A content push optimization method and system for analyzing user interests

Abstract

The present invention discloses a content push optimization method and system for analyzing user interests, which relates to the field of artificial intelligence and personalized recommendation technology, and comprises the following steps: obtaining user behavior data, constructing user interest vectors, and weighted fusion to generate a current interest portrait; vectorizing the content to be pushed, performing similarity matching with the current interest portrait, and obtaining a recommendation ranking score; determining the push content and order based on the recommendation ranking score in combination with user activity and push saturation; extracting user behavior path clustering and content similarity based on user feedback data, outputting response prediction results, and dynamically adjusting content recommendation priorities based on the response prediction results; the present application solves the problem of inaccurate content recommendation through multimodal fusion and feedback optimization methods.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and personalized recommendation technology, 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 adoption of smart devices, the amount of digital information users are exposed to daily is growing exponentially, leading to an increasingly serious problem of information overload. Against this backdrop, personalized content push technology has become a crucial tool for improving user experience, enhancing platform engagement, and increasing content distribution efficiency. Personalized push systems analyze user behavior data, interests, and historical interactions to deliver personalized content recommendations. These systems have been widely adopted across industries, including news and information, e-commerce, social media, and short video platforms.

[0003] Traditional recommendation systems are primarily based on collaborative filtering algorithms and content matching mechanisms, establishing associations based on similarities between users and content to infer content that users may be interested in. However, with the increasing complexity of user behavior and the diversification of content formats, traditional methods have gradually exposed numerous shortcomings. First, collaborative filtering methods are limited in effectiveness in cold-start scenarios, making it difficult to effectively recommend new users or new content. Second, static content matching ignores the dynamic changes in user interests and contextual factors, making it difficult to adapt to the shifting preferences brought about by changes in user behavior over time. Furthermore, some current systems lack accurate identification of user receptive capacity, active time periods, and information saturation during the content sorting process, which can easily lead to redundant push notifications or disrupt 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 multiple technologies, such as multimodal content features, dynamic user behavior, time series trends, and preference clustering, to achieve deep modeling of user interests and adaptive optimization of push strategies. This method should not only have the ability to accurately identify interests and match content, but also analyze user activity and content saturation, thereby dynamically adjusting the push order and frequency, improving the relevance, timeliness, and user acceptance of push notifications.

[0005] For example, the invention patent with publication number CN119128253A discloses an internet-based user interest intelligent recommendation method and system, which involves the field of internet technology. The system collects and integrates user behavior data on different platforms in real time to form a multi-source data set. Utilizing natural language processing and time series analysis technology, the system deeply explores user interests and interest change trends, identifying potential interest surges. Through a clustering algorithm, the system divides users with similar interest characteristics into dynamic user interest clusters and dynamically adjusts the parameters of user interest clusters based on instantaneous data. The system predicts users' future needs through a deep learning model, combines the time series analysis model with user interest cluster characteristics, generates personalized recommendation content, and pushes it to users through multiple channels. This improves the accuracy of pushing relevant content to users with sudden changes in interest characteristics.

[0006] For example, the invention patent with publication number CN119089398A discloses a content recommendation method and system based on semantic recognition. This method uses natural language processing technology to analyze user text, voice, or video input to generate user semantic representation data. It dynamically constructs a user interest graph based on historical user behavior data, and uses a multi-layer graph convolutional network to learn node and edge representations to generate time-series user interest graph representation data. It utilizes the Transformer model's self-attention and multi-head attention mechanisms, combined with a contrastive learning module, to optimize the match between user interest representations and recommended content, generating personalized recommended content data. It also collects user multimodal information and fuses it with personalized recommended content data, generating optimized multimodal recommended content data through a multimodal fusion algorithm. Finally, it collects user feedback feature data and dynamically adjusts and optimizes the recommended content generation strategy through an adaptive learning algorithm. This system achieves efficient, accurate, and personalized recommendations, improving the user experience.

[0007] The above disclosed technical solutions have at least the following technical problems:

[0008] Traditional recommendation systems often rely on static user tags or simple click behaviors, and are unable to dynamically portray user interest changes in multiple dimensions and across time periods, which can easily lead to biased or invalid recommendation results. In addition, most recommendation systems are only based on text content or historical click frequency, making it difficult to effectively handle multimodal content (such as images, text, and videos) and semantic matching differences, resulting in insufficient relevance of recommended content.

[0009] In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0010] In order to overcome the above-mentioned 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 multimodal fusion and feedback optimization methods.

[0011] To achieve the above object, the present invention provides the following technical solutions:

[0012] A content push optimization method for analyzing user interests includes the following steps: obtaining user behavior data, constructing user interest vectors, and weighted fusion to generate a current interest profile; vectorizing the content to be pushed, performing similarity matching with the current interest profile, and obtaining a recommendation ranking score; determining the push content and order based on the recommendation ranking score combined with user activity and push saturation; extracting user behavior path clustering and content similarity based on user feedback data, outputting response prediction results, and dynamically adjusting content recommendation priorities based on the response prediction results.

[0013] In a preferred embodiment, the user behavior data is obtained, the user interest vector is constructed, and weighted fusion is performed to generate the current interest portrait, specifically as follows: user behavior data in the past T hours is obtained to form a time series behavior set, and the time series behavior set is weightedly encoded through the attention mechanism to generate a short-term interest vector of fixed dimension; the user's long-term historical behavior data is obtained, the topic tags, semantic classifications and emotional tendencies of the corresponding content are extracted, and their occurrence frequencies are counted to form a statistical vector; a ternary graph structure is constructed based on knowledge graph technology, and node representation learning of the graph structure is performed based on a graph neural network to obtain the user's embedded features in the graph, and extract the embedding vector corresponding to the user node; the statistical vector is spliced ​​with the embedding vector, and nonlinear mapping is performed to output a long-term interest vector of unified dimension; the short-term interest vector and the long-term interest vector are weightedly fused to obtain the current interest portrait.

[0014] In a preferred embodiment, the content to be pushed is vectorized as follows: obtaining the content to be pushed and extracting the multimodal feature information of the content to be pushed; selecting a corresponding encoding model for vectorized encoding according to the type of multimodal feature information to obtain several sub-feature vectors; mapping the sub-feature vectors based on a fully connected mapping method to obtain a content feature vector, wherein the content feature vector has the same structural dimension as the user interest vector.

[0015] In a preferred embodiment, the similarity matching with the current interest portrait is performed to obtain a recommended ranking score, specifically as follows: based on the cosine similarity method, the cosine similarity between the current interest portrait and each content feature vector is obtained; all matching cosine similarity values ​​are normalized, and the cosine similarity values ​​are normalized based on linear scaling to obtain a recommended ranking score.

[0016] In a preferred embodiment, the push content and order are determined based on the recommendation ranking score combined with user activity and push saturation, as follows: based on the cosine similarity calculation method, the recommendation score of each candidate content is obtained to form an initial recommendation ranking list; the user's historical behavior frequency is obtained, and based on the big data clustering analysis method, the user's active patterns in different time periods are independently summarized to build an activity model; the amount of content push and response received by the user in the current period is obtained, and combined with the historical optimal push frequency, the user's information saturation status under different content types is identified through data-driven modeling technology, and a user push saturation model is built; the push saturation is derived according to the user push saturation model; the initial recommendation ranking list is dynamically adjusted according to the activity model and the push saturation model.

[0017] In a preferred embodiment, the method extracts user behavior path clustering and content similarity based on user feedback data, outputs response prediction results, and dynamically adjusts content recommendation priorities based on the response prediction results, specifically as follows: obtains user feedback data, uses user ID and content ID as a joint index for the feedback data, and records them in chronological order to construct a feedback sample set; performs sliding time window processing on the feedback sample set with a fixed time step, and obtains the frequency of occurrence and behavior change trend of each behavior type in each window to obtain change characteristics; obtains a sequence of behavior events in each time window, and combines the change characteristics to identify the logical association between adjacent behavior events to form a behavior path sequence; performs similarity calculation and cluster analysis on the behavior paths to obtain a set of user content preference paths; obtains the user behavior paths associated with the current recommended content in history, and obtains response data for the current recommended content in the preference path; if the recommended content is the first recommended content, then historical content with a similar structure to the recommended content is selected, and response data for similar content in similar paths is obtained for approximate estimation to obtain a response prediction result; and dynamically adjusts the recommendation list based on the response prediction result of the recommended content in the user preference path.

[0018] In a preferred embodiment, the similarity calculation and cluster analysis of the behavior paths are performed to obtain a set of user content preference paths, specifically as follows: based on the user's continuous interactive behavior, the content access sequence is extracted in chronological order; the behavior path is constructed according to the content access sequence to obtain an initial behavior path set; based on the content embedding model, each content node in each path is vectorized, 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; based on the path similarity matrix, the path vector set is classified by the K-means clustering algorithm to obtain a preference path set.

[0019] In a preferred embodiment, the response data of similar content in similar paths is obtained and approximately estimated to obtain a response estimation result, specifically as follows: a unique identifier is retrieved for each recommended content item to determine whether it is content that has appeared in historical interactions. If the content does not appear in the behavior path of any user, the system marks it as the first recommended content; the structural features of the first recommended content are extracted, and a content embedding vector is generated through the deployed content encoding network; the embedding vector is similarly calculated with all historical content vectors in the content database, and the top K most similar historical content items are selected; in the preferred path set, a path sample containing the top K most similar historical content items is screened out, and a similar path set is constructed; user response data for the corresponding content of each similar path is obtained; the user response data is normalized and weighted averaged in the path set to generate an approximate response estimate value to obtain a response estimation result.

[0020] A system for optimizing content push by analyzing user interests includes a data acquisition module, a matching module, a push module, and a feedback correction module, with 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 portrait through weighted fusion. The matching module is used to vectorize the content to be pushed, perform similarity matching with the current interest portrait, and obtain a recommendation ranking score. The push module is used to determine the push content and order based on the recommendation ranking score combined with user activity and push saturation. The feedback correction module is used to extract user behavior path clustering and content similarity based on user feedback data, output response estimation results, and dynamically adjust the content recommendation priority based on the response estimation results.

[0021] The technical effects and advantages of the content push optimization method and system for analyzing user interests of the present invention are as follows:

[0022] 1. This method dynamically depicts user interests by constructing user interest vectors and fusing them to generate a current interest profile. Specifically, by capturing the user's behavior sequence over a short period of time, an attention mechanism is introduced to extract short-term interests. Long-term interests are extracted by combining long-term historical behavior with knowledge graph-based node embedding. A weighted fusion strategy is then used to generate a multi-dimensional interest profile. This method overcomes the limitations of traditional tag-based or static profiles, offering advantages such as strong real-time performance, high personalization, and robust representation capabilities. It can more accurately reflect a user's true interests and preferences within a specific time period.

[0023] 2. This invention enhances the self-learning capabilities of the recommendation system by constructing a multi-objective optimization function through a feedback correction module. The system not only records user responses (clicks, browsing, comments, etc.), but also mines user preference paths based on behavioral event sequences. It introduces a similar path inference mechanism for first-time recommendations, enabling dynamic evaluation of recommendation effectiveness in scenarios lacking historical data. This ensures that the recommendation strategy is forward-looking and generalizable. This mechanism improves the coverage and rationality of new content recommendations, resolving the cold start problem in recommendation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a content push optimization method for analyzing user interests according to the present invention.

[0025] Figure 2 A schematic diagram of the system structure of a content push optimization method for analyzing user interests according to the present invention. DETAILED DESCRIPTION

[0026] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0027] Example 1, Figure 1 The present invention provides a content push optimization method for analyzing user interests, comprising the following steps:

[0028] S1, obtain user behavior data, construct user interest vectors, and perform weighted fusion to generate the current interest portrait.

[0029] In this embodiment, user behavior data is obtained, user interest vectors are constructed, and weighted fusion is performed to generate the current interest profile, as follows:

[0030] Obtain user behavior data within the past T hours (e.g., the last 24 hours) to form a time series behavior set. The behavior data includes click behavior, browsing time, search keywords, favorites, forwarding, and comment information;

[0031] The time-aware attention mechanism is used to weight the time series behavior set and generate a short-term interest vector of fixed dimension (such as 128 or 256 dimensions) to reflect the user's current focus of interest.

[0032] Obtain long-term historical user behavior data (e.g., the past 1 to 3 months), extract the corresponding content's topic tags, semantic classifications, and emotional tendencies, and count their occurrence frequencies to form the initial interest distribution features, i.e., statistical vectors;

[0033] Based on knowledge graph technology, we build a user-content-tag ternary graph structure, and learn node representations of the graph structure based on graph neural networks to obtain the embedded features of users in the graph and extract the embedding vectors corresponding to user nodes.

[0034] The statistical vector is concatenated with the embedding vector, and nonlinear mapping is performed to output a long-term interest vector of uniform dimension.

[0035] The short-term interest vector and the long-term interest vector are weightedly fused to obtain the current interest portrait.

[0036] The interest vector contains the following dimensions:

[0037] Topic preference vector: Identify user preferred topics such as finance, sports, entertainment, and technology using topic extraction technologies such as LDA, TF-IDF, and BERT.

[0038] Sentiment tendency features: Based on sentiment analysis models (such as Bi-LSTM + Attention), determine the positive or negative sentiment in comments and search content;

[0039] Content type preference: Such as pictures, videos, long articles, short videos, etc., combined with user stay time and interaction rate;

[0040] Time period preference: Statistics on user activity in different time periods of morning, noon and evening are collected to form a time dimension feature distribution;

[0041] Interaction frequency index: combines the click frequency, forwarding behavior, number of collections, etc. in a unit of time to measure the user's participation activity.

[0042] S2, vectorize the content to be pushed, match it with the current interest profile for similarity, and obtain the recommendation ranking score.

[0043] In this embodiment, the content to be pushed is vectorized as follows:

[0044] Obtaining the content to be pushed and extracting multimodal feature information of the content to be pushed, including text content (such as title, text summary, tag keywords), media information (such as pictures, videos, and voice), meta information (such as release time, source, topic tags, and field classification), and user interaction features (such as the overall click-through rate, comment popularity, and forwarding rate of the content);

[0045] Select the corresponding encoding model according to the type of multimodal feature information for vectorized encoding to obtain several sub-feature vectors;

[0046] The sub-feature vectors are mapped based on a fully connected mapping method to obtain a content feature vector, which has the same structural dimension as the user interest vector.

[0047] Choose the appropriate encoding model based on the characteristics of different types of content:

[0048] For text information, pre-trained language models (such as BERT and RoBERTa) are used to extract text semantic representations; for image information, convolutional neural networks (CNN) or visual transformers (such as ViT) are used for image embedding extraction;

[0049] For multimodal content, a cross-modal alignment model (such as CLIP and UNITER) is used to uniformly map image and text information into the same vector space.

[0050] For label and classification features, word vector models (such as Word2Vec, GloVe) or one-hot encoding combined with embedding matrix can be used for modeling.

[0051] In this embodiment, similarity matching is performed with the current interest profile to obtain the recommended ranking score, which is as follows:

[0052] Normalize the current interest profile and content feature vector to ensure that their dimensions are consistent and their values ​​are stable within a certain range;

[0053] Based on the cosine similarity method, the cosine similarity between the current interest profile and each content feature vector is obtained, and the degree of matching is measured according to the cosine similarity;

[0054] To facilitate the ranking and comparison of multiple content scores, the cosine similarity values ​​of all matches are normalized, and the cosine similarity values ​​are normalized based on linear scaling to obtain the recommendation ranking score.

[0055] It should be noted that this technical solution achieves a comprehensive vectorized representation of pushed content, covering not only textual information (such as titles, summaries, and tags), but also media modalities such as images, videos, and audio, as well as meta-information such as release time, hashtags, and categories. It even incorporates user interactions (such as click-through rate and comment popularity) as part of the content's dynamic popularity, thereby constructing a high-dimensional, multi-faceted, and semantically rich content feature space. This feature space breaks through the limitation of traditional recommendation systems that rely solely on a single content description field, providing a data foundation for measuring deep semantic similarity between content.

[0056] Secondly, multiple sub-feature vectors are fused using a fully connected mapping method to construct a unified content feature vector with a consistent structure. This vector fusion method maintains the same dimensions as the user interest vector, facilitating subsequent similarity calculations. Furthermore, this vector fusion method supports model-side training optimization, adjusting weight configuration based on historical recommendation performance to achieve content presentation that better aligns with user behavior.

[0057] S3, based on the recommendation ranking score combined with user activity and push saturation, determines the push content and order.

[0058] In this embodiment, the push content and order are determined based on the recommendation ranking score combined with user activity and push saturation, as follows:

[0059] For each candidate content, obtain its corresponding recommendation score based on the cosine similarity calculation method to form an initial recommendation ranking list;

[0060] Obtain the user's historical behavior frequency, and based on big data cluster analysis methods, autonomously summarize the user's activity patterns in different time periods to build an activity model. The historical behavior frequency includes but is not limited to the average daily clicks, page views, average usage time, push content response rate, and recent login frequency;

[0061] The activity model output value represents the user's sensitivity to content reception and expected response willingness level at the current time point;

[0062] Obtain the amount of content push and responses received by the user in the current period. Combined with the historical optimal push frequency, data-driven modeling techniques are used to identify the user's information saturation status for different content types and build a user push saturation model.

[0063] Push saturation is calculated based on the user push saturation model. A higher saturation value indicates that the receiving frequency is close to saturation, and push should be slowed down.

[0064] The initial recommendation ranking list is dynamically adjusted based on the activity model and push saturation model.

[0065] In this embodiment, the initial recommendation ranking list is dynamically adjusted based on the activity model and the push saturation model, as follows:

[0066] When user activity is high and push notifications are not yet saturated, retain the recommended ranking and increase the push frequency;

[0067] When user activity is low, reduce the number of pushed content and only push the top ranked content;

[0068] When user push saturation reaches saturation, some recommended content will be postponed or reordered to avoid information overload.

[0069] 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 limitations of existing recommendation systems that only focus on content similarity and ignore user status. Content recommendation is transformed from static, single-dimensional drive to dynamic, multi-dimensional regulation, improving the recommendation system's adaptability and response flexibility to changes in user status.

[0070] Secondly, the construction of the user activity model not only considers basic behavioral frequency (such as clicks, views, and login frequency), but also incorporates behavioral characteristics that reflect deeper usage intentions, such as response rate to pushed content and average usage duration. Using big data clustering technology, it mines user activity patterns across different time periods to achieve personalized modeling of user activity levels. This model supports the generation of activity tags at granular levels such as day, hour, and holiday, providing strong support for accurately matching recommended content with the push cadence.

[0071] S4, extracts user behavior path clustering and content similarity based on user feedback data, outputs response prediction results, and dynamically adjusts content recommendation priorities based on the response prediction results.

[0072] In this embodiment, user behavior path clustering and content similarity are extracted based on user feedback data, response prediction results are output, and content recommendation priorities are dynamically adjusted based on the response prediction results, as follows:

[0073] Obtain user feedback data, including click behavior, page browsing time, scrolling depth, content bounce, comment posting, like behavior, and forwarding and sharing;

[0074] The feedback data is indexed by user ID and content ID, and recorded in chronological order to construct a feedback sample set;

[0075] Perform sliding time window processing on the feedback sample set with a fixed time step, and obtain the frequency of occurrence and behavior change trend of each behavior type in each window to obtain change characteristics. The change characteristics include the relative proportion of different behaviors (such as the ratio of clicks to bounces), the growth rate of a certain behavior in a time period (such as the increasing trend of scrolling depth), and the user response intensity of content types in different time periods;

[0076] Obtain the behavioral event sequence within each time window, and combine the change characteristics to identify the logical relationship between adjacent behavioral events to form a behavioral path sequence;

[0077] Perform similarity calculation and cluster analysis on the behavior paths to obtain the user content preference path set;

[0078] Obtain the user behavior path associated with the current recommended content in history, and obtain the response data of the current recommended content in the preferred path (such as click rate, dwell time, forwarding rate);

[0079] If the recommended content is the first time recommended content, then historical content with similar structure to the recommended content is used, and the response data of similar content in similar paths is obtained for approximate estimation to obtain the response estimation result;

[0080] Dynamically adjust the recommendation list based on the estimated response results of the recommended content in the user's preference path.

[0081] In this embodiment, similarity calculation and cluster analysis are performed on the behavior paths to obtain a set of user content preference paths, as follows:

[0082] Based on the user's continuous interactive behavior, content access sequences are extracted in chronological order. Each sequence includes the user's click, browse, comment, and forward behaviors on multiple contents in one session or a short period of time.

[0083] Construct a behavior path based on the content access sequence to obtain an initial behavior path set. For example, if a user visits content A, B, and C in a certain time period, the path {A→B→C} is constructed.

[0084] Using a content embedding model (such as Text-CNN), each content node in each path is vectorized and weighted by behavior type (click, comment) and behavior intensity (stay duration, number of likes) to generate a path vector representation.

[0085] Based on the cosine similarity algorithm, the behavior path vectors of different users are calculated pairwise to obtain the path similarity matrix;

[0086] According to the path similarity matrix, the path vector set is classified by the K-means clustering algorithm to obtain the preference path set. Each cluster represents a class of highly similar user behavior paths with similar content evolution trajectories and preference characteristics.

[0087] In this embodiment, the response data of similar content in similar paths is obtained for approximate estimation to obtain the response estimation result, which is as follows:

[0088] A unique identifier is retrieved for each recommended content item to determine whether it 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.

[0089] Extracting structural features of the first recommended content and generating 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.

[0090] 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 structurally similar reference samples for recommended content;

[0091] In the preferred path set, the path samples containing the top K most similar historical content items are screened out, and a similar path set is constructed;

[0092] Obtain user response data for the corresponding content of each similar path, including but not limited to behavioral signals such as click-through rate, dwell time, interaction frequency, and bounce rate;

[0093] The user response data is normalized and weighted averaged in the path set to generate an approximate response estimate and obtain the response prediction result.

[0094] It's important to note that the sliding time window mechanism extracts dynamic trends in user behavior over continuous time periods and generates behavioral change features. These features, including the growth rate of behavior frequency, the relative proportions of different behaviors, and the changing patterns of user responses over different time periods, effectively reveal the regularity of user interest evolution over time. This provides time-sensitive input for dynamic recommendation strategies, enabling context-aware content recommendations.

[0095] Furthermore, in the user behavior path modeling process, rather than simply using a sequence of behavioral events, we incorporate behavioral change features extracted within a sliding time window to construct an enhanced behavioral path sequence. This path representation is based on content semantic embedding and a behavioral intensity weighting mechanism. Furthermore, through similarity calculation and cluster analysis, a set of preferred paths is derived. This modeling approach takes into account the characteristics of behavioral logic, temporal dynamics, and content semantics, resulting in clustering results with higher representation power and user discrimination capabilities, effectively identifying users' potential interests and preferences in different contexts.

[0096] Example 2, Figure 2 The present invention provides a system for analyzing user interests and optimizing content push methods, including a data acquisition module, a matching module, a push module, and a feedback correction module, wherein the modules are connected;

[0097] The data collection module is used to obtain user behavior data, construct user interest vectors, and perform weighted fusion to generate current interest portraits;

[0098] The matching module is used to vectorize the content to be pushed, match it with the current interest profile, and obtain the recommendation ranking score;

[0099] The push module is used to determine the push content and order based on the recommendation ranking score combined with user activity and push saturation;

[0100] The feedback correction module is used to extract user behavior path clustering and content similarity based on user feedback data, output response prediction results, and dynamically adjust content recommendation priorities based on the response prediction results.

[0101] The above formulas are all dimensionless and numerical calculations. The formula is a formula that is closest to the actual situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0102] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0103] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0104] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0105] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0106] Finally: The above description is only a preferred embodiment of the present invention and is not intended 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 scope of protection of the present invention.

Claims

1. A content push optimization method for analyzing user interests, characterized in that: The steps include: Obtain user behavior data, construct user interest vectors, and perform weighted fusion to generate current interest profiles; Vectorize the content to be pushed, perform similarity matching with the current interest profile, and obtain the recommendation ranking score; Determine the push content and order based on the recommendation ranking score, user activity, and push saturation; Extract user behavior path clustering and content similarity based on user feedback data, output response prediction results, and dynamically adjust content recommendation priorities based on the response prediction results; The method extracts user behavior path clustering and content similarity based on user feedback data, outputs response prediction results, and dynamically adjusts content recommendation priorities based on the response prediction results, as follows: Obtain user feedback data, use the user ID and content ID as the joint index for the feedback data, and record them in chronological order to construct a feedback sample set; Perform sliding time window processing on the feedback sample set with a fixed time step, and obtain the occurrence frequency and behavior change trend of each behavior type in each window to obtain the change characteristics; Obtain the behavioral event sequence within each time window, and combine the change characteristics to identify the logical relationship between adjacent behavioral events to form a behavioral path sequence; Perform similarity calculation and cluster analysis on the behavior paths to obtain the user content preference path set; Obtaining response data for the current recommended content in the preferred path, including click-through rate, dwell time, and forwarding rate; If the recommended content is the first recommended content, then obtain historical content with a similar structure to the recommended content, and obtain response data of the similar historical content in a similar path for approximate estimation to obtain a response estimation result; Dynamically adjust the recommendation list based on the estimated response results of the recommended content in the user's preference path.

2. The content push optimization method for analyzing user interests according to claim 1, characterized in that: The user behavior data is obtained, the user interest vector is constructed, and the weighted fusion is used to generate the current interest profile, as follows: Obtain user behavior data within the past T hours to form a time series behavior set. Use the attention mechanism to weight the time series behavior set and generate a short-term interest vector with a fixed dimension. Obtain long-term historical user behavior data, extract the corresponding content's topic tags, semantic classifications, and emotional tendencies, and count their occurrence frequencies to form a statistical vector; Build a ternary graph structure based on knowledge graph technology, learn node representations of the graph structure based on graph neural networks, obtain the user's embedded features in the graph, and extract the embedding vector corresponding to the user node; The statistical vector is concatenated with the embedding vector, and nonlinear mapping is performed to output a long-term interest vector of uniform dimension. The short-term interest vector and the long-term interest vector are weightedly fused to obtain the current interest portrait.

3. The content push optimization method for analyzing user interests according to claim 2, characterized in that: The content to be pushed is vectorized as follows: Obtain the content to be pushed and extract multimodal feature information of the content to be pushed; Select the corresponding encoding model according to the type of multimodal feature information for vectorized encoding to obtain several sub-feature vectors; The sub-feature vectors are mapped based on a fully connected mapping method to obtain a content feature vector, which has the same structural dimension as the user interest vector.

4. The content push optimization method for analyzing user interests according to claim 3, characterized in that: The similarity matching with the current interest profile is performed to obtain the recommended ranking score, which is as follows: Based on the cosine similarity method, the cosine similarity between the current interest profile and each content feature vector is obtained; The cosine similarity values ​​of all matches are normalized and normalized based on linear scaling to obtain the recommendation ranking score.

5. The content push optimization method for analyzing user interests according to claim 4, characterized in that: The recommended ranking score is combined with user activity and push saturation to determine the push content and order, as follows: Based on the cosine similarity calculation method, the recommendation score of each candidate content is obtained to form an initial recommendation ranking list; Obtain the user's historical behavior frequency, and based on big data cluster analysis methods, independently summarize the user's activity patterns in different time periods and build an activity model; Obtain the amount of content push and responses received by the user in the current period. Combined with the historical optimal push frequency, data-driven modeling techniques are used to identify the user's information saturation status for different content types and build a user push saturation model. Determine push saturation based on the user push saturation model; The initial recommendation ranking list is dynamically adjusted based on the activity model and push saturation model.

6. The content push optimization method for analyzing user interests according to claim 5, characterized in that: The similarity calculation and cluster analysis of the behavior paths are performed to obtain the user content preference path set, which is as follows: Based on the user's continuous interactive behavior, the content access sequence is extracted in chronological order; Construct a behavior path based on the content access sequence to obtain an initial behavior path set; Based on the content embedding model, each content node in each path is vectorized and weighted fusion is performed based on 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 the path similarity matrix; According to the path similarity matrix, the path vector set is classified by the K-means clustering algorithm to obtain the preferred path set.

7. The content push optimization method for analyzing user interests according to claim 6, characterized in that: The response data of the similar historical content in the similar path is obtained and approximated to obtain the response estimation result, which is specifically as follows: A unique identifier is retrieved for each recommended content item to determine whether it 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; 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 preferred path set, the path samples containing the top K most similar historical content items are screened out, and a similar path set is constructed; Obtain user response data for the corresponding content of each similar path; The user response data is normalized and weighted averaged in a set of similar paths to generate an approximate response estimate and obtain the response prediction result.

8. A system using the content push optimization method for analyzing user interests according to any one of claims 1 to 7, characterized in that: It includes data acquisition module, matching module, push module and feedback correction module, and there are connections between modules; The data collection module is used to obtain user behavior data, construct user interest vectors, and perform weighted fusion to generate the current interest portrait; The matching module is used to vectorize the content to be pushed, match it with the current interest profile, and obtain the recommendation ranking score; The push module is used to determine the push content and order based on the recommendation ranking score combined with user activity and push saturation; The feedback correction module is used to extract user behavior path clustering and content similarity based on user feedback data, output response prediction results, and dynamically adjust content recommendation priorities based on the response prediction results.

Images