Content recommendation method, device and computer readable storage medium

Abstract

The present application discloses a content recommendation method, device and computer-readable storage medium in the field of communication technology involved; the method can obtain a content multi-dimensional feature vector of the target content based on the target content and the user-related information providing the target content, the content multi-dimensional feature vector includes the feature information of the target content and the user-related information providing the target content; extract features of at least two interactive related information of the target content to obtain at least one interactive situation representation vector of the target content, wherein at least one interactive situation representation vector can be realized based on the natural language processing technology in the field of artificial intelligence; fuse the content multi-dimensional feature vector and at least one interactive situation representation vector to obtain a fused vector; classify the target content based on the fused vector to obtain a content classification result of the target content; and recommend the target content based on the content classification result. The present application can improve the accuracy of content recommendation.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a content recommendation method, device and computer-readable storage medium. Background Art

[0002] With the rapid development of information technology, more and more content is published in clients and websites, where users can browse various content, and clients and websites can also recommend various types of content to users for browsing. However, the accuracy of content recommendation is currently poor. Summary of the invention

[0003] The embodiments of the present application provide a content recommendation method, device, and computer-readable storage medium, which can improve the accuracy of content recommendation.

[0004] The present application provides a content recommendation method, including:

[0005] Based on the target content and the information associated with the user providing the target content, obtaining a content multi-dimensional feature vector of the target content, wherein the content multi-dimensional feature vector includes feature information of the target content and the information associated with the user providing the target content;

[0006] Performing feature extraction on at least two types of interaction-related information of the target content to obtain at least one interaction status representation vector of the target content;

[0007] Fusing the content multi-dimensional feature vector and the at least one interaction situation representation vector to obtain a fused vector;

[0008] Performing content classification of the target content based on the fused vector to obtain a content classification result of the target content;

[0009] Based on the content classification result, the target content is recommended.

[0010] Accordingly, an embodiment of the present application provides a content recommendation device, including:

[0011] an acquisition unit, configured to acquire a content multi-dimensional feature vector of the target content based on the target content and the user-providing information associated with the target content, wherein the content multi-dimensional feature vector includes feature information of the target content and the user-providing information associated with the target content;

[0012] A feature extraction unit, configured to extract features of at least two types of interaction-related information of the target content to obtain at least one interaction status representation vector of the target content;

[0013] A vector fusion unit, used to fuse the content multi-dimensional feature vector and the at least one interaction situation representation vector to obtain a fused vector;

[0014] A content classification unit, configured to classify the target content based on the fused vector to obtain a content classification result of the target content;

[0015] A recommendation unit is used to recommend the target content based on the content classification result.

[0016] In one embodiment, the acquisition unit includes:

[0017] an acquisition subunit, configured to acquire a first user representation vector of the target content providing user based on the user information of the target content providing user, the user information of the associated user, and the degree of association between the target content providing user and the associated user;

[0018] A first extraction subunit, configured to extract features from the target content to obtain a content feature vector of the target content;

[0019] The vector fusion subunit is used to fuse the first user representation vector and the content feature vector to obtain a multi-dimensional content feature vector of the target content.

[0020] In one embodiment, the acquisition subunit is further used to acquire a user relationship graph, the user relationship graph including: nodes of at least two content publishing users and at least two interactive users, a connection edge is set between the node of each content publishing user and the node of the corresponding interactive user, the connection edge is set with a weight, the weight represents the degree of association between the content publishing user and the corresponding interactive user, the content publishing user includes a target content providing user of the target content; according to a graph walking algorithm, a walking path from the node of the interactive user to the node of the target content providing user is determined, wherein the interactive users on the walking path are associated users of the content publishing user; based on user information of the associated users on the walking path, user information of the target content providing user, and the weight of the connection edge on the walking path, a node vector of the node of the target content providing user is determined, and the node vector is determined as a first user representation vector of the target content providing user.

[0021] In one embodiment, the feature extraction unit includes:

[0022] A second extraction subunit is used to extract features from at least two dimensions of the user interaction information to obtain information feature vectors of at least two dimensions of the user interaction information;

[0023] A third extraction subunit is used to extract user features from the interactive information publishing user-related information to obtain a second user representation vector of the publishing user of the user interactive information;

[0024] The processing subunit is used to process the information feature vectors of the at least two dimensions and the second user representation vector to obtain at least one interaction situation representation vector of the target content.

[0025] In one embodiment, the second extraction sub-unit is further used to perform feature extraction on the user interaction information from a time dimension to obtain a time feature vector of the user interaction information; perform feature extraction on the user interaction information from a content dimension to obtain a content feature vector of the user interaction information; and use the time feature vector and the content feature vector as information feature vectors of at least two dimensions of the user interaction information.

[0026] In one embodiment, the feature extraction unit includes:

[0027] A conversion subunit, configured to convert the words in the user interaction information into word vectors, and obtain an initial text vector of the user interaction information according to the order of the sentences in which the words in the user interaction information are located and the word vectors of the words;

[0028] An encoding subunit, configured to encode the interactive information publishing user association information to obtain an initial user representation vector of the publishing user of the user interactive information;

[0029] A fusion subunit, used for fusing the initial text vector and the initial user representation vector to obtain a fused initial vector;

[0030] The fourth extraction subunit is used to perform deep feature extraction on the fused initial vector to obtain an interaction situation representation vector.

[0031] In one embodiment, the content classification unit includes:

[0032] A mapping subunit, configured to map the fused vector according to a classification mapping matrix to obtain a predicted classification probability of the target content under at least two preset content types;

[0033] The result determination subunit is configured to determine the current content type of the target content as a content classification result based on the predicted classification probabilities of the target content under at least two preset content types.

[0034] In one embodiment, the recommendation unit includes:

[0035] An acquisition subunit, used to acquire historical content classification results of the target content within a preset historical time from the current moment, and the corresponding result acquisition time;

[0036] A weight determination subunit, configured to determine a recommendation reference weight of the content classification result and each of the historical content classification results based on the result acquisition time of the content classification result and the historical content classification result;

[0037] A score determination subunit is used to determine the content type representation score of each preset content type for the target content according to the number of occurrences of each preset content type in the content classification result and the historical content classification result, and the recommendation reference weight corresponding to each occurrence of the preset content type;

[0038] The recommendation subunit is used to determine the recommendation strength of the target content according to the content type representation score corresponding to the recommended reference content type in the preset content type, and recommend the target content based on the recommendation strength.

[0039] Correspondingly, an embodiment of the present application also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the steps in any content recommendation method provided in the embodiment of the present application.

[0040] Correspondingly, an embodiment of the present application further provides a computer-readable storage medium, which stores a plurality of instructions, and the instructions are suitable for loading by a processor to execute the steps in any content recommendation method provided in any embodiment of the present application.

[0041] The embodiment of the present application can obtain a content multi-dimensional feature vector of the target content based on the target content and the user-provided information associated with the target content, wherein the content multi-dimensional feature vector includes the feature information of the target content and the user-provided information associated with the target content; extract features of at least two interactive information associated with the target content to obtain at least one interactive situation representation vector of the target content; fuse the content multi-dimensional feature vector and the at least one interactive situation representation vector to obtain a fused vector; classify the content of the target content based on the fused vector to obtain a content classification result of the target content; and recommend the target content based on the content classification result. This solution can obtain a fused vector containing more feature information by fusing the content multi-dimensional feature vector including the target content and the user-provided information associated with the target content with at least one interactive situation representation vector, and then classify the target content based on the fused vector, which can then be used in the recommendation system to recommend the target content. It can quickly discover various negative content quality issues that were not discovered by humans and machines in the early stage of the review through content classification mining, effectively control the diffusion and dissemination of low-quality content, and enable high-quality content to be effectively weighted, quickly recommended, distributed and promoted, and improve the accuracy of content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 is a scenario diagram of a content recommendation method provided in an embodiment of the present application;

[0044] Figure 2a is a flow chart of a content recommendation method provided by an embodiment of the present application;

[0045] Figure 2b is a flow chart of a sentiment analysis system of a content recommendation method provided in an embodiment of the present application;

[0046] Figure 2c is a technical flow chart of the content recommendation method provided by an embodiment of the present application;

[0047] Figure 2d is a user relationship diagram of the content recommendation method provided in an embodiment of the present application;

[0048] Figure 2e is a node relationship diagram of the content recommendation method provided in an embodiment of the present application;

[0049] Figure 3 is another flow chart of the content recommendation method provided by an embodiment of the present application;

[0050] Figure 4a is a device diagram of a content recommendation method provided in an embodiment of the present application;

[0051] Figure 4b is another device diagram of the content recommendation method provided by an embodiment of the present application;

[0052] Figure 4c is another device diagram of the content recommendation method provided by an embodiment of the present application;

[0053] Figure 4d is another device diagram of the content recommendation method provided by an embodiment of the present application;

[0054] Figure 4e is another device diagram of the content recommendation method provided by an embodiment of the present application;

[0055] Figure 5 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0057] The embodiments of the present application provide a content recommendation method, device, computer equipment and computer-readable storage medium. Specifically, the embodiments of the present application provide a content recommendation device suitable for a computer device. The computer device may be a terminal or a server or other device, and the server may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. The terminal may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, which is not limited in the present application.

[0058] refer to Figure 1Taking a computer device as a server as an example, the server can obtain a content multi-dimensional feature vector of the target content based on the target content and the user-related information providing the target content, the content multi-dimensional feature vector including the feature information of the target content and the user-related information providing the target content; extract features of at least two types of interactive related information of the target content to obtain at least one interactive situation representation vector of the target content; fuse the content multi-dimensional feature vector and at least one interactive situation representation vector to obtain a fused vector; classify the target content based on the fused vector to obtain a content classification result of the target content; and recommend the target content based on the content classification result.

[0059] Among them, the interaction situation representation vector can be implemented based on the natural language processing technology in the field of artificial intelligence. For example, based on the natural language processing technology, feature extraction can be performed on at least two types of interaction-related information of the target content to obtain at least one interaction situation representation vector.

[0060] Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machine models controlled by digital computers to extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to achieve the best results. Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields and integrates both hardware-level and software-level technologies. Artificial intelligence software technology mainly includes natural language processing, machine learning / deep learning and other directions.

[0061] Natural Language Processing (NLP) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that can achieve effective communication between people and computers using natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field will involve natural language, that is, the language people use in daily life, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question answering, knowledge graph and other technologies.

[0062] From the above, it can be seen that the embodiment of the present application can obtain a fused vector containing more feature information by fusing a multi-dimensional feature vector including the target content and the content providing user-related information with at least one interaction situation representation vector, and then classify the target content based on the fused vector. The target content can then be used in a recommendation system to recommend the target content. It can quickly discover various negative content quality issues that have not been discovered by humans and machines in the early stage of the review through content classification mining, effectively control the spread and dissemination of low-quality content, and enable high-quality content to be effectively weighted, quickly recommended, distributed and improved, thereby improving the accuracy of content recommendations.

[0063] The present embodiments may be described in detail below. It should be noted that the description order of the following embodiments is not intended to limit the preferred order of the embodiments.

[0064] The present application embodiment provides a content recommendation method, which can be executed by a terminal or a server, or by both the terminal and the server. The present application embodiment takes the content recommendation method executed by the server as an example for explanation, and specifically, it is executed by a content recommendation device integrated in the server. Figure 2a As shown, the specific process of the content recommendation method can be as follows:

[0065] 201. Based on the target content and the target content providing user associated information, obtain a content multi-dimensional feature vector of the target content, where the content multi-dimensional feature vector includes feature information of the target content and the target content providing user associated information.

[0066] Among them, the user-related information provided may include the target content providing user of the target content, user information of the target content providing user (such as the account, nickname, avatar, etc. of the target content providing user), and user information of the associated users of the target content providing user. The target content providing user may include the target content publishing user and the target content producing user of the target content.

[0067] Among them, the target content publishing user refers to the user who publishes the target content, the target content producing user refers to the user who produces the target content, and the target content producing user may also be the target content publishing user.

[0068] Among them, associated users refer to target users associated with target content providing users, and target associated users of these target users. For example, there are user A and user B. User A is a target user associated with the target content providing user. User A may have browsed the same reading content with user B, or liked, commented, forwarded, deleted, collected, etc. on the same reading content, thus establishing an association between users. At this time, user B is the target associated user of user A.

[0069] The content multi-dimensional feature vector can represent the feature information of multiple dimensions of the target content. The content multi-dimensional feature vector can be obtained by extracting the content feature vector of the target content and the feature vector of the user provided by the target content, and then fusing the two vectors.

[0070] In one example, if Figure 2b As shown, the target content can be obtained from the content production end of the comment sentiment analysis system, wherein the content production end can first obtain the upload server interface address and then publish the target content by communicating with the upstream and downstream content interface servers. The upstream and downstream content interface servers communicate directly with the content production end. The target content submitted by the front end is usually the title, publisher, summary, cover image, and release time of the content. The target content can be stored in the content database.

[0071] In one embodiment, the user-associated information includes the target content providing user of the target content and the user information of the associated users of the target content providing user. In order to obtain the multi-dimensional feature information of the target content, the feature vectors of the user information of the target content providing user and the associated users may be extracted to obtain the multi-dimensional feature vectors of the target content containing the multi-dimensional feature information. The step of “obtaining the content multi-dimensional feature vector of the target content based on the target content and the target content providing user-associated information” may include:

[0072] Acquire a first user representation vector of the target content providing user based on user information of the target content providing user, user information of the associated user, and the degree of association between the target content providing user and the associated user;

[0073] Extracting features of the target content to obtain a content feature vector of the target content;

[0074] The first user representation vector and the content feature vector are fused to obtain a multi-dimensional content feature vector of the target content.

[0075] The feature extraction of the target content may include feature extraction of the title content, the body content, and the target content providing user (eg, content production user, content publishing user, etc.) of the target content.

[0076] In one example, if Figure 2cAs shown, the main text content of the target content and the user information of the content producing user (such as nickname, ID, etc.) can be passed through the BERT (Bidirectional Encoder Representation from Transformers) model to obtain the content feature vector of the target content providing user, and then it can be fused with the first user representation vector of the target content providing user to obtain a multi-dimensional content feature vector.

[0077] Among them, the first user representation vector (poster embedding) can be constructed based on the attention relationship between the consumer users of the target content (for example, the interactive information publisher of the target content, that is, the user who likes, reposts, comments, browses the target content, etc.) and the user who provides the target content, such as the self-media author, etc. This representation can be constructed

[0078] The embedding vector of the target content publisher can be used to obtain the associated feature vector that provides user association information.

[0079] In one embodiment, in order to obtain a first user characterization vector that characterizes user characterization information of a target content providing user, the first user characterization vector may be obtained based on a user relationship graph. The step of “obtaining a first user characterization vector of the target content providing user based on user information of the target content providing user, user information of associated users, and the degree of association between the target content providing user and the associated users” may include:

[0080] Acquire a user relationship graph, the user relationship graph including: nodes of at least two content publishing users and at least two interactive users, a connection edge is set between each content publishing user node and a corresponding interactive user node, the connection edge is set with a weight, the weight represents the degree of association between the content publishing user and the corresponding interactive user, the content publishing user includes a target content providing user of the target content;

[0081] According to the graph walking algorithm, a walking path from the node of the interactive user to the node of the target content providing user is determined, wherein the interactive users on the walking path are associated users of the content publishing user;

[0082] Based on the user information of the associated users on the walking path, the user information of the target content providing user, and the weights of the connecting edges on the walking path, a node vector of the node of the target content providing user is determined, and the node vector is determined as a first user representation vector of the target content providing user.

[0083] In one example, if Figure 2dAs shown in the figure, the author being followed by the user means that the author's self-media account is followed by the user's account, and the poster embedding (i.e., the first user representation vector mentioned above) of the author is constructed. The user u1 and the self-media author p1 (poster, there are many different users and different authors). An author will be followed by many different users. The attention relationship between them constitutes a graph, including nodes and edges. Similarly, for the positive behaviors of users and accounts (such as positive comments, collections, forwarding, sharing, and likes, if it is a video account, the average playback completion rate of the account's video content, and the average reading completion rate of the graphic content), the weights of these behaviors are different. Therefore, a weighted graph is constructed in the user's positive behavior. Based on the weight of the edge, the sampling (weighted walk) makes the sampling walk as much as possible in the direction of the hot nodes, so that the sampled samples have a higher confidence. In addition, in the graph structure, the poster's side-info (account attributes: authority, content quality, account activity) is used, which is very friendly to some long-tail sparse posters and solves the cold start problem to a certain extent. Finally, DeepWalk & Skip-gram are used to construct the poster embedding vector. DeepWalk is a combination of Random Walk and Skip-gram. Random Walk is responsible for sampling the graph structure to obtain the adjacency relationship between nodes in the graph, and Skip-gram trains the node embedding vector from the sampled sequence.

[0084] Among them, cold start refers to the case where the author account has no follow relationship or very few follow relationships, and the vector of the author account can be obtained through side-info. Long-tail sparseness refers to the case where a large number of author accounts have no follow information at the beginning.

[0085] In one example, if Figure 2e As shown in the figure, Deep Walk belongs to the Random Walk class of graph algorithms. Based on the original data of the user's behavior sequence, Random Walk generates a new sequence, and then uses word2vec training to obtain the behavior embedding. Due to different node types, there are two graph construction methods:

[0086] Isomorphic graph: Use behavior-type nodes to generate an isomorphic graph that only contains behaviors. Perform RandomWalk+Word2Vec training based on the isomorphic graph to obtain the embedding representation of the action. Then, the same pooling method as action2vec can be used to generate poster embeddings.

[0087] Heterogeneous graph: Use multiple types of nodes such as action, action attributes, poster, etc. to generate a heterogeneous graph containing multiple types of nodes. The addition of action attributes nodes can alleviate the problem of insufficient training of low-frequency items through side info; the addition of poster nodes allows the model to directly obtain poster embeddings through training. The process of account graph embedding is as follows. In order to obtain the feature representation of the poster, a data set is extracted from the relationship network of the publisher's followers, and the random walk method is used to sample ordered nodes, thereby converting the attention network into ordered nodes for learning. The Skip-gram neural language model is applied to these random walks to obtain graph embedding. The specific steps may include the following:

[0088] Sampling: Generate multiple random walks for each node.

[0089] Training Skip-gram: Treat these random walks as sentences in word2vec. Skip-gram neural network takes a node in the random walk as input and tries to predict its neighbor nodes.

[0090] Calculate the poster Embedding vector.

[0091] 202. Perform feature extraction on at least two types of interaction-related information of the target content to obtain at least one interaction status representation vector of the target content.

[0092] Among them, interactive related information may include information that is interactively related to the target content, such as interactive behavior (such as liking, commenting, deleting, forwarding, and collecting the target content), user interactive information, interactive time (the time when the interactive information is published), information of interactive users, and so on.

[0093] In one embodiment, the at least two types of interaction-related information of the target content include user interaction information and interaction information-posting user association information. In order to obtain interaction status feature information of the target content, feature extraction may be performed on the interaction-related information of the target content. The step of “feature extraction of at least two types of interaction-related information of the target content to obtain at least one interaction status representation vector of the target content” may include:

[0094] Extracting features from at least two dimensions of the user interaction information to obtain information feature vectors of at least two dimensions of the user interaction information;

[0095] Extracting user features from the interactive information publishing user-related information to obtain a second user representation vector of the publishing user of the user interactive information;

[0096] The information feature vectors and the second user representation vector of at least two dimensions are processed to obtain at least one interaction situation representation vector of the target content.

[0097] Among them, user interaction information may include time information on the time dimension and interactive content information on the content dimension. Feature information can be extracted from the time information on the time dimension and the interactive content information on the content dimension, and then processed to obtain at least one interaction situation representation vector of the target content.

[0098] In one embodiment, feature extraction is performed on the user interaction information from at least two dimensions, and feature extraction may be performed from a time dimension and a content dimension. The step of “feature extraction is performed on the user interaction information from at least two dimensions to obtain information feature vectors of at least two dimensions of the user interaction information” may include:

[0099] Extract features of user interaction information from the time dimension to obtain the time feature vector of the user interaction information;

[0100] Extract features of user interaction information from the content dimension to obtain a content feature vector of the user interaction information;

[0101] The time feature vector and the content feature vector are used as information feature vectors of at least two dimensions of user interaction information.

[0102] In one example, if Figure 2c As shown, at least one interaction representation vector of the target content can be obtained through the BERT model from the user interaction information of the target content and the information of the user who publishes the user interaction information, and the time feature vector obtained by encoding the timestamp corresponding to the user interaction information can also be used as an interaction representation information of the target content.

[0103] In one embodiment, the at least two types of interactive association information of the target content include user interactive information and interactive information publishing user association information, and the step of “extracting features from the at least two types of interactive association information of the target content to obtain at least one interactive situation representation vector of the target content” may include:

[0104] Convert the words in the user interaction information into word vectors, and obtain the initial text vector of the user interaction information according to the order of the sentences in which the words in the user interaction information are located and the word vectors of the words;

[0105] Encode the information associated with the user who posted the interactive information to obtain an initial user representation vector of the user who posted the user interactive information;

[0106] The initial text vector and the initial user representation vector are fused to obtain a fused initial vector;

[0107] Perform deep feature extraction on the fused initial vector to obtain the interaction situation representation vector.

[0108] In one example, the at least two types of interactive information of the target content may include user interactive information and interactive information publishing user associated information. For example, the user interactive information is the comment information of the target content, and the interactive information publishing user associated information is the user information of the publishing user of the comment information. Figure 2b As shown, in the comment sentiment analysis system, the target content, the comment information of the target content, the user information of the user who publishes the comment information, etc. can be obtained from the content database.

[0109] In one example, taking the at least two types of interactive relationship information of the target content including user interactive information and interactive information publishing user association information as an example, the feature vector of the target content and the information feature vector of the user interactive information can be extracted as follows:

[0110] The target content and user interaction information can be vectorized using the Bert-Pretrained method to obtain the feature vector of the target content and the information feature vector of the user interaction information. The core of BERT is the bidirectional TransformerEncoder. During the implementation process, the effect of the 12-layer BERT is obviously better than that of the 2-layer BERT, but the training speed and inference speed are also much slower than the 2-layer model. Considering both the model effect and efficiency, the 2-layer model can be selected as the actual model to greatly improve the inference speed under the premise of losing a small amount of accuracy. BERT is a deep bidirectional language representation model based on Transformer. In essence, it uses the Transformer structure to construct a multi-layer bidirectional Encoder network. Transformer is a deep model based on the self-attention mechanism proposed by Google in 2017. It has significant effects on many NLP tasks including machine translation, surpassing RNN and training faster. Pre-trained model Comment-Bert pre-trained model, in the information flow business scenario, a large number of comments are generated every day. By collecting the comment data generated in the business in the past year, there are about 500 million comments after cleaning and processing. This is a very important data resource. A large amount of these review business corpora are added to the general Chinese corpus for model pre-training to complete the domain transfer. In this way, the semantic information in the comments and text content can be well collected and represented.

[0111] 203. Fuse the content multi-dimensional feature vector and at least one interaction situation representation vector to obtain a fused vector.

[0112] The fused vector is a vector that contains more information than the content multi-dimensional feature vector and at least one interaction situation representation vector. The target content can be classified based on the fused vector to determine the content type of the target content.

[0113] In one example, the simplest vector direct joint concatenation method can be used to concatenate the content multi-dimensional feature vector and at least one interaction status representation vector to obtain a fused vector, for example, Figure 2c As shown, based on information processing of the target content, at least two types of interactive related information (e.g., at least two types of interactive related information may include first interactive related information and second interactive related information), and target content publishing user related information, a content multi-dimensional feature vector and at least one interactive situation representation vector (e.g., at least one interactive situation representation vector may include first interactive situation representation vector and second interactive situation representation vector) are obtained, and then vector splicing is performed to obtain a fused vector, and then the target content is classified based on the fused vector to obtain a content classification result.

[0114] 204. Perform content classification of the target content based on the fused vector to obtain a content classification result of the target content.

[0115] Among them, content classification refers to classifying the target content according to the type of content. For example, based on the analysis of the fused vector, the content sentiment type of the target content can be identified as a positive sentiment type, a negative sentiment type, or a neutral sentiment type, etc.

[0116] For example, the articles, videos, audios, etc. published by each content provider are classified according to the emotion type. If the content expression of each article, video, audio conveys positive and active thoughts, then the content emotion type of these articles, videos, audios, etc. can be classified as a positive emotion type. Similarly, if the content expression of each article, video, audio conveys negative and passive thoughts, then the content emotion type of these articles, videos, audios, etc. can be classified as a negative emotion type. If the content expression of each article, video, audio conveys thoughts between positive and negative emotions, then the content emotion type of these articles, videos, audios, etc. can be classified as a positive emotion type and a neutral emotion type. In one embodiment, in order to determine the content type of the target content, the target content can be content classified to determine the content type of the target content according to the content classification result of the target content. The step of "content classification of the target content based on the fused vector to obtain the content classification result of the target content" can include:

[0117] Mapping the fused vector according to the classification mapping matrix to obtain the predicted classification probability of the target content under at least two preset content types;

[0118] Determine the current content type of the target content as the content classification result based on the predicted classification probabilities of the target content under at least two preset content types.

[0119] In one example, a k-dimensional vector, that is, the fused vector, such as the fused vector being (a1, a2, a3, a4, …), can be mapped into a (b1, b2, b3, b4, …) through the softmax function, where bi is a constant between 0 and 1, and the sum of the output neurons is 1.0, so it is equivalent to a probability value. Then, the content classification of the target content can be performed according to the probability magnitude of bi.

[0120] In one embodiment, the content classification may further include content sentiment classification. The fused vector can be mapped according to a preset sentiment classification mapping matrix to obtain the predicted sentiment classification probabilities of the target content under at least two preset sentiment content types, and then determine the current content type of the target content as the content classification result of the target content based on the preset sentiment classification probabilities.

[0121] Among them, the processing of content sentiment classification user interaction information (comments) corpus may include: The sentiment classification system is divided into two major processes: online and offline. The online process performs feature mining, sentiment classification on the sentences output by users, and returns the sentiment analysis results. The offline process is responsible for corpus downloading, feature mining, etc. The accumulation of the corpus is the cornerstone of sentiment classification, and the corpus is divided into labeled corpus and unlabeled corpus. The labeled corpus includes comments on picture and text content, comments on short videos or small videos, comments and annotations on the author, etc. These corpora can determine the sentiment tendency of users through grading. When marking the final result, the account of the content and the text content need to be referred to; for the unlabeled corpus, it needs to be labeled by a classification model or manually before use. Polarity word mining (positive, negative, and neutral), one method is the "full word list method", that is, all words are regarded as polarity words (positive tendency, negative tendency, and neutral). Comment corpus preprocessing: includes related technologies such as word segmentation, denoising, and best match. Denoising is to remove irrelevant information in the content such as "@jjhuang", html tags, etc., and some function words and pronouns that do not have classification significance, such as "de", "a", "wo", etc. The best match is to ensure that the extracted features can correctly reflect the positive and negative tendencies. For example, for the word "at large", if the extracted word is "xiaoyao", it will be misidentified as a positive sentiment feature, while "at large" itself is a negative sentiment word. Generally, the longest match method can be used here.

[0122] Here, the cross-entropy loss is directly used as the final loss function. The cross-entropy cost function (Cross Entry) is as follows:

[0123]

[0124] Cross entropy is used to evaluate the difference between the probability distribution obtained from the current training and the true distribution. Reducing the cross entropy loss is to improve the prediction accuracy of the model. Where p(x) refers to the probability of the true distribution, and q(x) is the probability estimate calculated by the model through the data. Binary cross entropy cost function:

[0125]

[0126] Where f(x) can be a sigmoid function or other activation functions in deep learning. And y(i)∈0,1. It is usually used as the cost function for classification problems. For binary classification: sigmoid+cross entropy, which is the choice in this application. Finally, the value of this probability result is used to predict the sentiment tendency of the final article.

[0127] Among them, sentiment classification, also known as sentiment analysis, is a branch task in the field of natural language processing (NLP), which has emerged with the development of the Internet. In most cases, this task analyzes whether the information presented in a text is positive, negative or neutral. Some studies also make more detailed distinctions, such as further grading in the positive and negative polarities to distinguish different sentiment intensities.

[0128] In one example, in Figure 2b In the comment sentiment analysis system shown, the content production end module C1 first obtains the upload server interface address and then publishes the content through communication with the upstream and downstream content interface server C2. B1 indicates that the content production end uploads the content to be published to the upstream and downstream content interface server C2 based on the obtained upload server interface address. The upstream and downstream content interface server module C2 communicates directly with the content production end C1 to obtain the content submitted from the front end, usually the title, publisher, summary, cover image, and release time of the content, and stores this information in the content database module C6. B5 indicates that the information obtained by the upstream and downstream content interface server C2 is written into the content database module C6.

[0129] Among them, the content database module C6 is the core database of the content. The metadata of all content published by producers is stored in this business database, with the focus on the metadata of the content itself, such as content size, cover image link, title, release time, author account, content source channel, content entry time, as well as the classification results of content during manual review, and other information.

[0130] Among them, after the content production end module C1 communicates with the upstream and downstream content interface server module C2, the uploaded content can be published. The uploaded content can be the content producer of PGC or UGC, MCN or PUGC, which provides graphic content provided by the local or web publishing system or uploads video content including short videos and small videos through the mobile terminal or back-end interface API system.

[0131] Among them, the comment sentiment analysis service module C11 can obtain content from the content database module C6. For each new comment and comment content generated by the content that needs to be distributed, the comment sentiment analysis service module C11 can perform sentiment tendency analysis, and then the comment sentiment analysis model module C12 can process the analysis results to obtain sentiment analysis samples B15. Then the sentiment analysis samples and other information can be stored in the sentiment analysis sample database module C13. In addition, the comment sentiment analysis service module C11 can communicate with the recommendation sorting service module C8 at the same time to downgrade the low-quality content in the published content and distribute the high-quality content in a weighted manner. For example, the distribution weight of the positive content obtained after analysis can be accelerated, and the sorting weight of the negative sentiment content can be reduced to obtain the content distribution result. After that, the content distribution export service module C9 communicates with the recommendation sorting service C8 to obtain the recommendation distribution result, and sends the recommendation distribution result to the content consumption end module C10, and displays the distributed content in the user's Feeds list.

[0132] When performing sentiment analysis on content, it is possible to determine whether the user's behavior is positive or negative based on the user's viewing time ratio of the published content (for example, the time spent viewing the content ratio to the total time spent viewing the content in its entirety), cache time (for example, the time a user keeps the content in a terminal after browsing it), and various interactive behaviors of the user towards the content, and thus assist in determining whether the content type of the content is a positive sentiment type or a negative sentiment type, etc. For example, the user's comments, forwarding, sharing, favorites, likes, and attention to the content providing account can be considered positive behaviors, and the user's content reporting and unfollowing of the content providing account can be considered negative behaviors.

[0133] Among them, PGC (full name: Professional Generated Content) is an Internet term that refers to professionally produced content (video websites) and expert-produced content (Weibo). It is used to refer to content personalization, diversified perspectives, and virtual social relationships. It is also called PPC (Professionally-produced Content).

[0134] Among them, MCN (Multi-Channel Network) is a product form of a multi-channel network, which unites PGC content and, with the strong support of capital, ensures the continuous output of content, thereby ultimately achieving stable commercial monetization.

[0135] Among them, Feeds: news sources, also translated as source material, feed, information provision, feed, summary, source, news subscription, network source (English: web feed, news feed, syndicated feed) is a data format through which the website disseminates the latest information to users, usually arranged in a timeline. Timeline is the most primitive, direct and basic display form of Feed. The prerequisite for users to subscribe to a website is that the website provides a source of news. The convergence of feeds in one place is called aggregation, and the software used for aggregation is called an aggregator. For end users, an aggregator is software specifically used to subscribe to websites, and is generally also called an RSS reader, feed reader, news reader, etc.

[0136] 205. Recommend target content based on the content classification result.

[0137] Among them, the content classification result can indicate the content type of the target content. For example, the content classification result can indicate that the content type of the target content is a positive emotion type, a negative emotion type, or a neutral emotion type. For content of the positive emotion type, the weight of content distribution can be increased to increase the exposure of this type of content.

[0138] In one embodiment, in order to improve the accuracy of content recommendation, content may be recommended based on the content classification result. The step of “recommending target content based on the content classification result” may include:

[0139] Obtain the historical content classification results of the target content within a preset historical time from the current moment, and the corresponding result acquisition time;

[0140] Determine the recommendation reference weight of the content classification result and each historical content classification result based on the result acquisition time of the content classification result and the historical content classification result;

[0141] Determine the content type representation score of each preset content type for the target content according to the number of occurrences of each preset content type in the content classification results and the historical content classification results, and the recommendation reference weight corresponding to each occurrence of the preset content type;

[0142] According to the content type representation score corresponding to the recommended reference content type in the preset content type, the recommendation strength of the target content is determined, and the target content is recommended based on the recommendation strength.

[0143] In one example, the higher the content type representation score, the stronger the recommendation strength of the corresponding content. For a specific content, only the user who posted the comment and the current content of the comment change during the prediction. The other content and account embeddings do not change in the short term. They can be calculated in advance and saved in memory, which can speed up the prediction process. Finally, the comments and user embeddings are contacted to make the final prediction. In the process of combining the actual and recommended distribution sorting, a time window Wt is set (different types of content can set different Wt) and then three counters Cp positive, Cm neutral, and Cn negative are set. Then, the counts of the three values ​​corresponding to the content are counted in Wt respectively. The calculation weight of the most recent time window is higher. If Cp is greater than a certain value N, the weight is increased when sorting the corresponding article to increase exposure. The probability of increase is proportional to SUMCp (the accumulation of multiple Cp time windows); similarly, if Cn is greater than a certain value, the weight is reduced to reduce exposure. In this way, the changes in user comments will be directly related to the exposure probability, affecting the final content distribution efficiency.

[0144] In one example, by establishing a comment-content-account multimodal model, with comments and content as the main body, the sentiment tendency trend of content quality is judged in multiple dimensions, and the low-quality content is downgraded through linkage recommendation, and the high-quality content is weighted for distribution. If it is positive content, it can accelerate the increase of distribution weight, and if it is negative sentiment content, its ranking weight can be reduced. The sentiment analysis tendency of comments is negative, but the content itself is not necessarily negative. If you only look at the positive or negative sentiment tendency of the user's comment content itself, it will cause distortion of the results, so it is necessary to distinguish it in combination with the content corresponding to the comment. The sentiment analysis of user comments needs to distinguish whether it is the user's expression of opinions on the content itself or on other content, which is completely different in nature. If some comments are obviously negative comments, but their local corresponding content is not negative, these contents themselves are easy to arouse readers' discussion atmosphere, and the discussion often revolves around these stars or social events. A single comment is negative, but it can trigger extensive discussion and traffic in the distribution of information flow content. This should not be downgraded or restricted. Therefore, in order to distinguish these situations, the comment content and the user itself can be introduced.

[0145] In the process of research and practice of relevant technologies, the inventors of the present application found that in the current process of recommending content to users, most of the time, only the semantics of the content is considered to recommend content, and the accuracy of content recommendation is poor. However, through the embodiments of the present application, various negative quality issues that have not been discovered by humans and machines in the early stage of the review can be quickly discovered through user sentiment tendency mining, and the spread and dissemination of low-quality content can be effectively controlled; the weight of high-quality content distribution can be improved, and the user's comment behavior information can better promote the distribution of high-quality accounts. High-quality accounts can get targeted traffic first, and users can be more aware of the content released by high-quality accounts, so that the content ecology can enter a virtuous cycle and form a healthy content ecology, thereby improving the speed of cold start of high-quality content.

[0146] The embodiment of the present application can provide an analysis system for real-time mining of users' real-time comments on content, judge the positive and negative emotions of a single comment, and apply the results directly to the recommendation system based on the POP content of all comment results. The user's feedback and perception can be quickly responded to and fed back, allowing the system to cycle in a positive dimension. The core idea is: to classify the text characteristics of comments for sentiment: timeliness, short text, irregular expression, and large amount of information. The sentiment analysis system is divided into two major processes: online and offline. The online process performs feature mining and sentiment classification on the user's output sentences and returns the results. The offline process is responsible for corpus downloading and preprocessing. In the specific model of sentiment analysis, the comment-user-content multimodal joint modeling is performed, the text of the graphic content is pre-trained by BERT, and the comments, users, and accounts are Deep Walk vectorized to distinguish between comments and content information. Comments include comment content and user information, and content includes account and content information; finally, it is combined with the main process of content processing of the recommendation system.

[0147] From the above, it can be seen that the embodiment of the present application can obtain a fused vector containing more feature information by fusing a multi-dimensional feature vector including the target content and the content providing user-related information with at least one interaction situation representation vector, and then classify the target content based on the fused vector. The target content can then be used in a recommendation system to recommend the target content. It can quickly discover various negative content quality issues that have not been discovered by humans and machines in the early stage of the review through content classification mining, effectively control the spread and dissemination of low-quality content, and enable high-quality content to be effectively weighted, quickly recommended, distributed and improved, thereby improving the accuracy of content recommendations.

[0148] Based on the above introduction, the following example will be used to further illustrate the content recommendation method of this application. Figure 3 , a content recommendation method, the specific process can be as follows:

[0149] 301. The server determines a content multi-dimensional feature vector of the target content based on a content feature vector of the target content and a first user characterization vector of a user who provides the target content.

[0150] The first user representation vector of the target user can be obtained based on the user information of the target content provider, the user information of the associated user, and the degree of association between the target content provider and the associated user, for example. Figure 2d As shown, based on the graph walking algorithm, a walking path from the node of the interactive user to the node of the target content providing user is determined from the user relationship graph, and then based on the user information of the associated users on the walking path, the user information of the target content providing user, and the weight of the connecting edge on the walking path, the node vector of the node of the target content providing user is determined, and the node vector is determined as the first user representation vector of the target content providing user.

[0151] The content feature vector of the target content can be obtained by extracting features of the title content, the text content, the target content providing user (eg, content production user, content publishing user, etc.) and the like of the target content.

[0152] 302. The server extracts features of the user interaction information of the target content from at least two dimensions to obtain information feature vectors of at least two dimensions of the user interaction information.

[0153] When extracting features from at least two dimensions of user interaction information of target content, feature extraction may be performed from two dimensions: time dimension and content dimension, so as to obtain information feature vectors of at least two dimensions of user interaction information.

[0154] In one example, feature extraction can be performed on the user interaction information from the time dimension to obtain a time feature vector of the user interaction information. For example, the timestamp of the user interaction information can be encoded to obtain the time feature vector of the user interaction information. Then, feature extraction can be performed on the user interaction information from the content dimension to obtain a content feature vector of the user interaction information. Finally, the time feature vector and the content feature vector are used as information feature vectors of at least two dimensions of the user interaction information.

[0155] 303. The server extracts user features from the interactive information publishing user-related information to obtain a second user representation vector of the publishing user of the user interactive information.

[0156] In one example, user feature extraction is performed on the user-related information of the user who releases the interactive information. User feature extraction can be performed on the user information (such as nickname, ID, etc.) of the user who releases the interactive information. For example, the ID of the user who releases the interactive information can be One-hot encoded to obtain a second user representation vector of the user who releases the interactive information.

[0157] 304. The server processes the information feature vectors and the second user representation vector of at least two dimensions to obtain at least one interaction situation representation vector of the target content.

[0158] In one example, information feature vectors of at least two dimensions may include a time feature vector of user interaction information and a content feature vector of user interaction information. The content feature vector of the user interaction information and the second user representation vector may be fused, such as by vector splicing, to obtain an interaction situation representation vector of the target content. The time feature vector of the user interaction information may be processed to obtain another interaction situation representation vector of the target content.

[0159] 305. The server fuses the content multi-dimensional feature vector and at least one interaction situation representation vector to obtain a fused vector.

[0160] In one example, the simplest vector direct joint splicing method can be used to perform vector splicing on the content multi-dimensional feature vector and at least one interaction situation representation vector to obtain a fused vector, which contains more feature information than the content multi-dimensional feature vector and at least one interaction situation representation vector.

[0161] 306. The server performs content classification of the target content based on the fused vector to obtain a content classification result of the target content.

[0162] In one example, the target content can be classified by the softmax function, that is, the softmax function can map a k-dimensional fused vector, for example, the fused vector is (a1, a2, a3, a4, ...) into a (b1, b2, b3, b4, ...), where bi is a constant ranging from 0 to 1, and the sum of the output neurons is 1.0, so it is equivalent to a probability value. Then, the target content can be classified according to the probability of bi, and the content type of the target content can be determined as the content classification result of the target content.

[0163] 307. The server recommends target content based on the content classification result.

[0164] In one example, the content classification result can indicate the content type of the target content. The content type of the target content can be determined based on the content classification result. For positive and high-quality content, its distribution weight can be increased. The target content can be recommended based on the distribution weight, which can increase the exposure of the high-quality content.

[0165] Through this application, various negative quality issues that have not been discovered by humans and machines in the early stage of review can be quickly discovered through user sentiment tendency mining, effectively controlling the spread and dissemination of low-quality content; high-quality content with positive user evaluations can be quickly discovered, increasing the speed of cold start distribution of high-quality author content, and at the same time increasing the weight of high-quality content distribution, so that users' comment behavior information can better promote the distribution of high-quality accounts. High-quality accounts can get targeted traffic first, and users can be more aware of the content released by high-quality accounts, so that the content ecology enters a virtuous cycle and forms a healthy content ecology; the entire mining and processing expansion is driven by real-time interaction of users, which can effectively guarantee timeliness.

[0166] From the above, it can be seen that the embodiment of the present application can obtain a fused vector containing more feature information by fusing a multi-dimensional feature vector including the target content and the content providing user-related information with at least one interaction situation representation vector, and then classify the target content based on the fused vector. The target content can then be used in a recommendation system to recommend the target content. It can quickly discover various negative content quality issues that have not been discovered by humans and machines in the early stage of the review through content classification mining, effectively control the spread and dissemination of low-quality content, and enable high-quality content to be effectively weighted, quickly recommended, distributed and improved, thereby improving the accuracy of content recommendations.

[0167] In order to better implement the above method, accordingly, the embodiment of the present application also provides a content recommendation device, wherein the content recommendation device can be specifically integrated in a server, referring to Figure 4a , the content recommendation device may include an acquisition unit 401, a feature extraction unit 402, a vector fusion unit 403, a content classification unit 404 and a recommendation unit 405, as follows:

[0168] (1) Acquisition unit 401;

[0169] The acquisition unit 401 is used to acquire a content multi-dimensional feature vector of the target content based on the target content and the target content providing user associated information, wherein the content multi-dimensional feature vector includes feature information of the target content and the target content providing user associated information.

[0170] In one embodiment, if Figure 4b As shown, the acquisition unit 401 includes:

[0171] An acquisition subunit 4011 is used to acquire a first user representation vector of a target content providing user based on user information of the target content providing user, user information of associated users, and a degree of association between the target content providing user and the associated users;

[0172] The first extraction subunit 4012 is used to extract features from the target content to obtain a content feature vector of the target content;

[0173] The vector fusion subunit 4013 is used to fuse the first user representation vector and the content feature vector to obtain a multi-dimensional content feature vector of the target content.

[0174] In one embodiment, the acquisition subunit 4011 is also used to obtain a user relationship graph, which includes: nodes of at least two content publishing users and at least two interactive users, a connection edge is set between the node of each content publishing user and the node of the corresponding interactive user, and the connection edge is set with a weight, the weight represents the degree of association between the content publishing user and the corresponding interactive user, and the content publishing user includes a target content providing user of the target content; according to a graph walking algorithm, a walking path from the node of the interactive user to the node of the target content providing user is determined, wherein the interactive users on the walking path are associated users of the content publishing user; based on user information of the associated users on the walking path, user information of the target content providing user, and the weight of the connection edge on the walking path, a node vector of the node of the target content providing user is determined, and the node vector is determined as a first user representation vector of the target content providing user.

[0175] (2) feature extraction unit 402;

[0176] The feature extraction unit 402 is used to extract features of at least two types of interaction-related information of the target content to obtain at least one interaction status representation vector of the target content.

[0177] In one embodiment, if Figure 4c As shown, the feature extraction unit 402 includes:

[0178] The second extraction subunit 4021 is used to extract features from at least two dimensions of the user interaction information to obtain information feature vectors of at least two dimensions of the user interaction information;

[0179] The third extraction subunit 4022 is used to extract user features from the interactive information publishing user-related information to obtain a second user representation vector of the publishing user of the user interactive information;

[0180] The processing subunit 4023 is used to process the information feature vectors and the second user representation vector of at least two dimensions to obtain at least one interaction situation representation vector of the target content.

[0181] In one embodiment, the second extraction sub-unit 4021 is also used to perform feature extraction on the user interaction information from the time dimension to obtain the time feature vector of the user interaction information; perform feature extraction on the user interaction information from the content dimension to obtain the content feature vector of the user interaction information; and use the time feature vector and the content feature vector as information feature vectors of at least two dimensions of the user interaction information.

[0182] In one embodiment, if Figure 4c As shown, the feature extraction unit 402 includes:

[0183] The conversion subunit 4024 is used to convert the words in the user interaction information into word vectors, and obtain the initial text vector of the user interaction information according to the order of the sentences in which the words in the user interaction information are located and the word vectors of the words;

[0184] The encoding subunit 4025 is used to encode the interactive information publishing user association information to obtain an initial user representation vector of the publishing user of the user interactive information;

[0185] A fusion subunit 4026 is used to fuse the initial text vector and the initial user representation vector to obtain a fused initial vector;

[0186] The fourth extraction subunit 4027 is used to perform deep feature extraction on the fused initial vector to obtain an interaction situation representation vector.

[0187] (3) vector fusion unit 403;

[0188] The vector fusion unit 403 is used to fuse the content multi-dimensional feature vector and at least one interaction situation representation vector to obtain a fused vector.

[0189] (4) content classification unit 404;

[0190] The content classification unit 404 is used to classify the target content based on the fused vector to obtain a content classification result of the target content.

[0191] In one embodiment, if Figure 4d As shown, the content classification unit 404 includes:

[0192] A mapping subunit 4041 is used to map the fused vector according to the classification mapping matrix to obtain the predicted classification probability of the target content under at least two preset content types;

[0193] The result determination subunit 4042 is configured to determine the current content type of the target content as a content classification result based on the predicted classification probabilities of the target content under at least two preset content types.

[0194] (5) Recommendation unit 405;

[0195] The recommendation unit 405 is used to recommend target content based on the content classification result.

[0196] In one embodiment, if Figure 4e As shown, the recommendation unit 405 includes:

[0197] The acquisition subunit 4051 is used to acquire the historical content classification results of the target content within a preset historical time from the current moment, and the corresponding result acquisition time;

[0198] The weight determination subunit 4052 is used to determine the recommendation reference weight of the content classification result and each historical content classification result based on the result acquisition time of the content classification result and the historical content classification result;

[0199] The score determination subunit 4053 is used to determine the content type representation score of each preset content type for the target content according to the number of occurrences of each preset content type in the content classification result and the historical content classification result, and the recommendation reference weight corresponding to each occurrence of the preset content type;

[0200] The recommendation subunit 4054 is used to determine the recommendation strength of the target content according to the content type representation score corresponding to the recommended reference content type in the preset content type, and recommend the target content based on the recommendation strength.

[0201] As can be seen from the above, the acquisition unit 401 of the content recommendation device of the embodiment of the present application acquires the content multi-dimensional feature vector of the target content based on the target content and the user association information providing the target content, and the content multi-dimensional feature vector includes the feature information of the target content and the user association information providing the target content; then, the feature extraction unit 402 extracts features of at least two kinds of interactive association information of the target content to obtain at least one interactive situation representation vector of the target content; the vector fusion unit 403 fuses the content multi-dimensional feature vector and the at least one interactive situation representation vector to obtain a fused vector; the content classification unit 404 classifies the target content based on the fused vector to obtain a content classification result of the target content; the recommendation unit 405 recommends the target content based on the content classification result. This solution can obtain a fused vector containing more feature information by fusing a multi-dimensional feature vector including the target content and the content providing user-related information with at least one interaction situation representation vector, and then classify the target content based on the fused vector. The target content can then be used in a recommendation system to recommend the target content. Through content classification mining, it can quickly discover various negative content quality issues that have not been discovered by humans and machines in the early stages of the review, effectively control the spread and dissemination of low-quality content, and enable high-quality content to be effectively weighted, quickly recommended, distributed and promoted, thereby improving the accuracy of content recommendations.

[0202] In addition, the embodiment of the present application also provides a computer device, which can be a terminal or a server. Figure 5 As shown, it shows a schematic diagram of the structure of the computer device involved in the embodiment of the present application, specifically:

[0203] The computer device may include components such as a processor 501 with one or more processing cores, a memory 502 with one or more storage media, a power supply 503, and an input unit 504. Those skilled in the art will appreciate that Figure 5 The computer device structure shown in the figure does not constitute a limitation on the computer device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0204] in:

[0205] The processor 501 is the control center of the computer device. It uses various interfaces and lines to connect various parts of the entire computer device. By running or executing software programs and / or modules stored in the memory 502 and calling data stored in the memory 502, it executes various functions of the computer device and processes data, thereby performing overall detection of the computer device. Optionally, the processor 501 may include one or more processing cores; preferably, the processor 501 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 501.

[0206] The memory 502 can be used to store software programs and modules. The processor 501 executes various functional applications and data processing by running the software programs and modules stored in the memory 502. The memory 502 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 502 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory 502 may also include a memory controller to provide the processor 501 with access to the memory 502.

[0207] The computer device also includes a power supply 503 for supplying power to various components. Preferably, the power supply 503 can be logically connected to the processor 501 through a power management system, so as to manage charging, discharging, power consumption and other functions through the power management system. The power supply 503 can also include one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, power status indicators and other arbitrary components.

[0208] The computer device may further include an input unit 504, which may be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0209] Although not shown, the computer device may further include a display unit, etc., which will not be described in detail herein. Specifically in this embodiment, the processor 501 in the computer device will load the executable files corresponding to the processes of one or more application programs into the memory 502 according to the following instructions, and the processor 501 will run the application programs stored in the memory 502, thereby realizing various functions, as follows:

[0210] Based on the target content and the target content-providing user association information, a content multi-dimensional feature vector of the target content is obtained, the content multi-dimensional feature vector including the feature information of the target content and the target content-providing user association information; feature extraction is performed on at least two types of interaction association information of the target content to obtain at least one interaction situation representation vector of the target content; the content multi-dimensional feature vector and the at least one interaction situation representation vector are fused to obtain a fused vector; content classification of the target content is performed based on the fused vector to obtain a content classification result of the target content; and the target content is recommended based on the content classification result.

[0211] From the above, it can be seen that the embodiment of the present application can obtain a fused vector containing more feature information by fusing a multi-dimensional feature vector including the target content and the content providing user-related information with at least one interaction situation representation vector, and then classify the target content based on the fused vector. The target content can then be used in a recommendation system to recommend the target content. It can quickly discover various negative content quality issues that have not been discovered by humans and machines in the early stage of the review through content classification mining, effectively control the spread and dissemination of low-quality content, and enable high-quality content to be effectively weighted, quickly recommended, distributed and improved, thereby improving the accuracy of content recommendations.

[0212] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0213] To this end, an embodiment of the present application provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps in any content recommendation method provided in the embodiment of the present application. For example, the instructions can execute the following steps:

[0214] Based on the target content and the target content-providing user association information, a content multi-dimensional feature vector of the target content is obtained, the content multi-dimensional feature vector including the feature information of the target content and the target content-providing user association information; feature extraction is performed on at least two types of interaction association information of the target content to obtain at least one interaction situation representation vector of the target content; the content multi-dimensional feature vector and the at least one interaction situation representation vector are fused to obtain a fused vector; content classification of the target content is performed based on the fused vector to obtain a content classification result of the target content; and the target content is recommended based on the content classification result.

[0215] The computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0216] Since the instructions stored in the computer-readable storage medium can execute the steps in any content recommendation method provided in the embodiments of the present application, the beneficial effects that can be achieved by any content recommendation method provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0217] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the content recommendation method provided in the above-mentioned invention content and embodiments.

[0218] The above is a detailed introduction to a content recommendation method, device, computer equipment and computer-readable storage medium provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the ideas of the present application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A content recommendation method, characterized in that: include: Based on the target content and the information associated with the user providing the target content, obtaining a content multi-dimensional feature vector of the target content, wherein the content multi-dimensional feature vector includes feature information of the target content and the information associated with the user providing the target content; Extracting features of at least two types of interactive association information of the target content to obtain at least one interactive situation representation vector of the target content, wherein the at least two types of interactive association information include interactive users of the target content and interactive behaviors of the interactive users with respect to the target content; Fusing the content multi-dimensional feature vector and the at least one interaction situation representation vector to obtain a fused vector; Performing content classification of the target content based on the fused vector to obtain a content classification result of the target content includes: According to the interactive behavior of the interactive user with respect to the target content, determining a behavior corresponding to the interactive user from positive behaviors and negative behaviors; Performing content classification of the target content based on the behaviors corresponding to each interactive user and the fused vector to obtain a content classification result of the target content; Recommending the target content based on the content classification result includes: Obtaining historical content classification results of the target content within a preset historical time from the current moment, and the corresponding result acquisition time; determining, based on the content classification result and the result acquisition time of the historical content classification result, a recommendation reference weight of the content classification result and each of the historical content classification results; Determine a content type representation score of each preset content type for the target content according to the number of occurrences of each preset content type in the content classification result and the historical content classification result, and a recommendation reference weight corresponding to each occurrence of the preset content type; The recommendation strength of the target content is determined according to the content type representation score corresponding to the recommended reference content type in the preset content type, and the target content is recommended based on the recommendation strength.

2. The method according to claim 1, characterized in that The providing user associated information includes the target content providing user of the target content and user information of associated users of the target content providing user; The step of acquiring a content multi-dimensional feature vector of the target content based on the target content and the target content providing user association information includes: Acquire a first user representation vector of the target content providing user based on the user information of the target content providing user, the user information of the associated user, and the degree of association between the target content providing user and the associated user; Extracting features of the target content to obtain a content feature vector of the target content; The first user representation vector and the content feature vector are merged to obtain a multi-dimensional content feature vector of the target content.

3. The method according to claim 2, characterized in that The acquiring a first user representation vector of the target content providing user based on the user information of the target content providing user, the user information of the associated user, and the degree of association between the target content providing user and the associated user comprises: Acquire a user relationship graph, the user relationship graph comprising: nodes of at least two content publishing users and at least two interactive users, a connection edge is set between each content publishing user node and a corresponding interactive user node, the connection edge is set with a weight, the weight represents the degree of association between the content publishing user and the corresponding interactive user, the content publishing user includes a target content providing user of the target content; According to a graph walking algorithm, a walking path from a node of an interactive user to a node of a target content providing user is determined, wherein the interactive users on the walking path are associated users of the content publishing user; Based on the user information of the associated users on the walking path, the user information of the target content providing user, and the weights of the connecting edges on the walking path, a node vector of the node of the target content providing user is determined, and the node vector is determined as a first user representation vector of the target content providing user.

4. The method according to claim 1, characterized in that: The at least two types of interactive related information of the target content include user interactive information and interactive information publishing user related information; The extracting features of at least two types of interaction-related information of the target content to obtain at least one interaction status representation vector of the target content includes: Extracting features from at least two dimensions of the user interaction information to obtain information feature vectors of at least two dimensions of the user interaction information; Extracting user features from the interactive information publishing user-related information to obtain a second user representation vector of the publishing user of the user interactive information; The information feature vectors of the at least two dimensions and the second user representation vector are processed to obtain at least one interaction situation representation vector of the target content.

5. The method according to claim 4, characterized in that The extracting features of the user interaction information from at least two dimensions to obtain information feature vectors of at least two dimensions of the user interaction information includes: Extracting features of the user interaction information from a time dimension to obtain a time feature vector of the user interaction information; Extracting features from the content dimension of the user interaction information to obtain a content feature vector of the user interaction information; The time feature vector and the content feature vector are used as information feature vectors of at least two dimensions of the user interaction information.

6. The method according to claim 1, characterized in that The at least two types of interactive related information of the target content include user interactive information and interactive information publishing user related information; The extracting features of at least two types of interaction-related information of the target content to obtain at least one interaction status representation vector of the target content includes: Converting the words in the user interaction information into word vectors, and obtaining an initial text vector of the user interaction information according to the order of the sentences in which the words in the user interaction information are located and the word vectors of the words; Encoding the interactive information publishing user association information to obtain an initial user representation vector of the publishing user of the user interactive information; Fusing the initial text vector and the initial user representation vector to obtain a fused initial vector; Perform deep feature extraction on the fused initial vector to obtain an interaction situation representation vector.

7. The method according to claim 1, characterized in that The performing content classification of the target content based on the fused vector to obtain a content classification result of the target content includes: Mapping the fused vector according to a classification mapping matrix to obtain predicted classification probabilities of the target content under at least two preset content types; Based on the predicted classification probabilities of the target content under at least two preset content types, the current content type of the target content is determined as a content classification result.

8. A content recommendation device, characterized in that: include: an acquisition unit, configured to acquire a content multi-dimensional feature vector of the target content based on the target content and the user-providing information associated with the target content, wherein the content multi-dimensional feature vector includes feature information of the target content and the user-providing information associated with the target content; a feature extraction unit, configured to extract features of at least two types of interactive association information of the target content to obtain at least one interactive situation representation vector of the target content, wherein the at least two types of interactive association information include interactive users of the target content and interactive behaviors of the interactive users with respect to the target content; A vector fusion unit, used to fuse the content multi-dimensional feature vector and the at least one interaction situation representation vector to obtain a fused vector; A content classification unit, configured to classify the target content based on the fused vector to obtain a content classification result of the target content, comprising: According to the interactive behavior of the interactive user with respect to the target content, determining a behavior corresponding to the interactive user from positive behaviors and negative behaviors; Performing content classification of the target content based on the behaviors corresponding to each interactive user and the fused vector to obtain a content classification result of the target content; A recommendation unit, configured to recommend the target content based on the content classification result, comprising: Obtaining historical content classification results of the target content within a preset historical time from the current moment, and the corresponding result acquisition time; determining, based on the content classification result and the result acquisition time of the historical content classification result, a recommendation reference weight of the content classification result and each of the historical content classification results; Determine a content type representation score of each preset content type for the target content according to the number of occurrences of each preset content type in the content classification result and the historical content classification result, and a recommendation reference weight corresponding to each occurrence of the preset content type; The recommendation strength of the target content is determined according to the content type representation score corresponding to the recommended reference content type in the preset content type, and the target content is recommended based on the recommendation strength.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the content recommendation method according to any one of claims 1 to 7.

10. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the content recommendation method according to any one of claims 1 to 7.

Images