Content push optimization method and system for analyzing user interests

Abstract

The invention discloses a content push optimization method and system for analyzing user interests, and relates to the technical field of artificial intelligence and personalized recommendation, and the method comprises the following steps: obtaining user behavior data, constructing user interest vectors, and carrying out the weighted fusion to generate a current interest portrait; vectorizing the to-be-pushed content, and performing similarity matching on the to-be-pushed content and the current interested portrait to obtain a recommendation sorting score; according to the recommendation sorting score, in combination with the user activeness and the push saturation, determining the push content and sequence; extracting user behavior path clustering and content similarity according to the user feedback data, outputting a response estimation result, and dynamically adjusting a content recommendation priority according to the response estimation result; according to the method, the problem of inaccurate content recommendation is solved through the multi-modal fusion and feedback optimization method.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and personalized recommendation, and more specifically, to a content push optimization method and system for analyzing user interests. Background Art

[0002] With the rapid development of Internet technology and the widespread popularity of smart terminals, the digital information that users come into contact with every day has increased exponentially, and the problem of information overload has become increasingly serious. In this context, personalized content push technology has gradually become an important means to improve user experience, enhance platform stickiness, and improve content distribution efficiency. The personalized push system analyzes users' behavioral data, interest preferences, and historical interaction information to achieve personalized content recommendations for each user, and has been widely used in many industries such as news and information, e-commerce, social media, and short video platforms.

[0003] Traditional recommendation systems mainly rely on collaborative filtering algorithms and content matching mechanisms to establish associations based on the similarity between users and content, and infer the content that users may be interested in. However, with the complexity of user behavior and the diversification of content forms, traditional methods have gradually exposed many deficiencies. First, the collaborative filtering method has limited effectiveness in cold start scenarios and is difficult to effectively recommend new users or new content; second, static content matching ignores the dynamic changes of user interests and context factors, and is difficult to adapt to the preference migration brought about by the change of user behavior over time; in addition, some current systems lack accurate identification of users' reception ability, active time periods, and information saturation status during the content sorting process, which easily leads to redundant pushes or interference with the user experience.

[0004] Therefore, there is an urgent need for a more intelligent, efficient, and personalized content push optimization method and system that can integrate various technical means such as multi-modal content features, dynamic user behavior, time series trends, and preference clustering to achieve in-depth modeling of user interests and adaptive optimization of push strategies. This method should not only have accurate interest recognition and content matching capabilities, but also have the ability to analyze users' activity levels and content saturation levels, so as to dynamically adjust the push order and frequency, and improve the pertinence, timeliness, and user acceptance of the push.

[0005] For example, a method and system for intelligent recommendation of user interests based on the Internet disclosed in the invention patent with the publication number CN119128253A relates to the field of Internet technologies. The system forms a multi-source data set by collecting and integrating user behavior data on different platforms in real time; using natural language processing and time series analysis technologies, the system deeply mines user interests and interest change trends, and identifies potential sudden interest increase trends; through a clustering algorithm, the system divides users with similar interest characteristics into dynamic user interest clusters, and dynamically adjusts the parameters of the user interest clusters according to instantaneous data; the system predicts the user's future needs through a deep learning model, combines the time series analysis model and the characteristics of the user interest clusters, generates personalized recommendation content, and pushes it to the user through multiple channels. It improves the accurate push of relevant content when the user's interest characteristics suddenly change.

[0006] For example, a content recommendation method and system based on semantic recognition disclosed in the invention patent with the publication number CN119089398A parses the user's text, speech or video input through natural language processing technology to generate user semantic representation data; combines the user's historical behavior data to dynamically construct a user interest graph, and performs node and edge representation learning through a multi-layer graph convolutional network to generate time series user interest graph representation data; uses the self-attention mechanism and multi-head attention mechanism of the Transformer model, combines the contrast learning module to optimize the matching degree between the user interest representation and the recommended content, and generates personalized recommendation content data; collects the user's multi-modal information, fuses it with the personalized recommendation content data, and generates optimized multi-modal recommendation content data through a multi-modal fusion algorithm; collects user feedback feature data, and dynamically adjusts and optimizes the generation strategy of the recommended content through an adaptive learning algorithm. It realizes the efficient, accurate and personalized recommendation of the recommendation system and improves the user experience.

[0007] In the above disclosed technical solutions, there are at least the following technical problems: Traditional recommendation systems often rely on static user tags or simple click behaviors, and cannot dynamically depict the multi-dimensional and cross-time interest changes of users, which easily leads to deviation or invalidation of the recommendation results. Moreover, most recommendation systems are only based on text content or historical click frequencies, and it is difficult to effectively process multi-modal content (such as pictures and texts, videos) and semantic matching differences, resulting in insufficient relevance of the recommended content.

[0008] In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0009] In order to overcome the above defects of the prior art, an embodiment of the present invention provides a content push optimization method and system for analyzing user interests, which solves the problem of inaccurate content recommendation through a multi-modal fusion and feedback optimization method.

[0010] To achieve the above object, the present invention provides the following technical solutions: A content push optimization method for analyzing user interests, comprising the following steps: obtaining user behavior data, constructing a user interest vector, and weighted fusion to generate a current interest portrait; vectorizing the content to be pushed, performing similarity matching with the current interest portrait to obtain a recommended sorting score; determining the pushed content and order according to the recommended sorting score in combination with user activity and push saturation; extracting user behavior path clustering and content similarity from user feedback data, outputting a response prediction result, and dynamically adjusting the content recommendation priority according to the response prediction result.

[0011] In a preferred embodiment, the obtaining user behavior data, constructing a user interest vector, and weighted fusion to generate a current interest portrait are specifically as follows: obtaining user behavior data within the past T hours to form a time series behavior set, weighted encoding the time series behavior set through an attention mechanism to generate a short-term interest vector with a fixed dimension; obtaining user long-term historical behavior data, extracting the topic tags, semantic classifications, and emotional tendencies of the corresponding content, and counting their occurrence frequencies to form a statistical vector; constructing a triple graph structure based on knowledge graph technology, performing node representation learning on the graph structure based on a graph neural network to obtain the embedding features of the user in the graph, and extracting the embedding vector corresponding to the user node; splicing the statistical vector and the embedding vector, and performing a non-linear mapping to output a long-term interest vector with a unified dimension; performing weighted fusion on the short-term interest vector and the long-term interest vector to obtain the current interest portrait.

[0012] In a preferred embodiment, the vectorizing the content to be pushed is specifically as follows: obtaining the content to be pushed, and extracting multi-modal feature information of the content to be pushed; selecting a corresponding encoding model for vectorizing encoding according to the type of multi-modal feature information to obtain a number of sub-feature vectors; mapping the sub-feature vectors based on a fully connected mapping method to obtain a content feature vector, and the content feature vector has the same structural dimension as the user interest vector.

[0013] In a preferred embodiment, the performing similarity matching with the current interest portrait to obtain a recommended sorting score is specifically as follows: obtaining the cosine similarity of the current interest portrait matching each content feature vector based on the cosine similarity method; normalizing all the matched cosine similarity values, and normalizing the cosine similarity values based on linear scaling to obtain the recommended sorting score.

[0014] In a preferred embodiment, the push content and order are determined by combining the recommended sorting score with the user activity and push saturation as follows: Based on the cosine similarity calculation method, the recommended score of each candidate content is obtained to form an initial recommended sorting list; The historical behavior frequency of the user is obtained, and based on the big data clustering analysis method, the active patterns of the user in different time periods are autonomously summarized to construct an activity model; The amount of content pushed and the response volume received by the user within the current cycle are obtained. Combining with the historical optimal push frequency, the information saturation state of the user under different content types is identified through data-driven modeling technology, and a user push saturation model is constructed; The push saturation is obtained according to the user push saturation model; According to the activity model and the push saturation model, the initial recommended sorting list is dynamically adjusted.

[0015] In a preferred embodiment, the user behavior path clustering and content similarity are extracted from the user feedback data, the response prediction result is output, and the content recommendation priority is dynamically adjusted according to the response prediction result as follows: The user feedback data is obtained, and the feedback data is indexed by the user ID and the content ID as a joint index, and a feedback sample set is constructed by recording in chronological order; The feedback sample set is processed by a sliding time window with a fixed time step, and the occurrence frequency and behavior change trend of each behavior type within each window are obtained to obtain the change characteristics; The behavior event sequence within each time window is obtained, and combined with the change characteristics, the logical association between adjacent behavior events is identified to form a behavior path sequence; The similarity calculation and clustering analysis are performed on the behavior paths to obtain a set of preference paths for user content; The user behavior paths associated with the current recommended content in history are obtained, and the response data of the current recommended content in the preference paths is obtained; If the recommended content is a first-time recommended content, the historical content with a similar structure to the recommended content is used, and the response data of the similar content in the similar paths is obtained for approximate prediction to obtain the response prediction result; According to the response prediction result of the recommended content in the user preference path, the recommendation list is dynamically adjusted.

[0016] In a preferred embodiment, the similarity calculation and clustering analysis are performed on the behavior paths to obtain a set of preference paths for user content as follows: Based on the continuous interaction behavior of the user, the content access sequence is extracted in chronological order; The behavior paths are constructed according to the content access sequence to obtain an initial set of behavior paths; Each content node in each path is vectorized based on the content embedding model, and weighted fusion is performed in combination with the behavior type and behavior intensity to generate a path vector representation; Based on the cosine similarity algorithm, the behavior path vectors of different users are calculated pairwise to obtain a path similarity matrix; According to the path similarity matrix, the path vector set is classified by the K-means clustering algorithm to obtain a set of preference paths.

[0017] In a preferred embodiment, the response data of the obtained similar content in the similar path is approximately estimated to obtain a response estimation result, which is specifically as follows: perform a unique identifier retrieval for each recommended content item to determine whether it is the content that has appeared in the historical interaction. If the content has not appeared in any user's behavior path, the system marks it as the first recommended content; extract the structural features of the first recommended content, and generate a content embedding vector through the deployed content encoding network; calculate the similarity between the embedding vector and all historical content vectors in the content database, and select the top K most similar historical content items; in the set of preference paths, filter out the path samples containing the top K most similar historical content items, and construct a set of similar paths; obtain the user response data of the corresponding content of each similar path; normalize the user response data and perform weighted averaging in the path set to generate an approximate response estimation value, and obtain the response estimation result.

[0018] A system for a content push optimization method for analyzing user interests includes a data acquisition module, a matching module, a push module, and a feedback correction module, and there are connections between the modules; the data acquisition module is used to obtain user behavior data, construct a user interest vector, and perform weighted fusion to generate the current interest portrait; the matching module is used to vectorize the content to be pushed, perform similarity matching with the current interest portrait, and obtain a recommended sorting score; the push module is used to determine the pushed content and order according to the recommended sorting score in combination with the user activity and the push saturation; the feedback correction module is used to extract the user behavior path clustering and content similarity according to the user feedback data, output the response estimation result, and dynamically adjust the content recommendation priority according to the response estimation result.

[0019] The technical effects and advantages of the content push optimization method and system for analyzing user interests according to the present invention: 1. The present invention realizes the dynamic characterization of user interests by constructing a user interest vector and performing fusion to generate the current interest portrait. Specifically, by obtaining the behavior sequence of the user in the past short time, introducing the attention mechanism to extract short-term interests, and at the same time combining long-term historical behaviors and node embeddings based on the knowledge graph to extract the user's long-term interests, and adopting a weighted fusion strategy to generate a multi-dimensional interest portrait. This method overcomes the limitations of traditional methods based on tags or static portraits, and has the advantages of strong real-time performance, high personalization degree, and strong representation ability, and can more accurately reflect the true interest preferences of users in a specific time period.

[0020] 2. The present invention constructs a multi-objective optimization function through a feedback correction module, enhancing the self-learning ability of the recommendation system. The system not only records the user's response behaviors (such as clicks, browsing, comments, etc.), but also mines the user's preference paths based on the behavior event sequence, introduces a similar path reasoning mechanism for the first recommended content, and realizes the dynamic evaluation of the estimated recommendation effect in scenarios lacking historical data, thereby ensuring that the recommendation strategy has forward-looking and generalization capabilities. This mechanism improves the coverage and rationality of new content recommendations and solves the cold start problem in the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic flowchart of a method for optimizing content push for analyzing user interests according to the present invention.

[0022] Figure 2 It is a schematic structural diagram of a system for a method for optimizing content push for analyzing user interests according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] Embodiment 1 Figure 1 A method for optimizing content push for analyzing user interests according to the present invention is given, including the following steps: S1. Obtain user behavior data, construct a user interest vector, and generate a current interest portrait through weighted fusion.

[0025] In this embodiment, obtaining user behavior data, constructing a user interest vector, and generating a current interest portrait through weighted fusion are specifically as follows: Obtain the user behavior data within the past T hours (such as the most recent 24 hours) to form a time series behavior set, and the behavior data includes click behavior, browsing duration, search keywords, favorites, forwards, and comment information; Perform weighted encoding on the time series behavior set through a time-aware attention mechanism to generate a short-term interest vector with a fixed dimension (such as 128-dimensional or 256-dimensional) to reflect the user's current interest focus; Obtain the user's long-term historical behavior data (such as the recent 1 month to 3 months), extract the topic tags, semantic classifications, and emotional tendencies of the corresponding content, and count their occurrence frequencies to form an initial interest distribution feature, that is, a statistical vector; Construct a user-content-tag tripartite graph structure based on knowledge graph technology, and perform node representation learning on the graph structure based on a graph neural network to obtain the embedding features of users in the graph, and extract the embedding vectors corresponding to user nodes; Concatenate the statistical vector and the embedding vector, and perform a non-linear mapping to output a long-term interest vector of a unified dimension; Perform weighted fusion on the short-term interest vector and the long-term interest vector to obtain the current interest portrait.

[0026] The interest vector includes the following dimensions: Subject preference vector: Identify user-preferred topics such as finance, sports, entertainment, technology, etc. through topic extraction techniques such as LDA, TF-IDF, and BERT; Emotional tendency feature: Based on a sentiment analysis model (such as Bi-LSTM + Attention), judge the positive or negative emotions in comments and search content; Content type preference: Preference types such as pictures and texts, videos, long texts, short videos, etc., generated by combining the user's stay time and interaction rate; Time period preference: Statistically analyze the user's activity in different time periods of morning, noon, and evening to form a feature distribution in the time dimension; Interaction frequency index: Combine the click frequency, forwarding behavior, and collection times within a unit time to measure the user's participation activity.

[0027] S2, vectorize the content to be pushed, and perform similarity matching with the current interest portrait to obtain a recommended sorting score.

[0028] In this embodiment, vectorizing the content to be pushed is specifically as follows: Obtain the content to be pushed, and extract the multi-modal feature information of the content to be pushed. The multi-modal feature information includes text content (such as title, body abstract, tag keywords), media information (such as pictures, videos, voices), meta information (such as release time, source, topic tags, field classification), and user interaction features (such as the overall click-through rate, comment popularity, and forwarding rate of the content); Select a corresponding encoding model for vectorized encoding according to the type of multi-modal feature information to obtain a number of sub-feature vectors; Map the sub-feature vectors based on a fully connected mapping method to obtain a content feature vector, and the content feature vector has the same structural dimension as the user interest vector.

[0029] Select a suitable encoding model according to different types of content features: For text - type information, a pre - trained language model (such as BERT, RoBERTa) is used to extract text semantic representations; for image - type information, a convolutional neural network (CNN) or a vision Transformer (such as ViT) is used to extract image embeddings; For multi - modal content, a cross - modal alignment model (such as CLIP, UNITER) is used to uniformly map text - image information to the same vector space; For tags and classification features, a word vector model (such as Word2Vec, GloVe) or a combination of one - hot encoding and an embedding matrix can be used for modeling.

[0030] In this embodiment, similarity matching is performed with the current interest portrait to obtain a recommended sorting score, specifically as follows: The current interest portrait and the content feature vector are normalized to ensure that their respective dimensions are consistent and the values are stable within a certain range; Based on the cosine similarity method, the cosine similarity between the current interest portrait and each content feature vector is obtained, and the matching degree is measured according to the cosine similarity; For the convenience of sorting and comparing multi - content scores, all the matched cosine similarity values are normalized, and the cosine similarity values are normalized based on linear scaling to obtain the recommended sorting score.

[0031] It should be noted that this technical solution realizes a comprehensive vector representation of the content to be pushed. It not only covers text information (such as titles, abstracts, tags), but also extends to media modalities such as pictures, videos, and voices, as well as meta - information such as publication time, topic tags, and category. It even introduces user interaction behaviors (such as click - through rate, comment popularity) as part of the content's dynamic popularity, thereby constructing a high - dimensional, multi - angle, and semantically rich content feature space. The construction of this feature space breaks through the limitation of traditional recommendation systems that only use a single content description field, providing a data basis for measuring the deep semantic similarity between contents.

[0032] Secondly, multiple sub - feature vectors are fused through a fully - connected mapping method to uniformly construct a content feature vector with a consistent structure, which has the same dimension as the user interest vector, creating conditions for subsequent similarity calculations. At the same time, this vector fusion method supports model - side training optimization, and the weight configuration can be adjusted reversely according to historical recommendation effects to achieve a content expression that is more in line with user behavior.

[0033] S3. Determine the pushed content and order according to the recommended sorting score in combination with user activity and push saturation.

[0034] In this embodiment, according to the recommended sorting score in combination with user activity and push saturation, the pushed content and order are determined, specifically as follows: For each candidate content, according to the cosine similarity calculation method, obtain its corresponding recommended score to form an initial recommended sorted list; Obtain the historical behavior frequency of the user. Based on the big data clustering analysis method, independently summarize the active patterns of the user in different time periods and construct an activity model. The historical behavior frequency includes but is not limited to the average daily click volume, view volume, average usage duration, push content response rate, and recent login frequency; The output value of the activity model represents the sensitivity of the user to content reception and the expected response willingness level at the current time point; Obtain the content push volume and response volume received by the user within the current period. Combine the historical optimal push frequency. Through data-driven modeling technology, identify the information saturation state of the user under different content types and construct a user push saturation model; Obtain the push saturation according to the user push saturation model. The higher the saturation value, the closer the reception frequency is to saturation, and the push should be slowed down; According to the activity model and the push saturation model, dynamically adjust the initial recommended sorted list.

[0035] In this embodiment, according to the activity model and the push saturation model, the initial recommended sorted list is dynamically adjusted as follows: When the user activity is high and the push saturation is not saturated, retain the recommended sorting and increase the push frequency; When the user activity is low, reduce the number of push contents and only push the contents in the front row of the sorting; When the user push saturation is in a saturated state, postpone or re-sort some recommended contents to avoid information overload.

[0036] It should be noted that two regulatory factors, user activity and push saturation, are introduced on the basis of traditional content recommendation, breaking through the limitation of the existing recommendation system that only focuses on content similarity and ignores the user state, making the content recommendation shift from static and single-dimensional drive to dynamic and multi-dimensional regulation, and improving the adaptability and response flexibility of the recommendation system to user state changes.

[0037] Secondly, the construction of the user activity model not only considers basic behavior frequencies (such as click volume, view volume, login frequency, etc.), but also integrates behavior characteristics reflecting deep usage willingness such as push content response rate and average usage duration, and mines the active patterns of users in different time periods through big data clustering technology, realizing personalized modeling of the user's active level. This model supports generating active labels at granularities such as daily, hourly, and holidays, providing strong support for the precise matching of recommended content and push rhythm.

[0038] S4. Extract the user behavior path clustering and content similarity based on the user feedback data, output the response prediction result, and dynamically adjust the content recommendation priority according to the response prediction result.

[0039] In this embodiment, the user behavior path clustering and content similarity are extracted based on the user feedback data, the response prediction result is output, and the content recommendation priority is dynamically adjusted according to the response prediction result, as follows: Obtain the user feedback data, where the feedback data includes click behavior, page browsing time, scroll depth, content bounce situation, comment publishing, like behavior, and forward sharing; Use the user ID and content ID as a joint index for the feedback data, and record and construct a feedback sample set in chronological order; Perform a sliding time window process on the feedback sample set with a fixed time step, and obtain the occurrence frequency and behavior change trend of each behavior type within each window to obtain change features. The change features include the relative ratio between different behaviors (such as the ratio of clicks to bounces), the growth rate of a certain behavior within a time period (such as the increasing trend of scroll depth), and the user response intensity of content types in different time periods; Obtain the behavior event sequence within each time window, and combine the change features to identify the logical association between adjacent behavior events to form a behavior path sequence; Perform similarity calculation and clustering analysis on the behavior paths to obtain a set of preferred paths for user content; Obtain the user behavior paths associated with the current recommended content in history, and obtain the response data (such as click-through rate, dwell time, forward rate) of the current recommended content in the preferred paths; If the recommended content is the first recommended content, then use the historical content similar to the recommended content structure, and obtain the response data of the similar content in the similar paths for approximate estimation to obtain the response prediction result; Dynamically adjust the recommendation list according to the response prediction result of the recommended content in the user preferred paths.

[0040] In this embodiment, perform similarity calculation and clustering analysis on the behavior paths to obtain a set of preferred paths for user content, as follows: Based on the continuous interaction behavior of the user, extract the content access sequence in chronological order. Each sequence contains the user's click, browse, comment, and forward behaviors on multiple contents during a session or in a short period; Construct behavior paths according to the content access sequence to obtain an initial set of behavior paths. For example, if a user accesses contents A, B, and C successively within a certain time period, then construct the path {A→B→C}; Vectorize each content node in each path based on a content embedding model (such as Text-CNN), and perform weighted fusion by combining the behavior type (click, comment) and behavior intensity (stay duration, number of likes) to generate a path vector representation; Calculate the pairwise similarity of the behavior path vectors of different users based on the cosine similarity algorithm to obtain a path similarity matrix; According to the path similarity matrix, classify the path vector set through the K-means clustering algorithm to obtain a set of preferred paths. Each cluster represents a type of user behavior path with high similarity, having a similar content evolution trajectory and preference characteristics.

[0041] In this embodiment, obtain the response data of similar content in similar paths for approximate estimation to obtain a response estimation result, as follows: Perform a unique identifier search for each recommended content item to determine whether it is the content that has appeared in historical interactions. If the content does not appear in any user's behavior path, the system marks it as the first recommended content; Extract the structural features of the first recommended content and generate a content embedding vector through the deployed content encoding network. The structural features include but are not limited to text summary, tag set, media type, content length, and release time; Calculate the similarity between the embedding vector and all historical content vectors in the content database, and select the top K most similar historical content items as the structural similarity reference samples for the recommended content; In the set of preferred paths, filter out the path samples that contain the top K most similar historical content items and construct a set of similar paths; Obtain the user response data of the corresponding content of each similar path. The user response data includes but is not limited to behavior signals such as click-through rate, stay time, interaction frequency, and bounce rate; Normalize the user response data and perform weighted averaging in the path set to generate an approximate response estimate value to obtain a response estimation result.

[0042] It should be noted that the dynamic change trend of user behavior in a continuous time period is extracted through a sliding time window mechanism to generate behavior change characteristics. These change characteristics include the growth rate of behavior frequency, the relative proportion of different behaviors, and the change pattern of user response at different time periods, which can effectively reveal the regularity of the evolution of user interests over time, thereby providing a time-sensitive input basis for dynamic recommendation strategies and realizing time-context-aware content recommendation.

[0043] Furthermore, in the process of user behavior path modeling, instead of simply adopting the behavior event sequence, it integrates the behavior change features extracted within a sliding time window, constructs an enhanced behavior path sequence, and represents the path based on the content semantic embedding and behavior intensity weighting mechanism. Then, through similarity calculation and clustering analysis, a set of preference paths is obtained. This modeling method takes into account the characteristics of behavior logic, time dynamics, and content semantics, making the clustering results have higher representation ability and user discrimination ability, and can effectively identify the potential interest preferences of users in different contexts.

[0044] Embodiment 2 Figure 2 A system for an optimized content push method for analyzing user interests according to the present invention is provided, including a data acquisition module, a matching module, a push module, and a feedback correction module, and there are connections between the modules; The data acquisition module is used to obtain user behavior data, construct a user interest vector, and generate a current interest profile through weighted fusion; The matching module is used to vectorize the content to be pushed, perform similarity matching with the current interest profile, and obtain a recommended sorting score; The push module is used to determine the pushed content and order according to the recommended sorting score in combination with user activity and push saturation; The feedback correction module is used to extract user behavior path clustering and content similarity from user feedback data, output a response prediction result, and dynamically adjust the content recommendation priority according to the response prediction result.

[0045] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain a formula closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0046] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product.

[0047] Those of ordinary skill in the art can realize that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Skilled professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0048] In addition, in each embodiment of the present application, each functional module may be integrated into one processing module, may exist separately as individual physical modules, or two or more modules may be integrated into one module.

[0049] As described above, only the specific embodiments of the present application are provided, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0050] Finally: The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A content push optimization method for analyzing user interests, characterized in that It includes the following steps: Obtain user behavior data, construct a user interest vector, and perform weighted fusion to generate the current interest portrait; Vectorize the content to be pushed, perform similarity matching with the current interest portrait, and obtain the recommended sorting score; Determine the pushed content and order according to the recommended sorting score in combination with user activity and push saturation; Extract the user behavior path clustering and content similarity from the user feedback data, output the response prediction result, and dynamically adjust the content recommendation priority according to the response prediction result.

2. The content push optimization method for analyzing user interests according to claim 1, wherein The obtaining of user behavior data, constructing a user interest vector, and performing weighted fusion to generate the current interest portrait are specifically as follows: Obtain the user behavior data within the past T hours to form a time series behavior set, and perform weighted encoding on the time series behavior set through an attention mechanism to generate a short-term interest vector with a fixed dimension; Obtain the user's long-term historical behavior data, extract the topic tags, semantic classifications, and emotional tendencies of the corresponding content, and count their occurrence frequencies to form a statistical vector; Construct a triple graph structure based on knowledge graph technology, perform node representation learning on the graph structure based on a graph neural network, obtain the embedding features of the user in the graph, and extract the embedding vector corresponding to the user node; Concatenate the statistical vector and the embedding vector, and perform a non-linear mapping to output a long-term interest vector with a unified dimension; Perform weighted fusion on the short-term interest vector and the long-term interest vector to obtain the current interest portrait.

3. The content push optimization method for analyzing user interests according to claim 2, wherein The vectorization of the content to be pushed is specifically as follows: Obtain the content to be pushed, and extract the multi-modal feature information of the content to be pushed; Select the corresponding encoding model for vectorization encoding according to the type of multi-modal feature information to obtain several sub-feature vectors; Map the sub-feature vectors based on the fully connected mapping method to obtain a content feature vector, and the content feature vector has the same structural dimension as the user interest vector.

4. A method for optimizing content push for analyzing user interests according to claim 3, characterized in that The similarity matching with the current interest portrait to obtain the recommended sorting score is specifically as follows: Obtain the cosine similarity of the current interest portrait matching with each content feature vector based on the cosine similarity method; Normalize all the matching cosine similarity values, and normalize the cosine similarity values based on linear scaling to obtain the recommended sorting score.

5. The content push optimization method for analyzing user interests according to claim 4, wherein The determination of the pushed content and order according to the recommended sorting score in combination with user activity and push saturation is specifically as follows: Based on the cosine similarity calculation method, obtain the recommended score of each candidate content to form an initial recommended sorting list; Obtain the historical behavior frequency of the user, and based on the big data clustering analysis method, autonomously induce the active patterns of the user in different time periods to construct an activity model; Obtain the content push volume and response volume received by the user within the current period, combine the historical optimal push frequency, and identify the information saturation state of the user under different content types through data-driven modeling technology to construct a user push saturation model; Obtain the push saturation according to the user push saturation model; Dynamically adjust the initial recommended sorting list according to the activity model and the push saturation model.

6. The content push optimization method for analyzing user interests according to claim 5, characterized in that, Extract the user behavior path clustering and content similarity based on the user feedback data, output the response prediction result, and dynamically adjust the content recommendation priority according to the response prediction result, as follows: Obtain the user feedback data, use the user ID and content ID as the joint index for the feedback data, and record and construct the feedback sample set in chronological order; Perform a sliding time window process on the feedback sample set with a fixed time step, and obtain the occurrence frequency and behavior change trend of each behavior type within each window to obtain the change characteristics; Obtain the behavior event sequence within each time window, and combine the change characteristics to identify the logical association between adjacent behavior events to form a behavior path sequence; Perform similarity calculation and clustering analysis on the behavior paths to obtain the set of preferred paths for user content; Obtain the user behavior paths associated with the current recommended content in history, and obtain the response data of the current recommended content in the preferred paths; If the recommended content is the first recommended content, obtain the historical content similar to the structure of the recommended content, and obtain the response data of the similar content in the similar paths for approximate estimation to obtain the response prediction result; Dynamically adjust the recommendation list according to the response prediction result of the recommended content in the user preferred path; 7. A method for optimizing content push for analyzing user interests according to claim 6, characterized in that, The similarity calculation and clustering analysis of the behavior paths to obtain the set of preferred paths for user content are as follows: Based on the continuous interaction behavior of the user, extract the content access sequence in chronological order; Construct behavior paths according to the content access sequence to obtain the initial set of behavior paths; Vectorize each content node in each path based on the content embedding model, and perform weighted fusion in combination with the behavior type and behavior intensity to generate a path vector representation; Based on the cosine similarity algorithm, calculate the pairwise behavior path vectors of different users to obtain a path similarity matrix; According to the path similarity matrix, classify the set of path vectors through the K-means clustering algorithm to obtain the set of preferred paths; 8. The content push optimization method for analyzing user interests according to claim 7, characterized in that The obtaining of the response data of the similar content in the similar paths for approximate estimation to obtain the response prediction result is as follows: Perform a unique identifier search on each recommended content item to determine whether it is the content that has appeared in the historical interaction. If the content does not appear in the behavior path of any user, the system marks it as the first recommended content; Extract the structural features of the first recommended content, and generate a content embedding vector through the deployed content encoding network; Calculate the similarity between the embedding vector and all historical content vectors in the content database, and select the top K most similar historical content items; In the set of preferred paths, filter out the path samples containing the top K most similar historical content items, and construct a set of similar paths; Obtain the user response data of the corresponding content of each similar path; Normalize the user response data and perform weighted averaging in the path set to generate an approximate response estimate value to obtain the response prediction result; 9. A system using a content push optimization method for analyzing user interests as described in any one of claims 1-8, characterized in that, It includes a data collection module, a matching module, a push module, and a feedback correction module, and there are connections between the modules; The data collection module is used to obtain user behavior data, construct a user interest vector, and perform weighted fusion to generate the current interest profile; A matching module, which is used to vectorize the content to be pushed, match the similarity with the current interest profile, and obtain a recommended sorting score; A pushing module, which is used to determine the pushed content and order according to the recommended sorting score in combination with user activity and pushing saturation; A feedback correction module, which is used to extract user behavior path clustering and content similarity from user feedback data, output a response prediction result, and dynamically adjust the content recommendation priority according to the response prediction result.

Images