Method and System for Personalized Content Recommendation and Behavior Analysis of Media Convergence Users

Abstract

The present invention provides a method and system for personalized content recommendation and behavior analysis of media convergence users, which relates to the field of big data technology. It includes obtaining historical browsing data of users to construct a behavior feature vector containing content preferences and time preferences, processing the feature vector by using a dual attention mechanism to obtain a fused feature vector, obtaining a user interest prediction result based on the fused feature vector through a learnable transformation matrix combined with a gated recurrent unit, establishing a multi-objective balanced recommendation strategy in combination with the propagation trend of real-time hot events, and screening and pushing the content to be recommended from the content library. The present invention can improve the timely push effect of hot events while ensuring the personalization of the recommended content, and enhance the user experience.

Description

Technical Field

[0001] The present invention relates to the technology of media convergence, and particularly to a method and system for personalized content recommendation and behavior analysis of media convergence users. Background Art

[0002] With the rapid development of Internet technology and mobile terminals, media convergence platforms have become an important channel for users to obtain information. In order to improve the user experience and the operation efficiency of the platform, personalized content recommendation systems play an increasingly important role in media convergence platforms. Traditional media convergence recommendation systems mainly rely on methods such as collaborative filtering and content feature matching. By analyzing the historical behavior data of users, they predict user interests and make content recommendations accordingly.

[0003] First of all, the existing technologies often adopt simple feature splicing or weighted average methods when dealing with user behavior characteristics, and are unable to effectively capture the correlation between different types of content and the dynamic change characteristics of user interests over time, resulting in insufficient accuracy and personalization of the recommendation results.

[0004] Secondly, when traditional recommendation methods build user interest models, they rarely consider the influence of temporal information and fail to fully utilize the behavior pattern characteristics of users at different time periods. This makes it difficult for the system to accurately grasp the evolution law of user interests and affects the recommendation effect.

[0005] Finally, when existing recommendation systems deal with hot events, it is often difficult to achieve a good balance between personalized recommendation and hot event dissemination. Either they overemphasize personalization and ignore the timely push of hot content, or they blindly pursue hot event dissemination at the expense of the personalized experience of the recommendation, lacking an effective multi-objective balance mechanism. Summary of the Invention

[0006] Embodiments of the present invention provide a method and system for personalized content recommendation and behavior analysis of media convergence users, which can solve the problems in the existing technologies.

[0007] In the first aspect of the embodiments of the present invention,

[0008] A method for personalized content recommendation and behavior analysis of media convergence users is provided, including:

[0009] Obtain the historical browsing data of users on the media convergence platform. Based on the historical browsing data, construct a user behavior feature vector, where the user behavior feature vector includes content preference features and time preference features. The content preference features are obtained by performing word vector encoding on content classification, and the time preference features are calculated based on the temporal distribution of the viewing time period;

[0010] The user behavior feature vector is processed using a dual attention mechanism. The importance weights of different content types are calculated based on the content-level attention mechanism, and the dynamic weights of different time windows are calculated based on the temporal-level attention mechanism. The attention weights at the two levels are fused with the user behavior feature vector to obtain a fused feature vector;

[0011] Based on the fused feature vector, a hidden state feature representing the dynamic change of the user's interest is obtained through a learnable transformation matrix combined with a gated recurrent unit, and the user interest prediction result is obtained according to the hierarchical prediction probability;

[0012] According to the user interest prediction result, combined with the propagation trend of real-time hot events, a multi-objective balanced recommendation strategy is established. The multi-objective balanced recommendation strategy improves the timely push effect of hot events while ensuring the personalization of the recommended content by dynamically adjusting the content timeliness and user interest matching degree;

[0013] According to the multi-objective balanced recommendation strategy, the content to be recommended is screened from the media integration content library, and the content to be recommended is pushed to the user according to the preset display rules.

[0014] Based on the historical browsing data, constructing a user behavior feature vector includes:

[0015] The content preference feature is obtained by word vector encoding of content classification, including: constructing a multi-level content classification system, hierarchically dividing the content according to theme, domain, and specific category; using a hierarchical attention mechanism to calculate semantic features at different levels respectively, and adaptively fusing the features at each level based on the attention weights to generate the content preference feature;

[0016] The time preference feature is calculated based on the temporal distribution of the viewing time period, including: constructing multiple temporal distribution matrices based on the viewing time period, and using a time decay function to perform weighted calculation on the temporal distribution matrices of different time windows to generate the time preference feature.

[0017] Calculating the importance weights of different content types based on the content-level attention mechanism, calculating the dynamic weights of different time windows based on the temporal-level attention mechanism, and fusing the attention weights at the two levels with the user behavior feature vector to obtain a fused feature vector, including:

[0018] Construct the content preference feature into a content feature matrix, perform dimension mapping on the content feature matrix to obtain the content hidden layer feature, calculate the attention score based on the content hidden layer feature, and convert the attention score into the importance weight of the content type through the softmax function. The importance weight is used to identify the content types that have an impact on the user interest expression;

[0019] Divide the user's behavior sequence into multiple time windows at preset time intervals, construct a time series feature matrix, calculate the attention coefficient of each time window by combining the hidden state information of historical behaviors, and convert the attention coefficient into the dynamic weight of the time window through the softmax function. The dynamic weight is used to characterize the influence degree of different time windows on the user's current interest;

[0020] Perform weighted fusion on the importance weight of the content type and the content feature matrix to obtain content weighted features, and perform weighted fusion on the dynamic weight of the time window and the time series feature matrix to obtain time series weighted features;

[0021] Perform feature concatenation on the content weighted features and the time series weighted features, and perform residual connection on the concatenated features through a learnable fusion weight matrix to obtain a fusion feature vector.

[0022] Based on the fusion feature vector, obtain the hidden state feature representing the dynamic change of the user's interest through a learnable transformation matrix combined with a gated recurrent unit, and obtain the user interest prediction result according to the hierarchical prediction probability, including:

[0023] Calculate the interaction matrix between different dimensions of the fusion feature vector, and the interaction matrix is used to characterize the correlation strength between different feature dimensions;

[0024] Map the interaction matrix to a query matrix, a key matrix, and a value matrix through a learnable transformation matrix, calculate the similarity between the query matrix and the key matrix to obtain the attention distribution, and weight the attention distribution with the value matrix to obtain the attention enhanced feature;

[0025] Input the attention enhanced feature into the gated recurrent unit, control the forgetting degree of historical information through the reset gate, and adjust the update ratio of the current state through the update gate to obtain the hidden state feature representing the dynamic change of the user's interest;

[0026] Calculate the coarse-grained prediction probability based on the hidden state feature through the sigmoid function, calculate the fine-grained prediction probability based on the combination of the hidden state feature and the attention enhanced feature through the sigmoid function, and calculate the fusion weight coefficient based on the predicted confidence;

[0027] Perform weighted combination on the coarse-grained prediction probability and the fine-grained prediction probability through the fusion weight coefficient to obtain the final user interest prediction result.

[0028] Before establishing a multi-objective balanced recommendation strategy according to the user interest prediction result and combining the propagation trend of real-time hot events, the method further includes determining the propagation trend of real-time hot events:

[0029] Obtain hot event data from multiple monitoring platforms, where the hot event data includes platform access data, user interaction data, and propagation path data; extract the access volume at each time point, multiply the access volume by the preset weight coefficient of the corresponding monitoring platform, and accumulate the weighted access volumes of all monitoring platforms to obtain the access volume feature of the hot event data. The preset weight coefficient of the monitoring platform is determined based on the platform user scale and platform activity;

[0030] Construct a propagation hierarchy tree based on the propagation path data, identify the propagation level to which each propagation node belongs, count the number of propagation nodes in each propagation level, multiply the number of propagation nodes by the corresponding level influence factor of that level, and perform normalization processing on the calculation results of all levels to obtain the propagation depth feature of the hot event data;

[0031] Calculate the normalized values of the access volume feature and the propagation depth feature respectively, substitute each normalized value into an exponential function to obtain the corresponding adaptive weight coefficient, and perform weighted combination on the access volume feature and the propagation depth feature based on the adaptive weight coefficient to obtain the propagation potential energy of the hot event data;

[0032] Obtain the occurrence time of the hot event data, calculate the time difference between the current time and the occurrence time, calculate the change rate of the propagation potential energy per unit time, dynamically adjust the preset initial decay coefficient according to the change rate to obtain the real-time decay coefficient, and multiply the propagation potential energy by the real-time decay coefficient to obtain the final propagation potential energy considering the time decay effect.

[0033] According to the user interest prediction result, combined with the propagation trend of real-time hot events, establish a multi-objective balanced recommendation strategy including:

[0034] Obtain the data of the propagation trend of real-time hot events and the recommendation delay data, input the data of the propagation potential energy and the recommendation delay data into a time decay function, calculate to obtain the event timeliness score, and use the product of the event timeliness score and the preset timeliness weight as the timeliness feature;

[0035] Extract the user semantic vector, user content feature vector, and user topic label set from the user historical behavior data; extract the event semantic vector, event content feature vector, and event topic label set from the real-time hot event;

[0036] Calculate the first cosine similarity between the user semantic vector and the event semantic vector, the second cosine similarity between the user content feature vector and the event content feature vector, and the overlap degree between the user topic label set and the event topic label set respectively, and perform weighted combination on the first cosine similarity, the second cosine similarity, and the overlap degree to obtain the user interest matching degree;

[0037] Obtain the real-time click data of the user on the recommended content, calculate the real-time click-through rate, input the difference between the real-time click-through rate and the preset benchmark click-through rate into the sigmoid function to obtain a dynamic weight coefficient, and the dynamic weight coefficient is used to balance the timeliness feature and the user interest matching degree;

[0038] Calculate the content similarity between the event to be recommended and the historical recommended events based on the user's historical recommendation records, construct the content similarity as a diversity constraint term, and combine the diversity constraint term with the timeliness feature and the user interest matching degree weighted by the dynamic weight coefficient to establish a multi-objective balanced recommendation strategy.

[0039] The method further includes:

[0040] Obtain the user feedback data, extract the feedback timestamp and the feedback intensity value, calculate the feedback influence factor based on the time decay function, adaptively adjust the optimization learning rate using the feedback influence factor, and use the stochastic gradient descent method to iteratively optimize the parameters of the multi-objective balanced recommendation strategy to obtain the final recommendation result.

[0041] In the second aspect of the embodiments of the present invention,

[0042] Provide a personalized content recommendation and behavior analysis system for media convergence users, including:

[0043] A first unit for obtaining the historical browsing data of the user on the media convergence platform, and constructing a user behavior feature vector based on the historical browsing data. The user behavior feature vector includes a content preference feature and a time preference feature, where the content preference feature is obtained by performing word vector encoding on the content classification, and the time preference feature is calculated based on the temporal distribution of the viewing time period;

[0044] A second unit for processing the user behavior feature vector by using a dual attention mechanism, calculating the importance weights of different content types based on the content-level attention mechanism, calculating the dynamic weights of different time windows based on the temporal-level attention mechanism, and fusing the two levels of attention weights with the user behavior feature vector to obtain a fused feature vector;

[0045] The third unit is used to obtain a hidden state feature representing the dynamic change of the user's interest through a learnable transformation matrix combined with a gated recurrent unit based on the fused feature vector, and obtain a user interest prediction result according to the hierarchical prediction probability; according to the user interest prediction result, combined with the propagation trend of real-time hot events, a multi-objective balanced recommendation strategy is established, where the multi-objective balanced recommendation strategy improves the timely push effect of hot events while ensuring the personalization of the recommended content by dynamically adjusting the content timeliness and user interest matching degree; according to the multi-objective balanced recommendation strategy, the content to be recommended is screened from the media convergence content library, and the content to be recommended is pushed to the user according to the preset display rules.

[0046] In the third aspect of the embodiments of the present invention,

[0047] A kind of electronic device is provided, including:

[0048] A processor;

[0049] A memory for storing instructions executable by the processor;

[0050] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0051] In the fourth aspect of the embodiments of the present invention,

[0052] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0053] The beneficial effects of this application are as follows:

[0054] By constructing a user behavior feature vector containing content preference and time preference, and adopting a dual attention mechanism for feature fusion, the present invention can more comprehensively depict the user's interest characteristics and improve the accuracy of personalized recommendation.

[0055] The present invention introduces a gated recurrent unit to capture the dynamic change characteristics of the user's interest, and conducts interest prediction through hierarchical prediction probability, enabling the recommendation system to timely perceive the evolution law of the user's interest and improving the timeliness of the recommended content and the user experience.

[0056] By establishing a multi-objective balanced recommendation strategy, the present invention takes into account the timely push of hot events while ensuring the personalized recommendation effect, realizes the dynamic balance between personalization and hot event push, and effectively improves the information distribution efficiency and user satisfaction of the media convergence platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic flowchart of the method for personalized content recommendation and behavior analysis of media convergence users in the embodiments of the present invention;

[0058] Figure 2 Schematic diagram for comparing the user interest recognition effect of the embodiments of the present invention;

[0059] Figure 3 Schematic diagram for comparing the performance of double attention feature fusion of the embodiments of the present invention;

[0060] Figure 4 Schematic diagram for multi-dimensional evaluation of the user interest prediction performance of the embodiments of the present invention;

[0061] Figure 5 Schematic diagram of the multi-objective balanced recommendation system of the embodiments of the present invention;

[0062] Figure 6 Schematic diagram of the structure of the personalized content recommendation and behavior analysis system for media convergence users of the embodiments of the present invention. Detailed implementation manners

[0063] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to 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. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0064] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0065] Figure 1 Schematic flow chart of the personalized content recommendation and behavior analysis method for media convergence users of the embodiments of the present invention, as Figure 1 shown, the method includes:

[0066] Obtain the historical browsing data of the user on the media convergence platform, and based on the historical browsing data, construct a user behavior feature vector, where the user behavior feature vector includes content preference features and time preference features, and the content preference features are obtained by word vector encoding of content classification, and the time preference features are calculated based on the temporal distribution of the viewing time period;

[0067] Process the user behavior feature vector by using a dual attention mechanism, calculate the importance weights of different content types based on the content-level attention mechanism, calculate the dynamic weights of different time windows based on the temporal-level attention mechanism, and fuse the two levels of attention weights with the user behavior feature vector to obtain a fused feature vector;

[0068] Based on the fused feature vector, a hidden state feature representing the dynamic change of the user's interest is obtained through a learnable transformation matrix combined with a gated recurrent unit, and a user interest prediction result is obtained according to the hierarchical prediction probability; according to the user interest prediction result, combined with the propagation trend of real-time hot events, a multi-objective balanced recommendation strategy is established, where the multi-objective balanced recommendation strategy dynamically adjusts the content timeliness and user interest matching degree, and while ensuring the personalization of the recommended content, improves the timely push effect of hot events; according to the multi-objective balanced recommendation strategy, the content to be recommended is screened from the media integration content library, and the content to be recommended is pushed to the user according to the preset display rules.

[0069] In an optional implementation manner, constructing a user behavior feature vector based on the historical browsing data includes:

[0070] The content preference feature is obtained by performing word vector encoding on the content classification, including: constructing a multi-level content classification system, and hierarchically dividing the content according to the theme, field, and specific category; adopting a hierarchical attention mechanism to calculate the semantic features of different levels respectively, and adaptively fusing the features of each level based on the attention weights to generate the content preference feature;

[0071] The time preference feature is calculated based on the temporal distribution of the viewing time period, including: constructing multiple temporal distribution matrices based on the viewing time period, and using a time decay function to perform weighted calculation on the temporal distribution matrices of different time windows to generate the time preference feature.

[0072] First, obtain the user's historical browsing data, including browsing content information and browsing time information. For the content information, construct a multi-level content classification system: the first layer is the theme classification, such as technology, culture, entertainment, etc.; the second layer is the field classification, such as under technology, it is subdivided into artificial intelligence, Internet, electronics, etc.; the third layer is the specific category, such as under artificial intelligence, it is subdivided into machine learning, deep learning, computer vision, etc.

[0073] For each layer of classification, use the word vector encoding method to convert the category name into a word vector. Taking "artificial intelligence - machine learning - neural network" as an example, obtain the word vectors of the three layers respectively. Adopt a hierarchical attention mechanism to calculate the user's attention to the content of different levels. Specifically, the proportion of the browsing duration of the user in each level category is statistically used as the attention weight, such as the proportion of technology in the theme layer is 0.6, the proportion of artificial intelligence in the field layer is 0.3, and the proportion of machine learning in the specific category layer is 0.1. Based on the attention weights, the word vectors of the three layers are weighted and fused to obtain the content preference feature vector.

[0074] For time information, divide 24 hours of a day into multiple time windows, such as one window every 2 hours. Count the browsing behaviors of users in each time window and construct a time series distribution matrix. Use a time decay function to assign different weights to different time windows, with the weight being larger for time windows closer to the current time. For example, the weight of data in the most recent week is 0.5, the weight of data in the previous week is 0.3, and the weight of earlier data is 0.2. Weight the distribution matrix of each time window to obtain a time preference feature vector.

[0075] Finally, concatenate the content preference feature vector and the time preference feature vector to construct a complete user behavior feature vector. This feature vector can be used for subsequent personalized recommendation and other applications.

[0076] Figure 2 This is a schematic diagram for comparing the user interest recognition effects of the embodiments of the present invention:

[0077] This figure shows the performance comparison of three different methods. Among them, the circle (●) represents the technical solution of the present invention that combines multi-level content classification and hierarchical attention mechanism, the square (■) represents the traditional collaborative filtering algorithm based on user historical behavior similarity, and the triangle (▲) represents the basic deep learning recommendation method using a simple neural network structure. During the 0-10 iteration process, the accuracy of the technical solution of the present invention increases from 0.82 to 0.95, with a total increase of 0.13, and reaches a high accuracy of 0.89 at the 4th iteration; the traditional collaborative filtering method increases from 0.75 to 0.82, with an increase of 0.07; the basic deep learning method increases from 0.78 to 0.85, also with an increase of 0.07. The technical solution of the present invention not only has the highest initial accuracy but also maintains a leading advantage throughout the iteration process. Especially in the 6-8 iteration stage, the accuracy rapidly increases from 0.92 to 0.94, showing excellent performance improvement ability. From the overall trend of the curve, the performance gap between the technical solution of the present invention and the other two methods gradually expands with the increase of the iteration times, fully demonstrating that this solution has significant advantages in capturing users' initial interest characteristics and continuously optimizing the model recognition ability, and has higher practical value.

[0078] Existing user interest recognition methods are mainly based on single behavior similarity calculation or simple deep learning models, and these methods have obvious deficiencies in dealing with the diversity and dynamic evolution characteristics of user interests. Traditional collaborative filtering methods only focus on the historical behavior similarity of users and cannot deeply understand content features; although the basic deep learning method introduces a neural network structure, it fails to fully utilize the hierarchical features of content, resulting in poor accuracy and adaptability of interest recognition.

[0079] Starting from improving the accuracy and generalization ability of user interest recognition, this application innovatively proposes a technical solution that combines multi-level content classification with a hierarchical attention mechanism. By constructing a multi-level content classification system, the semantic features at different levels such as the theme, domain, and specific category of the content are systematically characterized; at the same time, a hierarchical attention mechanism is introduced to adaptively calculate the importance weights of different-level features, realizing the dynamic fusion of features. This method of hierarchical recognition and adaptive fusion significantly improves the depth of the model's understanding of user interests and the accuracy of recognition.

[0080] The experimental results show that compared with the existing technologies, this solution exhibits better recognition effects at the initial stage, and with the progress of the iterative optimization process, both the improvement amplitude and the convergence speed of the recognition accuracy have obvious advantages. Especially after multiple rounds of iteration, the performance gap between this solution and the existing methods further expands, fully verifying the significant effects of this technical solution in capturing user interest features and continuously optimizing the model performance. This advantage mainly benefits from the refined characterization of user interests by the multi-level content classification system and the dynamic adjustment ability of the hierarchical attention mechanism in the feature fusion process, enabling the model to more accurately identify and predict the interest preferences of users, and having important application value.

[0081] In an optional implementation manner, based on the content-level attention mechanism, calculate the importance weights of different content types, and based on the time-series-level attention mechanism, calculate the dynamic weights of different time windows. Fuse the attention weights at the two levels with the user behavior feature vector to obtain a fused feature vector, including:

[0082] Construct the content preference features into a content feature matrix, perform dimensional mapping on the content feature matrix to obtain content hidden layer features, calculate attention scores based on the content hidden layer features, and convert the attention scores into the importance weights of content types through the softmax function. The importance weights are used to identify the content types that have an impact on user interest expression;

[0083] Divide the user's behavior sequence into multiple time windows at a preset time interval, construct a time-series feature matrix, calculate the attention coefficient of each time window in combination with the hidden state information of historical behaviors, and convert the attention coefficient into the dynamic weight of the time window through the softmax function. The dynamic weight is used to characterize the influence degree of different time windows on the user's current interest;

[0084] Perform weighted fusion of the importance weights of the content types with the content feature matrix to obtain content weighted features, and perform weighted fusion of the dynamic weights of the time windows with the time-series feature matrix to obtain time-series weighted features;

[0085] Concatenate the weighted content features and the weighted temporal features, and perform residual connection on the concatenated features through a learnable fusion weight matrix to obtain a fused feature vector.

[0086] First, obtain the user's behavior data on the content platform, including interaction behaviors such as clicks, views, and collections of different types of content. Preprocess the original behavior data to extract key information such as user ID, content ID, content type, behavior type, and timestamp.

[0087] In the content-level attention mechanism, construct a feature matrix for different content types. Taking news content as an example, extract text features such as the title, body text, and keywords, and convert the text into a dense vector representation through methods such as word2vec to obtain a content feature vector with a dimension of 100. For video content, extract features such as the video title, description text, and video duration. Concatenate the feature vectors of different types of content to construct a content feature matrix.

[0088] Perform dimensionality mapping on the content feature matrix. Use a two-layer fully connected network with the input layer dimension the same as the original feature dimension, the hidden layer dimension set to 64, and the output layer dimension to 32 to obtain a more compact representation of the content hidden layer features. Calculate the attention scores based on the content hidden layer features, specifically obtaining the attention coefficients through a non-linear transformation. Normalize the attention coefficients through the softmax function to obtain the importance weights of different content types.

[0089] In the temporal-level attention mechanism, divide the user's behavior sequence in the past 3 months into time windows of 7 days each, obtaining 12 time windows. For the behavior sequence within each time window, extract features such as behavior type and content type to construct a temporal feature matrix. Combine the historical behavior hidden state information obtained by the GRU network to calculate the attention coefficients for each time window. Normalize the attention coefficients through the softmax function to obtain the dynamic weights of different time windows.

[0090] Multiply the content type importance weights by the content feature matrix to obtain the weighted content features. Multiply the time window dynamic weights by the temporal feature matrix to obtain the weighted temporal features. Concatenate the two weighted features to obtain a 256-dimensional concatenated feature. Perform residual connection on the concatenated feature through a learnable fusion weight matrix to obtain the final fused feature vector.

[0091] Taking 100 pieces of content recently viewed by a certain user as an example, among which there are 40 news items, 30 videos, and 30 picture-and-text items. For news content, features such as the title word vector (100 dimensions), the body keyword vector (100 dimensions), and the news classification vector (20 dimensions) are extracted; for video content, the video title vector (100 dimensions), the video description vector (100 dimensions), the duration feature (1 dimension), etc. are extracted. These features are concatenated to construct a 220-dimensional content feature matrix.

[0092] The importance weights of content types calculated through the attention mechanism are: news 0.5, video 0.3, picture-and-text 0.2, indicating that the user is more interested in news content. The dynamic weights of the 12 time windows of the time series features are in order from far to near as follows:

[0093] 0.02, 0.03, 0.05, 0.06, 0.07, 0.08, 0.09, 0.1, 0.12, 0.13, 0.14, 0.11, reflecting the time series evolution characteristics of the user's interests.

[0094] Figure 3 This is a schematic diagram for comparing the performance of double attention feature fusion in the embodiments of the present invention:

[0095] The figure shows the performance of three different feature fusion methods. The circle (●) represents the proposed technical solution, which adopts a two-layer attention mechanism at the content level and the temporal level. It obtains the content hidden layer features by performing dimensional mapping on the content feature matrix and calculates the importance weights of content types. At the same time, the user behavior sequence is divided into multiple time windows to calculate the dynamic weights, and finally, feature fusion is achieved through residual connection. The square (■) represents the single-layer attention method, which only uses the content-level attention mechanism to calculate the content type weights and does not consider the dynamic changes of temporal features. The triangle (▲) represents the simple concatenation method, which directly concatenates the content features and the temporal features without introducing an attention mechanism for feature selection and weighting. During the 0-10 iteration process, the accuracy of the proposed technical solution increases from 0.85 to 0.95, with a total increase of 0.10, and reaches a high accuracy of 0.90 at the 4th iteration. The single-layer attention method increases from 0.78 to 0.85, with an increase of 0.07. The simple concatenation method increases from 0.72 to 0.80, with an increase of 0.08. By combining the two-layer attention mechanism at the content level and the temporal level, the proposed technical solution not only has the highest initial accuracy but also maintains a significant leading advantage throughout the iteration process. Especially in the 6-8 iteration stage, the accuracy rapidly increases from 0.92 to 0.94, demonstrating excellent feature fusion capabilities. From the overall trend of the curve, the performance gap between the proposed technical solution and the other two methods gradually expands as the number of iterations increases, indicating that the two-layer attention mechanism has a significant effect on improving the feature expression ability and model performance. Compared with the single-layer attention and simple concatenation methods, it can better capture the multi-dimensional information of user interest features and achieve more accurate feature fusion.

[0096] Existing feature fusion methods mainly adopt a single-layer attention mechanism or simple feature concatenation. The single-layer attention method only focuses on calculating the importance weights of content features and cannot effectively capture the temporal evolution features of user interests. The simple concatenation method directly concatenates the content features and the temporal features, lacking a mechanism for distinguishing and selecting the importance of different features, resulting in unsatisfactory feature fusion effects and difficulty in accurately expressing the multi-dimensional interest features of users.

[0097] Starting from improving the accuracy of feature fusion and the ability of temporal modeling, this application innovatively proposes a double-layer attention feature fusion scheme. At the content-level attention layer, hidden features are obtained by dimension mapping of the content feature matrix, and the importance weights of content types are calculated based on this, realizing accurate modeling of user content preferences; at the temporal-level attention layer, the user behavior sequence is divided into multiple time windows, and the dynamic weights of the time windows are calculated in combination with the hidden state information of historical behaviors, effectively capturing the temporal evolution law of user interests. Finally, through the residual connection mechanism, the weighted features of the two levels are adaptively fused, which not only retains the original feature information but also introduces non-linear transformation to enhance the feature expression ability.

[0098] Experimental results show that compared with existing technologies, this scheme has significant advantages in feature fusion effects. This advantage is mainly reflected in three aspects: First, the double-layer attention mechanism can more accurately identify important features, improving the accuracy of feature fusion; second, the introduction of temporal-level attention enables the model to effectively model the dynamic changes of user interests; finally, the fusion method of residual connection enhances the feature expression ability of the model. These improvements enable this scheme to show better performance and faster convergence speed in the user interest recognition task, and have important application value.

[0099] In an optional implementation manner, based on the fusion feature vector, a hidden state feature representing the dynamic changes of user interests is obtained through a learnable transformation matrix combined with a gated recurrent unit, and the user interest prediction result is obtained according to the hierarchical prediction probability, including:

[0100] Calculate the interaction matrix between different dimensions of the fusion feature vector, and the interaction matrix is used to represent the correlation strength between different feature dimensions;

[0101] Map the interaction matrix to a query matrix, a key matrix, and a value matrix through a learnable transformation matrix, calculate the similarity between the query matrix and the key matrix to obtain an attention distribution, and weight the attention distribution with the value matrix to obtain an attention-enhanced feature;

[0102] Input the attention-enhanced feature into the gated recurrent unit, control the forgetting degree of historical information through the reset gate, and adjust the update ratio of the current state through the update gate to obtain a hidden state feature representing the dynamic changes of user interests;

[0103] Calculate the coarse-grained prediction probability based on the hidden state feature through the sigmoid function, calculate the fine-grained prediction probability based on the combination of the hidden state feature and the attention-enhanced feature through the sigmoid function, and calculate the fusion weight coefficient based on the predicted confidence;

[0104] The coarse-grained prediction probability and the fine-grained prediction probability are weighted and combined through the fusion weight coefficient to obtain the final user interest prediction result.

[0105] First, a fused feature vector is obtained, which contains multi-dimensional feature information such as user historical behavior and item attributes. By calculating the interaction relationship between different dimensional features, an interaction matrix is generated. Specifically, for any two-dimensional features in the fused feature vector, the dot product similarity between them is calculated to obtain an interaction matrix representing the correlation strength between feature dimensions. For example, for a fused feature vector containing 128 dimensions, a 128×128 interaction matrix is generated.

[0106] Next, a learnable transformation matrix is introduced to perform a mapping transformation on the interaction matrix. The transformation matrix contains three independent matrices, which are used to generate a query matrix, a key matrix, and a value matrix respectively. Through matrix multiplication operations, the interaction matrix is mapped into three matrices of the same dimension. The dot product similarity between the query matrix and the key matrix is calculated and normalized to obtain an attention distribution. The attention distribution and the value matrix are weighted and summed to obtain an attention-enhanced feature.

[0107] The attention-enhanced feature is input into a gated recurrent unit. The gated recurrent unit contains two gating mechanisms: a reset gate and an update gate. The reset gate calculates a forgetting coefficient based on the current input feature and the historical state, controlling the retention degree of historical information. The update gate calculates an update ratio, which is used to adjust the update amplitude of the current state. Through the adjustment of the gating mechanism, a hidden state feature that can represent the dynamic change of user interest is obtained.

[0108] Based on the hidden state feature, the coarse-grained user interest prediction probability is calculated through the sigmoid function. At the same time, the hidden state feature and the attention-enhanced feature are concatenated and combined, and then the fine-grained prediction probability is calculated through the sigmoid function. The fusion weight coefficient is calculated according to the confidence of the prediction result, and the prediction result with higher confidence obtains a greater weight. Finally, the coarse-grained and fine-grained prediction probabilities are weighted and combined through the weight coefficient to obtain the final user interest prediction result.

[0109] Taking specific data as an example: Assume that the dimension of the fused feature vector is 128, and a 128×128 interaction matrix is generated. It is mapped through a 64×128-dimensional transformation matrix to obtain 64-dimensional query, key, and value matrices. A 64-dimensional attention distribution vector is calculated and weighted with the value matrix to obtain a 64-dimensional attention-enhanced feature. The gated recurrent unit outputs a 32-dimensional hidden state feature. Finally, a prediction probability value between 0 and 1 is output.

[0110] Figure 4 This is a schematic diagram for multi-dimensional evaluation of the user interest prediction performance in the embodiments of the present invention:

[0111] The figure shows the performance comparison of three different prediction methods. The circle (●) represents the technical solution of the present invention, which maps interaction features using a learnable transformation matrix, combines a gated recurrent unit to achieve dynamic modeling, and obtains results through hierarchical prediction; the square (■) represents Comparative Method 1, which only uses a single-layer attention mechanism to calculate feature weights for prediction; the triangle (▲) represents Comparative Method 2, which uses simple feature splicing and a basic neural network structure. From the evaluation indicators in six dimensions, the technical solution of the present invention is significantly better than Comparative Method 1 (0.85, 0.83, 0.84, 0.86, 0.82, 0.81 respectively) and Comparative Method 2 (0.75, 0.74, 0.76, 0.77, 0.73, 0.72 respectively) in terms of accuracy (0.95), recall rate (0.92), F1 value (0.93), AUC (0.94), stability (0.91) and convergence speed (0.90). From the overall shape of the radar chart, it can be seen that the technical solution of the present invention has obvious advantages in all dimensions, forming the outermost hexagon, and the performance in each dimension is balanced, fully verifying the effectiveness of the double-layer attention mechanism and the hierarchical prediction strategy.

[0112] By calculating the interaction matrix between feature dimensions and introducing the attention mechanism, the correlation relationship between different features can be effectively captured, and the feature expression ability can be improved. The gated recurrent unit is used to model the user interest. Through the adjustment mechanism of the reset gate and the update gate, both the historical interest information is retained and the change of the user interest can be responded to in a timely manner. Combining the coarse-grained and fine-grained hierarchical prediction methods, and adaptively adjusting the weights through confidence, the user interest can be characterized at different granularity levels, and the accuracy and robustness of the prediction can be improved.

[0113] In an alternative embodiment, before establishing a multi-objective balanced recommendation strategy based on the predicted result of the user interest and in combination with the propagation trend of real-time hot events, the method further includes determining the propagation trend of the real-time hot events:

[0114] Obtain the hot event data of multiple monitoring platforms, where the hot event data includes platform access data, user interaction data, and propagation path data; extract the access volume at each time point, multiply the access volume by the preset weight coefficient of the corresponding monitoring platform, and accumulate the weighted access volumes of all monitoring platforms to obtain the access volume feature of the hot event data. The preset weight coefficient of the monitoring platform is determined based on the platform user scale and platform activity;

[0115] Based on the propagation path data, construct a propagation hierarchy tree, identify the propagation hierarchy to which each propagation node belongs, count the number of propagation nodes in each propagation hierarchy, multiply the number of propagation nodes by the corresponding hierarchy influence factor of the hierarchy, and perform normalization processing on the calculation results of all hierarchies to obtain the propagation depth feature of the hot event data;

[0116] Calculate the normalized values of the access volume feature and the propagation depth feature respectively, substitute each normalized value into the exponential function to obtain the corresponding adaptive weight coefficient, and perform a weighted combination of the access volume feature and the propagation depth feature based on the adaptive weight coefficient to obtain the propagation potential of the hot event data;

[0117] Obtain the occurrence time of the hot event data, calculate the time difference between the current time and the occurrence time, calculate the change rate of the propagation potential per unit time, dynamically adjust the preset initial attenuation coefficient according to the change rate to obtain the real-time attenuation coefficient, and multiply the propagation potential by the real-time attenuation coefficient to obtain the final propagation potential considering the time decay effect.

[0118] Obtain hot event data from multiple mainstream content platforms, including interactive data such as access volume, user comments, and forwards, as well as the content propagation path. For different platforms, set the weight coefficient based on their user scale and daily active user number. The weight of platforms with a user scale of over 10 million and a daily activity of over 30% is 0.4, the weight of platforms with a user scale of 5 - 10 million and a daily activity of 20 - 30% is 0.3, and the weight of other platforms is 0.2.

[0119] For each hot event, collect the real-time access volume data of each platform at 10 - minute intervals. Multiply the access volume at each time point by the weight coefficient of the corresponding platform and accumulate to obtain the weighted access volume. Perform normalization processing on the weighted access volume sequence for 24 consecutive hours to obtain the access volume feature reflecting the event attention.

[0120] Construct an event propagation hierarchy tree, with the starting release node as layer 0, direct forwards as layer 1, forwards of forwards as layer 2, and so on. Count the number of propagation nodes in each layer, and set the layer influence factor: 1.0 for layer 0, 0.8 for layer 1, 0.6 for layer 2, 0.4 for layer 3, and 0.2 for layer 4 and above. Multiply the number of nodes in each layer by the corresponding influence factor and normalize to obtain the propagation depth feature.

[0121] Perform 0 - 1 normalization on the access volume feature and the propagation depth feature respectively. Substitute the normalized values into the negative exponential function to calculate the adaptive weight: the larger the feature value, the larger the corresponding weight. Perform a weighted sum of the two features based on the adaptive weight to obtain the propagation potential of the hot event.

[0122] Record the first release time of the event and calculate the time difference from the current time. Set the initial attenuation coefficient to 0.9. When the change rate of the propagation potential is positive, increase the attenuation coefficient by 10%; when the change rate is negative, decrease the attenuation coefficient by 10%. Multiply the propagation potential by the real-time attenuation coefficient to obtain the final propagation potential considering time decay.

[0123] Taking a certain hot event as an example, it is spread on three major platforms. Platform A has a user scale of 12 million, a daily activity rate of 35%, and a weight of 0.4; Platform B has a user scale of 8 million, a daily activity rate of 25%, and a weight of 0.3; Platform C has a user scale of 4 million, a daily activity rate of 15%, and a weight of 0.2.

[0124] At a certain time point, the visit volumes of the three platforms are 100,000, 80,000, and 50,000 respectively. After weighting, the total visit volume is 85,000. The distribution of the propagation levels is as follows: 1 node at layer 0, 100 nodes at layer 1, 500 nodes at layer 2, 2,000 nodes at layer 3, and 5,000 nodes at layer 4. The calculated propagation depth eigenvalue is 0.75.

[0125] Two hours after the event is released, the normalized value of the visit volume feature is 0.85, and the normalized value of the propagation depth feature is 0.75. The corresponding adaptive weights are calculated to be 0.6 and 0.4 respectively. The propagation potential energy after weighted combination is 0.81. Since the propagation potential energy continues to rise, the real-time decay coefficient is adjusted from the initial 0.9 to 0.99, and the final propagation potential energy is 0.80.

[0126] Through multi-platform data fusion and differential weight allocation, the propagation data of hot events can be obtained comprehensively and accurately, improving the reliability and representativeness of the propagation trend analysis. By introducing the propagation level tree and the level influence factor, the propagation depth and influence range of events are deeply characterized, making the calculation of the propagation potential energy more reasonable and three-dimensional. By adopting the adaptive weight and the dynamic decay mechanism, the feature combination method and the time decay effect can be flexibly adjusted, enabling the propagation potential energy to better reflect the real-time propagation state and development trend of events.

[0127] In an optional implementation manner, according to the user interest prediction result and combined with the propagation trend of real-time hot events, establishing a multi-objective balanced recommendation strategy includes:

[0128] Obtaining the data of the propagation trend of real-time hot events and the recommendation delay data, inputting the propagation potential energy data and the recommendation delay data into a time decay function, calculating the event timeliness score, and taking the product of the event timeliness score and a preset timeliness weight as the timeliness feature;

[0129] Extracting the user semantic vector, the user content feature vector, and the user topic label set from the user historical behavior data; extracting the event semantic vector, the event content feature vector, and the event topic label set from the real-time hot events;

[0130] Calculate the first cosine similarity between the user semantic vector and the event semantic vector, the second cosine similarity between the user content feature vector and the event content feature vector, and the overlap degree between the user topic label set and the event topic label set, and perform weighted combination on the first cosine similarity, the second cosine similarity and the overlap degree to obtain the user interest matching degree;

[0131] Obtain the real-time click data of the user on the recommended content, calculate the real-time click-through rate, and input the difference between the real-time click-through rate and the preset benchmark click-through rate into the sigmoid function to obtain the dynamic weight coefficient, which is used to balance the timeliness feature and the user interest matching degree;

[0132] Calculate the content similarity between the event to be recommended and the historical recommended events based on the user's historical recommendation records, construct the content similarity as a diversity constraint item, and combine the diversity constraint item with the timeliness feature and the user interest matching degree weighted by the dynamic weight coefficient to establish a multi-objective balanced recommendation strategy.

[0133] Obtain the propagation potential data of real-time hot events and the latency data of the recommendation system. The recommendation latency includes system processing latency and network transmission latency, usually in the range of 0.1 - 1 second. Input the propagation potential and latency data into the exponential decay function to calculate the event timeliness score. Set the timeliness weight to 0.4, and multiply the timeliness score by the weight to obtain the timeliness feature.

[0134] From the user's historical behavior data, use the BERT model to extract the semantic vector of the user's recently browsed content, with a dimension of 768. Extract the content feature vectors of multi-modal content such as pictures, texts, videos, etc. through a deep learning model, with a dimension of 512. Based on the content classification system, construct a user interest topic label set, including multiple domain labels such as entertainment, sports, and technology.

[0135] Similarly extract the semantic vector, content feature vector and topic label set for real-time hot events. Calculate the cosine similarity of the user-event semantic vectors, with a value range of 0 - 1. Calculate the cosine similarity of the content feature vectors. Count the proportion of the number of coincidences between the user topic labels and the event topic labels to obtain the overlap degree. Set the weights of the three features to 0.4, 0.3, and 0.3 respectively, and perform weighted summation to obtain the user interest matching degree.

[0136] Real-time count the click data of the user on the recommended content, and calculate the click-through rate in the most recent 1 hour. Set the benchmark click-through rate to 5%, and input the difference between the actual click-through rate and the benchmark value into the sigmoid function to obtain a dynamic weight coefficient between 0 and 1. When the click-through rate is higher than the benchmark, the weight coefficient increases, and vice versa.

[0137] Compare the event to be recommended with the recommended content received by the user in the last 7 days, and calculate their cosine similarity in the semantic space. Take the negative of the similarity to obtain the diversity constraint term, where the higher the similarity, the stronger the constraint. Combine the diversity constraint term with the weighted timeliness feature and interest matching degree to construct the final recommendation strategy.

[0138] Take the recommendation process of a certain user for a certain hot event as an example. The event propagation potential is 0.85, the system delay is 0.2 seconds, the calculated timeliness score is 0.82, and multiplying it by the weight of 0.4 gives the timeliness feature of 0.328.

[0139] The user has browsed 100 pieces of content recently. The cosine similarity between the average semantic vector extracted and the event semantic vector is 0.75. The content feature vector similarity is 0.68. The user's topic tag set contains 8 tags, the event contains 5 tags, and 3 tags overlap, with an overlap degree of 0.6. The three features are weighted and combined to obtain the user interest matching degree of 0.686.

[0140] The recommendation click-through rate of the user in the last 1 hour is 8%, which is higher than the benchmark click-through rate of 5%. The dynamic weight coefficient of 0.65 is calculated through the sigmoid function. Searching the user's recommendation records in the last 7 days finds 20 relevant events, with an average content similarity of 0.3, which is converted into a diversity constraint term of 0.7.

[0141] Taking into account the timeliness feature (0.328), interest matching degree (0.686) and diversity constraint (0.7) comprehensively, the recommendation score of this event is finally obtained as 0.571.

[0142] Figure 5 The following is a schematic diagram of the multi-objective balanced recommendation system according to the embodiment of the present invention:

[0143] This figure shows the operation panel of the multi-objective balanced recommendation system, intuitively presenting the core functions and data analysis results. The real-time system metrics area shows the system delay of 0.2 seconds, click-through rate of 8% and dynamic weight coefficient of 0.65, indicating that the system performance is good and inclined towards the user's interests. The user interest analysis part shows the semantic similarity of 0.75, feature similarity of 0.68 and tag overlap degree of 0.6, as well as tags such as technology and entertainment that the user is concerned about. The hot event analysis area shows the propagation potential of 0.85, timeliness score of 0.82 and diversity constraint of 0.7.

[0144] The weight configuration area shows the timeliness weight of 0.4, semantic vector weight of 0.4, and content feature and theme tag weights of 0.3 each. The recommendation score composition shows the timeliness feature of 0.328, interest matching degree of 0.686, diversity constraint of 0.7, and the final score of 0.571.

[0145] The recommended list area displays three sorted events: "A domestic technology company releases a new generation of AI chips" (recommendation score 0.571), "The global technology conference is about to be held" (recommendation score 0.532), and "The country issues a new round of scientific and technological innovation support policies" (recommendation score 0.498). Each shows the details of three indicators: timeliness, interest matching degree, and diversity. The overall design visualizes the complex algorithm process to help operators understand the recommendation mechanism and optimize the content distribution effect.

[0146] By introducing the timeliness feature and the dynamic weight mechanism, the recommendation strategy can respond in a timely manner to the changes in the spread trend of hot events, avoiding over-recommending hot content while ensuring the timeliness of information. The multi-dimensional feature extraction and similarity calculation method comprehensively depict the matching degree between user interests and event content, improving the accuracy of personalized recommendations and the user experience. The diversity constraint based on historical recommendation records effectively avoids the problem of content homogenization, making the recommendation results more diverse, and enhancing the exploration ability and user stickiness of the recommendation system.

[0147] In an alternative implementation, the method further includes:

[0148] Obtain user feedback data, extract the feedback timestamp and feedback intensity value, calculate the feedback influence factor based on the time decay function, adaptively adjust the optimization learning rate using the feedback influence factor, and use the stochastic gradient descent method to iteratively optimize the parameters of the multi-objective balanced recommendation strategy to obtain the final recommendation result.

[0149] First, obtain user feedback data, including user behavior data such as clicks, collections, and comments on the recommended content. Extract the feedback timestamp information from the feedback data to record the specific time point when each feedback occurs. At the same time, extract the feedback intensity value. For example, a click behavior is assigned a value of 0.2, a collection behavior is assigned a value of 0.5, and a comment behavior is assigned a value of 0.8. The larger the value, the higher the feedback intensity.

[0150] Next, calculate the feedback influence factor based on the time decay function. Adopt an exponential decay method, where the weight of feedback data farther from the current time is smaller. For example, for a user's click behavior 2 days ago, its influence factor is 0.2 multiplied by the square of 0.8; for a collection behavior 7 days ago, the influence factor is 0.5 multiplied by the 7th power of 0.8. Through time decay, it is ensured that the most recent user feedback has greater reference value.

[0151] Then, adaptively adjust the optimization learning rate using the calculated feedback influence factor. When the feedback influence factor is large, it indicates that the user's recent activity is high, and at this time, increase the learning rate to accelerate model convergence; when the influence factor is small, the user's activity is low, and accordingly, decrease the learning rate to maintain stability. In specific implementation, if the influence factor is greater than 0.5, the learning rate is increased by 20%, and if it is less than 0.2, it is decreased by 30%.

[0152] Finally, the random gradient descent method is used to iteratively optimize the parameters of the multi-objective balanced recommendation strategy. In each iteration, training samples are randomly selected to calculate the gradient of the objective function, and the model parameters are updated according to the adaptive learning rate. After multiple iterations, the optimal parameter configurations for multiple objectives such as balanced accuracy, diversity, and novelty are obtained, thus generating the final personalized recommendation results.

[0153] By introducing a time-decaying feedback influence factor, the importance of the user's historical behavior data is effectively balanced, making the recommendation results more in line with the user's current interest preferences. Based on the feedback influence factor, the learning rate is adaptively adjusted, improving the model training efficiency, accelerating the convergence speed while ensuring the stability of the optimization process. The multi-objective balanced recommendation strategy is adopted, taking into account content diversity and novelty while ensuring the recommendation accuracy, thus enhancing the overall recommendation effect and user satisfaction.

[0154] Figure 6 This is a schematic structural diagram of the personalized content recommendation and behavior analysis system for the media convergence user in the embodiment of the present invention, as Figure 6 shown, the system includes:

[0155] The first unit is used to obtain the historical browsing data of the user on the media convergence platform, and based on the historical browsing data, construct a user behavior feature vector, where the user behavior feature vector includes content preference features and time preference features, and the content preference features are obtained by word vector encoding of content classification, and the time preference features are calculated based on the temporal distribution of the viewing time period;

[0156] The second unit is used to process the user behavior feature vector by using a dual attention mechanism, calculate the importance weights of different content types based on the content-level attention mechanism, calculate the dynamic weights of different time windows based on the temporal-level attention mechanism, and fuse the two levels of attention weights with the user behavior feature vector to obtain a fused feature vector;

[0157] The third unit is used to obtain the hidden state features representing the dynamic changes of the user's interest through a learnable transformation matrix combined with a gated recurrent unit based on the fused feature vector, and obtain the user interest prediction result according to the hierarchical prediction probability; according to the user interest prediction result, combined with the propagation trend of real-time hot events, establish a multi-objective balanced recommendation strategy, where the multi-objective balanced recommendation strategy improves the timely push effect of hot events while ensuring the personalization of the recommended content by dynamically adjusting the content timeliness and user interest matching degree; according to the multi-objective balanced recommendation strategy, screen the content to be recommended from the media convergence content library, and push the content to be recommended to the user according to the preset display rules.

[0158] In a third aspect of the embodiments of the present invention,

[0159] there is provided an electronic device, comprising:

[0160] a processor;

[0161] a memory for storing instructions executable by the processor;

[0162] wherein the processor is configured to call the instructions stored in the memory to execute the method described above.

[0163] In a fourth aspect of the embodiments of the present invention,

[0164] there is provided a computer-readable storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0165] The present invention may be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions for performing various aspects of the present invention stored thereon.

[0166] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for personalized content recommendation and behavior analysis for integrated media users, characterized in that: include: Obtaining historical browsing data of the user on the integrated media platform, and constructing a user behavior feature vector based on the historical browsing data, wherein the user behavior feature vector includes a content preference feature and a time preference feature, wherein the content preference feature is obtained by encoding the content classification by word vectors, and the time preference feature is obtained by calculating the temporal distribution of the viewing time period; The user behavior feature vector is processed using a dual attention mechanism, the importance weights of different content types are calculated based on the content-level attention mechanism, the dynamic weights of different time windows are calculated based on the time-level attention mechanism, and the attention weights of the two levels are fused with the user behavior feature vector to obtain a fused feature vector; Based on the fusion feature vector, a hidden state feature representing the dynamic change of user interest is obtained through a learnable transformation matrix combined with a gated recurrent unit, and the user interest prediction result is obtained according to the hierarchical prediction probability; according to the user interest prediction result, combined with the propagation trend of real-time hot events, a multi-objective balanced recommendation strategy is established, wherein the multi-objective balanced recommendation strategy dynamically adjusts the content timeliness and user interest matching degree, while ensuring the personalization of recommended content, and improves the timely push effect of hot events; according to the multi-objective balanced recommendation strategy, the content to be recommended is screened from the integrated media content library, and the content to be recommended is pushed to the user according to the preset display rules; Calculating an interaction matrix between different dimensions of the fused feature vector, wherein the interaction matrix is ​​used to characterize the correlation strength between different feature dimensions; The interaction matrix is ​​mapped into a query matrix, a key matrix and a value matrix through a learnable transformation matrix, the similarity between the query matrix and the key matrix is ​​calculated to obtain an attention distribution, and the attention distribution is weighted with the value matrix to obtain an attention enhancement feature; The attention enhancement feature is input into the gated recurrent unit, the forgetting degree of historical information is controlled by resetting the gate, and the update ratio of the current state is adjusted by updating the gate, so as to obtain the hidden state feature representing the dynamic change of user interest; Calculating a coarse-grained prediction probability through a sigmoid function based on the hidden state feature, calculating a fine-grained prediction probability through a sigmoid function based on a combination of the hidden state feature and the attention enhancement feature, and calculating a fusion weight coefficient based on the prediction confidence; The coarse-grained prediction probability and the fine-grained prediction probability are weighted and combined by the fusion weight coefficient to obtain a final user interest prediction result.

2. The method according to claim 1, characterized in that Based on the historical browsing data, constructing a user behavior feature vector includes: The content preference feature is obtained by encoding the word vector of the content classification, including: constructing a multi-level content classification system, dividing the content into layers according to themes, fields, and specific categories; using a hierarchical attention mechanism, respectively calculating semantic features at different levels, and adaptively fusing features at each level based on attention weights to generate content preference features; The time preference feature is obtained based on the temporal distribution calculation of the viewing time period, including: constructing multiple temporal distribution matrices based on the viewing time period, using a time decay function to perform weighted calculation on the temporal distribution matrices of different time windows, and generating the time preference feature.

3. The method according to claim 1, characterized in that The importance weights of different content types are calculated based on the content-level attention mechanism, and the dynamic weights of different time windows are calculated based on the time-level attention mechanism. The attention weights of the two levels are fused with the user behavior feature vector to obtain a fused feature vector including: Constructing content preference features into a content feature matrix, performing dimension mapping on the content feature matrix to obtain content hidden features, calculating attention scores based on the content hidden features, and converting the attention scores into importance weights of content types through a softmax function, wherein the importance weights are used to identify content types that have an impact on user interest expression; The user's behavior sequence is divided into multiple time windows according to preset time intervals, and a time series feature matrix is ​​constructed. The attention coefficient of each time window is calculated in combination with the hidden state information of historical behavior, and the attention coefficient is converted into a dynamic weight of the time window through a softmax function. The dynamic weight is used to characterize the influence of different time windows on the user's current interest; The importance weight of the content type is weighted and fused with the content feature matrix to obtain content weighted features, and the dynamic weight of the time window is weighted and fused with the time series feature matrix to obtain time series weighted features; The content weighted feature and the temporal weighted feature are concatenated, and a fusion feature vector is obtained by performing residual connection on the concatenated features through a learnable fusion weight matrix.

4. The method according to claim 1, characterized in that: According to the user interest prediction result, combined with the propagation trend of the real-time hot events, before establishing the multi-objective balanced recommendation strategy, the method further includes determining the propagation trend of the real-time hot events: Obtain hot event data of multiple monitoring platforms, wherein the hot event data includes platform access data, user interaction data, and propagation path data; extract the number of visits at each time point, multiply the number of visits by the weight coefficient preset by the corresponding monitoring platform, and accumulate the weighted number of visits of all monitoring platforms to obtain the number of visits characteristics of the hot event data, wherein the weight coefficient preset by the monitoring platform is determined based on the platform user scale and platform activity; Constructing a propagation hierarchy tree based on the propagation path data, identifying the propagation hierarchy to which each propagation node belongs, counting the number of propagation nodes in each propagation hierarchy, multiplying the number of propagation nodes by the hierarchy impact factor corresponding to the hierarchy, normalizing the calculation results of all hierarchies, and obtaining the propagation depth characteristics of the hot event data; Calculating normalized values ​​of the visit volume feature and the propagation depth feature respectively, substituting each normalized value into an exponential function to obtain a corresponding adaptive weight coefficient, and performing a weighted combination of the visit volume feature and the propagation depth feature based on the adaptive weight coefficient to obtain the propagation potential energy of the hot event data; Obtain the occurrence time of the hot event data, calculate the time difference between the current time and the occurrence time, calculate the rate of change of the propagation potential energy per unit time, dynamically adjust the preset initial attenuation coefficient according to the rate of change to obtain the real-time attenuation coefficient, multiply the propagation potential energy by the real-time attenuation coefficient to obtain the final propagation potential energy considering the time attenuation effect.

5. The method according to claim 4, characterized in that Based on the user interest prediction results and the dissemination trend of real-time hot events, a multi-objective balanced recommendation strategy is established, including: Obtain the data of the propagation trend and the recommended delay data of the real-time hot event, input the data of the propagation potential and the recommended delay data into the time decay function, calculate the event timeliness score, and use the product of the event timeliness score and the preset timeliness weight as the timeliness feature; Extract user semantic vectors, user content feature vectors and user topic tag sets from user historical behavior data; extract event semantic vectors, event content feature vectors and event topic tag sets from real-time hot events; Calculate the first cosine similarity between the user semantic vector and the event semantic vector, the second cosine similarity between the user content feature vector and the event content feature vector, and the overlap between the user topic tag set and the event topic tag set, and perform weighted combination of the first cosine similarity, the second cosine similarity, and the overlap to obtain the user interest matching degree; Obtaining real-time click data of users on recommended content, calculating the real-time click rate, inputting the difference between the real-time click rate and the preset benchmark click rate into the sigmoid function to obtain a dynamic weight coefficient, which is used to balance the timeliness feature and the matching degree of user interest; Based on the user's historical recommendation records, the content similarity between the event to be recommended and the historical recommended events is calculated, the content similarity is constructed as a diversity constraint item, the diversity constraint item is combined with the timeliness feature weighted by the dynamic weight coefficient and the user interest matching degree, and a multi-objective balanced recommendation strategy is established.

6. The method according to claim 5, characterized in that The method further comprises: Obtain user feedback data, extract feedback timestamps and feedback strength values, calculate feedback influence factors based on time decay functions, use the feedback influence factors to adaptively adjust the optimization learning rate, and use the stochastic gradient descent method to iteratively optimize the parameters of the multi-objective balanced recommendation strategy to obtain the final recommendation results.

7. A system for personalized content recommendation and behavior analysis for integrated media users, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain historical browsing data of users on the integrated media platform, and construct a user behavior feature vector based on the historical browsing data, wherein the user behavior feature vector includes a content preference feature and a time preference feature, wherein the content preference feature is obtained by encoding the content classification by word vectors, and the time preference feature is obtained by calculating the temporal distribution of the viewing time period; The second unit is used to process the user behavior feature vector using a dual attention mechanism, calculate the importance weights of different content types based on the content-level attention mechanism, calculate the dynamic weights of different time windows based on the time-level attention mechanism, and fuse the attention weights of the two levels with the user behavior feature vector to obtain a fused feature vector; The third unit is used to obtain hidden state features that characterize the dynamic changes of user interests based on the fused feature vector through a learnable transformation matrix combined with a gated recurrent unit, and obtain user interest prediction results based on hierarchical prediction probabilities; based on the user interest prediction results, combined with the propagation trend of real-time hot events, a multi-objective balanced recommendation strategy is established, wherein the multi-objective balanced recommendation strategy dynamically adjusts the timeliness of content and the matching degree of user interests to ensure the personalization of recommended content while improving the timely push effect of hot events; based on the multi-objective balanced recommendation strategy, the content to be recommended is screened from the integrated media content library, and the content to be recommended is pushed to the user according to preset display rules.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images