Topic recommendation method and system based on interaction and popularity prediction of highly influential figures

Abstract

The present invention relates to topic recommendation technology, and provides a topic recommendation method and system based on the interaction and popularity prediction of highly influential people. The method includes: constructing a dynamic user-topic knowledge graph based on user information, social media topic libraries, and ontology models, and using graph embedding model training to obtain interaction preference values; extracting user relationship networks based on the knowledge graph, identifying highly influential users, obtaining their historical blog posts and clustering them into topics, performing sentiment analysis on comments, calculating interest levels based on interaction intensity, and decomposing the interest matrix to obtain latent vectors of users and blog posts; fusing user historical comments and blog posts through an attention mechanism to obtain a fusion vector, and jointly predicting interest scores with the latent vectors to obtain implicit interest scores and a set of potential preferred topics; obtaining a heat prediction value based on the topic heat index; and obtaining a comprehensive score based on each score to complete personalized topic recommendation. The present invention can characterize user potential interests and predict heat trends, thereby improving recommendation effectiveness.

Description

Technical Field

[0001] The present invention belongs to the field of topic recommendation, and in particular relates to a topic recommendation method and system based on interaction with highly influential people and popularity prediction. Background Art

[0002] With the rapid development of the internet and social media, the speed and reach of topic information are increasing. This rapid growth and diverse sources of information have led to topic information overload, placing immense pressure on users to sift through information. Personalized topic recommendations can push relevant topic information based on user interests and needs, avoiding information overload and redundancy. Compared to traditional "information bombardment" push notifications, personalized recommendations expose users to more content tailored to their needs, increasing user engagement and satisfaction. However, finding a technical approach to comprehensively consider both user and topic characteristics within this massive amount of topic data, effectively balancing user needs and topic characteristics to accurately recommend personalized hot topics to users, remains a major challenge.

[0003] In related technologies, topic recommendation mainly adopts the following methods: processing massive Internet data to obtain hot topics and corresponding keywords, expanding them to generate a library of short sentences in hot areas, using a text semantic similarity model to analyze the proportion of hot topics in articles, and completing hot area recommendations; screening news data with high popularity ratings and Weibo data with high popularity values ​​to complete news and Weibo recommendations; then, based on a personalized user dictionary, training a text semantic similarity model to analyze the proportion of personalized content in articles, and completing personalized customized recommendations; combining all the above recommendations, pushing hot articles in real time.

[0004] Although the above technology can complete recommendations based on current popularity data, it lacks the prediction and analysis of the changing trends of topic popularity. Therefore, it cannot effectively deal with the volatility or rapid changes in topic popularity, nor can it fully capture potential hot spots or topics that are about to break out, making the recommended content untimely and unforesighted, reducing the user experience. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a topic recommendation method and system based on the interaction and popularity prediction of highly influential people, which can predict the popularity change trend in the future, thereby recommending more accurate topic information to users and improving user experience.

[0006] A topic recommendation method based on interactions between highly influential people and popularity prediction, including:

[0007] Obtaining a social media topic library and user information, wherein the user information includes operation information, personal information, and social relationship information;

[0008] Constructing a user-topic knowledge graph based on the user information, social media topic library, and a preset ontology model;

[0009] Initialize the initial embedding vector of each node and edge in the user-topic knowledge graph using the graph embedding model, where the nodes include user nodes and topic nodes;

[0010] Performing contrastive loss training on the initial embedding vector to obtain a first loss;

[0011] Aggregate the nodes in the user-topic knowledge graph to obtain a node aggregation vector, where the node aggregation vector includes a user aggregation vector and a topic aggregation vector;

[0012] Calculating an interaction preference value based on the user aggregation vector and the topic aggregation vector, calculating a second loss based on the interaction preference value, summing the first loss and the second loss to obtain a total loss, and training the graph embedding model based on the total loss until the total loss is minimized, thereby obtaining an optimized graph embedding model;

[0013] Obtaining an optimized user aggregation vector and an optimized topic aggregation vector based on the optimized graph embedding model and the user-topic knowledge graph, and obtaining an optimized interaction preference value based on the optimized user aggregation vector and the optimized topic aggregation vector;

[0014] All user aggregation vectors form a user vector collection, select any vector in the user vector collection as the target vector, calculate the similarity between the target vector and other vectors in the user vector collection, and select the vector with the highest similarity. k The users are taken as a collection of similar users, and the historical interaction topics of each user in the collection of similar users within a preset time period are obtained as a set of potential preferred topics;

[0015] Obtain the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows, sort the initial topic set according to the implicit interest score, and select the top topic with the highest implicit interest score in the initial topic set. k The initial topics are used as the supplementary preference topic set, and the completed preference topic set is obtained according to the potential preference topic set and the supplementary preference topic set;

[0016] Obtaining topic popularity indicators within a first preset time period from the social media topic library, screening a set of complementary preferred topics based on the topic popularity indicators to obtain a set of candidate topics, and predicting corresponding popularity prediction values ​​of topics within a second future time period based on a time series of topics in the candidate topic set;

[0017] According to the optimized interaction preference value, implicit interest score and heat prediction value, the user's comprehensive score for each topic in the candidate topic set is calculated, and each topic in the candidate topic set is sorted according to the comprehensive score, and the top topics are selected from the largest to the smallest. N topics as recommendation results.

[0018] Optionally, constructing a user-topic knowledge graph based on the user information, the social media topic library, and a preset ontology model includes:

[0019] Using the ontology model, user information and topic information in the social media topic library are converted into triples;

[0020] Users and topics are used as nodes in the user-topic knowledge graph. Each node carries attribute information. Edges between nodes are generated based on the triples and the weights of the edges are initialized.

[0021] Obtaining interaction behavior between the user and the topic based on the user information;

[0022] According to the interaction behavior, a historical interaction edge is generated, and the interaction weight of the historical interaction edge is calculated, which is expressed as:

[0023]

[0024] in, Represents a user and topic The weight of the edge between It is a collection of user interaction behaviors, including search, click, forward, like and comment. For each interaction type The weight coefficient is used to adjust the influence of different behaviors. For users On topic Interaction behavior type the number of For users On topic The intensity of emotional polarity, , is the time attenuation factor, which controls the influence of time on the weight. Represents a user Last interaction topic The time difference from the current time, For users and topic Initialized weights of edges between them;

[0025] Construct a user-topic knowledge graph based on nodes, edge weights, interaction edges, and interaction weights.

[0026] Optionally, the calculating an interaction preference value based on the user aggregation vector and the topic aggregation vector, calculating a second loss based on the interaction preference value, summing the first loss and the second loss to obtain a total loss, and training the graph embedding model based on the total loss until the total loss is minimized, thereby obtaining the optimized graph embedding model, including:

[0027] The interaction preference value calculated based on the user aggregation vector and the topic aggregation vector is expressed as:

[0028]

[0029] in, is the interaction preference value, Aggregate vectors for users, is the topic aggregation vector, T is the transposed matrix;

[0030] The second loss is calculated according to the interaction preference value and is expressed as:

[0031]

[0032] in, , Represents the user's real historical topic interaction record, and Represents the interaction set in which the user has not interacted, that is, the interaction set that is positively sampled and negatively sampled based on the user's historical topic interaction records. express sigmoid function, The topics in the interaction set that the user has not interacted with;

[0033] The first loss and the second loss are summed to obtain the total loss, which is expressed as:

[0034]

[0035] in, For the first loss, For the second loss, Represents the parameter set of the model, E refers to the embedding matrix of all entities and relationships in the knowledge graph, Representing relationships The weight of For the Layer attention information aggregation weight matrix, For the Layer attention information propagation weight matrix, For the Layer neighbor node information aggregation stage, is the set of all relationship types in the knowledge graph, is the regularization coefficient, L The maximum number of layers for aggregation.

[0036] The graph embedding model is trained according to the total loss until the total loss is minimized, thereby obtaining an optimized graph embedding model.

[0037] Optionally, obtaining a topic popularity index within a first preset time period from the social media topic library, screening a set of potential preferred topics based on the topic popularity index to obtain a set of candidate topics, and predicting a popularity prediction value corresponding to a topic within a second future time period based on a time series of topics in the candidate topic set includes:

[0038] The topic popularity index within the first preset time period is obtained from the social media topic library, and a topic popularity time series is constructed according to the time sequence, which is expressed as:

[0039]

[0040] in, For the The topic popularity index of the day;

[0041] According to the topic popularity time series, the damped cumulative series of the corresponding topic popularity is obtained, which is expressed as:

[0042]

[0043] in, For the The accumulated damping value of the day, is the damping accumulation parameter, , For the sky, For the The topic popularity index of the day;

[0044] Obtaining a prediction result according to the damped cumulative sequence and the whitening equation of the grey prediction model;

[0045] The prediction results are restored by an inverse accumulation generation operator to obtain a topic popularity prediction sequence as the popularity prediction value corresponding to the topic in the second time period;

[0046] The whitening equation is expressed as:

[0047]

[0048] in, and are the model parameters estimated by the least squares method, is the first item in the time series and is the indicator farthest from the current time. For the predicted The damped cumulative forecast value of the day, that is, the forecast result.

[0049] Optionally, the step of restoring the prediction result by using an inverse accumulation generation operator to obtain a topic popularity prediction sequence as a popularity prediction value corresponding to the topic in the second time period includes:

[0050] Get the reduction formula;

[0051] Input the prediction result into the restoration formula to obtain the initial popularity prediction value corresponding to the topic;

[0052] The reduction formula is expressed as:

[0053]

[0054] in, hour, For the Damped cumulative forecast value for the day, For the Damped cumulative forecast value for the day, The restored The initial heat forecast value for the day;

[0055] Perform residual correction on the initial popularity prediction value to obtain the popularity prediction value corresponding to the topic in the second time period.

[0056] Optionally, the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows are obtained, the initial topic set is sorted according to the implicit interest score, and the topic with the highest implicit interest score in the initial topic set is selected. k The initial topics as supplementary preference topic sets include:

[0057] Analyze the relationship network based on user information;

[0058] Analyze the relationship network based on the user-topic knowledge graph to screen out a set of highly influential users who have influence on the user;

[0059] Obtaining historical blog posts of each user in the set of highly influential users to obtain a blog post set, performing text clustering on the blog post set to obtain a topic set, merging and removing duplicates from the topic sets of all users in the set of highly influential users to obtain an initial topic set;

[0060] The sentiment polarity analysis is performed on the comment text under each historical blog post to obtain the sentiment score. The interest degree is calculated based on the sentiment score, which is expressed as:

[0061]

[0062] in, Represents a user For high-impact users Published blog posts The sentiment score, Indicates the interaction intensity of this comment. Represents a user About the blog interest level;

[0063] Constructing an interest matrix based on the interest, and decomposing the interest matrix to obtain a user latent vector and a blog post latent vector;

[0064] Obtain the user's historical comment sequence and the corresponding commented blog post sequence to form a blog post comment pair;

[0065] Perform cross-attention calculation on the blog post comment pair to obtain a weighted comment vector and a weighted blog post vector, and perform weighted averaging on the weighted comment vector and the weighted blog post vector to obtain a fusion vector;

[0066] Select one of the blog posts in the collection as the blog post to be predicted, and calculate the cosine similarity between the embedding vectors of the blog post to be predicted and the blog post in the commented blog sequence as the attention weight;

[0067] Calculate an aggregate vector based on the fusion vector and the attention weight;

[0068] Obtaining a concatenated vector based on the aggregated vector, the user latent vector, and the blog post latent vector, and inputting the concatenated vector into a classifier to obtain an interest prediction score;

[0069] Get the topic of the blog post to be predicted;

[0070] The interest prediction scores of all blog posts under the same topic are weighted and summed to obtain the user's implicit interest score for the topic;

[0071] The initial topic set is sorted according to the implicit interest score, and the topic with the highest implicit interest score in the initial topic set is selected. k The initial topics are used as the supplementary preference topic set.

[0072] Optionally, calculating the user's comprehensive score for each topic in the candidate topic set based on the interaction preference value, the implicit interest score, and the popularity prediction value includes:

[0073] Obtain optimized interaction preference value weights, implicit interest score weights, and popularity prediction value weights;

[0074] According to the optimized interaction preference value weight, optimized interaction preference value, implicit interest score weight, implicit interest score, popularity prediction value weight and popularity prediction value, the user's comprehensive score for each topic in the candidate topic set is calculated and expressed as:

[0075]

[0076] in, To optimize the interaction preference value weight, is the implicit interest score weight, The weight of the heat prediction value, Represents a user and topic The optimized interaction preference value of Represents a user and topic The implicit interest score of Indicates topic The predicted value of heat.

[0077] A topic recommendation system based on knowledge modeling and popularity prediction, including:

[0078] A first acquisition module is used to acquire a social media topic library and user information, wherein the user information includes operation information and personal information;

[0079] A knowledge graph construction module is used to construct a user-topic knowledge graph based on the user information, the social media topic library and the preset ontology model;

[0080] An initialization module is used to initialize the initial embedding vector of each node and edge in the user-topic knowledge graph using a graph embedding model, wherein the nodes include user nodes and topic nodes;

[0081] A training module, configured to perform contrastive loss training on the initial embedding vector to obtain a first loss;

[0082] An aggregation module is used to aggregate nodes in the user-topic knowledge graph to obtain a node aggregation vector, wherein the node aggregation vector includes a user aggregation vector and a topic aggregation vector;

[0083] a calculation module, configured to calculate an interaction preference value based on the user aggregation vector and the topic aggregation vector, calculate a second loss based on the interaction preference value, sum the first loss and the second loss to obtain a total loss, and train the graph embedding model based on the total loss until the total loss is minimized, thereby obtaining an optimized graph embedding model;

[0084] an optimization module, configured to obtain an optimized user aggregation vector and an optimized topic aggregation vector based on the optimized graph embedding model and the user-topic knowledge graph, and obtain an optimized interaction preference value based on the optimized user aggregation vector and the optimized topic aggregation vector;

[0085] The second acquisition module is used to aggregate all user vectors into a user vector set, select any vector in the user vector set as the target vector, calculate the similarity between the target vector and other vectors in the user vector set, and select the vector with the highest similarity. k The users are taken as a collection of similar users, and the historical interaction topics of each user in the collection of similar users within a preset time period are obtained as a set of potential preferred topics;

[0086] The third acquisition module is used to obtain the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows, sort the initial topic set according to the implicit interest score, and select the top topic with the highest implicit interest score in the initial topic set. k The initial topics are used as the supplementary preference topic set, and the completed preference topic set is obtained according to the potential preference topic set and the supplementary preference topic set;

[0087] a prediction module configured to obtain a topic popularity index within a first preset time period from the social media topic library, filter a set of complementary preferred topics based on the topic popularity index to obtain a set of candidate topics, and predict a predicted popularity value corresponding to a topic within a second future time period based on a time series of topics in the candidate topic set;

[0088] The recommendation module is used to calculate the user's comprehensive score for each topic in the candidate topic set based on the optimized interaction preference value, implicit interest score and heat prediction value, sort each topic in the candidate topic set according to the comprehensive score, and select the most popular topic from the largest to the smallest. N topics as recommendation results.

[0089] A terminal device includes a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, a topic recommendation method based on interaction with highly influential people and popularity prediction is adopted.

[0090] A computer-readable storage medium stores a computer program. When the computer program is loaded and executed by a processor, a topic recommendation method based on interaction with highly influential people and popularity prediction is adopted.

[0091] The beneficial effects of the present invention are:

[0092] 1. Based on the social media topic library and user information, a user-topic knowledge graph is constructed, and the initial embedding vector is obtained by combining the graph embedding model. The first loss is obtained through comparative training, and the nodes in the user-topic knowledge graph are aggregated to obtain a node aggregation vector. The node aggregation vector includes a user aggregation vector and a topic aggregation vector, and the interaction preference value is calculated by the user aggregation vector and the topic aggregation vector. The second loss is calculated based on the interaction preference value. The first loss and the second loss are summed to obtain the total loss. The graph embedding model is trained according to the total loss until the total loss is minimized to obtain an optimized graph embedding model. The optimized graph embedding model is used to obtain the optimized user aggregation vector and the optimized topic aggregation vector. The topic aggregation vector is obtained, and the optimized interaction preference value is calculated to obtain the user's implicit interest score for the topic, and a potential preference topic set is obtained based on the implicit interest score. The topic heat index within the first preset time period is obtained from the social media topic library, and the potential preference topic set is screened according to the topic heat index to obtain a candidate topic set. The heat prediction value corresponding to the topic in the second time period in the future is predicted based on the time series of the topics in the candidate topic set. The user's comprehensive score for each topic in the candidate topic set is calculated based on the optimized interaction preference value, the implicit interest score and the heat prediction value. Each topic in the candidate topic set is sorted according to the comprehensive score, and the top topics are selected from the largest to the smallest. N Compared with the existing technology, this application predicts the topic popularity of topics in the future based on the existing topic popularity index and obtains the popularity prediction value, so as to accurately identify the topics that users are most likely to be interested in and have development potential, significantly improving the accuracy and foresight of recommendations.

[0093] 2. We leverage the weights of edges and the distances between nodes in the knowledge graph to capture the graph's structural information. We also leverage the attention mechanism and graph convolutional neural networks to effectively aggregate high-level semantic features and accurately identify complex relationships within the context of topic recommendations. During this process, we optimize the weight calculation of interaction edges, making the constructed knowledge graph more accurate, and thus the recommendation results, more precise.

[0094] 3. Analyze the relationship network based on the user-topic knowledge graph, screen out the set of high-influence users who have influence on users, and perform text aggregation on the historical blog posts of each user in the high-influence user set to obtain a topic set, and then merge and remove duplicates to obtain the initial topic set. Perform sentiment polarity analysis on the comment text under each historical blog post to obtain the sentiment score, calculate the interest degree based on the sentiment score, construct the interest degree matrix based on the interest degree, and decompose the interest degree matrix to obtain the user latent vector and the blog post latent vector, obtain the user's historical comment sequence and the corresponding commented blog post sequence to form a blog post comment pair; perform cross-attention calculation on the blog post comment pair to obtain the weighted comment vector and the weighted blog post vector, and add the weighted comment to the The weighted average of the weighted blog vector and the weighted blog vector is used to obtain the fusion vector; one of the blog posts in the blog post collection is selected as the blog post to be predicted, and the cosine similarity between the embedding vector of the blog post to be predicted and the blog post in the commented blog post sequence is calculated as the attention weight; the aggregation vector is calculated based on the fusion vector and the attention weight; the splicing vector is obtained based on the aggregation vector, the user latent vector and the blog post latent vector, and the splicing vector is input into the classifier to obtain the interest prediction score; the topic of the blog post to be predicted is obtained; the interest prediction scores of all blog posts under the same topic are weightedly summed to obtain the user's implicit interest score for the topic; the initial topic set is sorted according to the implicit interest score, and the top blog posts with the highest implicit interest score in the initial topic set are selected k Compared with the existing technology, this application calculates the implicit interest score to more comprehensively and accurately describe the user's deep interests, thereby significantly improving the accuracy and timeliness of topic recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 The figure is a flowchart of a topic recommendation method based on interaction with highly influential people and popularity prediction according to the present invention. DETAILED DESCRIPTION

[0096] A topic recommendation method based on interaction and popularity prediction of highly influential people, such as Figure 1 As shown, the present invention includes:

[0097] S1. Obtain social media topic database and user information, including operation information, personal information and social relationship information;

[0098] Specifically, the data in the social media topic database can be obtained from the Internet. When the system's set 24-hour timer expires, the crawlers deployed in the system begin working, crawling topic data from the designated network platform, including information such as topics, event descriptions, involved fields, and topic popularity indicators. If topic data already exists in the system, it will be updated. If data for topics that do not exist in the system is crawled, it will be newly created and stored to form the social media topic database.

[0099] User information includes personal information such as gender, occupation, date of birth, place of residence, and areas of interest provided by users during registration; historical topic interaction information, such as clicks, searches, likes, and comments (operation information); and information about users' social network relationships. Collected user data is collectively categorized as user information.

[0100] Social network information is used to identify users with high influence on the target user.

[0101] S2. Build a user-topic knowledge graph based on user information, social media topic library, and preset ontology model;

[0102] Based on user information, social media topic library and preset ontology model, the user-topic knowledge graph is constructed, including:

[0103] Use the ontology model to convert user information and topic information in the social media topic library into triples;

[0104] Specifically, data is preprocessed to obtain topic terms, event summaries, release dates, and popularity metrics from social media topic repositories. Simultaneously, basic user attributes (such as gender, occupation, age, and areas of interest) and historical topic interaction records (such as clicks, searches, likes, and comments, along with their timing) are extracted from user profiles. The extracted data is then cleaned to remove redundant, incomplete, or inconsistent records to ensure data quality. The collected data is then formatted uniformly, including time formatting, field naming, and value normalization.

[0105] Ontology is a core concept in knowledge representation and semantic technology, used to structure the knowledge framework within a domain. It builds a machine-understandable semantic model by clarifying concepts, attributes, relationships, and constraints.

[0106] Users and topics are treated as nodes in the user-topic knowledge graph. Each node carries attribute information. Edges between nodes are generated based on triples, and the weights of the edges are initialized.

[0107] Specifically, user information and topic information are converted into triple representations based on the semantic rules of the ontology model. During this conversion process, entity extraction is used to further enrich the semantic expression of the triples. Specifically, Named Entity Recognition (NER) technology is first applied to the topics and event descriptions to extract relevant entity information (such as location, person, time, etc.). For example, location entities (such as "Beijing" or "Shanghai") are extracted from the topic or event description, along with other possible entities (such as involved people or specific time). Each triple describes the semantic association in the data in the form of "subject-verb-object". For example, "User A searches for topic X" can be expressed as (User A, search, topic X). By generating a large number of triples, a knowledge network describing the relationship between users and topics is initially constructed.

[0108] First, users and topics are treated as nodes in the graph, with each node carrying corresponding attribute information. Second, edges between nodes are generated based on triples, and weights are initialized for the edges (for example, all edges are assigned a weight of 1).

[0109] Based on user information, obtain the interaction behavior between users and topics;

[0110] Specifically, during edge generation, if user A has multiple historical interactions with topic X (e.g., multiple searches, likes, reposts, and comments), the information from these interactions is combined and represented as a historical interaction edge. Subsequently, the weight of the historical interaction edge is optimized, taking into account factors such as search, click, repost, and like frequency, time decay, and the sentiment intensity of comments or posts, as follows:

[0111] According to the interaction behavior, historical interaction edges are generated, and the interaction weights of the historical interaction edges are calculated, which can be expressed as:

[0112]

[0113] in, Represents a user and topic The weight of the edge between It is a collection of user interaction behaviors, including search, click, forward, like and comment. For each interaction type The weight coefficient is used to adjust the influence of different behaviors. For users On topic Interaction behavior type the number of For users On topic The intensity of emotional polarity, , is the time attenuation factor, which controls the influence of time on the weight. Represents a user Last interaction topic The time difference from the current time, For users and topic Initialized weights of edges between them;

[0114] Construct a user-topic knowledge graph based on nodes, edge weights, interaction edges, and interaction weights.

[0115] Specifically, nodes and edges are organized into a graph data structure to form a network representation of the knowledge graph. This knowledge graph not only includes information such as hot topics and historical interaction data, but also covers semantic relationships and social network relationships between various topics, such as event relevance, person relationships, geographic location, and time dimension.

[0116] Because social media topic repositories and user information are regularly updated, the generated user-topic knowledge graph is also updated. Based on the social media topic repositories and user information, new triples are generated and added to the knowledge graph. Simultaneously, redundant or irrelevant nodes and edges in the graph are regularly checked and removed to ensure real-time and accuracy. Furthermore, to better reflect changes in user interests, edge weights in the knowledge graph are updated to match users' latest interaction preferences.

[0117] S3. Use the graph embedding model to initialize the initial embedding vectors of each node and edge in the user-topic knowledge graph, including user nodes and topic nodes;

[0118] Specifically, the embedding vectors of each node and edge in the knowledge graph are initialized by the TransR embedding algorithm. The TransR algorithm formula is as follows:

[0119]

[0120] in, Represents the triples in the social media topic library, 、 and Head entity , tail entity and relationships The embedding vector representation of for The transformation matrix of and from dimensional space is mapped to dimensional space, is the norm.

[0121] S4. Perform contrastive loss training on the initial embedding vector to obtain the first loss;

[0122] Specifically, the first loss is calculated as follows:

[0123]

[0124] in, , represents a randomly generated triple that does not exist in the "user-topic knowledge graph", represents the sigmoid activation function, Represents the tail entity in the triple that does not exist in the "user-topic knowledge graph".

[0125] S5. Aggregate the nodes in the user-topic knowledge graph to obtain a node aggregation vector, where the node aggregation vector includes a user aggregation vector and a topic aggregation vector;

[0126] Specifically, the attention mechanism and graph convolutional neural network are introduced to aggregate high-level semantic information in the knowledge graph to update the embedding representation of each node and edge. with it The information aggregation process of the first-order neighbor nodes is as follows:

[0127]

[0128]

[0129]

[0130]

[0131] in, is a trainable matrix used to extract effective features in the process of information propagation. Indicates aggregation Nodes with layer neighbor node information The embedding vector of Representation node No. The embedding vectors of the neighbor nodes of the layer, Representing relationships The weight of Representation node Through relationships How much information can be passed to the node .

[0132] S6. Calculate an interaction preference value based on the user aggregation vector and the topic aggregation vector, calculate a second loss based on the interaction preference value, sum the first loss and the second loss to obtain a total loss, and train the graph embedding model based on the total loss until the total loss is minimized, thereby obtaining an optimized graph embedding model.

[0133] The interaction preference value is calculated based on the user aggregation vector and the topic aggregation vector. The second loss is calculated based on the interaction preference value. The first loss and the second loss are summed to obtain the total loss. The graph embedding model is trained based on the total loss until the total loss is minimized. The optimized graph embedding model includes:

[0134] Specific, user-based Embedding vector of the node and topics Embedding vector of the node , predict the interaction preference value between users and topics.

[0135] The interaction preference value calculated based on the user aggregation vector and topic aggregation vector is expressed as:

[0136]

[0137] in, is the interaction preference value, Aggregate vectors for users, is the topic aggregation vector, T is the transposed matrix;

[0138] The second loss is calculated based on the interaction preference value and is expressed as:

[0139]

[0140] in, , Represents the user's real historical topic interaction record, and Represents the interaction set in which the user has not interacted, that is, the interaction set that is positively sampled and negatively sampled based on the user's historical topic interaction records. represents the sigmoid function, The topics in the interaction set that the user has not interacted with;

[0141] The total loss is obtained by summing the first loss and the second loss, which is expressed as:

[0142]

[0143] in, For the first loss, For the second loss, Represents the parameter set of the model, E refers to the embedding matrix of all entities and relationships in the knowledge graph, Representing relationships The weight of For the Layer attention information aggregation weight matrix, For the Layer attention information propagation weight matrix, For the Layer neighbor node information aggregation stage, is the set of all relationship types in the knowledge graph, is the regularization coefficient, and L is the maximum number of aggregation layers.

[0144] S7. Obtain an optimized user aggregation vector and an optimized topic aggregation vector based on the optimized graph embedding model and the user-topic knowledge graph, and obtain an optimized interaction preference value based on the optimized user aggregation vector and the optimized topic aggregation vector;

[0145] S8. All user aggregation vectors form a user vector collection, select any vector in the user vector collection as the target vector, calculate the similarity between the target vector and other vectors in the user vector collection, and select the vector with the highest similarity. k The users are taken as a similar user collection, and the historical interaction topics of each user in the similar user collection within a preset time period are obtained as a potential preference topic set.

[0146] Specifically, the similarity between the target vector and other vectors in the user vector collection is calculated as:

[0147]

[0148] in, and Represents users and users The embedding vector of For users , Represents a user .

[0149] S9. Obtain the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows, sort the initial topic set according to the implicit interest score, and select the top topic with the highest implicit interest score in the initial topic set. k The initial topics are used as the supplementary preference topic set, and the completed preference topic set is obtained according to the potential preference topic set and the supplementary preference topic set;

[0150] Obtain the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows, sort the initial topic set according to the implicit interest score, and select the top topic with the highest implicit interest score in the initial topic set. k The initial topics as supplementary preference topic sets include:

[0151] Analyze the relationship network based on user information;

[0152] Analyze the relationship network based on the user-topic knowledge graph and select the set of highly influential users who have an impact on users;

[0153] Specifically, based on the social relationship information in the user information, analyze the relationship network on the social platform , to identify their frequently interacted objects and follow lists. Based on the dynamically updated user-topic knowledge graph G , using high-influence user identification algorithms (such as those based on PageRank or a comprehensive scoring mechanism that introduces communication capabilities and interactive activity) to The users in the group can score and select a group of The set of highly influential users with significant influence is denoted as ,in, For the High-impact user.

[0154] Obtain the historical blog posts of each user in the high-influence user set to obtain a blog post set, perform text clustering on the blog post set to obtain a topic set, merge and remove duplicate topic sets of all users in the high-influence user set to obtain an initial topic set;

[0155] Specifically, for each high-influence user , collect its historical publication collection , and use topic modeling methods (such as LDA or text clustering) to classify its content into topics, and generate a set of topics involved by the user in the time window By aggregating the topic sets of all highly influential users and removing and merging them, we can form the initial topic set of the target user:

[0156]

[0157] in, is the initial topic set, For high-impact users The set of topics involved in the time window.

[0158] The sentiment polarity analysis is performed on the comment text under each historical blog post to obtain the sentiment score. The interest degree is calculated based on the sentiment score, which is expressed as:

[0159]

[0160] in, Represents a user For high-impact users Published blog posts The sentiment score, Indicates the interaction intensity of the comment, such as the normalized value of the number of likes and replies. Represents a user About the blog interest level;

[0161] Specifically, the BERT model is used to analyze each historical blog post. Comment text below Perform sentiment polarity analysis to get the sentiment of each comment user Sentiment score for the blog post .

[0162] Construct an interest matrix based on the interest level and decompose the interest matrix to obtain the user latent vector and blog post latent vector;

[0163] Specifically, the interest matrix is ​​expressed as:

[0164]

[0165] in, The number of all comment users, The number of historical posts by high-influence users.

[0166] FunkSVD is used to perform matrix decomposition on the comment user-blog post interest matrix M. The goal is to decompose it into the product of two low-rank latent vector matrices to explore the potential correlation structure between user interests and blog posts.

[0167]

[0168] In the above formula, , represents the user's latent vector matrix, each row Represents a user Representation in d-dimensional latent semantic space; , represents the latent vector matrix of the blog post, each row Indicates blog post Representation in d-dimensional latent semantic space.

[0169] The comment user-blog post interest matrix M is decomposed through matrix decomposition to extract the comment user latent vector V and the blog post latent vector P. This process can effectively capture the potential patterns of user interests and provide support for subsequent interest prediction.

[0170] Obtain the user's historical comment sequence and the corresponding commented blog post sequence to form a blog post comment pair;

[0171] Perform cross-attention calculation on the blog post and comment pairs to obtain weighted comment vectors and weighted blog post vectors, and perform weighted average of the weighted comment vectors and weighted blog post vectors to obtain a fusion vector;

[0172] Specifically, to further explore target users Based on the potential interests of users, a historical interaction modeling and semantic enhancement method based on attention mechanism is proposed. This method achieves fine modeling of historical interaction information and interest feature extraction by modeling the deep semantic association between user historical comments and the blog posts they commented on, thereby enhancing the ability to express users' potential interests. The historical comment sequence is: , the corresponding commented blog post sequence is: Each comment-blog pair are encoded as token sequence embeddings: , Subsequently, the cross attention weights are calculated using comments as queries and blog posts as keys, and blog posts as queries and comments as keys, and used to perform weighted summation on the representations on the other side to generate a fused context representation. and Finally, the context-enhanced representations of the comment side and the blog post side are weighted averaged to obtain the fused comment-blog post pair embedding vector:

[0173]

[0174] In the above formula, represents the comment side context representation generated by blog post guidance, Represents the blog post context representation generated through comment guidance.

[0175] Select one of the blog posts in the collection as the blog post to be predicted, and calculate the cosine similarity between the embedding vectors of the blog post to be predicted and the blog post in the commented blog sequence as the attention weight;

[0176] Specifically, for the focus and current forecast blog post The most relevant historical behavior, introducing its blog post embed . Calculate the current predicted blog post embedding separately The cosine similarity with each historical blog post embedding is used as the attention weight:

[0177]

[0178] in, Indicates the Blog post embed, Indicates the Blog post embed, Indicates the The attention weight of a historical blog post.

[0179] According to the fusion vector and attention weight, the aggregation vector is calculated;

[0180] Specifically, the aggregation vector is expressed as:

[0181]

[0182] According to the aggregation vector, user latent vector and blog post latent vector, a concatenated vector is obtained, which is input into the classifier to obtain the interest prediction score;

[0183] Specifically, for a specific prediction task, the latent vector of the target user, the latent vector of the blog post to be predicted, and the embedding vector of the comment-blog post pair related to the task are further fused, and the concatenated feature vector is input into the multi-layer perceptron (MLP) to predict the target user. Interest score for this blog post .

[0184] Finally, according to the target users For high-impact users The interest prediction score of the published blog posts is combined with the topic of each blog post, and the interest scores are aggregated by topic to obtain the target user For each initial topic set implicit interest score.

[0185] Get the topic of the blog post to be predicted;

[0186] The interest prediction scores of all blog posts under the same topic are weighted and summed to obtain the user's implicit interest score for the topic.

[0187] Specifically, each blog post Categorized under a topic , then the user Conversational Topic The "implicit interest score" is the weighted sum of the interest scores of all blog posts under the topic, recorded as:

[0188]

[0189] In the above formula, Represents a user About the blog interest, For blog posts The influence weight of the post (such as the normalized value of the number of reposts and likes).

[0190] The initial topic set is sorted according to the implicit interest score, and the topic with the highest implicit interest score in the initial topic set is selected. k The initial topics are used as the supplementary preference topic set.

[0191] The potential preference topic set and the supplementary preference topic set are merged to obtain the supplementary preference topic set.

[0192] S10, obtaining topic popularity indicators within a first preset time period from a social media topic database, screening a set of complementary preferred topics based on the topic popularity indicators to obtain a set of candidate topics, and predicting corresponding popularity prediction values ​​for topics within a second future time period based on a time series of topics in the candidate topic set;

[0193] Obtaining topic popularity indexes within a first preset time period from a social media topic database, screening a set of potential preferred topics based on the topic popularity indexes to obtain a set of candidate topics, and predicting corresponding popularity prediction values ​​for topics within a second time period in the future based on the time series of topics in the candidate topic set includes:

[0194] Specifically, read from the "topic database" The module contains recent topic popularity indicators (such as the Baidu Index) for topics, such as the past 5 days (the first preset time period). First, the module sets a popularity threshold. By taking a weighted sum of recent topic popularity indicators and comparing it with the popularity threshold, it screens out topics that meet the conditions and removes outdated or low-popularity topics to obtain a "candidate topic set." Based on the time series topic popularity data of each topic in the "candidate topic set," the module uses a gray prediction hybrid model to predict its future popularity changes, for example, over the next 2 days (the second time period).

[0195] The topic popularity index within the first preset time period is obtained from the social media topic database, and the topic popularity time series is constructed according to the chronological order, which is expressed as:

[0196]

[0197] in, For the Daily topic heat index;

[0198] According to the accumulation of topic popularity time series, the corresponding damped accumulation series of topic popularity is obtained, which is expressed as:

[0199]

[0200] in, For the The accumulated damping value of the day, is the damping accumulation parameter, , For the sky, For the The topic popularity index of the day;

[0201] According to the damped cumulative sequence and the whitening equation of the grey prediction model, the prediction results are obtained;

[0202] The prediction results are restored through the inverse accumulation generation operator to obtain the topic popularity prediction sequence as the popularity prediction value corresponding to the topic in the second time period.

[0203] The whitening equation is expressed as:

[0204]

[0205] in, and are the model parameters estimated by the least squares method, is the first item in the time series and is the indicator farthest from the current time. For the predicted The damped cumulative forecast value of the day, that is, the forecast result.

[0206] The prediction results are restored by the inverse accumulation generation operator to obtain the topic popularity prediction sequence as the popularity prediction value corresponding to the topic in the second time period, including:

[0207] Get the reduction formula;

[0208] Input the prediction results into the reduction formula to obtain the initial heat prediction value corresponding to the topic;

[0209] The reduction formula is expressed as:

[0210]

[0211] in, hour, For the Damped cumulative forecast value for the day, For the Damped cumulative forecast value for the day, The restored The initial heat forecast value for the day;

[0212] Perform residual correction on the initial heat prediction value to obtain the heat prediction value corresponding to the topic in the second time period.

[0213] Specifically, in order to improve the adaptability of the prediction model, a metabolic mechanism is introduced to dynamically update model parameters and realize the replacement of old and new topic popularity data. The specific implementation method is as follows: the latest prediction results are incorporated into the historical data through a rolling window, and the data within the window is updated. In each round of prediction, the model is trained using the latest historical data, and the damping parameters are adjusted using the particle swarm algorithm. To improve the model's adaptability to new data.

[0214] The error back-propagation neural network is used to correct the residuals of the gray prediction model's predictions to improve the accuracy of the predictions. The corrected prediction sequence more accurately reflects the changing trends in the topic's popularity. Finally, the average of the predictions for the next two days in the prediction sequence is taken as the predicted popularity value for the corresponding topic.

[0215] S11. Calculate the user's comprehensive score for each topic in the candidate topic set based on the optimized interaction preference value, implicit interest score, and heat prediction value, and sort each topic in the candidate topic set based on the comprehensive score, selecting the top topics from largest to smallest. N topics as recommendation results.

[0216] Based on the optimized interaction preference value, implicit interest score, and popularity prediction value, the user's comprehensive score for each topic in the candidate topic set is calculated, including:

[0217] Obtain optimized interaction preference value weights, implicit interest score weights, and popularity prediction value weights;

[0218] According to the optimized interaction preference value weight, optimized interaction preference value, implicit interest score weight, implicit interest score, popularity prediction value weight and popularity prediction value, the user's comprehensive score for each topic in the candidate topic set is calculated and expressed as:

[0219]

[0220] in, To optimize the interaction preference value weight, is the implicit interest score weight, The weight of the heat prediction value, Represents a user and topic The optimized interaction preference value of Represents a user and topic The implicit interest score of Indicates topic The predicted value of heat.

[0221] Finally, according to the comprehensive score, select the top TOP- Ntopics to form a personalized hot topic recommendation list for target users.

[0222] A topic recommendation system based on knowledge modeling and popularity prediction, including:

[0223] The first acquisition module is used to obtain the social media topic library and user information, where the user information includes operation information and personal information;

[0224] The knowledge graph construction module is used to build a user-topic knowledge graph based on user information, social media topic library and preset ontology model;

[0225] The initialization module is used to initialize the initial embedding vectors of each node and edge in the user-topic knowledge graph using the graph embedding model. The nodes include user nodes and topic nodes.

[0226] The training module is used to perform contrastive loss training on the initial embedding vector to obtain the first loss;

[0227] The aggregation module is used to aggregate the nodes in the user-topic knowledge graph to obtain the node aggregation vector, which includes the user aggregation vector and the topic aggregation vector;

[0228] A calculation module is used to calculate an interaction preference value based on the user aggregation vector and the topic aggregation vector, calculate a second loss based on the interaction preference value, sum the first loss and the second loss to obtain a total loss, and train the graph embedding model based on the total loss until the total loss is minimized to obtain an optimized graph embedding model;

[0229] An optimization module is used to obtain an optimized user aggregation vector and an optimized topic aggregation vector based on the optimized graph embedding model and the user-topic knowledge graph, and to obtain an optimized interaction preference value based on the optimized user aggregation vector and the optimized topic aggregation vector;

[0230] The second acquisition module is used to aggregate all user vectors into a user vector set, select any vector in the user vector set as the target vector, calculate the similarity between the target vector and other vectors in the user vector set, and select the vector with the highest similarity. k The users are taken as a collection of similar users, and the historical interaction topics of each user in the collection of similar users within a preset time period are obtained as a set of potential preferred topics;

[0231] The third acquisition module is used to obtain the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows, sort the initial topic set according to the implicit interest score, and select the top topic with the highest implicit interest score in the initial topic set. kThe initial topics are used as the supplementary preference topic set, and the completed preference topic set is obtained according to the potential preference topic set and the supplementary preference topic set;

[0232] A prediction module is configured to obtain a topic popularity index within a first preset time period from a social media topic library, filter a set of completion preference topics based on the topic popularity index to obtain a set of candidate topics, and predict a predicted popularity value corresponding to a topic within a second future time period based on a time series of topics in the candidate topic set;

[0233] The recommendation module is used to calculate the user's comprehensive score for each topic in the candidate topic set based on the optimized interaction preference value, implicit interest score and heat prediction value, and sort each topic in the candidate topic set according to the comprehensive score, selecting the top topics from largest to smallest. N topics as recommendation results.

[0234] An embodiment of the present application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, it adopts a topic recommendation method based on interaction with highly influential people and popularity prediction.

[0235] Among them, the terminal device can be a computer device such as a desktop computer, a laptop computer or a cloud server, and the terminal device includes but is not limited to a processor and a memory. For example, the terminal device can also include input and output devices, network access devices and buses, etc.

[0236] Among them, the processor can adopt a central processing unit (CPU). Of course, according to actual usage, other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be adopted. The general-purpose processor can adopt a microprocessor or any conventional processor, etc., and this application does not impose any restrictions on this.

[0237] Among them, the memory can be an internal storage unit of the terminal device, such as the hard disk or memory of the terminal device, or it can be an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD) or flash memory card (FC) equipped on the terminal device, etc., and the memory can also be a combination of the internal storage unit and the external storage device of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or is to be output. This application does not impose any restrictions on this.

[0238] Among them, through this terminal device, a topic recommendation method based on interaction with highly influential people and popularity prediction in the above embodiment is stored in the memory of the terminal device, and is loaded and executed on the processor of the terminal device for easy use.

[0239] An embodiment of the present application further discloses a computer-readable storage medium, and the computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, a topic recommendation method based on interaction with highly influential people and popularity prediction in the above embodiment is adopted.

[0240] Among them, the computer program can be stored in a computer-readable medium, the computer program includes computer program code, the computer program code can be in the form of source code, object code, executable file or certain middleware, etc. The computer-readable medium includes any entity or device that can carry computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that computer-readable medium includes but is not limited to the above-mentioned components.

[0241] Among them, through this computer-readable storage medium, a topic recommendation method based on interaction with highly influential people and popularity prediction in the above embodiment is stored in a computer-readable storage medium, and is loaded and executed on a processor to facilitate the storage and application of the above method.

[0242] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as above, which are not provided in detail for the sake of simplicity.

[0243] The one or more embodiments of this application are intended to encompass all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of this application should be included in the scope of protection of this application.

Claims

1. A topic recommendation method based on interaction and popularity prediction of highly influential people, characterized by: include: Obtaining a social media topic library and user information, wherein the user information includes operation information, personal information, and social relationship information; Constructing a user-topic knowledge graph based on the user information, social media topic library, and a preset ontology model; Initialize the initial embedding vector of each node and edge in the user-topic knowledge graph using the graph embedding model, where the nodes include user nodes and topic nodes; Performing contrastive loss training on the initial embedding vector to obtain a first loss; Aggregate the nodes in the user-topic knowledge graph to obtain a node aggregation vector, where the node aggregation vector includes a user aggregation vector and a topic aggregation vector; Calculating an interaction preference value based on the user aggregation vector and the topic aggregation vector, calculating a second loss based on the interaction preference value, summing the first loss and the second loss to obtain a total loss, and training the graph embedding model based on the total loss until the total loss is minimized, thereby obtaining an optimized graph embedding model; Obtaining an optimized user aggregation vector and an optimized topic aggregation vector based on the optimized graph embedding model and the user-topic knowledge graph, and obtaining an optimized interaction preference value based on the optimized user aggregation vector and the optimized topic aggregation vector; All user aggregation vectors form a user vector collection, select any vector in the user vector collection as the target vector, calculate the similarity between the target vector and other vectors in the user vector collection, and select the vector with the highest similarity. k The users are taken as a collection of similar users, and the historical interaction topics of each user in the collection of similar users within a preset time period are obtained as a set of potential preferred topics; Obtain the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows, sort the initial topic set according to the implicit interest score, and select the top topic with the highest implicit interest score in the initial topic set. k The initial topics are used as the supplementary preference topic set, and the completed preference topic set is obtained according to the potential preference topic set and the supplementary preference topic set; Obtaining topic popularity indicators within a first preset time period from the social media topic library, screening a set of complementary preferred topics based on the topic popularity indicators to obtain a set of candidate topics, and predicting corresponding popularity prediction values ​​of topics within a second future time period based on a time series of topics in the candidate topic set; Based on the optimized interaction preference value, implicit interest score and popularity prediction value, the user's comprehensive score for each topic in the candidate topic set is calculated, each topic in the candidate topic set is sorted according to the comprehensive score, and the top N topics are selected from large to small as recommendation results.

2. The topic recommendation method based on high-influence person interaction and popularity prediction according to claim 1, characterized in that: The constructing of a user-topic knowledge graph based on the user information, the social media topic library, and the preset ontology model includes: Using the ontology model, user information and topic information in the social media topic library are converted into triples; Users and topics are used as nodes in the user-topic knowledge graph. Each node carries attribute information. Edges between nodes are generated based on the triples and the weights of the edges are initialized. Obtaining interaction behavior between the user and the topic based on the user information; According to the interaction behavior, a historical interaction edge is generated, and the interaction weight of the historical interaction edge is calculated, which is expressed as: in, Represents a user and topic The weight of the edge between It is a collection of user interaction behaviors, including search, click, forward, like and comment. For each interaction type The weight coefficient is used to adjust the influence of different behaviors. For users On topic Interaction behavior type the number of For users On topic The intensity of emotional polarity, , is the time attenuation factor, which controls the influence of time on the weight. Represents a user Last interaction topic The time difference from the current time, For users and topic Initialized weights of edges between them; Construct a user-topic knowledge graph based on nodes, edge weights, interaction edges, and interaction weights.

3. The topic recommendation method based on interaction and popularity prediction of highly influential people according to claim 1, characterized in that: The step of calculating an interaction preference value based on the user aggregation vector and the topic aggregation vector, calculating a second loss based on the interaction preference value, summing the first loss and the second loss to obtain a total loss, and training the graph embedding model based on the total loss until the total loss is minimized to obtain an optimized graph embedding model includes: The interaction preference value calculated based on the user aggregation vector and the topic aggregation vector is expressed as: in, is the interaction preference value, Aggregate vectors for users, is the topic aggregation vector, T is the transposed matrix; The second loss is calculated according to the interaction preference value and is expressed as: in, , Represents the user's real historical topic interaction record, and Represents the interaction set in which the user has not interacted, that is, the interaction set that is positively sampled and negatively sampled based on the user's historical topic interaction records. represents the sigmoid function, The topics in the interaction set that the user has not interacted with; The first loss and the second loss are summed to obtain the total loss, which is expressed as: in, For the first loss, For the second loss, Represents the parameter set of the model, E refers to the embedding matrix of all entities and relationships in the knowledge graph, for The transformation matrix, is the embedding vector representation of relation r, For the Layer attention information aggregation weight matrix, For the Layer attention information propagation weight matrix, For the Layer neighbor node information aggregation stage, is the set of all relationship types in the knowledge graph, is the regularization coefficient, L is the maximum number of layers of aggregation; The graph embedding model is trained according to the total loss until the total loss is minimized, thereby obtaining an optimized graph embedding model.

4. The topic recommendation method based on high-influence person interaction and popularity prediction according to claim 1, characterized in that: Obtaining a topic popularity index within a first preset time period from the social media topic database, screening a potential preferred topic set based on the topic popularity index to obtain a candidate topic set, and predicting a popularity prediction value corresponding to a topic within a second time period in the future based on a time series of topics in the candidate topic set includes: The topic popularity index within the first preset time period is obtained from the social media topic library, and a topic popularity time series is constructed according to the time sequence, which is expressed as: in, For the The topic popularity index of the day; According to the topic popularity time series, the damped cumulative series of the corresponding topic popularity is obtained, which is expressed as: in, For the The accumulated damping value of the day, is the damping accumulation parameter, , For the sky, For the The topic popularity index of the day; Obtaining a prediction result according to the damped cumulative sequence and the whitening equation of the grey prediction model; The prediction results are restored by an inverse accumulation generation operator to obtain a topic popularity prediction sequence as the popularity prediction value corresponding to the topic in the second time period; The whitening equation is expressed as: in, and are the model parameters estimated by the least squares method, is the first item in the time series and is the indicator farthest from the current time. For the predicted The damped cumulative forecast value of the day, that is, the forecast result.

5. The topic recommendation method based on high-influence person interaction and popularity prediction as claimed in claim 4 is characterized in that: The step of restoring the prediction result by an inverse accumulation generation operator to obtain a topic popularity prediction sequence as a popularity prediction value corresponding to the topic in the second time period includes: Get the reduction formula; Input the prediction result into the restoration formula to obtain the initial popularity prediction value corresponding to the topic; The reduction formula is expressed as: in, hour, For the Damped cumulative forecast value for the day, For the Damped cumulative forecast value for the day, The restored The initial heat forecast value for the day; Perform residual correction on the initial popularity prediction value to obtain the popularity prediction value corresponding to the topic in the second time period.

6. The topic recommendation method based on high-influence person interaction and popularity prediction according to claim 1, characterized in that: The method obtains the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows, sorts the initial topic set according to the implicit interest score, and selects the top topic with the highest implicit interest score in the initial topic set. k The initial topics as supplementary preference topic sets include: Analyze the relationship network based on user information; Analyze the relationship network based on the user-topic knowledge graph to screen out a set of highly influential users who have influence on the user; Obtaining historical blog posts of each user in the set of highly influential users to obtain a blog post set, performing text clustering on the blog post set to obtain a topic set, merging and removing duplicates from the topic sets of all users in the set of highly influential users to obtain an initial topic set; Perform sentiment polarity analysis on the comment text under each historical blog post to obtain the sentiment score. Calculate the interest level based on the sentiment score, which is expressed as: in, Represents a user For high-impact users Published blog posts The sentiment score, Indicates the interaction intensity of the comment text. Represents a user About the blog interest level; Constructing an interest matrix based on the interest, and decomposing the interest matrix to obtain a user latent vector and a blog post latent vector; Obtain the user's historical comment sequence and the corresponding commented blog post sequence to form a blog post comment pair; Perform cross-attention calculation on the blog post comment pair to obtain a weighted comment vector and a weighted blog post vector, and perform weighted averaging on the weighted comment vector and the weighted blog post vector to obtain a fusion vector; Select one of the blog posts in the collection as the blog post to be predicted, and calculate the cosine similarity between the embedding vectors of the blog post to be predicted and the blog post in the commented blog sequence as the attention weight; Calculate an aggregate vector based on the fusion vector and the attention weight; Obtaining a concatenated vector based on the aggregated vector, the user latent vector, and the blog post latent vector, and inputting the concatenated vector into a classifier to obtain an interest prediction score; Get the topic of the blog post to be predicted; The interest prediction scores of all blog posts under the same topic are weighted and summed to obtain the user's implicit interest score for the topic; The initial topic set is sorted according to the implicit interest score, and the topic with the highest implicit interest score in the initial topic set is selected. k The initial topics are used as the supplementary preference topic set.

7. The topic recommendation method based on high-influence person interaction and popularity prediction according to claim 1 is characterized in that: The step of calculating the user's comprehensive score for each topic in the candidate topic set based on the optimized interaction preference value, the implicit interest score, and the heat prediction value includes: Obtain optimized interaction preference value weights, implicit interest score weights, and popularity prediction value weights; According to the optimized interaction preference value weight, optimized interaction preference value, implicit interest score weight, implicit interest score, popularity prediction value weight and popularity prediction value, the user's comprehensive score for each topic in the candidate topic set is calculated and expressed as: in, To optimize the interaction preference value weight, is the implicit interest score weight, The weight of the heat prediction value, Represents a user and topic The optimized interaction preference value of Represents a user and topic The implicit interest score of Indicates topic The predicted value of heat.

8. A topic recommendation system based on knowledge modeling and popularity prediction, characterized by: include: A first acquisition module is used to acquire a social media topic library and user information, wherein the user information includes operation information and personal information; A knowledge graph construction module is used to construct a user-topic knowledge graph based on the user information, the social media topic library and the preset ontology model; An initialization module is used to initialize the initial embedding vector of each node and edge in the user-topic knowledge graph using a graph embedding model, wherein the nodes include user nodes and topic nodes; A training module, configured to perform contrastive loss training on the initial embedding vector to obtain a first loss; An aggregation module is used to aggregate nodes in the user-topic knowledge graph to obtain a node aggregation vector, wherein the node aggregation vector includes a user aggregation vector and a topic aggregation vector; a calculation module, configured to calculate an interaction preference value based on the user aggregation vector and the topic aggregation vector, calculate a second loss based on the interaction preference value, sum the first loss and the second loss to obtain a total loss, and train the graph embedding model based on the total loss until the total loss is minimized, thereby obtaining an optimized graph embedding model; an optimization module, configured to obtain an optimized user aggregation vector and an optimized topic aggregation vector based on the optimized graph embedding model and the user-topic knowledge graph, and obtain an optimized interaction preference value based on the optimized user aggregation vector and the optimized topic aggregation vector; The second acquisition module is used to aggregate all user vectors into a user vector set, select any vector in the user vector set as the target vector, calculate the similarity between the target vector and other vectors in the user vector set, and select the vector with the highest similarity. k The users are taken as a collection of similar users, and the historical interaction topics of each user in the collection of similar users within a preset time period are obtained as a set of potential preferred topics; The third acquisition module is used to obtain the user's implicit interest score for the topic and the initial topic set involved in the posts of the highly influential people the user follows, sort the initial topic set according to the implicit interest score, and select the top topic with the highest implicit interest score in the initial topic set. k The initial topics are used as the supplementary preference topic set, and the completed preference topic set is obtained according to the potential preference topic set and the supplementary preference topic set; a prediction module configured to obtain a topic popularity index within a first preset time period from the social media topic library, filter a set of complementary preferred topics based on the topic popularity index to obtain a set of candidate topics, and predict a predicted popularity value corresponding to a topic within a second future time period based on a time series of topics in the candidate topic set; The recommendation module is used to calculate the user's comprehensive score for each topic in the candidate topic set based on the optimized interaction preference value, implicit interest score and popularity prediction value, sort each topic in the candidate topic set according to the comprehensive score, and select the top N topics from large to small as the recommendation results.

9. A terminal device comprising a memory and a processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, the method according to any one of claims 1 to 7 is adopted.

10. A computer-readable storage medium storing a computer program, wherein: When the computer program is loaded and executed by a processor, the method according to any one of claims 1 to 7 is adopted.

Images