obtain the corresponding maximum likelihood processing result; determine the second residual between each content feature of the candidate media account and the second cluster center through a local clustering core network; and perform weighted summation processing on the second residual corresponding to each content feature of the candidate media account using the corresponding maximum likelihood processing result as weight to obtain the second residual distribution of multiple content features corresponding to the second cluster center.

[0178] In some embodiments, the media account learning model further includes a squeeze activation network; the vector module 2552 is further configured to: after performing overall normalization processing on the normalization result of each second cluster center through a normalization network to obtain the content vector of the candidate media account, perform channel-based average pooling processing on the content vector of the candidate media account through the squeeze activation network to obtain the global content features of each channel corresponding to the content vector of the candidate media account; perform fully connected processing on the global content features of each channel corresponding to the content vector of the candidate media account through the squeeze activation network to obtain the activation value of each channel corresponding to the content vector of the candidate media account; perform dot product processing on the activation value of each channel corresponding to the content vector of the candidate media account and the original content features of each channel in the content vector of the candidate media account, and update the content vector of the candidate media account based on the dot product processing result.

[0179] In some embodiments, the similarity module 2553 is further configured to: determine the account similarity between the account vector of each interactive media account and the account vectors of multiple candidate media accounts; fuse the content similarity and account similarity between each interactive media account and multiple candidate media accounts to obtain the similarity between each interactive media account and multiple candidate media accounts; sort the similarity between each interactive media account and multiple candidate media accounts in descending order, and select at least one candidate media account ranked first in the descending order as a candidate media account to be recommended.

[0180] In some embodiments, the similarity module 2553 is further configured to: perform the following processing for each candidate media account: extract multiple account features from the account information of the candidate media account, compress the multiple account features of the candidate media account to obtain an account vector of the candidate media account; perform the following processing for each interactive media account: extract multiple account features from the account information of the interactive media account, compress the multiple account features of the interactive media account to obtain an account vector of the interactive media account; determine the account similarity between the account vector of each interactive media account and the account vectors of multiple candidate media accounts; perform the following processing for each interactive media account: determine the account similarity between the account vector of the interactive media account and the account vector of each candidate media account, and determine the content similarity between the content vector of the interactive media account and the content vector of each candidate media account; and average the account similarity and the content similarity.

[0181] In some embodiments, the similarity module 2553 is further configured to: embed multiple account features of the candidate media account to obtain the weights of the multiple account features of the candidate media account; perform weighted summation on the multiple account features of the candidate media account based on the weights of the multiple account features of the candidate media account to obtain the account vector of the corresponding candidate media account; embed multiple account features of the interactive media account to obtain the weights of the multiple account features of the interactive media account; and perform weighted summation on the multiple account features of the interactive media account based on the weights of the multiple account features of the interactive media account to obtain the account vector of the corresponding interactive media account.

[0182] In some embodiments, the similarity module 2553 is further configured to: embed multiple account features and multiple content features of the candidate media account to obtain the weights of the multiple account features and multiple content features of the candidate media account; perform weighted summation on the multiple account features and multiple content features of the candidate media account based on the weights of the multiple account features and multiple content features of the candidate media account to obtain the account vector of the corresponding candidate media account; embed multiple account features and multiple content features of the interactive media account to obtain the weights of the multiple account features and multiple content features of the interactive media account; and perform weighted summation on the multiple account features and multiple content features of the interactive media account based on the weights of the multiple account features and multiple content features of the interactive media account to obtain the account vector of the corresponding interactive media account.

[0183] In some embodiments, the content vector of an interactive media account is determined by a media account learning model, which includes a local clustering core network and a normalization network. The apparatus further includes a training module 2555, configured to: before acquiring multiple content features of at least one interactive media account of a user account and acquiring multiple content features corresponding to multiple candidate media accounts respectively, train the media account learning model in the following manner: acquire multiple media account samples and construct multiple first-class triplet samples based on the number of associated users of the multiple media account samples; decompose the multiple first-class triplet samples to obtain multiple media account samples, so as to predict the sample content vector of each media account sample through the media account learning model; determine the second-class triplet samples that meet the training conditions based on the sample content vectors; substitute the sample content vector corresponding to each media account sample in the second-class triplet samples into the triplet loss function to determine the parameters of the media account learning model when the triplet loss function reaches its minimum value.

[0184] In some embodiments, the training module 2555 is further configured to: take any media account sample as the first media account sample, obtain a second media account sample with the same label as the first media account sample from multiple media account samples, and determine a first content distance between the sample content vector of the second media account sample and the sample content vector of the first media account sample; obtain a third media account sample with a different label from the first media account sample from multiple media account samples, and determine a second content distance between the sample content vector of the third media account sample and the sample content vector of the first media account sample; extract second media account samples and third media account samples that satisfy the following training conditions from multiple second media account samples and multiple third media account samples: the second content distance of the corresponding third media account sample is greater than the first content distance of the corresponding second media account sample; the difference between the second content distance of the corresponding third media account sample and the first content distance of the corresponding second media account sample is less than a first threshold; and form a second type of triplet sample by combining the first media account sample, the second media account sample that satisfies the training conditions, and the third media account sample.

[0185] In some embodiments, the training module 2555 is further configured to: obtain the associated users of each remaining media account sample; wherein, the remaining media account sample is a media account sample that is different from the first media account sample among multiple media account samples; determine the user intersection between the associated users of each remaining media account sample and the associated users of the first media account sample; determine the remaining media account samples whose number of elements in the user intersection exceeds a second threshold as second media account samples with the same label as the first media account sample; and determine the remaining media account samples whose number of elements in the user intersection is less than a third threshold as third media account samples with different labels from the first media account sample.

[0186] This application 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 and executes the computer instructions, causing the computer device to perform the AI-based media account recommendation method described above in this application.

[0187] This application provides a computer-readable storage medium storing executable instructions. When these executable instructions are executed by a processor, the processor performs the AI-based media account recommendation method provided in this application. For example... Figures 5A-5C The example shown is an AI-based media account recommendation method.

[0188] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EP ROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0189] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0190] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0191] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0192] In summary, by obtaining the content features of media accounts through the embodiments of this application and performing clustering and residual processing on the content features, the feature distribution differences of the content features themselves can be hidden, and only the distribution differences between the content features and the cluster centers can be retained. This allows for a more efficient learning of the feature distribution of media accounts in the content dimension, thereby establishing relationships between accounts in the content dimension. Since the interaction between accounts and users in the recommendation system is largely based on content, this more accurately characterizes the account vectors of media accounts, thereby improving the accuracy of media account recommendations based on the similarity of account vectors.

[0193] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A media account recommendation method based on artificial intelligence, characterized in that, include: Obtain multiple content features of at least one interactive media account of a user account, and obtain multiple content features corresponding to multiple candidate media accounts respectively; Clustering and residual processing are performed on multiple content features of each interactive media account to obtain the content vector of the interactive media account; Clustering and residual processing are performed on multiple content features of each candidate media account to obtain the content vector of each candidate media account; For each candidate media account, the following processing is performed: multiple account features are extracted from the account information of the candidate media account, and the multiple account features of the candidate media account are compressed to obtain the account vector of the candidate media account; For each of the interactive media accounts, the following processing is performed: multiple account features are extracted from the account information of the interactive media account, and the multiple account features of the interactive media account are compressed to obtain the account vector of the interactive media account; Determine the account similarity between the account vector of each interactive media account and the account vector of the plurality of candidate media accounts, and determine the content similarity between the content vector of the interactive media account and the content vector of each candidate media account; The account similarity and content similarity are averaged to obtain the similarity between each interactive media account and the plurality of candidate media accounts; The similarity between each interactive media account and the plurality of candidate media accounts is sorted in descending order, and at least one candidate media account with the highest ranking in the descending order is selected as the candidate media account to be recommended. Based on the candidate media accounts to be recommended, a recommendation operation is performed on the corresponding user accounts.

2. The method according to claim 1, characterized in that, The acquisition of multiple content features of at least one interactive media account of a user account includes: For each of the interactive media accounts, the following processing is performed: extracting corresponding content features from multiple messages published by the interactive media account, and using them as the content features of the interactive media account; The acquisition of multiple content features corresponding to multiple candidate media accounts includes: For each candidate media account, the following processing is performed: extracting corresponding content features from multiple pieces of information published by the candidate media account, and using them as the content features of the candidate media account; The information includes at least one of the following types: video information, text information, and image information.

3. The method according to claim 1, characterized in that, The process of clustering and residual processing on multiple content features of each interactive media account to obtain the content vector of the interactive media account includes: Perform the following processing for each of the aforementioned interactive media accounts: Clustering is performed on multiple content features of the interactive media account to obtain at least one first cluster center; Based on the at least one first cluster center, residual processing is performed on multiple content features of the interactive media account to obtain the content vector of the interactive media account; The step of performing clustering and residual processing on multiple content features of each candidate media account to obtain a content vector for each candidate media account includes: For each of the aforementioned candidate media accounts, the following processing is performed: Clustering is performed on multiple content features of the candidate media accounts to obtain at least one second cluster center; Based on the at least one second clustering center, residual processing is performed on multiple content features of the candidate media account to obtain the content vector of the candidate media account.

4. The method according to claim 3, characterized in that, The content vector of the interactive media account is determined by a media account learning model, which includes a local clustering core network and a normalization network. The step of performing residual processing on multiple content features of the interactive media account based on at least one first cluster center to obtain the content vector of the interactive media account includes: For each of the first cluster centers, the following process is performed: through the local clustering core network, the first residual distribution of multiple content features of the interactive media account corresponding to the first cluster center is determined; For each of the first cluster centers, the following process is performed: the first residual distribution corresponding to the first cluster center is normalized through the normalization network to obtain the normalization result corresponding to the first cluster center; The normalization results of each of the first cluster centers are normalized as a whole through the normalization network to obtain the content vector of the interactive media account.

5. The method according to claim 4, characterized in that, The media account learning model also includes a squeeze activation network; After performing overall normalization processing on the normalization results of each first cluster center through the normalization network to obtain the content vector of the interactive media account, the method further includes: The squeeze activation network is used to perform channel-based average pooling on the content vector of the interactive media account to obtain the global content features of each channel corresponding to the content vector of the interactive media account. By using the squeeze activation network, the global content features of each channel corresponding to the content vector of the interactive media account are processed by full connection to obtain the activation value of each channel corresponding to the content vector of the interactive media account. The content vector of the interactive media account is multiplied by the activation value of each channel corresponding to the content vector of the interactive media account and the original content features of each channel in the content vector of the interactive media account. The content vector of the interactive media account is then updated based on the result of the multiplication.

6. The method according to claim 4, characterized in that, The step of determining the first residual distribution of multiple content features of the interactive media account corresponding to the first cluster center through the local clustering core network includes: The local clustering core network is used to perform convolution processing on each content feature of the interactive media account to obtain the corresponding convolution result. The maximum likelihood function is applied to the convolution result of each content feature of the interactive media account through the local clustering core network to obtain the corresponding maximum likelihood processing result. The first residual between each content feature of each interactive media account and the first cluster center is determined through the local clustering core network; Using the maximum likelihood processing result corresponding to each content feature of the interactive media account as the weight, the first residual corresponding to each content feature of the interactive media account is weighted and summed to obtain the first residual distribution of multiple content features of the interactive media account corresponding to the first cluster center.

7. The method according to claim 1, characterized in that, The step of compressing multiple account features of the candidate media accounts to obtain the account vector of the candidate media accounts includes: Multiple account features of the candidate media account are embedded to obtain the account embedding features of the multiple account features of the candidate media account. Based on the weights of multiple account features of the candidate media account, a weighted summation is performed on multiple account embedding features of the candidate media account to obtain the account vector corresponding to the candidate media account. The compression process of multiple account features of the interactive media account to obtain the account vector of the interactive media account includes: Multiple account features of the interactive media account are embedded to obtain the account embedding features of the multiple account features of the interactive media account. Based on the weights of multiple account features of the interactive media account, a weighted summation is performed on the multiple account embedding features of the interactive media account to obtain the account vector corresponding to the interactive media account.

8. The method according to claim 1, characterized in that, The content vector of the interactive media account is determined through a media account learning model; Before obtaining multiple content features of at least one interactive media account of a user account, and obtaining multiple content features corresponding to multiple candidate media accounts respectively, the method includes: The media account learning model was trained using the following method: Multiple media account samples are obtained, and multiple first-type triplet samples are constructed based on the number of associated users of the multiple media account samples; Multiple triplet samples of the first type are decomposed to obtain multiple media account samples, so as to predict the sample content vector of each media account sample through the media account learning model; Based on the sample content vector, determine the second type of triplet samples that meet the training conditions; Substitute the sample content vector corresponding to each media account sample in the second type of triplet samples into the triplet loss function to determine the parameters of the media account learning model when the triplet loss function reaches its minimum value.

9. The method according to claim 8, characterized in that, The step of determining the second type of triplet samples that meet the training conditions based on the sample content vector includes: Using any one of the media account samples as the first media account sample, obtain a second media account sample with the same label as the first media account sample from multiple media account samples, and determine a first content distance between the sample content vector of the second media account sample and the sample content vector of the first media account sample; Obtain a third media account sample with a different label from the first media account sample from multiple media account samples, and determine a second content distance between the sample content vector of the third media account sample and the sample content vector of the first media account sample; From multiple second media account samples and multiple third media account samples, extract second media account samples and third media account samples that meet the following training conditions: The second content distance corresponding to the third media account sample is greater than the first content distance corresponding to the second media account sample; The difference between the second content distance corresponding to the third media account sample and the first content distance corresponding to the second media account sample is less than the first threshold. The first media account sample, the second media account sample that meets the training conditions, and the third media account sample are combined to form the second type of triplet sample.

10. The method according to claim 9, characterized in that, The step of obtaining a second media account sample with the same tag as the first media account sample from a plurality of media account samples includes: Obtain the associated users for each remaining media account sample; The remaining media account samples are media account samples that are different from the first media account sample among the multiple media account samples; Determine the user intersection between the associated users of each of the remaining media account samples and the associated users of the first media account sample; The remaining media account samples whose number of elements in the user intersection exceeds the second threshold are identified as the second media account samples with the same label as the first media account samples. The step of obtaining a third media account sample with different tags from the first media account sample from the plurality of media account samples includes: The remaining media account samples with the number of elements in the user intersection less than the third threshold are determined as third media account samples with different labels from the first media account samples. Wherein, the second threshold is greater than the third threshold.

11. A media account recommendation device based on artificial intelligence, characterized in that, include: The feature module is used to obtain multiple content features of at least one interactive media account of a user account, and to obtain multiple content features corresponding to multiple candidate media accounts respectively. The vector module is used to perform clustering and residual processing on multiple content features of each interactive media account to obtain the content vector of the interactive media account. The vector module is also used to perform clustering and residual processing on multiple content features of each candidate media account to obtain the content vector of each candidate media account. The similarity module is used to perform the following processing for each candidate media account: extract multiple account features from the account information of the candidate media account, compress the multiple account features of the candidate media account, and obtain the account vector of the candidate media account; For each of the interactive media accounts, the following processing is performed: multiple account features are extracted from the account information of the interactive media account, and the multiple account features of the interactive media account are compressed to obtain the account vector of the interactive media account; Determine the account similarity between the account vector of each interactive media account and the account vector of the plurality of candidate media accounts, and determine the content similarity between the content vector of the interactive media account and the content vector of each candidate media account; The account similarity and content similarity are averaged to obtain the similarity between each interactive media account and the plurality of candidate media accounts; The similarity between each interactive media account and the plurality of candidate media accounts is sorted in descending order, and at least one candidate media account with the highest ranking in the descending order is selected as the candidate media account to be recommended. The recommendation module is used to perform a recommendation operation on the corresponding user account based on the candidate media accounts to be recommended.

12. The apparatus according to claim 11, characterized in that, The feature module is further configured to: perform the following processing for each of the interactive media accounts: extract corresponding content features from multiple messages published by the interactive media account, and use them as the content features of the interactive media account; For each candidate media account, the following processing is performed: extracting corresponding content features from multiple pieces of information published by the candidate media account to serve as the content features of the candidate media account; wherein, the type of information includes at least one of the following: video information, text information, and image information.

13. An electronic device, characterized in that, include: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the AI-based media account recommendation method according to any one of claims 1 to 10.

14. A computer-readable storage medium, characterized in that, It stores executable instructions for implementing the AI-based media account recommendation method as described in any one of claims 1 to 10 when executed by a processor.

15. A computer program product comprising computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, the artificial intelligence-based media account recommendation method according to any one of claims 1 to 10 is implemented.

Images