Personalized content recommendation and behavior analysis method and system for convergence media user

Abstract

The invention provides a fusion media user personalized content recommendation and behavior analysis method and system, and relates to the technical field of big data, and the method comprises the steps: obtaining user historical browsing data, constructing a behavior feature vector containing content preference and time preference, employing a dual attention mechanism to process the feature vector, and obtaining a fusion feature vector; and obtaining a user interest prediction result by combining a learnable transformation matrix with a gating loop unit based on the fusion feature vector, establishing a multi-target balance recommendation strategy by combining a real-time hot event propagation trend, and screening and pushing to-be-recommended contents from a content library. According to the method, the timely pushing effect of the hot events can be improved while the individuation of the recommended content is ensured, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to converged media technology, and in particular to a method and system for personalized content recommendation and behavior analysis for converged media users. Background Art

[0002] With the rapid development of Internet technology and mobile terminals, converged media platforms have become an important channel for users to obtain information. In order to improve user experience and platform operation efficiency, personalized content recommendation systems play an increasingly important role in converged media platforms. Traditional converged media recommendation systems are mainly based on collaborative filtering and content feature matching methods. They predict user interests by analyzing users' historical behavior data and recommend content accordingly.

[0003] First, existing technologies often use simple feature splicing or weighted averaging to process user behavior characteristics, which cannot effectively capture the correlation between different types of content and the dynamic changes in user interests over time, resulting in insufficient accuracy and personalization of recommendation results.

[0004] Secondly, traditional recommendation methods pay little attention to the impact of temporal information when modeling user interests and fail to fully utilize the behavioral pattern characteristics of users in different time periods. This makes it difficult for the system to accurately grasp the evolution of user interests and affects the recommendation effect.

[0005] Finally, when dealing with hot events, existing recommendation systems often find it difficult to strike a good balance between personalized recommendations and hot topic dissemination. They either overemphasize personalization and ignore the timely push of hot content, or blindly pursue hot topic dissemination and damage the personalized experience of recommendations. There is a lack of an effective multi-objective balancing mechanism. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for personalized content recommendation and behavior analysis for converged media users, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provide personalized content recommendation and behavior analysis methods for converged media users, including:

[0009] 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;

[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, 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;

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

[0012] According to the user interest prediction results, combined with the dissemination 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, thereby ensuring the personalization of recommended content and improving the timely push effect of hot events;

[0013] 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.

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

[0015] 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;

[0016] 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.

[0017] 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:

[0018] 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;

[0019] 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;

[0020] 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;

[0021] 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.

[0022] Based on the fused feature vector, a learnable transformation matrix is ​​combined with a gated recurrent unit to obtain a hidden state feature that characterizes the dynamic change of user interest, and a user interest prediction result is obtained according to the hierarchical prediction probability, including:

[0023] 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;

[0024] 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;

[0025] 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;

[0026] 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;

[0027] 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.

[0028] 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:

[0029] 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;

[0030] 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;

[0031] 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;

[0032] 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.

[0033] 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:

[0034] 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;

[0035] 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;

[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 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;

[0037] 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;

[0038] 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.

[0039] The method further comprises:

[0040] 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.

[0041] According to a second aspect of the embodiments of the present invention,

[0042] Provides personalized content recommendation and behavior analysis system for converged media users, including:

[0043] 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;

[0044] 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;

[0045] 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.

[0046] According to a third aspect of the embodiments of the present invention,

[0047] An electronic device is provided, comprising:

[0048] processor;

[0049] a memory for storing processor-executable instructions;

[0050] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0051] A fourth aspect of the embodiments of the present invention is:

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

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

[0054] The present invention constructs a user behavior feature vector that includes content preference and time preference and adopts a dual attention mechanism for feature fusion, which can more comprehensively characterize the user's interest characteristics and improve the accuracy of personalized recommendations.

[0055] The present invention introduces a gated recurrent unit to capture the dynamic change characteristics of user interests, and performs interest prediction through hierarchical prediction probability, so that the recommendation system can timely perceive the evolution law of user interests and improve the timeliness of recommended content and user experience.

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

[0057] Figure 1 A flowchart of a method for personalized content recommendation and behavior analysis for converged media users according to an embodiment of the present invention;

[0058] Figure 2 A schematic diagram showing a comparison of user interest recognition effects according to an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the performance comparison of the dual-layer attention feature fusion according to an embodiment of the present invention;

[0060] Figure 4 A schematic diagram of multi-dimensional evaluation of user interest prediction performance according to an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram of a multi-objective balanced recommendation system according to an embodiment of the present invention;

[0062] Figure 6 Schematic diagram of the structure of the personalized content recommendation and behavior analysis system for converged media users according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

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

[0065] Figure 1 FIG. 1 is a flow chart of a method for personalized content recommendation and behavior analysis for integrated media users according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] 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;

[0067] 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;

[0068] Based on the fused feature vector, a learnable transformation matrix is ​​combined with a gated recurrent unit to obtain hidden state features that characterize the dynamic changes of user interests, and the user interest prediction results are obtained according to the hierarchical prediction probability; 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 content timeliness and user interest matching degree to ensure the personalization of recommended content while improving 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 preset display rules.

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

[0070] 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;

[0071] 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.

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

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

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

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

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

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

[0078] Existing methods for identifying user interests are mainly based on single behavioral similarity calculations or simple deep learning models. These methods have obvious deficiencies in dealing with the diversity and dynamic evolution of user interests. Traditional collaborative filtering methods only focus on the similarity of users' historical behaviors and cannot deeply understand content features; although basic deep learning methods introduce neural network structures, they fail to fully utilize the hierarchical features of content, resulting in poor accuracy and adaptability in interest identification.

[0079] This application innovatively proposes a technical solution that combines multi-level content classification with a hierarchical attention mechanism in order to improve the accuracy and generalization of user interest recognition. By constructing a multi-level content classification system, the semantic features of different levels such as the subject, field, 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 features at different levels, thus achieving dynamic fusion of features. This method of hierarchical recognition and adaptive fusion significantly improves the model's depth of understanding of user interests and recognition accuracy.

[0080] Experimental results show that compared with existing technologies, this solution shows better recognition effects in the initial stage, and as the iterative optimization process progresses, the recognition accuracy rate and convergence speed have obvious advantages. Especially after multiple rounds of iterations, the performance gap between this solution and existing methods has further widened, which fully verifies the significant effect of this technical solution in capturing user interest characteristics and continuously optimizing model performance. This advantage is mainly due to the multi-level content classification system's refined characterization of user interests, and the dynamic adjustment ability of the hierarchical attention mechanism in the feature fusion process, which enables the model to more accurately identify and predict user interest preferences, and has important application value.

[0081] In an optional implementation, 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 including:

[0082] 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;

[0083] 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;

[0084] 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;

[0085] 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.

[0086] First, we obtain the user behavior data on the content platform, including the user's interactive behaviors such as clicking, browsing, and collecting different types of content. We preprocess the original behavior data and extract key information such as user ID, content ID, content type, behavior type, and timestamp.

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

[0088] The content feature matrix is ​​dimensionally mapped, and a two-layer fully connected network is used. The input layer dimension is the same as the original feature dimension, the hidden layer dimension is set to 64, and the output layer dimension is 32, so as to obtain a more compact content hidden layer feature representation. The attention score is calculated based on the content hidden layer features, and the attention coefficient is obtained through nonlinear transformation. The attention coefficient is normalized by the softmax function to obtain the importance weights of different content types.

[0089] In the time-level attention mechanism, the user's behavior sequence in the past three months is divided into a time window of 7 days, resulting in 12 time windows. For the behavior sequence in each time window, the behavior type, content type and other features are extracted to construct a time-series feature matrix. Combined with the historical behavior hidden state information obtained by the GRU network, the attention coefficient of each time window is calculated. The attention coefficient is normalized by the softmax function to obtain the dynamic weights of different time windows.

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

[0091] Take the 100 pieces of content that a user has browsed recently as an example, of which 40 are news, 30 are videos, and 30 are pictures and texts. For news content, we extract features such as title word vector (100 dimensions), text keyword vector (100 dimensions), and news classification vector (20 dimensions); for video content, we extract video title vector (100 dimensions), video description vector (100 dimensions), duration feature (1 dimension), etc. These features are spliced ​​to construct a 220-dimensional content feature matrix.

[0092] The importance weights of content types calculated by the attention mechanism are: news 0.5, video 0.3, 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 feature are as follows from far to near:

[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, which reflects the temporal evolution characteristics of user interests.

[0094] Figure 3 This is a schematic diagram of the performance comparison of the dual-layer attention feature fusion according to an embodiment of the present invention:

[0095] The figure shows the performance of three different feature fusion methods. The circle (●) represents the technical solution, which adopts a dual-layer attention mechanism at the content level and the time-series level. The content hidden features are obtained by dimension mapping the content feature matrix and the content type importance weight is calculated. At the same time, the user behavior sequence is divided into multiple time windows to calculate the dynamic weight, and finally the feature fusion is realized through residual connection; the square (■) represents the single-layer attention method, which only uses the content-level attention mechanism to calculate the content type weight without considering the dynamic changes of the time-series features; the triangle (▲) represents the simple splicing method, which directly splices the content features and the time-series features without introducing the attention mechanism for feature selection and weighting. In the process of 0-10 iterations, the accuracy of the technical solution increased from 0.85 to 0.95, with a total improvement of 0.10, and reached a high accuracy of 0.90 after 4 iterations; the single-layer attention method increased from 0.78 to 0.85, with an improvement of 0.07; the simple splicing method increased from 0.72 to 0.80, with an improvement of 0.08. This technical solution combines the content-level and time-level dual-layer attention mechanisms, which not only has the highest initial accuracy, but also maintains a significant leading advantage throughout the entire iteration process, especially in the 6-8 iteration stage, the accuracy rate quickly increased from 0.92 to 0.94, showing excellent feature fusion capabilities. From the overall trend of the curve, the performance gap between this technical solution and the other two methods gradually widens with the increase in the number of iterations, indicating that the dual-layer attention mechanism has a significant effect in improving feature expression capabilities and model performance. Compared with single-layer attention and simple splicing 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 use single-layer attention mechanism or simple feature splicing. The single-layer attention method only focuses on the importance weight calculation of content features, and cannot effectively capture the temporal evolution characteristics of user interests; the simple splicing method directly splices content features and temporal features, lacking a mechanism to distinguish and select the importance of different features, resulting in unsatisfactory feature fusion effects and difficulty in accurately expressing the multi-dimensional interest characteristics of users.

[0097] This application innovatively proposes a dual-layer attention feature fusion solution from the perspective of improving the accuracy of feature fusion and the ability of temporal modeling. At the content-level attention level, the hidden features are obtained by dimensionally mapping the content feature matrix, and the importance weights of the content types are calculated based on this, thereby achieving accurate modeling of user content preferences; at the temporal level attention level, 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 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 nonlinear transformations to enhance feature expression capabilities.

[0098] Experimental results show that compared with existing technologies, this solution has significant advantages in feature fusion effect. This advantage is mainly reflected in three aspects: first, the two-layer attention mechanism can more accurately identify important features and improve 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 solution to show better performance and faster convergence speed in user interest recognition tasks, and have important application value.

[0099] In an optional implementation, obtaining a hidden state feature representing the dynamic change of user interest based on the fused feature vector through a learnable transformation matrix combined with a gated recurrent unit, and obtaining a user interest prediction result according to a hierarchical prediction probability includes:

[0100] 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;

[0101] 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;

[0102] 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;

[0103] 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;

[0104] 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.

[0105] First, obtain the fused feature vector, which contains multi-dimensional feature information such as user historical behavior and item attributes. Generate an interaction matrix by calculating the interaction relationship between features of different dimensions. Specifically, for any two dimensions of features in the fused feature vector, calculate the dot product similarity between them to obtain an interaction matrix that represents the strength of association between feature dimensions. For example, for a fused feature vector containing 128 dimensions, generate a 128×128 interaction matrix.

[0106] Next, a learnable transformation matrix is ​​introduced to map the interaction matrix. The transformation matrix contains three independent matrices, which are used to generate the query matrix, key matrix and value matrix respectively. Through matrix multiplication, 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 the attention distribution. The attention distribution is weighted and summed with the value matrix to obtain the attention enhancement feature.

[0107] The attention enhancement feature is input into the gated recurrent unit. The gated recurrent unit contains two gating mechanisms: reset gate and update gate. The reset gate calculates the forgetting coefficient based on the current input feature and historical state to control the degree of retention of historical information. The update gate calculates the update ratio to adjust the update amplitude of the current state. Through the adjustment of the gating mechanism, the hidden state features that can characterize the dynamic changes of user interests are obtained.

[0108] Based on the hidden state features, the coarse-grained user interest prediction probability is calculated through the sigmoid function. At the same time, the hidden state features are combined with the attention enhancement features, and 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. The prediction result with higher confidence has 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] Take specific data as an example: Assuming that the dimension of the fused feature vector is 128, a 128×128 interaction matrix is ​​generated. The 64-dimensional query, key, and value matrices are mapped through the 64×128-dimensional transformation matrix. The 64-dimensional attention distribution vector is calculated and weighted with the value matrix to obtain the 64-dimensional attention enhancement feature. The gated recurrent unit outputs a 32-dimensional hidden state feature. Finally, the predicted probability value between 0 and 1 is output.

[0110] Figure 4 This is a schematic diagram of multi-dimensional evaluation of user interest prediction performance according to an embodiment of the present invention:

[0111] This figure shows the performance comparison of three different prediction methods. The circle (●) represents this technical solution, which uses a learnable transformation matrix to map interactive features, combines gated recurrent units to achieve dynamic modeling, and obtains results through hierarchical prediction; the square (■) represents the comparison method 1, which only uses a single-layer attention mechanism to calculate feature weights for prediction; the triangle (▲) represents the comparison method 2, which uses simple feature concatenation and basic neural network structure. From the evaluation indicators of six dimensions, this technical solution is significantly better than the comparison method 1 (0.85, 0.83, 0.84, 0.86, 0.82, 0.81) and the comparison 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 this technical solution has obvious advantages in all dimensions, forming the outermost hexagon, and the performance of each dimension is balanced, which fully verifies the effectiveness of the dual-layer attention mechanism and hierarchical prediction strategy.

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

[0113] In an optional implementation, before establishing a multi-objective balanced recommendation strategy based on the user interest prediction result and the propagation trend of the real-time hot events, the method further includes determining the propagation trend of the real-time hot events:

[0114] 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;

[0115] 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;

[0116] 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;

[0117] 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.

[0118] Obtain hot event data from multiple mainstream content platforms, including visit volume, user comments, forwarding and other interactive data as well as content dissemination paths. For different platforms, weight coefficients are set based on their user scale and number of daily active users. The weight of platforms with a user scale of more than 10 million and a daily activeness of more than 30% is 0.4, the weight of platforms with a user scale of 5-10 million and a daily activeness of 20-30% is 0.3, and the weight of other platforms is 0.2.

[0119] For each hot event, real-time visit data of each platform is collected at intervals of 10 minutes. The visit volume at each time point is multiplied by the weight coefficient of the corresponding platform and accumulated to obtain the weighted visit volume. The weighted visit volume sequence for 24 consecutive hours is normalized to obtain the visit volume characteristics that reflect the attention of the event.

[0120] Construct an event propagation hierarchical tree, with the starting publishing node as the 0th layer, direct forwarding as the 1st layer, forwarding of forwarding as the 2nd layer, and so on. Count the number of propagation nodes at each layer, and set the hierarchical impact factor: 1.0 for the 0th layer, 0.8 for the 1st layer, 0.6 for the 2nd layer, 0.4 for the 3rd layer, and 0.2 for the 4th layer and above. Multiply the number of nodes at each layer by the corresponding impact factor and normalize them to obtain the propagation depth feature.

[0121] The visit volume feature and the propagation depth feature are normalized to 0-1 respectively. The normalized value is substituted into the negative exponential function to calculate the adaptive weight: the larger the feature value, the larger the corresponding weight. The two features are weighted and summed based on the adaptive weight to obtain the propagation potential of the hot event.

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

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

[0124] At a certain point in time, the visits to the three platforms were 100,000, 80,000, and 50,000 respectively, and the total visits after weighting were 85,000. The propagation level distribution is: 1 node in the 0th layer, 100 nodes in the 1st layer, 500 nodes in the 2nd layer, 2,000 nodes in the 3rd layer, and 5,000 nodes in the 4th layer. The propagation depth characteristic value is calculated to be 0.75.

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

[0126] Through multi-platform data fusion and differentiated weight allocation, the dissemination data of hot events are comprehensively and accurately obtained, which improves the reliability and representativeness of dissemination trend analysis. The introduction of dissemination hierarchical tree and hierarchical influence factors deeply depicts the dissemination depth and impact range of events, making the calculation of dissemination potential more reasonable and three-dimensional. The use of adaptive weights and dynamic attenuation mechanisms can flexibly adjust the feature combination mode and time attenuation effect, so that the dissemination potential can better reflect the real-time dissemination status and development trend of events.

[0127] In an optional implementation, based on the user interest prediction results and in combination with the propagation trend of real-time hot events, a multi-objective balanced recommendation strategy is established, including:

[0128] 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;

[0129] 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;

[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 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;

[0131] 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;

[0132] 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.

[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, which is usually in the range of 0.1-1 seconds. 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, the BERT model is used to extract the semantic vector of the user's recently browsed content, with a dimension of 768. The deep learning model is used to extract the feature vector of multimodal content such as pictures, texts, and videos, with a dimension of 512. A user interest topic label set is constructed based on the content classification system, including labels in multiple fields such as entertainment, sports, and technology.

[0135] Semantic vectors, content feature vectors, and topic tag sets are also extracted for real-time hot events. The cosine similarity of user-event semantic vectors is calculated, with a value range of 0-1. The cosine similarity of content feature vectors is calculated. The overlap ratio of user topic tags and event topic tags is calculated. The weights of the three features are set to 0.4, 0.3, and 0.3 respectively, and the weighted sum is used to obtain the user interest matching degree.

[0136] Real-time statistics of user click data on recommended content are collected to calculate the click rate in the last hour. The benchmark click rate is set to 5%, and the difference between the actual click rate and the benchmark value is input into the sigmoid function to obtain a dynamic weight coefficient between 0 and 1. When the click 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. Invert the similarity to get the diversity constraint item. The higher the similarity, the stronger the constraint. Combine the diversity constraint item with the weighted timeliness feature and interest matching to construct the final recommendation strategy.

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

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

[0140] The user's recommendation click rate in the last hour was 8%, which was 5% higher than the benchmark click rate. The dynamic weight coefficient was calculated by the sigmoid function to be 0.65. Retrieving the user's recommendation records in the past 7 days found 20 related events with an average content similarity of 0.3, which was converted into a diversity constraint of 0.7.

[0141] Taking into account the timeliness feature (0.328), interest matching (0.686) and diversity constraints (0.7), the final recommendation score for this event is 0.571.

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

[0143] This figure shows the operation panel of the multi-objective balanced recommendation system, which intuitively presents the core functions and data analysis results. The real-time system indicator area shows a system delay of 0.2 seconds, a click-through rate of 8%, and a dynamic weight coefficient of 0.65, indicating that the system performance is good and is positively inclined towards user interests. The user interest analysis section shows a semantic similarity of 0.75, a feature similarity of 0.68, and a label overlap of 0.6, as well as labels such as technology and entertainment that users are concerned about. The hot event analysis area shows a communication potential of 0.85, a timeliness score of 0.82, and a diversity constraint of 0.7.

[0144] The weight configuration area shows that the timeliness weight is 0.4, the semantic vector weight is 0.4, and the content feature and topic tag weights are 0.3 each. The recommendation score consists of 0.328 display timeliness features, 0.686 interest matching, 0.7 diversity constraints, and a final score of 0.571.

[0145] The recommendation list area displays three ranking 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 releases a new round of scientific and technological innovation support policies" (recommendation score 0.498), each of which shows the details of three indicators: timeliness, interest matching and diversity. The overall design visualizes the complex algorithm process, helping operators understand the recommendation mechanism and optimize the content distribution effect.

[0146] By introducing timeliness features and dynamic weight mechanisms, the recommendation strategy can respond to changes in the dissemination trend of hot events in a timely manner, ensuring the timeliness of information while avoiding over-recommendation of hot content. Multi-dimensional feature extraction and similarity calculation methods comprehensively characterize the degree of match between user interests and event content, improving the accuracy of personalized recommendations and user experience. Diversity constraints based on historical recommendation records effectively avoid content homogeneity problems, making recommendation results more diverse and richer, and improving the exploratory nature and user stickiness of the recommendation system.

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

[0148] 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.

[0149] First, we obtain user feedback data, including user behavior data such as clicks, favorites, and comments on recommended content. We extract feedback timestamp information from the feedback data and record the specific time point when each feedback occurred. At the same time, we extract feedback strength values, such as assigning 0.2 to click behavior, 0.5 to favorite behavior, and 0.8 to comment behavior. The larger the value, the higher the feedback strength.

[0150] Then, the feedback impact factor is calculated based on the time decay function. Using exponential decay, the weight of feedback data farther away from the current time is smaller. For example, the impact factor of a user's click behavior 2 days ago is 0.2 times 0.8 squared; the impact factor of the collection behavior 7 days ago is 0.5 times 0.8 to the 7th power. Time decay ensures that the most recent user feedback has greater reference value.

[0151] Then, the calculated feedback influence factor is used to adaptively adjust the optimization learning rate. When the feedback influence factor is large, it means that the user has been active recently. At this time, increasing the learning rate will speed up the convergence of the model; when the influence factor is small, the user is not active, and the learning rate is reduced accordingly to maintain stability. In the 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 reduced by 30%.

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

[0153] By introducing the time-decayed 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, which improves the model training efficiency, accelerates the convergence speed and ensures the stability of the optimization process. The multi-objective balanced recommendation strategy is adopted to ensure the accuracy of the recommendation while taking into account the diversity and novelty of the content, thus improving the overall recommendation effect and user satisfaction.

[0154] Figure 6 FIG. 1 is a structural diagram of a system for personalized content recommendation and behavior analysis for integrated media users according to an embodiment of the present invention. Figure 6 As shown, the system comprises:

[0155] 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;

[0156] 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;

[0157] 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.

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

[0159] An electronic device is provided, comprising:

[0160] processor;

[0161] a memory for storing processor-executable instructions;

[0162] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0163] A fourth aspect of the embodiments of the present invention is:

[0164] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

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

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, 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 fused feature vector, a learnable transformation matrix is ​​combined with a gated recurrent unit to obtain hidden state features that characterize the dynamic changes of user interests, and the user interest prediction results are obtained according to the hierarchical prediction probability; according to 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 content timeliness and user interest matching degree to ensure the personalization of recommended content while improving 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 preset display rules.

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: Based on the fused feature vector, a learnable transformation matrix is ​​combined with a gated recurrent unit to obtain a hidden state feature that characterizes the dynamic change of user interest, and a user interest prediction result is obtained according to the hierarchical prediction probability, including: 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.

5. 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.

6. The method according to claim 1, 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.

7. The method according to claim 6, 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.

8. 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 7, 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.

9. 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 described in any one of claims 1 to 7.

10. 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 7 is implemented.

Images